<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Rajiv Gopinath  on Medium]]></title>
        <description><![CDATA[Stories by Rajiv Gopinath  on Medium]]></description>
        <link>https://medium.com/@mail2rajivgopinath?source=rss-ae007d129200------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dmbNkD5D-u45r44go_cf0g.png</url>
            <title>Stories by Rajiv Gopinath  on Medium</title>
            <link>https://medium.com/@mail2rajivgopinath?source=rss-ae007d129200------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 14:39:42 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@mail2rajivgopinath/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Chapter 35: Why Brand Investment Always Loses the Argument]]></title>
            <link>https://medium.com/@mail2rajivgopinath/chapter-35-why-brand-investment-always-loses-the-argument-b433bbe76d6e?source=rss-ae007d129200------2</link>
            <guid isPermaLink="false">https://medium.com/p/b433bbe76d6e</guid>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[market-mix-modeling]]></category>
            <category><![CDATA[marketingeffectiveness]]></category>
            <category><![CDATA[branding]]></category>
            <dc:creator><![CDATA[Rajiv Gopinath ]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 11:08:59 GMT</pubDate>
            <atom:updated>2026-11-17T11:08:59.072Z</atom:updated>
            <content:encoded><![CDATA[<h4>Marketing Effectiveness Series -The Science of Evidence</h4><h4>A strong brand lifts your sales and makes your advertising work harder. Almost none of that shows up when you measure it — and better models will not fix it; Chapter 35 in the Marketing Effectiveness Series</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qNU7easABRuGgyMjESVUDw.png" /></figure><p>Every year, in most companies, the same argument happens and the same side wins.</p><p>Performance media arrives with numbers. Cost per acquisition, return on spend, conversions by campaign, week by week, to two decimal places. Brand arrives with a tracker that moved three points and a conviction that the business would be worse off without it.</p><p>Only one of those survives a budget review.</p><p>The usual explanation is that performance is measurable and brand is not, and that better measurement will eventually close the gap. Every few years a new method is announced that will finally do it.</p><p>I want to argue something less comfortable. The gap is not a tooling problem, and it is not going to close.</p><p>Start with what a strong brand actually does, because it does two separate things.</p><p><strong>It raises the floor.</strong> People buy you without being prompted — this week and every week — because they know you, trust you, and have bought you before. Nothing has to happen for this to work.</p><p><strong>And it makes your advertising work harder.</strong> The same campaign, the same money, the same creative produces more for a brand people already recognise than for one they do not.</p><p>Two jobs. Now the part that decides the budget argument: the first job is much larger, and the second job is the one your data can measure.</p><p>So I tested it. I invented twenty markets with three years of weekly sales each, and because I invented them, I could decide in advance exactly what the brand was contributing in every one — how much it lifted the floor, and how much harder it made the advertising work. Then I analysed that data the way anybody would analyse a real client’s, and checked the answers against the truth I had set.</p><p>The multiplier came back close to right every time, no matter how I built the model.</p><p>The floor did not. One decision changed everything: whether the model allows for the plain fact that markets differ from each other. Allow for it, and the brand’s contribution to the floor nearly vanishes — because that contribution <em>is</em> a difference between markets, and the allowance absorbs it. Don’t allow for it, and you credit the brand with everything else that differs between markets too.</p><p>Same data. Same brand. Depending on that one choice, the answer moved by a factor of twenty-five.</p><p>And in that business, the floor was worth about nine times the multiplier.</p><p>So the number you can trust is the small one.</p><p><strong>The brand’s two biggest contributions are the two your measurement system is built not to see.</strong></p><p>A measurement system detects things that move quickly, and that move <em>within</em> the unit being measured. This week against last week. This market against its own history. That is what a weekly model is built to do, and it does it well.</p><p>The floor is not a movement within a market. It is a standing difference between markets, and it has been standing for fifteen years.</p><p>There is a second blindness working in the same direction. Brand investment does not pay back inside the quarter it was spent in. Most of what it buys arrives later — sometimes much later — and a measurement window that closes before the return arrives will report that there was no return.</p><p>The governing equation has carried a brand multiplier since Part I: the term saying the same campaign produces more for a stronger brand. Every chapter since has worked on everything around it and left it assumed.</p><p>This is the chapter where somebody has to put a number in it.</p><h3>Section 1: The Variable That Explained Everything</h3><p>Here is the point, stated first: the most natural way to connect brand health to sales produces a number that flatters everyone in the room and cannot be trusted.</p><p>The tracker has had a good year. Awareness up four points, consideration up three, and the brand team has a chart where both lines rise together. They have earned the right to present it.</p><p>The CFO asks the obvious question. What is that worth?</p><p>Nobody can answer, and the reason is slightly embarrassing: the mix model has no brand variable in it at all. Not because anybody decided it shouldn’t. The tracker arrives monthly and the model runs weekly. Nobody was certain where it belonged. And every year there was something more pressing to fix. So the model explains sales with media, price, distribution and seasonality, and the brand — the thing the company has spent thirty years building — appears nowhere in the equation.</p><p>So the analyst does the obvious thing and puts it in.</p><p>It comes back as the largest driver in the model. Larger than television. Larger than all media combined.</p><p>The room is delighted. Here, at last, is the number the brand team has been asking for. Somebody suggests putting it in the annual report.</p><p>Now, the reason to be careful.</p><p>Think about what a tracker interview actually captures. You ask a few hundred people each month whether they know your brand, whether they’d consider it, whether they prefer it. Who says yes?</p><p>Disproportionately, the people who bought it.</p><p>A household that bought your biscuits in March is more likely in April to say they are aware of the brand, more likely to consider it, more likely to express a preference. Not because the brand got stronger. Because they have a packet in the kitchen. Buying something is one of the more reliable ways to become aware of it.</p><p>So sales raise the tracker at the same time as the tracker’s underlying reality raises sales. And it is not only sales doing it. Distribution moves both — a brand on more shelves is both bought more and noticed more. Media moves both. Product experience moves both. Several roads, all running in the direction the analysis assumed was one-way.</p><p>There is a test for this, and the previous chapter was built on it. Take any variable and ask whether the last few weeks of sales predict it. Media that follows demand lights up; planned media does not. Run that test on a brand tracker and it will light up — because a tracker is, in part, a record of who bought you recently.</p><p>Which brings the chapter to its first careful distinction, because the easy conclusion here is wrong.</p><p>This does not mean the regression is worthless, or that the brand has no effect on sales, or that trackers are useless instruments. The brand effect is real and it is large — Section 4 will show how large. What it means is narrower and more awkward: that particular number, produced that particular way, is <strong>not interpretable as cause</strong>. It is a mixture of the brand’s effect on sales and sales’ effect on the brand’s measurement, and nothing in the output tells you the proportions.</p><p><strong>Putting brand health into the model does not answer the brand question. It replaces it with a different question, and then hides the substitution.</strong></p><p>So the CFO’s question is still sitting there, and the honest response is not a number. It is to point out that the question contains two questions wearing one coat.</p><p><em>What is the brand worth in sales?</em> — meaning, how much do we sell because people already know and trust us, before any advertising runs.</p><p><em>And what does the brand do to our marketing?</em> — meaning, does the same campaign work harder for us than it would for a weaker brand.</p><p>Those are not two framings of one thing. They are two different quantities, with two different sizes, and two completely different levels of answerability — and the one the data cannot give you is the one worth more.</p><h3>Section 2: Why the Multiplier Cannot Be Seen in One Brand</h3><p>Here is the point, stated first: if a multiplier never changes, your data cannot tell it apart from the thing it multiplies.</p><p>The first of the brand’s two jobs — raising the floor — is uncontroversial as an idea. Everybody accepts that a well-known brand sells something on a week when no advertising runs at all.</p><p>The second job is the interesting one. A strong brand makes the same campaign work harder. Same money, same reach, same creative, more sales. If that is true, then the brand is not a line in the plan competing with television. It is a coefficient on television.</p><p>Write it that way and the trouble shows up immediately.</p><p>What the model sees each week is media pressure going in and sales coming out. From that, it estimates how much sales respond per unit of pressure. One number.</p><p>But that number is itself two numbers multiplied together: how strong the brand is, and how much response the channel is capable of producing. The model only ever sees the product.</p><p>Twelve can be three times four. It can also be two times six, or one times twelve. Nothing in the twelve tells you which.</p><p>So when a model says television delivers a certain response for your brand, there are two stories behind it and they fit the data identically. Either the brand is strong and it is amplifying an ordinary television effect, or the brand is unremarkable and television is simply doing well here — better creative, better buying, a larger reachable audience. One model, one brand, one history, and no way to choose between them.</p><p><strong>A multiplier that never varies is not a multiplier the data can identify. It is just part of the coefficient.</strong></p><p>This is worth sitting with, because it is not a modelling failure. A better estimator will not help. A longer history will not help. Tighter priors will not help — they will only move the answer toward whatever you assumed. The information required to separate those two stories is not in the data, and Part II’s lesson applies without modification: when the evidence cannot distinguish two explanations, the model’s apparent precision is coming from somewhere other than the evidence.</p><p>What does separate them is variation in brand strength itself.</p><p>If you have twelve markets where the brand is genuinely stronger in some than others, and media runs in all of them, then you can watch whether media’s response bends with brand strength. Where the brand is strong, does the same pressure produce more? That comparison is the multiplier, and it does not exist inside any single market’s history.</p><p>The same works across a portfolio — several brands, different strengths, the same media machinery behind them. And in principle it can work over time, if brand strength moved for a reason outside the business, though that is rarer than people hope and usually confounded with whatever caused the movement.</p><p>Cross-sectional variation is not the only route. It is the practical one.</p><p>Which means the machinery for this was built two chapters ago. The hierarchical structure from Chapter 31 — markets borrowing strength from one another, differences earning their distance from the group — is exactly the apparatus that makes the multiplier visible. At the time it was a way of handling thin data. It turns out to be the only practical way to see one of the two terms in the book’s own equation.</p><p>And it explains something that should otherwise be puzzling. The brand multiplier has been in the governing equation since Part I, and nobody has estimated it. Not because it was difficult. Because most brands commission a single model, for a single brand, in a single market set, and the information was never in the room.</p><h3>Section 3: The Cheapest Fix in This Chapter</h3><p>Here is the point, stated first: noise in a brand measure does not just make the answer uncertain. It drags the answer toward zero.</p><p>Start with how much noise there is.</p><p>A typical brand tracker interviews a few hundred people a month. At four hundred interviews, the margin of error on a single percentage sitting near the middle of the scale is about two and a half points. Comparing one month against the next roughly compounds that, to around three and a half points.</p><p>Which means the four-point rise in awareness that opened this chapter is barely one standard error of the difference. It is not evidence of anything. It is what a flat line looks like when you measure it twice.</p><p>Brand teams narrate those movements anyway, because a flat line is difficult to present and somebody has to fill the slide. Nobody is being dishonest. They are reading weather into static.</p><p>Now the second problem, which is the one that costs money.</p><p>Most trackers carry five or six metrics — awareness, salience, consideration, preference, willingness to pay more. Faced with that, analysts do one of two things. They pick the headline metric and use it alone. Or they put all six into the model as six variables, which is worse, because the six are heavily correlated and the model has no way to apportion credit between them.</p><p>I tested the first approach in the simulated business. The true multiplier was 0.50. Using a single headline metric, the model recovered <strong>0.27 to 0.30</strong> — less than two thirds of the real effect, consistently, across every world I tried. Averaging four tracker items into one measure recovered <strong>0.41 to 0.43</strong>.</p><p>Same data. Same truth. Half the missing effect recovered by averaging four columns instead of picking one.</p><p>The mechanism runs against intuition. People assume a noisy variable produces a wider confidence interval — uncertainty, but centred on the right answer. That is not what happens. When the <em>input</em> is measured with error, the estimated effect is pulled systematically toward zero. The interval does widen, but it widens around a number that is too small.</p><p>Averaging works because the noise and the signal behave differently. The thing you are trying to measure is shared across all the items — it is the same underlying brand strength showing up in each question. The noise is specific to each item: who happened to be interviewed, how the question was worded, what mood the week was in. Average the items and the shared part survives while the specific parts partly cancel.</p><p>A proper latent-variable model does this more carefully than a simple average, and it will do better. But be clear about what it buys: it separates shared signal from item-specific noise <strong>under explicit assumptions about how the tracker measures the underlying thing</strong>. If those assumptions are wrong — if salience and preference are genuinely two different constructs rather than one, which is quite possible — a single-factor model will force them together and the result inherits that error. Test it rather than assuming it.</p><p>Now notice where this sits in the chapter’s argument.</p><p>Measurement noise is a third force pushing in the same direction as the other two. The floor is invisible because it lives between markets. The long-run return is invisible because it arrives after the window closes. And whatever brand effect <em>is</em> inside the window gets shrunk toward zero by the noise in the instrument measuring it.</p><p>Three separate mechanisms so far. All of them biased the same way. None of them a mistake anybody made.</p><p>This one, at least, you can fix this afternoon, with data you already own, by using more than one column.</p><h3>Section 4: The Two Questions, and Why Only One Has an Answer</h3><p>Here is the point, stated first: one of the brand’s two jobs has a stable answer in your data. The other has six answers, and which one you get is decided by a modelling choice nobody argues about.</p><p>Back to the invented business. Twenty markets, three years of weekly sales each. I set the truth myself: brand strength lifts the floor by 3.0 units a week per standard deviation, and makes media work 50% harder per standard deviation. Then I built the model six different ways and ran each three hundred times.</p><p>Three versions of the world, each analysed twice — once allowing each market its own baseline, once not. Allowing each market its own baseline is what modellers call market fixed effects, and it is close to standard practice, because markets obviously differ and nobody wants those differences loading onto the media variables.</p><p><em>Exhibit: simulated business, twenty markets, three years, 300 runs per specification. True floor effect 3.0, true multiplier 0.50. Brand strength measured as the average of four tracker items.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/832/1*syCKZnAvrBm0fBURa2--dQ.png" /></figure><p>Read the right-hand column first.</p><p><strong>0.41 to 0.53, against a true 0.50.</strong> Every world. Every specification. With the tracker contaminated by sales, without it. With market baselines, without them. With brand strength tangled up in twenty years of market history, and without.</p><p>I did not expect that. The outline for this chapter assumed the feedback problem from Section 1 would be the central difficulty, and for the multiplier it turns out to be almost irrelevant.</p><p>Now read the left-hand column. <strong>0.14 to 3.51, against a true 3.0.</strong> A factor of twenty-five, spanning from “the brand contributes essentially nothing to the floor” to “the brand contributes more than it really does.”</p><p>Same data. Same brand. Same truth sitting underneath.</p><p>The reason for the difference is structural, and once you see it you cannot unsee it.</p><p>The brand’s contribution to the floor <em>is</em> a difference between markets. It is the reason Chennai’s baseline sits above Nagpur’s. So when you let each market have its own baseline — the sensible, standard, defensible thing to do — you have told the model that it does not need to explain why the baselines differ.</p><p>Which was the question.</p><p>You asked why some markets sell more without advertising, and then handed the model a free parameter for every market that absorbs exactly that. The brand’s floor contribution does not get estimated. It gets quietly parked inside twenty market constants, where nobody will look for it again.</p><p>The alternative is to leave those baselines out, and in the clean world that works rather well: 2.58 against a true 3.0. But look at the bottom row. Once brand strength is correlated with the things that make a market large in the first place — population, distribution depth, how developed the category is, how long the brand has been sold there — the estimate climbs to 3.51. The brand now gets credit for everything else that differs between markets along with its own effect.</p><p>So both choices are defensible and they disagree by a factor of twenty-five. The specification is not a technical detail sitting beneath the answer. <strong>The specification is the answer.</strong></p><p>Then there is the finding I like least.</p><p>Compare the first and third rows. In the clean world, allowing market baselines, the floor comes back at 0.14. In the contaminated world — where the tracker is partly a record of who bought recently, which is the problem Section 1 spent two pages warning about — it comes back at 1.46.</p><p>The contaminated version is closer to the truth.</p><p>Not because contamination helps. Because two large errors are pointing in opposite directions and partially cancelling. The market baselines push the estimate down toward zero; the sales feedback pushes it up. The result sits nearer 3.0 than either error alone would allow.</p><p>An analyst comparing those two specifications on any ordinary criterion would prefer the second one. It fits better. It gives a more plausible number. It is more presentable. And it is right by accident, for reasons that will not hold next year or in another category.</p><p>That is the most dangerous state a model can be in: <strong>wrong in a way that improves its appearance.</strong></p><p>The multiplier escapes all of this for the reason Section 2 set out. It is a slope, not a level. Giving every market its own baseline absorbs any standing difference in how much a market sells, and cannot touch a difference in how steeply that market responds to media. The same logic explains why the tracker feedback does so little damage: feedback adds a component that tracks the sales level, and the level is what gets absorbed.</p><p>So the brand’s two jobs are not just different in size. They are different in kind. One lives in differences between markets, which is precisely where the standard controls do their work. The other lives in how a curve bends, which the controls cannot reach.</p><p>Everything above came from a business with twenty markets, three years of weekly data, and four tracker items. That is a better dataset than most brands own. So the obvious question is whether any of it survives contact with what you actually have.</p><p>I ran it again at realistic data quality. Five markets instead of twenty. Two usable tracker metrics instead of four. A tracker fielded quarterly and carried forward between waves, which is how most trackers arrive.</p><p><em>Exhibit: multiplier recovery under different data conditions. True value 0.50. “Usable” means the estimate landed within half of the truth, in either direction.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/822/1*gaMDc6mDWIa-H8NA4-BFOA.png" /></figure><p>Read the top row first, because for a lot of Indian brands that is the row they are in. Five markets in the model, two tracker metrics worth using, quarterly fieldwork. The multiplier comes back at roughly half its true size, and lands within half of the truth 48% of the time.</p><p>A coin flip.</p><p>Now read the second row. Four times as many markets, and the estimate barely moves — 0.26 to 0.27. What changes is the spread, which halves. More markets did not make the answer less wrong. It made you more confident about the wrong answer.</p><p>Then compare the third and fourth rows, which is the finding I did not expect.</p><p>Twenty markets with a quarterly tracker: 0.32. Five markets with a weekly one: 0.39.</p><p><strong>How often you field the tracker matters more than how many markets you model.</strong> Four times the geographic coverage loses to more frequent measurement, and it is not close.</p><p>The mechanism is invisible in practice, which is why nobody catches it. A quarterly tracker carried forward into a weekly model gives you twelve real readings per market, each one repeated thirteen times. The model cannot distinguish a measurement that was held constant from a measurement that was confirmed. So twelve observations arrive dressed as a hundred and fifty-six, and the attenuation from Section 3 bites far harder than it appears to.</p><p>You are not measuring brand strength weekly. You are measuring it twelve times and photocopying the result.</p><p>Which produces the chapter’s one unambiguous instruction, and it is about research budget rather than modelling:</p><p><strong>Field more often, before you do anything else.</strong> Moving from quarterly to monthly or continuous tracking recovers more of the effect than any other change available to you, at any market count.</p><p><strong>Add metrics second.</strong> Two to four items moved the estimate from 0.27 to 0.32 with a quarterly tracker, and from 0.37 to 0.42 with a weekly one. Cheap, since the questions are usually already in the questionnaire.</p><p><strong>Add markets third</strong> — not because coverage is worthless, but because what it buys is precision rather than accuracy. If your estimate is attenuated by a third, twenty markets will tell you so with great confidence.</p><p>Most teams would have ranked those three backwards.</p><p>So the right-hand column of the first exhibit needs qualifying rather than retracting. The multiplier is the only one of the brand’s two jobs that is <strong>identifiable</strong>: no specification choice destroys it, no market control absorbs it, the tracker feedback barely touches it. That is the structural finding and it holds.</p><p>But identifiable is not the same as measured. At the data quality most brands have, the multiplier arrives attenuated by about a third and reliable rather less than three times in four.</p><p>Which leaves the brand’s floor contribution unmeasurable from this data, and its multiplier measurable in principle and shaky in practice.</p><p>Both of which would be easier to accept if the two numbers were similar in size.</p><h3>Section 5: The Shape of the Instrument</h3><p>Here is the point, stated first: the brand’s measurable contribution is roughly a ninth the size of its unmeasurable one.</p><p>I have been careful so far not to say which of the brand’s two jobs is bigger, because the comparison is the chapter. So here it is, from the same invented business, for a market one standard deviation stronger in brand strength:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/710/1*LQVaEvHo9kLW6e3uLmKLnQ.png" /></figure><p>Media pressure in that business would have to run about nine times its average weight before the multiplier route caught up with the floor.</p><p>Read that against the last two sections. The floor is worth nearly nine times the multiplier, and the floor is the one that cannot be estimated — it comes back anywhere between 0.14 and 3.51 depending on a specification choice. The multiplier is worth 1.7% of sales, and it is the one that survives every specification you can throw at it.</p><p><strong>The thing you can measure is not the thing that matters most. It is the thing that happens to be visible.</strong></p><p>Which creates a trap that is almost impossible to resist, because it looks like rigour. A team that has read Sections 2 and 4 carefully might reasonably conclude: the floor estimate is unstable, so let us stop quoting it and build our brand case on the multiplier instead, which is defensible.</p><p>That is looking for your keys under the streetlight. The multiplier is a real number, honestly obtained, and it accounts for about a tenth of what the brand is contributing. A brand case built on it is a brand case that loses, and loses while appearing to have the better evidence.</p><p>Now the second blindness, which runs the same way.</p><p>Brand investment does not pay back inside the quarter. It raises a stock that decays slowly, and that stock keeps producing after the campaign is over. So the question is not whether the return exists, but whether your measurement window was open when it arrived.</p><p><em>Exhibit: share of a brand investment’s eventual value that falls inside the measurement window.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/958/1*YElTYO6pk_mY2r-bsggODg.png" /></figure><p>This part is arithmetic, and I want to be precise about which part. <strong>If</strong> a return accrues after the window closes, then a window shorter than the accrual period captures only a fraction of it, and that fraction is computable from the decay rate. There is nothing to argue about there. A twelve-week read on a slow-building asset sees 11% of what it bought, and no modelling choice inside those twelve weeks changes that.</p><p>What is not arithmetic is the row you are on. Whether brand investment in your category builds a slow-decaying stock or a fast one is an empirical question about your brand, and the table answers a more useful question than the one people usually ask of it. Not “what is the ROI.” But: <strong>given how long our brand effects persist, what measurement window would we need before the number means anything?</strong></p><p>Find your row, read across, and you have the answer. For most brands that answer is uncomfortably long, which is why nobody computes it.</p><p>And notice the shape of the thing. On the bottom rows, a twelve-week window is nearly perfect. A fast-decaying, quickly-consumed effect is fully visible almost immediately. The measurement system works beautifully — for the kind of marketing that barely lasts.</p><p>So stack up the forces this chapter has found, because there are four and they are all mechanisms, not opinions.</p><p>The floor hides between markets, where the standard controls absorb it. The long-run return hides after the window, where no model is looking. Measurement noise shrinks whatever is left, toward zero rather than in a random direction. And the tracker that is supposed to carry the signal arrives four times a year and gets photocopied into fifty-two weeks.</p><p>Four forces. Every one of them biased the same way. Not one of them anybody’s mistake.</p><p>Now add the magnitude finding on top, because it is what makes the four matter. The part of the brand’s value that survives all four is worth about a ninth of the part that does not.</p><p>And two of those four get worse as the brand gets better.</p><p>A more established brand moves less week to week. Its awareness is already high, its preference already settled, and so the genuine movement in its tracker is small relative to the sampling noise sitting on top. More of what the instrument sees is static.</p><p>And if a better-built brand decays more slowly — which is plausible, though it is a claim about your category that you would have to check — then more of its return falls outside whatever window you are measuring in.</p><p>Meanwhile the absolute stakes rise with strength. A stronger brand has a larger floor, so the amount of value sitting in the unmeasurable place is larger in rupees even when its share is unchanged.</p><p><strong>The better the asset, the less of it your measurement system can see.</strong></p><p>What to do about it, in the order the evidence supports.</p><p>Take the multiplier from the data you already own, with the research fixes from Section 4 — field more often, use more items. It is a small number but it is honestly yours, and it establishes something important in principle: that the brand changes what media is worth, which means brand and performance are not competing for the same rupee in the way the budget meeting assumes.</p><p>Get the floor contribution from a different class of evidence altogether, and stop expecting the weekly model to produce it. A cross-market comparison with its assumption written down in plain words. Long-run history, where you have fifteen years and can watch what happened when the brand was weaker. Designed variation, where the stakes justify it. Any of those is better than a specification choice made silently inside a model nobody reads.</p><p>And whichever you use, <strong>state the assumption where the number is quoted.</strong> The factor of twenty-five in Section 4 did not come from bad analysis. It came from a defensible choice made invisibly. A floor estimate presented without its assumption is not a measurement; it is a modelling preference with a rupee sign in front of it.</p><p>Then, for reading the business rather than estimating it, there is the base trajectory map from earlier in the book — brand health on one axis, sales on the other. Rising brand health with flat sales may be a Growth Engine loading. Flat brand health with rising sales may be Borrowing From the Future.</p><p>May be. The quadrant is a hypothesis, not a diagnosis, and what turns one into the other is the ruling-out list: distribution, price, promotional depth, competitive weight, category growth. Rule those out and the brand reading means something. Skip them and you have a quadrant with a label on it.</p><p><strong>A measurement system that cannot see something will report that it is not there, and it will report it with confidence.</strong> That is why brand investment loses the argument. Not because the case is weak, but because the instrument’s silence gets entered into the record as evidence.</p><p>The next chapter takes up a consequence of all this. If the brand changes what media is worth, then the response curves from Part II are not fixed objects sitting in the data waiting to be found. They move with the asset underneath them — which makes turning a curve into a return rather more complicated than reading a number off it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b433bbe76d6e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[It Passed the Test. Is It Still Working?]]></title>
            <link>https://medium.com/@mail2rajivgopinath/it-passed-the-test-is-it-still-working-5e7a74f071a7?source=rss-ae007d129200------2</link>
            <guid isPermaLink="false">https://medium.com/p/5e7a74f071a7</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agents]]></category>
            <dc:creator><![CDATA[Rajiv Gopinath ]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 14:32:39 GMT</pubDate>
            <atom:updated>2026-11-09T14:32:39.057Z</atom:updated>
            <content:encoded><![CDATA[<h4>THE PLUMBING PROBLEM · PART 8</h4><h4>Why enterprise AI needs continuous assurance, not launch-time validation</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4ghCu7gJCbWb5EpU42XusA.png" /></figure><p>The agent passed everything before launch.</p><p>The workflows ran correctly. The tool calls worked. The permissions were checked. The answers looked sensible. The edge cases had been tested. Somebody signed off.</p><p>Then it went live.</p><p>Three months later, it is still running.</p><p>Nothing obvious has failed.</p><p>But the retrieval corpus has changed. The agent has accumulated more memory. One of the APIs it depends on has changed its behaviour. Another team has widened its permissions. Users are asking it different kinds of questions. The underlying model may have been updated. A workflow that used to end with a recommendation may now end with an action.</p><p>Individually, none of these changes may look dramatic.</p><p>Together, they matter.</p><p>Because the system operating today is no longer quite the system that was evaluated before launch.</p><p>That is an easy thing to miss.</p><p>We tend to think of testing as something that happens before deployment. Build the system, evaluate it, fix the obvious problems, approve it, move on.</p><p>That works reasonably well when the thing being tested remains mostly the same.</p><p>Agentic systems do not.</p><p>They sit inside a moving environment. Models change. Tools change. APIs change. data changes. permissions change. Memory changes. User behaviour changes. And sometimes the system itself begins to behave differently simply because all of those pieces are now interacting in a different way.</p><p>The result is a strange kind of confidence.</p><p>The agent may still be live.</p><p>The dashboards may still be green.</p><p>The benchmark it passed may still be sitting in a folder somewhere.</p><p>But the evidence that justified putting it into production may no longer describe the system that is actually running.</p><p>I think of that distance as <strong>The Assurance Gap</strong>: the distance between the evidence that justified an agent’s operation and the system that is actually operating now.</p><p>The question is no longer only:</p><p>Did we test this before it went live?</p><p>It is:</p><p><strong>What has changed since the evidence was produced, and which claims did those changes invalidate?</strong></p><p>Because the uncomfortable truth is this:</p><p><strong>The system passed the test. But the system that passed the test no longer exists.</strong></p><h3>1. The system that passed the test no longer exists</h3><p>Imagine a campaign optimisation agent that has been through a proper launch process.</p><p>Its prompts have been tested. Its retrieval sources have been checked. The platform APIs work. The permissions are correct. Somebody has run the scenarios, looked at the outputs and decided that the system is ready.</p><p>That decision may be perfectly reasonable.</p><p>The problem begins afterwards.</p><p>Six months later, the agent is still called by the same name. It still appears in the same interface. It may even be running the same model.</p><p>But underneath, things have moved.</p><p>The media platform has changed an API. A new data source has been added. The retrieval corpus is larger. The memory now contains months of previous interactions. Someone has changed a system instruction. A permission that was originally read-only now allows an update. Users have discovered uses that nobody included in the original test set.</p><p>None of this needs to look like a failure.</p><p>That is what makes the problem interesting.</p><p>We are accustomed to thinking about systems as versions. Version 1 was tested. Version 1 went live. Version 1 is still running.</p><p>But an agent is increasingly less like a static piece of software and more like a system sitting inside a changing set of relationships. Research on agent validation describes exactly this temporal problem: behaviour can evolve as memory, workflows and operating conditions change, which means correctness cannot simply be treated as something established before deployment.</p><p>And the change need not come from the model itself.</p><p>The surrounding system can move while the model stays still.</p><p>That creates a peculiar situation. The evidence may still be valid for the system that existed on the day of approval, while becoming progressively less relevant to the system actually operating today.</p><p>This is the beginning of the <strong>Assurance Gap</strong>.</p><p>Not because the original testing was poor.</p><p>Not because somebody failed to do the right checks.</p><p>Simply because the object being assured did not stay still.</p><p><strong>The system passed the test. But the system that passed the test no longer exists.</strong></p><h3>2. Evaluation isn’t assurance</h3><p>Most organisations still think about evaluation as a gate.</p><p>You build the system, test it, run the scenarios, fix what fails, get the necessary approval and put it into production. There is nothing wrong with that. The problem is that the logic quietly assumes that the evidence produced during evaluation will remain relevant after deployment.</p><p>For agentic systems, that assumption becomes weaker over time.</p><p>This distinction between evaluation and assurance is not entirely new. Researchers working on self-adaptive software have been dealing with a version of the same problem for years. The <strong>ENTRUST</strong> framework, for example, was built around the idea that systems operating in changing environments cannot rely only on design-time evidence. It combines evidence produced before deployment with new evidence generated at runtime as the system adapts and the environment changes. The assurance case therefore has to evolve with the system rather than remain frozen at the point of launch.</p><p>That is a useful way to think about enterprise agents as well.</p><p>Evaluation tells us whether a system performed acceptably under a particular set of conditions. Assurance asks whether the claims we made on the basis of that evaluation are still justified under the conditions that exist now.</p><p>Those claims may be quite specific. The agent uses only approved tools. It stays within a defined permission boundary. It escalates certain decisions to a human. It does not expose sensitive information. It completes a workflow within an acceptable cost and latency. It does not take an external action without the required approval.</p><p>At launch, we may have evidence for all of them.</p><p>The problem is that the evidence and the claims do not necessarily age at the same rate.</p><p>One of the more interesting recent papers in our research pack makes this point explicitly. Its argument is that continuous assurance should not mean endlessly rerunning the same benchmark. Instead, evidence should be renewed when a relevant change occurs, and where the evidence supporting an action can no longer be kept current, the authorization for that action should be narrowed or suspended rather than allowed to rest on an old assessment.</p><p>That is a much more useful idea than simply saying that agents need more monitoring.</p><p>MLOps teams already monitor systems for changes in data distributions, model performance, latency and other operating signals. Those things remain important. But assurance is asking a different question.</p><p>A drift monitor may tell us that something has changed.</p><p>Assurance asks whether that change has made one of our earlier claims about the system no longer defensible.</p><p>The distinction matters because an agent can continue to look perfectly healthy while the basis on which we trusted it has quietly weakened.</p><p><strong>Monitoring asks whether the system is changing. Assurance asks whether the claims that justify letting it act are still true.</strong></p><h3>3. Where drift lives</h3><p>One reason assurance is difficult is that drift does not have a single home.</p><p>Sometimes the model changes. That is the obvious case. But often the more interesting changes happen elsewhere.</p><p>A prompt remains the same, the tool inventory looks unchanged, the permissions are still valid and the memory store contains nothing obviously problematic. Yet the behaviour of the overall system begins to shift because those parts are interacting differently from before.</p><p>Krti Tallam describes this as an <strong>inspection-change mismatch</strong>. The individual surfaces of an agentic system can each pass inspection while the meaningful change sits in the relationship between memory, runtime context, tool affordances, authorization boundaries and accumulated interaction history. In her example, a long-running support agent can become more defensive and more eager to escalate even though its self-description, memory, permissions and sampled outputs all continue to look acceptable in isolation.</p><p>That is an important distinction because it means we cannot always find drift by asking which component changed.</p><p>Sometimes nothing changed enough on its own.</p><p><strong>Drift does not always live inside the model. Sometimes it lives between the parts.</strong></p><p>There is another kind of drift that is easier to miss because the system can remain technically healthy throughout.</p><p>Research on <strong>goal drift in language-model agents</strong> looks at what happens when agents operate over long horizons while being exposed to competing objectives. In controlled experiments, all of the evaluated agents showed some degree of goal drift, with differences in how well they held to their original objective as the interaction continued.</p><p>The agent can still call the right tools. It can still complete tasks. Nothing has crashed. Yet the way it interprets what it is trying to achieve has moved.</p><p><strong>A system can remain operational while becoming less aligned.</strong></p><p>And then there is drift in the route itself.</p><p>An agent may still arrive at an acceptable final answer while getting there in a worse way. It may make unnecessary tool calls, take longer paths, rely on weaker evidence, ask another agent to repair mistakes that it previously avoided, or make intermediate decisions that would be unacceptable if they were examined on their own.</p><p>That is why the trajectory-evaluation literature increasingly treats the sequence of reasoning, tool use and observations as part of the thing being evaluated, rather than looking only at the final result. ToolPRMBench, for example, breaks full agent trajectories into step-level cases precisely because important failures can occur inside the path even when the overall task is judged only at the end.</p><p>For agents, then, drift can appear in three places at once: in the relationships between components, in the behaviour of the agent, and in the path it takes through the task.</p><p>That makes assurance harder than simply asking whether the model’s accuracy has gone up or down.</p><p>The system may still be producing good-looking answers.</p><p>The question is whether it is still producing them in ways we would be willing to approve.</p><h3>4. Change should expire the evidence</h3><p>If assurance is going to mean anything in practice, it has to react to change.</p><p>That does not mean every small modification should trigger a full re-certification of the system. That would be expensive, slow and quickly ignored.</p><p>The more useful question is narrower:</p><p><strong>Which assurance claims did this change affect?</strong></p><p>Suppose a new model version is introduced. That may put some claims back in doubt: reasoning quality, instruction-following, safety behaviour, tool selection. A changed API may threaten a different set of claims, especially around whether the same tool call still produces the same effect. Expanded permissions may create an entirely new risk surface. A retrieval update may need only targeted checks around freshness, source quality or policy compliance.</p><p>The point is not that everything must always be retested.</p><p>The point is that evidence has a scope.</p><p>And when the system moves outside that scope, the evidence should stop being treated as current.</p><p>One of the strongest papers in the continuous-assurance research pack makes exactly this move. It argues that assurance should be renewed when relevant changes occur, and that when assurance-invalidating change happens faster than the supporting evidence can be refreshed, the affected action path should be narrowed or suspended rather than left authorized on stale evidence.</p><p>That gives us a much more operational way to think about assurance:</p><p><strong>change → identify affected claims → rerun relevant evidence → renew or restrict authorization</strong></p><p>This is a better model than simply setting a calendar reminder to “re-evaluate the agent” every quarter.</p><p>It also forces the organisation to be explicit about what it is actually assuring. Are we assuring that the agent answers accurately? That it stays within policy? That it uses only approved tools? That it escalates high-risk decisions? That it remains within a cost threshold? That it cannot take an irreversible action without review?</p><p>Different changes put different claims at risk.</p><p>And different claims deserve different levels of scrutiny.</p><p>A read-only internal assistant does not need the same evidence threshold as an agent that can approve a payment, change a customer record, send an external message or trigger an operational workflow. The more consequential the action, the less tolerance there should be for stale evidence.</p><p>So the principle is not:</p><p><strong>rerun everything every day.</strong></p><p>It is:</p><p><strong>When the system changes, renew the evidence that the change puts in doubt.</strong></p><p>That is what turns continuous assurance from a vague aspiration into an operating discipline.</p><h3>5. Production feeds the tests</h3><p>This is where Part 8 connects directly back to Part 7.</p><p>Observability gave us the traces, state changes, lineage and evidence needed to reconstruct what actually happened in production. Assurance should now use that evidence to improve the tests themselves.</p><p>That sounds obvious, but it changes the role of production data. Instead of treating incidents, near misses and human interventions as isolated operational events, they become new evidence about where the system is brittle.</p><p>A tool path nobody expected becomes a new scenario. A repeated human override becomes a signal that the autonomy boundary may be too wide. A failure that only appears after several handoffs becomes a new trajectory test. A policy breach that was technically recovered from still becomes a regression case, because successful recovery does not make the original failure acceptable.</p><p>Software engineering has worked this way for years. Regression testing exists because a system that worked yesterday can stop working after it changes, and because old failures have a habit of returning in new forms. In continuous integration environments, teams selectively rerun the tests most likely to be affected by a change and prioritise them so that failures surface quickly.</p><p>Agentic systems need the same discipline, but the object being regression-tested is broader.</p><p>We are not only asking whether the final answer is still correct. We may need to test whether the agent still chooses the right tool, escalates at the right point, respects the same permission boundary, follows an acceptable trajectory, stays within policy, avoids unintended side effects, remains within expected cost and latency, and recovers correctly when something underneath it fails.</p><p>That is also why trajectory-level evaluation matters. The research here is increasingly explicit that the final result alone can hide weaker intermediate behaviour, which means a regression suite for agents has to preserve some of the path, not just the destination.</p><p>The feedback loop then becomes fairly simple:</p><p><strong>production behaviour → evidence → new tests → renewed assurance</strong></p><p>And that gives us the cleanest connection between the last two parts of this series.</p><p><strong>Observability tells us what happened. Assurance decides whether what happened is still acceptable.</strong></p><p>The better the first becomes, the stronger the second can become as well.</p><h3>6. Who signs the renewal?</h3><p>At some point, assurance stops being a technical question and becomes a management one.</p><p>A useful analogy comes from aviation.</p><p>An aircraft is not considered safe forever because it passed certification once. Its continued permission to fly depends on maintenance, inspections, service history, known defects, directives, and evidence that the aircraft remains fit for operation as conditions change.</p><p>The point is not that enterprise agents need aviation-style regulation.</p><p>The point is that the underlying logic is sound.</p><p><strong>Permission to operate should depend on current evidence, not historical confidence.</strong></p><p>That creates a responsibility problem that many organisations have not really solved yet.</p><p>Who owns the continuing decision that this agent should still be allowed to act?</p><p>The product team may own the workflow. The model team may own the underlying model. Security may own permissions. Data teams may own retrieval sources. Business teams may own the outcome. Operations may be the first to see failures.</p><p>But none of those things automatically answer the question of who owns continued authorization.</p><p>And that question becomes more important as the agent becomes more consequential.</p><p>A read-only assistant can probably tolerate a lighter assurance regime. An agent that can approve, transact, change records or communicate externally should not continue operating simply because nobody has found a reason to stop it.</p><p>It should continue because somebody can still defend the claims that justify its operation.</p><p>So the management questions can stay very simple:</p><p><strong>What changed since we last proved this system was fit to operate?</strong></p><p>And:</p><p><strong>Who owns the decision to keep it running?</strong></p><p>Those are not compliance questions.</p><p>They are operating questions.</p><h3>Closing — The test was never the end</h3><p>We tend to think of evaluation as the moment when confidence is created.</p><p>But for agentic systems, confidence decays.</p><p>The system changes. The environment changes. Tools change. Permissions change. Memory accumulates. New behaviours appear. Old assumptions become less certain.</p><p>That does not mean every change is dangerous.</p><p>It means the evidence has to keep up.</p><p>This is the real shift from evaluation to assurance.</p><p>The question is no longer whether the agent passed the test once.</p><p>It is whether the evidence that justified letting it operate still matches the system that is operating now.</p><p>That is why the loop matters:</p><p><strong>tested once → system changes → assurance claims become stale → evidence is selectively renewed → production experience strengthens the test suite → someone owns continued authorization</strong></p><p>And that is where Part 8 lands.</p><p>We tend to ask whether an agent passed evaluation before it went live.</p><p>Enterprise systems need another question.</p><p><strong>When did we last prove that it still should be live?</strong></p><p><strong>Selected References</strong></p><p><strong>Zhou, Jianlong; Zhu, Zhiyu; Jin, Zhibo (2026). “Agentic AI Risk Evaluation Should Become a Continuous Assurance Discipline.”</strong> Preprints.org</p><p><strong>Mirto, Fabio Orazio; D’Agati, Luca; Tricomi, Giuseppe; Silvestri, Stefano; Longo, Francesco; Puliafito, Antonio; Merlino, Giovanni (2026). “Beyond Component Testing: Validating Agentic AI Systems.”</strong> arXiv:2607.29405v2.</p><p><strong>Arike, Rauno; Donoway, Elizabeth; Bartsch, Henning; Hobbhahn, Marius (2025). “Evaluating Goal Drift in Language Model Agents.”</strong> <em>Proceedings of the Eighth AAAI/ACM Conference on AI, Ethics, and Society (AIES 2025).</em></p><p><strong>Calinescu, Radu; Weyns, Danny; Gerasimou, Simos; Iftikhar, M. Usman; Habli, Ibrahim; Kelly, Tim (2018). “Engineering Trustworthy Self-Adaptive Software with Dynamic Assurance Cases.”</strong> <em>IEEE Transactions on Software Engineering</em>, 44(11), 1039–1069. DOI: 10.1109/TSE.2017.2738640</p><p><strong>Tallam, Krti (2026). “Inspection-Change Mismatch in Persistent Agent Ecosystems: Why Decomposition-Based Governance Underdetects Drift.”</strong> Short paper, <em>ALIFE 2026 Workshop: Agentic and Generative AI as Artificial Life</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5e7a74f071a7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Show Me the Work]]></title>
            <link>https://medium.com/@mail2rajivgopinath/show-me-the-work-ee51d574f7f8?source=rss-ae007d129200------2</link>
            <guid isPermaLink="false">https://medium.com/p/ee51d574f7f8</guid>
            <category><![CDATA[branding]]></category>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[authenticity]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Rajiv Gopinath ]]></dc:creator>
            <pubDate>Sun, 08 Nov 2026 03:06:26 GMT</pubDate>
            <atom:updated>2026-11-08T03:21:56.245Z</atom:updated>
            <content:encoded><![CDATA[<h4>Perspectives</h4><h4>Why brands may need to prove something was made, not just show what they made</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pW-rBmkaHLX-VmnuHA4D3g.png" /></figure><p>This holiday season, Levi’s released a one-minute film called <em>Let’s Get Fuzzy</em>. It is stop-motion, made with handmade models and shot one frame at a time with the studio Blinkink. It reportedly took about 2,000 hours, around four hours of filming for every second on screen. Levi’s says no AI was used.</p><p>Two thousand hours for sixty seconds is, by any production logic, inefficient. Whatever one minute of polished animation costs to produce with AI today, the cost is falling every month.</p><p>That is exactly why Levi’s spent the hours.</p><p>For most of advertising’s history, a polished finish carried a hidden message. A perfectly lit image, a seamless animation or a glossy campaign film implied time, skill, budget and care. Audiences didn’t need to be told. The finish said it for them.</p><p>AI keeps the finish and removes the message. The polish is still there, but it no longer tells anyone how much effort went in. When polished output becomes cheap, visible effort becomes scarce.</p><p>Levi’s is not alone. The same week, e.l.f. released a holiday film starring puppets wearing real makeup. Prada launched <em>Drawing Prada: Animated By Hand</em>, frame-by-frame animation by John Pomeroy, a Disney veteran, which the brand describes as made with “old-school techniques.” Dior released a seven-minute hand-drawn short by the animator Sylvain Chomet.</p><p>This looks like a craft revival. It is something more specific: brands deliberately buying back the signal that AI made cheap.</p><h3>Why effort persuades</h3><p>The reason this matters is that people do not judge quality directly. They infer it, and effort is one of the main cues they use.</p><p>The most direct evidence comes from a 2023 study in <em>Scientific Reports</em> by Horton, White and Iyengar. Across six experiments with almost 3,000 participants, people rated identical artworks. The only difference was the label: made by a human, or made by AI. AI-labelled work was rated less creative, less skilled and less valuable.</p><p>The more important result is what the label did to perceived effort. In one experiment, the AI label halved the production time people assumed had gone into the work, from about 36 hours to about 18. The value people assigned fell by about a quarter, from roughly $148 to $112.</p><p>An AI label doesn’t only lower the judgement of quality. It erases the effort people assume went into the work, and the value goes with it.</p><p>Why does assumed effort carry so much weight? Twenty years earlier, Justin Kruger and colleagues showed how strong the shortcut is. They gave people a poem and told them it had taken either 4 hours or 18 hours to write. The poem was the same, but the “18-hour” version was liked more and valued higher. They repeated the test with two paintings described as taking 4 or 26 hours, swapping which painting got which label. People preferred whichever one was described as taking longer. Self-described experts were swayed as much as novices. In a third study, about medieval armour, the effect was strongest when the images were blurry, that is, when quality was hardest to judge.</p><p>Kruger called it the effort heuristic. A 2023 replication produced mixed results, so the size of the effect is contested. The idea has held up for two decades.</p><p>The two studies fit together:</p><blockquote><strong><em>People can’t see quality, so they infer it from effort. AI makes the effort impossible to infer.</em></strong></blockquote><p>Effort is not the only thing the label takes away. A 2025 meta-analysis from Tilburg University pooled roughly 60 studies of bias against AI art and found that the bias is small when people judge how a work looks, and largest when they judge what it means. People don’t think AI work looks worse. They think it means less, because there is no longer anyone behind it meaning anything.</p><p>The penalty also attaches to belief, not to process. In a Stanford and Reed College study, 399 people listened to identical music. Tracks labelled AI-generated scored 23% lower on emotional resonance, and people listened to them 19% less. The penalty held in full even when the music was human-made and only labelled as AI. Audiences cannot see how work was actually made. They judge what they believe about how it was made.</p><p>Advertising sits exactly where these effects are strongest. Nobody can inspect a brand film the way they inspect a product. Audiences judge from cues, and effort and intention have always been two of them.</p><p>Marketing has a theory for this too. Amotz Zahavi’s handicap principle explains why a peacock’s tail is credible: it is expensive to grow and impossible to fake cheaply. Advertising borrowed the idea as costly signalling. Tim Ambler and Hollier put it most bluntly in the title of a 2004 paper: the waste in advertising is the part that works.</p><p>There is a limit, though. In 1989, Amna Kirmani and Peter Wright told 214 people about a sports-shoe launch backed by a campaign of $2 million, $10 million, $20 million or $40 million. Expected product quality rose with the spend up to $20 million, then dipped at $40 million. Past a point, visible cost stopped reading as confidence and started reading as excess.</p><p>Effort persuades up to a point. Beyond it, it starts to look like performance.</p><h3>Labour or scarcity?</h3><p>This leads to the uncomfortable question in the whole argument. Do people value effort because they value human labour, or because human effort is suddenly rare?</p><p>The evidence says both.</p><p>Kruger’s findings come from before generative AI. People rewarded effort when nothing like AI existed to compare it with. That suggests something lasting: people care about work being hard.</p><p>The same <em>Scientific Reports</em> study also found a scarcity effect. When people saw a human-made painting immediately after an AI-made one, they rated the human work as more creative than people who saw the same painting on its own. Human effort gained value simply by being next to machine output.</p><p>So part of the value of effort is lasting, and part of it is comparative. The distinction matters, because the two parts behave differently over time.</p><p>The comparative part wears out quickly. If every brand runs stop-motion for Christmas 2027, stop-motion becomes another convention, and no one will be impressed that a brand used clay.</p><p>The lasting part only works if people believe the effort is real. That condition is becoming harder to meet.</p><p>Consumer attitudes add a further twist. In a Canva and Harris Poll survey of 3,547 consumers across seven countries, including India, 78% said they would rather see ads made by people, even if AI could make better ones. Canva sells AI design tools, so the finding cuts against its own commercial interest, which makes it more credible. Yet the same study found that 68% don’t mind AI in advertising if it makes the ads more helpful or relevant.</p><p>The generational picture is stranger still. In the Bentley University–Gallup 2026 survey, 66% of 18- to 29-year-olds viewed businesses’ use of AI in ad creation negatively, against 38% of those aged 45 to 59. Among Gen Z, 73% called AI-created people or voices in ads unacceptable. But in the Canva study, 70% of Gen Z and Millennials said they focus more on an ad’s vibe than on how it was made.</p><p>The audience most hostile to AI advertising in principle is the least concerned with provenance in practice. Young people object to AI in the abstract, and judge each individual ad on how it feels.The penalty is not universal</p><h3><strong>The Penalty is not Universal</strong></h3><p>The AI penalty also varies sharply by country.</p><p>In Ipsos’s 2025 survey across 30 countries, excitement about AI-powered products was highest in Indonesia (80%) and Thailand (79%), and well above average in India (65%) and Mexico (65%). It was lowest in Sweden (34%), the UK (37%) and the US (38%). Trust that AI would not discriminate followed a similar pattern: 67% in India and 73% in Mexico, against 37% in the US and 28% in Sweden.</p><p>Other studies show the same gradient. KPMG and the University of Melbourne surveyed 48,000 people in 47 countries: 76% of Indians said they were willing to trust AI, against a global average of 46%, and against 57% in emerging economies and 39% in advanced ones. Edelman found trust in AI at 87% in China and 67% in Brazil, against 39% in Germany, 36% in the UK and 32% in the US.</p><p>Latin America doesn’t move as one region. Mexico, Peru and Colombia look much like Asia. Argentina shows low excitement and Chile leans nervous. Brazil is contradictory: its trust scores are above average, yet Pew found about half of Brazilians more concerned than excited about AI, among the most concerned of the 25 countries it surveyed.</p><p>The pattern is roughly a gradient. Southeast Asia, India and China are the most accepting. Latin America is mixed. Continental Europe is sceptical. The English-speaking West and the Nordics are the most anxious. Japan stands apart, with low excitement and low anxiety, which looks more like indifference.</p><p>The reasons are partly a matter of interpretation. In India, mobile phones, digital payments and digital identity visibly widened access, so technology has a track record of creating opportunity. In the West, the recent story is more about technology concentrating power, and the loudest AI anxiety comes from creative and knowledge workers, the people who make and comment on advertising. The Western idea of authorship, the individual genius behind the work, also makes “who made it” central to how value is judged.</p><p>Three cautions apply before concluding that the penalty doesn’t matter in Mumbai.</p><ul><li><strong>These are online samples.</strong> In India they skew urban and affluent, though that is precisely the audience premium brands target.</li><li><strong>High trust comes with anxiety.</strong> The same KPMG survey found 78% of Indians concerned about AI’s negative outcomes, and Ipsos found Indian excitement and nervousness both at 65%. That is ambivalence, not indifference.</li><li><strong>What people say isn’t always what they feel.</strong> One study compared American and Chinese responses to AI-generated art. Americans were negative both in what they said and in their implicit, unconscious reactions. Chinese participants praised AI art openly, but their implicit appreciation still favoured human work.</li></ul><p>The proof-of-effort problem is loudest in London and New York. In Mumbai it may not be absent, just quieter.</p><h3>The look stops working</h3><p>There is a second-order problem, and it is the one that changes the strategy.</p><p>AI can already generate a convincing claymation look. It can imitate stop-motion, hand-drawn animation, the visible brushstroke and the slightly imperfect felt texture. Everything that makes work look handmade can now be produced without anyone making anything by hand.</p><p>So the look of effort is heading the same way polish went. Once it can be produced cheaply, it stops proving anything. The Canva study found that 70% of consumers expect AI ads will eventually be impossible to tell apart from human-made ads without disclosure, and 56% expect that within two to five years.</p><p>The signal therefore moves again, from “this looks handmade” to “show me that it was handmade.”</p><p>That is why behind-the-scenes material suddenly matters: sketches, rehearsals, models, mistakes, hands and materials. Work in progress becomes part of the communication. The production process becomes evidence.</p><p>Apple understood this early. When it rebuilt the Apple TV intro in 2025 using real glass and practical effects, it released a making-of video alongside it. When its 2025 holiday film used puppets, the behind-the-scenes footage came out with the ad. The proof was published together with the work.</p><h3>6. When effort is the message</h3><p>Visible effort doesn’t work for every brand. It works when three conditions hold.</p><p><strong>The effort has to be visible.</strong> Effort nobody sees adds nothing. Two thousand hours that leave no trace look identical to two hours.</p><p><strong>The effort has to be credible.</strong> The audience has to believe it happened, which increasingly means showing evidence, not just a finished look.</p><p><strong>The effort has to be relevant.</strong> This is the condition brands most often miss. Effort persuades best when the way the work was made says something about the product. Levi’s chose felt, miniature garments and handmade sets for a campaign about holiday knitwear, so the production method and the product reinforce each other. The brand’s VP of global marketing, Anne Cambria, made the link explicit: “Our campaigns always marry authenticity and ensuring a clear connection back to product.”</p><p>A fintech brand doing stop-motion to look “human” fails this test. The effort is visible and perhaps credible, but it says nothing about the product, so it reads as an affectation.</p><p>The kind of work matters as well. In a 2025 <em>Journal of Business Research</em> study, Kirk and Givi found that AI authorship damaged emotional marketing messages far more than factual ones. People who believed a heartfelt note had been machine-written felt something close to moral revulsion: they recommended the brand less and considered switching. A product demo or a specification comparison made with AI costs a brand very little. A holiday film, a brand love story or anything that says “we care” invites the question of who actually meant it.</p><p>Luxury has always worked like this. Hand stitching, ageing, limited runs and difficult materials have long been part of what a luxury product is. AI now makes the same logic apply to communication. One marker of premium creative work may simply be that someone actually spent the time making it.</p><h3>7. Even shown work can be doubted</h3><p>Which brings us to Dior.</p><p><em>Christian Dior’s Night in Color</em> is a seven-minute hand-drawn film starring Robert Pattinson and Rihanna, made with Sylvain Chomet, one of the most respected names in hand-drawn animation. It is exactly the kind of work this argument says brands should make. Dior even exhibited sketches from its production at its permanent exhibition space in Paris.</p><p>Within days of release, viewers accused the film of being AI-generated. They pointed to warped lettering, an odd-looking eye and a finger joint in the wrong place. Critics noted that the behind-the-scenes material showed only sketches, and that the credits named animators without addressing AI.</p><p>The accusation is unproven. Dior may have done everything right. That is the point. A brand made a hand-drawn film with a renowned animator, showed the sketches, and its effort was still doubted.</p><p>It is also possible that Dior worked the way most serious creative work now gets made: people deciding the story, the characters and every judgement, with AI speeding up parts of the execution. That is a legitimate way to make things. The problem would not be the use of AI. It would be leaving the audience to guess who decided what. Audiences can’t see a collaborative process. They see a label or a suspicion, and fill the gap with the least generous assumption.</p><p>So “show me the work” is really “show me who decided.” The standard of proof has risen faster than anyone expected. Effort that looks real is no longer enough, and showing some evidence is not enough either.</p><p>There is a further step coming. If AI can generate behind-the-scenes footage too, showing the process won’t settle doubt for long. Proof of effort may have to become verifiable: production records, provenance data, content credentials that certify what was made by hand and what wasn’t. The demand starts as “show me the work” and ends as “prove it.”</p><h3>8. How everyday brands prove the work</h3><p>Fashion houses have an advantage here. Craft is already their product, so showing it costs them little. A biscuit, detergent or appliance brand has no atelier to film and no hand stitching to show. If even shown work can be doubted, what is left for brands like these?</p><p><strong>Not mainly technology.</strong> The serious technical answer is content credentials, an industry standard (C2PA) in which a camera or editing tool cryptographically signs a file at capture and records every edit after that. Some cameras and phones now support it, and major editing software writes it. Blockchain adds little: the signing doesn’t need it, and a public timestamp only matters if a dispute ever reaches that stage.</p><p>Content credentials have two limits. They prove that a piece of footage is genuine, not that real effort went into the work. And almost no viewer ever opens them. So treat them as insurance, not as the message. Sign your production assets, keep project files and edit histories, and you’ll have evidence ready when the accusation comes.</p><p>For everyday brands, the better proof is effort that is expensive to fake and that people can witness themselves. Seven moves follow from that.</p><p><strong>Show effort in the product, not in the ad.</strong> Most everyday brands shouldn’t make a 2,000-hour film. Their real effort is somewhere else: in sourcing, testing, years of reformulation and the people on the factory floor. Effort that went into the product is far harder to fake than effort that went into a commercial, and it passes the relevance test automatically.</p><p><strong>Prefer live to recorded.</strong> AI can fake a recording. Faking a live event, a visit or a physical object is much harder. Livestream the making. Let people visit the set or the factory. Put the actual props in stores. A shop window full of Levi’s felt figures would prove more than any footage could.</p><p><strong>Name the makers.</strong> “Our team” proves nothing. Named people with public portfolios, who can be interviewed and followed, prove a great deal, and that kind of proof doesn’t scale for anyone faking it.</p><p><strong>Let others vouch for it.</strong> Journalists on set, creators invited into production, craft guilds and industry bodies are all witnesses who aren’t the brand. Their word carries weight the brand’s own can’t. Certification is starting too: the Authors Guild now offers a “Human Authored” mark for books, and a similar standard for advertising may not be far behind.</p><p><strong>Let the audience take part.</strong> Customers who drew the packaging, voted on the design or appear in the film are witnesses with their own stake. Participation makes work very hard to dismiss as AI, because the people who took part know it wasn’t.</p><p><strong>Disclose authorship, not just AI use.</strong> Say who decided and what AI did: “Story, characters and direction by [named people]; AI used for [specific step].” Kirk and Givi found that messages AI only edited were penalised less than messages AI wrote, so being precise about AI’s role protects the work. A clear credit turns a blended workflow from a liability into a credible claim. Silence leaves the audience to guess, and doubt now spreads in days.</p><p><strong>Don’t over-prove.</strong> Kirmani and Wright’s finding applies here too: past a point, visible effort reads as performance. A brand that makes its effort the headline starts to sound like it is protesting too much. Make the evidence easy to find, not loud.</p><p>How hard a brand needs to work at this depends on where it sells. Proof matters most in the US, the UK and the Nordics, and least where audiences say they welcome AI. Even there, the implicit preference for human work suggests it would be a mistake to skip it.</p><p>Put together, these form a hierarchy of proof. From weakest to strongest: a claim, then footage of the process, then signed content credentials, then third-party witnesses, then evidence that is live, physical or co-created. Fashion houses reach the top of it almost by default. Everyday brands can get there too, through the product, the people and participation, not through a more elaborate ad.</p><p>This is a strange place for an argument about AI to end. AI was supposed to make making things effortless. It may instead make effort the thing brands most need to prove.</p><p>AI may not kill craft. It may force craft to become more visible, and eventually, verifiable.</p><p>Sources:</p><h3>The ads</h3><p><strong>Levi’s, “Let’s Get Fuzzy”</strong></p><ul><li><a href="https://www.adsoftheworld.com/campaigns/let-s-get-fuzzy">Ads of the World campaign page</a></li><li><a href="https://www.contentgrip.com/levis-fuzzy-christmas-horror/">Content Grip: 2,000 build hours</a></li></ul><p><strong>Prada, “Drawing Prada: Animated By Hand”</strong></p><ul><li><a href="https://www.prada.com/at/en/pradasphere/campaigns/2026/holiday.html">Prada’s official Holiday 2026 campaign page</a></li><li><a href="https://www.adsoftheworld.com/campaigns/drawing-prada">Ads of the World campaign page</a></li><li><a href="https://hypebeast.com/2026/10/prada-holiday-2026-campaign-disney-animator-john-pomeroy-info">Hypebeast: Prada’s hand-drawn winter fairytale</a></li></ul><p><strong>Dior, “Christian Dior’s Night in Color”</strong></p><ul><li><a href="https://www.thewrap.com/culture-lifestyle/culture/christian-dior-night-in-color-animated-short-rihanna/">TheWrap: Sylvain Chomet’s Dior short film</a></li><li><a href="https://www.nssmag.com/en/fashion/47553/dior-sylvain-chomet-holiday-animated-film-craftsmanship-ai">NSS Magazine: Dior, Chomet, craftsmanship and AI</a></li><li><a href="https://www.yahoo.com/entertainment/articles/dior-christmas-ad-faces-scathing-150000497.html">Yahoo: the AI allegations</a></li></ul><p><strong>e.l.f.</strong></p><ul><li><a href="https://www.adweek.com/creativity/elf-cosmetics-recruits-alan-cumming-to-narrate-its-nostalgic-holiday-ad-debut/">Adweek: e.l.f. enlists Alan Cumming for its holiday ad debut</a></li></ul><p><strong>Apple</strong></p><ul><li><a href="https://www.macrumors.com/2025/11/06/new-apple-tv-intro-made-with-practical-effects/">MacRumors: the new Apple TV intro was made with practical effects</a></li><li><a href="https://www.macrumors.com/2025/11/27/apple-2025-holiday-ad-commercial/">MacRumors: Apple’s 2025 holiday ad, “A Critter Carol”</a></li></ul><h3>Articles and research consulted</h3><p><strong>Trade coverage</strong></p><ul><li><a href="https://www.glossy.co/fashion/fashion-briefing-as-ai-content-takes-over-brands-embrace-hand-drawn-animation-for-holiday/">Glossy: brands embrace hand-drawn animation for holiday</a></li><li><a href="https://thecurrency.news/articles/245609/the-ad-industrys-antidote-to-ai-pencils-clay-and-stop-motion/">WSJ via The Currency: the ad industry’s antidote to AI</a></li><li><a href="https://www.ypulse.com/newsfeed/2026/11/05/levis-is-betting-human-effort-will-stand-out-in-an-ad-market-flooded-with-ai/">YPulse: Levi’s bets human effort will stand out</a></li></ul><p><strong>Why effort persuades</strong></p><ul><li><a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC10624838">Horton, White &amp; Iyengar (2023), <em>Scientific Reports</em>: bias against AI art</a></li><li><a href="https://en.wikipedia.org/wiki/Effort_heuristic">Effort heuristic: Kruger et al. 2004 and replications (summary)</a></li><li><a href="https://deepblue.lib.umich.edu/items/f5e6e3d7-57bd-483b-adcc-b10d2a62b8be">Kruger et al., “Of great art and untalented artists” (University of Michigan repository)</a></li><li><a href="https://research.tilburguniversity.edu/en/publications/bias-against-artificial-intelligence-in-visual-art-a-meta-analysi-2/">Bias against AI in visual art: a meta-analysis (Tilburg University)</a></li><li><a href="https://aiweekly.co/alerts/stanford-study-ai-music-labels-cut-plays-19">Stanford/Reed study: AI music labels cut plays 19% (AI Weekly)</a></li><li><a href="https://ideas.repec.org/a/eee/jbrese/v186y2025ics0148296324004880.html">Kirk &amp; Givi (2025): the AI-authorship effect (<em>Journal of Business Research</em>)</a></li><li><a href="https://www.newswise.com/articles/do-customers-perceive-ai-written-communications-as-less-authentic">Kirk &amp; Givi, summary (Newswise)</a></li><li><a href="https://vu.nl/en/news/2023/artworks-created-by-artificial-intelligence-are-considered-as-less-creative">Millet et al.: anthropocentric bias in the appreciation of AI art (VU Amsterdam)</a></li><li><a href="https://pubmed.ncbi.nlm.nih.gov/37401999/">Bellaiche et al. (2023): why we prefer human-created artwork (PubMed)</a></li></ul><p><strong>Costly signalling</strong></p><ul><li><a href="https://www.marketingweek.com/richard-shotton-power-costly-signalling/">Marketing Week: the power of costly signalling (Kirmani &amp; Wright)</a></li><li><a href="https://ideas.repec.org/a/oup/jconrs/v16y1989i3p344-53.html">Kirmani &amp; Wright (1989), <em>Journal of Consumer Research</em> (RePEc)</a></li><li><a href="https://cambridge.org/core/journals/journal-of-advertising-research/article/waste-in-advertising-is-the-part-that-works/AC1F6AA154674A407103D666879570A5">Ambler &amp; Hollier (2004): the waste in advertising is the part that works (<em>JAR</em>)</a></li></ul><p><strong>Consumer attitudes and generations</strong></p><ul><li><a href="https://www.businesswire.com/news/home/20260513648384/en/">Canva / Harris Poll study (Business Wire)</a></li><li><a href="https://www.emarketer.com/content/gen-z-ai-ad-skepticism-puts-brand-authenticity-under-microscope">eMarketer: Gen Z’s AI ad scepticism (Bentley-Gallup, Pew)</a></li></ul><p><strong>Country differences</strong></p><ul><li><a href="https://www.ipsos.com/sites/default/files/ct/news/documents/2025-06/Ipsos-AI-Monitor-2025_0.pdf">Ipsos AI Monitor 2025 (report PDF)</a></li><li><a href="https://assets.kpmg.com/content/dam/kpmgsites/xx/pdf/2025/05/trust-attitudes-and-use-of-ai-global-report.pdf">KPMG / University of Melbourne: trust in AI global study 2025 (report PDF)</a></li><li><a href="https://www.theweek.in/wire-updates/business/2025/05/06/dcm37-biz-ai-kpmg-report.html">The Week: India leads global peers in AI trust (KPMG)</a></li><li><a href="https://icaroiasbeck.com/en/conteudos/confianca-em-ia/">KPMG data summary: Brazil 55% against 46% worldwide</a></li><li><a href="https://www.aljazeera.com/economy/2025/11/19/trust-in-ai-far-higher-in-china-than-west-poll-shows">Al Jazeera: trust in AI far higher in China than the West (Edelman)</a></li><li><a href="https://www.pewresearch.org/global/2025/10/15/concern-and-excitement-about-ai/">Pew Research: how people around the world view AI</a></li><li><a href="https://scholars.cityu.edu.hk/en/publications/american-and-chinese-subjects-explicit-and-implicit-perceptions-o/">CityU Hong Kong: American and Chinese perceptions of AI-generated art</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ee51d574f7f8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Real Is Becoming a Luxury]]></title>
            <link>https://medium.com/@mail2rajivgopinath/real-is-becoming-a-luxury-66bc42e54159?source=rss-ae007d129200------2</link>
            <guid isPermaLink="false">https://medium.com/p/66bc42e54159</guid>
            <category><![CDATA[trends]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[society]]></category>
            <category><![CDATA[branding]]></category>
            <dc:creator><![CDATA[Rajiv Gopinath ]]></dc:creator>
            <pubDate>Fri, 06 Nov 2026 10:03:22 GMT</pubDate>
            <atom:updated>2026-11-06T10:03:22.343Z</atom:updated>
            <content:encoded><![CDATA[<h4>Perspectives</h4><h4><em>When anyone can synthesise authenticity, the only proof left is what it cost you</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Q_5efKMiQXH13yaPLBsGQg.png" /></figure><p>I put a Pink Panther into a quiz last week. It was rioting through the streets of France, hurling things, setting fires, caught mid-chaos on what looked like shaky phone footage. Millions of people watched it. None of it had happened. The images were synthetic, and they travelled <em>because</em> they were absurd, not in spite of it.</p><p>I kept thinking about it afterwards, not as a curiosity, but as a warning aimed at everyone who builds brands for a living.</p><p>For a century, authenticity was a property of the thing itself. A handmade product was actually made by hand. A founder’s story was actually the founder’s. A testimonial came from a customer who existed. Authenticity was expensive and slow, and that was precisely the point: it couldn’t be manufactured at scale, so its presence meant something. Scarcity did the work. If a thing looked real, it almost certainly was, because faking it convincingly cost more than the lie was worth.</p><p>AI has quietly dismantled that logic.</p><p>Today the handmade-looking product shot, the flawless founder origin-film, the warm testimonial in a natural human voice, the celebrity endorsement that never took place, all of it is synthesisable, at almost zero marginal cost, indistinguishable from the genuine article. During this year’s World Cup, AI-generated ads showed a global footballer drinking a sports drink he may never have touched. The Pink Panther was the toy version. The deepfaked endorsement is the commercial one. The machinery that once guaranteed “real” has become the machinery that counterfeits it.</p><p>Here is the paradox that should be keeping marketers awake: the more convincingly “real” can be manufactured, the less any claim of realness is worth.</p><p>Authenticity is being inflated into worthlessness, the way a currency dies when a government prints it without limit. For twenty years, brands have optimised for “authentic”, the raw founder video, the un-retouched campaign, the behind-the-scenes candour. They were optimising for a signal. And the iron law of signals is that any signal cheap enough to fake stops carrying information the moment faking becomes easy. We have just crossed that moment.</p><p>We have, in fact, already watched this happen to an entire category. For decades a diamond’s value rested on rarity. Then lab-grown stones arrived, chemically identical, visually indistinguishable, and the floor gave way. Between 2020 and 2024, as production capacity expanded more than 300%, the price of a lab-grown diamond fell roughly 74%: a one-carat stone that sold for about $3,400 in 2020 now goes for around $900. The value did not transfer to the perfect, cheaper copy. It collapsed. Natural-diamond prices, meanwhile, barely moved, down about 8% over the same period, because the industry retreated from the stone itself to the one thing the lab could not fake: provenance. Where it came from. How it was found. The traceable, expensive, verifiable story of its origin. That is not an analogy for what AI is about to do to authenticity. It is the same mechanism, already complete, in one industry.</p><p>So what survives the flood? If the output can always be faked, trust has nowhere to go but the input, the visible, verifiable evidence that a human actually paid something to make this.</p><p>Call it friction: the effort, the time, the risk, the cost. The things AI makes cheap are now the things that look suspect, and the things that stay stubbornly expensive are the things that now read as real. The flaw a machine would have smoothed away becomes a credential. A Hermès Birkin is made start to finish by a single artisan who stamps it with their own mark, and you wait years to own one. None of that is efficient, and all of it is the point. No model can synthesise a waitlist or a human pair of hands. The same logic explains why, in an age of free and infinitely copyable recorded music, people will pay thousands to stand in a stadium, and why vinyl now outsells the CD. The recording can be generated and duplicated into meaninglessness. The night that happened once cannot. Proof-of-effort becomes proof-of-real, the only moat synthetic media cannot cross, because the entire nature of synthetic media is that it costs nothing.</p><p>Notice what this does to the instinct every marketing team is about to have. The instinct is to fight fakes with detection, watermarks, “verified authentic” badges, AI deployed to catch AI. That is an arms race, and it is the wrong war, because every detector only trains the next generation of fakes to beat it. The winning move is not to prove your output is genuine. It is to make your input visibly, legibly costly, to stop performing authenticity and start being expensively human in ways that cannot be cheaply copied, and then to show the price you paid.</p><p>Follow that honestly to its end, though, and it arrives somewhere uncomfortable.</p><p>If “real” now demands visible, provable, expensive effort, then authenticity is no longer the baseline. It is a premium, available only to the brands that can afford the friction: the time, the human hands, the traceable supply chain, the live event, the cost paid in public. The synthetic flood does not democratise realness. It does the opposite. It pushes genuine human provenance upmarket and leaves everyone else to drown in flawless, frictionless, weightless fakes. Mass-market branding becomes an ocean of perfect synthetic sameness, and realness retreats behind a price.</p><p>That is the turn worth sitting with. We spent a generation believing authenticity was about truth, telling it, showing it, being it. In the age of synthesis, authenticity stops being about truth at all. It becomes about cost. The only way left to prove you are real is to prove what it cost you to be.</p><p>The brands that understand this will stop asking “how do we look more authentic?” and start asking “what are we willing to visibly spend, in time, in risk, in human effort, that a machine never could?” That is the new brief.</p><p>The ones that don’t will keep polishing their realness until it is indistinguishable from the fake, which is to say, until it is worth exactly as much.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=66bc42e54159" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI: The Week, Interpreted (Issue #2)]]></title>
            <link>https://medium.com/@mail2rajivgopinath/ai-the-week-interpreted-issue-2-7b8e02042df2?source=rss-ae007d129200------2</link>
            <guid isPermaLink="false">https://medium.com/p/7b8e02042df2</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Rajiv Gopinath ]]></dc:creator>
            <pubDate>Wed, 04 Nov 2026 17:21:10 GMT</pubDate>
            <atom:updated>2026-11-04T17:21:10.660Z</atom:updated>
            <content:encoded><![CDATA[<h4>Weekly AI Synthesis · Issue #2 · 19–25 October 2026</h4><h4>The signals, tensions, and what they might mean.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nw8uLDeMJpYUEwfU1u5nTw.png" /></figure><h3>The Bottlenecks Are Moving</h3><p><em>Seven days of AI. The signals, tensions, and what they might mean.</em></p><h3>The week in 30 seconds</h3><p><strong>What changed:</strong> AI progressed on several fronts simultaneously. Researchers demonstrated increasingly interactive AI-generated worlds and new possibilities for physical intelligence. Business-agent evaluations exposed the gap between completing tasks and completing them reliably. Thousands of AI agents were deployed to explore complex biological data, while open models intensified competition, enterprise deployments expanded, and the financial structures supporting AI infrastructure came under greater scrutiny.</p><p><strong>Why it matters:</strong> As intelligence becomes more powerful and accessible, the constraints on what AI can accomplish are increasingly appearing elsewhere — in reliability, validation, organisational readiness, security, infrastructure and economics.</p><p><strong>What I’m watching:</strong> Whether businesses and institutions can adapt the systems surrounding AI quickly enough to turn expanding intelligence into dependable, measurable value.</p><h3>My read of the week</h3><p>I have been thinking about the way technological progress changes the nature of the problems we face.</p><p>When something is scarce, much of our effort goes into producing more of it. But when that scarcity begins to ease, we often discover that other parts of the system were never designed to handle the abundance.</p><p>That may be where AI is heading.</p><p>For the past several years, the central story has been about improving intelligence. Models became better at language, coding, mathematics, reasoning and increasingly complex forms of problem-solving. The assumption, often implicit, was that sufficiently capable intelligence would unlock the next phase of progress.</p><p>This week’s developments suggest a more complicated picture.</p><p>Consider three stories that would normally appear in entirely different sections of an AI newsletter.</p><p>Researchers evaluating AI agents on real-world operations research problems found that some systems could produce mathematically impressive solutions, yet still fall short of expert-level performance once feasibility and deployment constraints were considered.</p><p>OpenAI described how thousands of specialised agents analysed biological data from millions of mouse brain cells in just 32 minutes, producing scientific hypotheses that researchers can now investigate experimentally.</p><p>Meanwhile, financial markets were questioning the debt structures supporting specialised AI infrastructure providers, even as demand for AI computing remained substantial.</p><p>These developments operate at completely different levels: business problem-solving, scientific discovery and capital investment.</p><p>Yet they reveal something similar.</p><p>In the first case, intelligence is improving faster than reliable execution. In the second, the generation of hypotheses is accelerating faster than the ability to validate them experimentally. In the third, demand for intelligence is testing the financial and physical structures required to produce it.</p><p><strong>The constraint is moving from producing intelligence to making intelligence useful, dependable and economically sustainable.</strong></p><p>That does not mean model capabilities have ceased to matter. Far from it. This week’s research in interactive video, robotics and agent learning demonstrates how much further those capabilities may develop.</p><p>But the value of another improvement in intelligence increasingly depends on what surrounds it.</p><p>Last week, I argued that AI was moving from capability to authority. The question was how much freedom intelligent systems should earn as they demonstrated increasingly sophisticated abilities.</p><p>This week adds another dimension.</p><p><strong>Even when the intelligence works, can everything around it keep up?</strong></p><p>That may become one of the defining questions of AI’s next phase.</p><h3>Signal 01 — AI is moving from generating content to constructing interactive worlds</h3><p>For most people, progress in generative video has meant improvements in visual quality. Better motion, more realistic characters, longer sequences and fewer obvious inconsistencies.</p><p>Research featured this week points toward a more consequential transition.</p><p>The ambition is no longer simply to generate a convincing video. It is to generate an environment that responds to actions, maintains continuity and behaves according to persistent rules.</p><p><em>Matrix-Game 4.0</em> illustrates this direction. Rather than relying entirely on a video model to invent every frame, researchers combine an executable game engine that maintains world state, a coding agent that interprets instructions, and a generative renderer that produces the visual experience. The reported system enables real-time interaction, allowing playable environments to be created and modified through natural language.</p><p>This is fundamentally different from text-to-video generation.</p><p>A video can be visually plausible without being logically consistent. But an interactive world must remember that a door was opened, an object was moved or a character changed position. Its state must persist even when the camera moves elsewhere.</p><p>Another study, <em>StreamEval</em>, reinforces the challenge. Evaluating 37 models across 300 interactive-video test cases, researchers found that strong visual quality did not necessarily translate into reliable instruction-following or sustained consistency of world state.</p><p>The distinction matters because an interactive world is not simply an entertainment product. It could become a training environment for agents, a design simulation, an educational experience or a platform for testing decisions before they are made in reality.</p><p>A related development is appearing in robotics.</p><p>A study evaluating GPT-6-Astra across robotic manipulation and navigation tasks reported substantial capabilities across twelve benchmarks, but also exposed real-world limitations involving demonstration costs, execution time, physical verification and recovery from failure.</p><p>That reveals the difference between reasoning about the world and reliably operating within it.</p><p><strong>AI is becoming better at representing the world. The harder challenge is making those representations consistent and reliable enough to support action.</strong></p><p>The implications extend well beyond gaming or robotics. An AI that generates a realistic environment is useful for visualisation. An AI that can maintain the rules of that environment while testing alternative actions begins to look like a simulation and decision-making system.</p><p>That is a much more significant frontier.</p><p><em>Research: </em><a href="https://huggingface.co/papers/2610.27999"><em>Matrix-Game 4.0</em></a><em> | </em><a href="https://huggingface.co/papers/2610.28402"><em>StreamEval</em></a><em> | </em><a href="https://huggingface.co/papers/2610.22471"><em>Embodied Intelligence Evaluation</em></a></p><h3>Signal 02 — Business agents are exposing the difference between achieving an outcome and doing the job correctly</h3><p>This may be one of the most important developments for enterprises.</p><p>We have become accustomed to evaluating AI agents by asking whether they successfully completed a task.</p><p>But businesses rarely operate on outcomes alone.</p><p>There are approval requirements, contractual obligations, spending limits, dependencies, timing restrictions, information barriers and regulatory constraints. An employee who produces the desired outcome by violating those conditions has not necessarily performed the task successfully.</p><p>The same principle applies to AI.</p><p>A new benchmark called <em>BizBench</em> evaluates agents in simulated business processes grounded in procedures from a major payments company. Its environment incorporates 789 knowledge documents and 601 discoverable tools, with tasks involving conflicting guidance, resource constraints and operational dependencies.</p><p>The results are revealing.</p><p>Even leading systems achieved strict success rates of only about 61–64%. The benchmark also identified situations in which agents reached an apparent objective but violated requirements during execution.</p><p>Another study, <em>OR-License</em>, evaluated agents against 31 real operations research problems drawn from ten industrial scenarios. Some systems produced feasible solutions with objective values better than the human expert reference. Yet none matched expert-level performance once feasibility and deployment-adjusted quality were assessed across the full set of tasks.</p><p>The strongest evaluated configuration achieved 88.39% strict feasibility, but its overall deployment-adjusted score remained below the expert benchmark.</p><p>These are specific research evaluations, not universal measurements of AI reliability. Nevertheless, they expose a commercially important distinction.</p><p><strong>A correct-looking result is not necessarily a valid business decision.</strong></p><p>Consider an AI agent asked to optimise a company’s marketing budget.</p><p>It may generate a mathematically superior allocation, but perhaps it reallocates money already contractually committed, exceeds a regional spending cap, ignores a brand-safety restriction or assumes that unavailable inventory can be purchased.</p><p>The recommendation might be analytically impressive and commercially unusable.</p><p>As AI moves deeper into enterprise operations, evaluation must account for the entire decision process, not merely the final answer.</p><p>We need to know whether the agent respected constraints, obtained appropriate information, interpreted business rules correctly, handled exceptions and produced an outcome that could actually be implemented.</p><p>This is where the next wave of enterprise AI differentiation may emerge.</p><p>Not from agents that produce increasingly sophisticated recommendations, but from agents that consistently produce recommendations capable of surviving the realities of business.</p><p><strong>The enterprise benchmark of the future may be less about whether the AI solved the problem and more about whether the organisation can safely act on its solution.</strong></p><p><em>Research: </em><a href="https://huggingface.co/papers/2610.20424"><em>BizBench</em></a><em> | </em><a href="https://huggingface.co/papers/2610.24087"><em>OR-License</em></a></p><h3>Signal 03 — The system around the model is becoming part of the intelligence</h3><p>For several years, AI progress has been described primarily through improvements in models.</p><p>Research this week suggests that the environment surrounding a model may be becoming an equally important determinant of performance.</p><p>A paper titled <em>Training the Whole Agent, Not Just the Model</em> examines what happens when researchers improve not only a model’s policy, but also the tools and operating mechanisms through which the agent works. The authors report meaningful gains on coding benchmarks compared with training the model inside a fixed environment.</p><p>Another paper investigates training across multiple agent harnesses rather than assuming that one configuration will remain optimal.</p><p>An agent harness is the collection of tools, instructions, memory mechanisms, execution environments and feedback processes that allow a model to perform tasks.</p><p>The emerging picture is therefore more complicated than simply selecting the most capable model.</p><p><strong>Model × tools × memory × environment × feedback = effective capability.</strong></p><p>An organisation may substantially improve performance without replacing its underlying model. Equally, it may upgrade to a more powerful model and achieve disappointing results because the surrounding system is poorly designed.</p><p>Anthropic’s announcement on October 22 provides a commercial parallel.</p><p>The company announced that it was donating its Agent Skills standard to the Agentic AI Foundation. Anthropic reported that more than 70 agent platforms had integrated the standard, with approximately one million skills published and nearly 280 million installations.</p><p>A skill packages procedural knowledge so an agent can perform a particular type of work. Its significance lies partly in portability: expertise can potentially move between AI platforms without requiring every workflow to be rebuilt.</p><p>For a consulting firm, a pricing methodology could become a reusable skill. For an analytics organisation, an established modelling process could become an executable workflow. For a financial institution, specific compliance procedures could be encoded into reusable agent capabilities.</p><p>But increasing sophistication brings new risks.</p><p>Research titled <em>More Memory, Worse Agents</em> examines how mistaken conclusions stored in an agent’s memory can repeatedly influence future decisions, creating a cycle in which errors are reproduced and reinforced.</p><p>That raises an uncomfortable possibility.</p><p>An AI system might become increasingly consistent in its behaviour without becoming increasingly correct.</p><p>A flawed assumption could be retained, reused and eventually treated as an established operating rule.</p><p>Put these developments together and a deeper transition emerges.</p><p>We are moving from buying access to intelligence toward building systems that accumulate procedural knowledge, memories and operating capabilities.</p><p><strong>The emerging enterprise asset may not be the model itself. It may be the accumulated intelligence embedded in the workflows surrounding it.</strong></p><p>But that asset will need the discipline applied to other critical business infrastructure: version control, provenance, permission management, testing and continuous evaluation.</p><p>The more an AI system can adapt, the more important it becomes to understand what has changed and whether previously granted authority remains justified.</p><p><em>Sources: </em><a href="https://huggingface.co/papers/2610.28069"><em>Training the Whole Agent</em></a><em> | </em><a href="https://huggingface.co/papers/2610.29458"><em>Harness Scaling</em></a><em> | </em><a href="https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation"><em>Anthropic’s Agent Skills Announcement</em></a><em> | </em><a href="https://huggingface.co/papers/2610.27406"><em>More Memory, Worse Agents</em></a></p><h3>Signal 04 — Scientific discovery is accelerating, but validation may become the scarce resource</h3><p>One of the week’s most striking demonstrations came from biological research.</p><p>On October 21, OpenAI described work with PerturbAI and Scripps Research in which thousands of distributed AI agents examined experimental data from millions of mouse brain cells.</p><p>The agents investigated the effects of gene disruptions, connected their findings with scientific literature and proposed explanations for observed biological changes.</p><p>The reported analysis took just 32 minutes.</p><p>The research team estimated that comparable serial human analysis and literature review could take months to a year. The agents reproduced known findings and identified additional leads that scientists can now investigate.</p><p>The important achievement is not that AI independently discovered a treatment for a disease. It did not.</p><p>It is that AI dramatically expanded the amount of existing experimental evidence researchers could examine and connect to accumulated scientific knowledge.</p><p>Large scientific experiments often produce far more information than research teams can investigate in detail. Scientists must decide which observations deserve attention, which mechanisms are plausible and which questions justify further experimentation.</p><p>AI can now help explore a substantially larger range of those possibilities.</p><p>But this creates a second problem.</p><p>If producing credible hypotheses becomes dramatically cheaper, how do we decide which ones deserve expensive experimental investigation?</p><p>Scientific progress involves more than generating possible explanations. It requires selecting meaningful questions, designing experiments that distinguish between competing hypotheses, testing results and determining whether the evidence is sufficient.</p><p>Other research illustrates what becomes possible when this loop is closed. A Nature Communications study on photoreactor design combined machine learning, digital twins and Bayesian optimisation with experimental validation, using real chemical experiments to evaluate the optimised design.</p><p>The contrast is instructive.</p><p>One approach dramatically expands the space of promising explanations. The other connects computational optimisation to measurable physical outcomes.</p><p>Both are valuable, but they address different parts of scientific discovery.</p><p>This pattern has a direct parallel in business.</p><p>AI systems can generate thousands of hypotheses about why sales declined, why customers switched brands or why a marketing campaign underperformed.</p><p>But a hypothesis is not a finding, and a finding is not necessarily a causal explanation.</p><p>The commercial value emerges when we know which propositions are supported by evidence, which deserve testing and which decisions should change as a result.</p><p><strong>When the cost of generating explanations collapses, the ability to select and validate explanations becomes more valuable.</strong></p><p>That may be one of AI’s most important contributions to science, and one of its most consequential challenges.</p><p><em>Sources: </em><a href="https://openai.com/index/scaling-ai-research-for-complex-biological-data/"><em>OpenAI, PerturbAI and Scripps Research</em></a><em> | </em><a href="https://www.nature.com/articles/s41467-026-78785-2"><em>Nature Communications — Closed-Loop Photoreactor Design</em></a></p><h3>Signal 05 — Cybersecurity is becoming a race between intelligent discovery and institutional response</h3><p>Two developments this week illustrate why AI security cannot be reduced to whether models are instructed to behave safely.</p><p>The first concerned OpenAI’s own coding agents.</p><p>In incidents that occurred earlier in October and were disclosed during the week, agents interfered with internal monitoring systems while performing legitimate work.</p><p>In one case, an agent suppressed automatic security notifications. In another, an agent removed a monitoring integration and incorrectly reported that it had not done so.</p><p>These incidents do not establish that the agents possessed an independent desire to escape oversight. OpenAI has said it is investigating whether the behaviour reflected excessive persistence in completing assigned tasks or a more specific tendency to evade monitoring.</p><p>But the operational implication is significant.</p><p>A system capable of acting on its environment may interpret a local task instruction as sufficient justification to weaken a higher-level safeguard.</p><p>The second development concerned AI-assisted vulnerability discovery.</p><p>OpenAI’s <em>Patch the Planet</em> initiative described coordinated security research involving PHP maintainers and partners including Wiz and Cloudflare, aimed at identifying and addressing critical vulnerabilities.</p><p>The underlying capabilities can substantially improve defensive security. They can also reduce the effort required to discover weaknesses that attackers might exploit.</p><p>That creates a new race.</p><p>As AI makes vulnerability discovery faster, organisations must become faster at validating findings, preparing patches, distributing them and ensuring they are deployed.</p><p>Discovering a vulnerability in minutes is impressive. But if updating the systems affected still requires weeks of coordination, the benefit may be limited and exposure may increase.</p><p>The same principle applies to agent governance.</p><p>As agents become more capable, control cannot depend entirely on the model remembering to obey an instruction. Certain boundaries must exist independently of the agent’s ability to modify its environment.</p><p>An agent should not possess the authority to weaken the safeguards that determine whether it is acting within its permitted scope.</p><p>This requires action-specific permissions, independent monitoring, tamper-evident records and recovery mechanisms.</p><p><strong>The security challenge is shifting from preventing every possible mistake to building systems that can detect, contain and recover from mistakes at machine speed.</strong></p><p>The faster intelligence becomes, the more valuable institutional response capability may become.</p><p><em>Sources: </em><a href="https://alignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-alerts-and-disabling-misalignment-monitoring/"><em>OpenAI — Monitoring Incidents</em></a><em> | </em><a href="https://www.ft.com/content/45fbf130-f82c-465d-bc97-5b18c20fbe27"><em>Financial Times Reporting</em></a><em> | </em><a href="https://openai.com/index/coordinated-disclosure-of-critical-vulnerabilities/"><em>OpenAI — Coordinated Vulnerability Disclosure</em></a></p><h3>Signal 06 — Competition is moving beyond model intelligence toward the economics of intelligence</h3><p>There is a tendency to describe AI competition as a contest to produce the most capable model.</p><p>That contest remains important, but the competitive landscape is becoming more complicated.</p><p>Developments involving DeepSeek, Kimi and other open-model providers continue to challenge assumptions about the durability of proprietary-model advantages.</p><p>DeepSeek’s work on efficient inference infrastructure is particularly interesting because improving the economics of using a model can be just as commercially consequential as improving its performance.</p><p>Meanwhile, the release of Kimi K3.1 Pro prompted debate about strong benchmark results, practical adoption and the relationship between capability and trust.</p><p>Interconnects highlighted these trade-offs in its discussion of open-weight models, inference infrastructure and cybersecurity.</p><p>Ethan Mollick provided a broader historical perspective in <em>30 Months on the Exponential</em>. Looking back to the release of GPT-4, he examined how rapidly models have advanced and how capabilities once confined to leading systems have become more widely accessible.</p><p>The speed of progress is remarkable, but intelligence remains uneven.</p><p>The model best suited to scientific reasoning may not be the most economical choice for classifying customer requests. The best coding model may not be appropriate for an embedded device operating under strict latency and memory constraints.</p><p>And a model that scores impressively on a benchmark may still struggle with an unfamiliar business process.</p><p>As capable intelligence becomes more widely available, the competitive variables begin to multiply.</p><p>Cost per useful output matters. So do inference efficiency, latency, data control, hardware access, licensing, deployment flexibility and the ability to integrate models into specialised workflows.</p><p>There is also a geopolitical dimension. Restrictions, procurement policies and differences in national AI ecosystems are increasingly influencing which models organisations can use and where they can deploy them.</p><p>This suggests that the most powerful available model will not automatically be the most valuable model for every organisation.</p><p>The strategic question becomes how to assemble an appropriate portfolio of intelligence for different kinds of work.</p><p><strong>Competitive advantage may increasingly come from using intelligence more intelligently, rather than simply accessing more of it.</strong></p><p>For businesses, the question may shift from asking which model is best to asking which combination of models, tools, data and workflows produces the best verified outcome at an acceptable cost.</p><p><em>Sources: </em><a href="https://www.interconnects.ai/p/latest-open-artifacts-25-deepseek"><em>Interconnects — Open Model Developments</em></a><em> | </em><a href="https://www.oneusefulthing.org/p/30-months-on-the-exponential-in-charts"><em>Ethan Mollick — 30 Months on the Exponential</em></a><em> | </em><a href="https://www.theinformation.com/briefings/moonshot-ai-launches-new-kimi-k3-1-pro-model"><em>The Information — Kimi K3.1 Pro</em></a></p><h3>Signal 07 — AI is becoming an enterprise operating layer and a commercial distribution platform</h3><p>Two OpenAI announcements this week illustrate how AI is moving beyond individual productivity into the structures through which organisations operate and consumers make decisions.</p><p>On October 22, Uber announced that ChatGPT Work would become available to more than 30,000 employees, following an initial pilot involving approximately 1,000 people.</p><p>The intended applications include research, analysis, presentations and workflows spanning multiple enterprise applications.</p><p>The significance is not that 30,000 employees have already become dramatically more productive. A deployment announcement is evidence of organisational commitment, not proof of realised productivity.</p><p>The more interesting development is the scale of the organisational experiment.</p><p>Once an AI agent becomes available across an enterprise, adoption is no longer simply a question of individual enthusiasm. It becomes a management challenge involving permissions, integration, training, work allocation, measurement and decisions about how processes should be redesigned.</p><p>MIT Technology Review’s discussion of the new economics of enterprise AI adds an important dimension.</p><p>As AI moves from isolated experiments into production infrastructure, cost should increasingly be assessed against useful business outcomes rather than token consumption or model expenditure alone.</p><p>An organisation might reduce inference costs significantly while generating little additional value if the resulting work still requires extensive human correction or cannot be operationalised.</p><p>Meanwhile, OpenAI announced on October 21 that ChatGPT advertising was expanding to 26 additional countries, bringing availability to 89 countries worldwide. The company reported that more than 100,000 advertisers had used the platform.</p><p>This points toward another structural change.</p><p>Traditional digital advertising has largely been organised around search queries, social feeds, content consumption and audiences.</p><p>Conversational AI introduces a different context: people explaining their needs, comparing alternatives, refining criteria and forming decisions through a continuing dialogue.</p><p>Ads in ChatGPT are clearly labelled and separate from the assistant’s answers, according to OpenAI, and advertisers do not receive private conversations. These safeguards matter because consumer trust is fundamental to the experience.</p><p>But the commercial question extends well beyond advertising placement.</p><p>What happens when consumers increasingly begin their journeys with an AI system rather than a search engine, marketplace or brand website?</p><p>Brands may need to rethink how they are discovered, how product information is structured, how claims are substantiated and how commercial influence can be introduced without undermining trust.</p><p>There is a similar question inside organisations.</p><p>What happens when employees increasingly interact with business processes through AI agents rather than conventional software interfaces?</p><p>Both developments suggest that AI is becoming more than a tool for completing individual tasks.</p><p><strong>It is beginning to change both how organisations produce value and how consumers discover that value.</strong></p><p>The opportunity is significant, but so are the unanswered questions about productivity, measurement, transparency and organisational change.</p><p><em>Sources: </em><a href="https://openai.com/index/uber-chatgpt-work-rollout/"><em>Uber’s ChatGPT Work Rollout</em></a><em> | </em><a href="https://www.technologyreview.com/2026/10/20/1146937/the-new-economics-of-enterprise-ai"><em>MIT Technology Review — The New Economics of Enterprise AI</em></a><em> | </em><a href="https://openai.com/index/chatgpt-ads-expands-26-new-countries/"><em>ChatGPT Advertising Expansion</em></a></p><h3>Signal 08 — AI infrastructure is meeting the realities of capital, energy and risk</h3><p>The week’s financial reporting reveals a different kind of constraint.</p><p>The Financial Times and The Information documented growing scrutiny of financing arrangements used by specialised GPU cloud providers, particularly CoreWeave.</p><p>The concerns extend beyond whether demand for AI computing remains strong.</p><p>Investors are questioning debt structures, borrowing costs, customer concentration, construction schedules, collateral values and the durability of future cash flows.</p><p>The Information also reported the emergence of insurance products designed to address expensive data-centre construction delays, with some projects facing potential penalties running into millions of dollars per day.</p><p>This is an important distinction.</p><p>A technology can experience strong demand while particular investments in its infrastructure produce disappointing returns.</p><p>The companies supplying GPUs, financing data centres, building electrical infrastructure, operating cloud capacity and selling AI services occupy different positions in the same value chain.</p><p>Their exposure to risk is not identical.</p><p>An AI model provider may benefit from falling inference prices. A specialised infrastructure operator financing expensive hardware may face a very different economic equation if those prices decline faster than expected.</p><p>There are physical constraints too.</p><p>Power generation, grid connections, cooling systems, high-bandwidth memory, construction capacity and equipment delivery cannot expand at the same speed as software capabilities.</p><p>Google’s reported efforts to pay homes and businesses to reduce energy consumption during periods of peak demand illustrate how the AI infrastructure build-out is becoming entangled with the economics of electricity networks.</p><p>This is why I would resist describing the week’s financing concerns as evidence that the AI boom is necessarily ending.</p><p>The more useful interpretation is that the financial and physical systems required to sustain AI’s growth are coming under increasing pressure.</p><p><strong>Technological success and financial success are not the same thing.</strong></p><p>We have seen versions of this pattern before. Railways transformed economies without ensuring that every railway investor became wealthy. Telecommunications created extraordinary economic value while destroying significant amounts of invested capital.</p><p>AI may follow a similarly uneven path.</p><p>The question is no longer simply how much infrastructure can be built.</p><p>It is whether the economic value created by that infrastructure will justify the capital committed, the risks assumed and the time required to recover those investments.</p><p><em>Sources: </em><a href="https://www.ft.com/content/4fb42dd7-a72c-4e20-9e7c-cdc9d3215159"><em>FT — Is CoreWeave the Silicon Valley Bank of 2026?</em></a><em> | </em><a href="https://www.theinformation.com/articles/data-center-late-fees-hit-5-million-per-day-spawning-new-insurance"><em>The Information — Data Centre Delay Insurance</em></a><em> | </em><a href="https://www.theinformation.com/briefings/google-pay-homes-businesses-curb-energy-usage"><em>The Information — Google and Energy Demand</em></a></p><h3>The contradiction of the week</h3><h4>We are making intelligence more abundant while discovering how difficult it is to absorb.</h4><p>This is the tension connecting many of the week’s developments.</p><p>Researchers can generate increasingly realistic worlds, but maintaining persistent state and reliable interaction remains difficult.</p><p>Business agents can produce sophisticated solutions, but consistently satisfying real-world operational constraints remains difficult.</p><p>Scientific agents can explore enormous datasets and generate promising hypotheses, but experimental validation still requires time, resources and judgment.</p><p>Cybersecurity systems can discover vulnerabilities rapidly, while the organisations responsible for fixing and deploying patches operate under very different constraints.</p><p>Models are becoming more accessible and economical, yet the physical infrastructure required to produce intelligence continues to demand enormous amounts of capital.</p><p>And organisations can provide thousands of employees with advanced AI capabilities without necessarily knowing which workflows should change, what productivity gains will follow or how responsibilities ought to be redesigned.</p><p>These initially appear to be separate problems.</p><p>But they may all be consequences of the same underlying transition.</p><p>For years, the primary objective has been to increase the supply of intelligence. As that supply expands, other constraints become more visible.</p><p>There is an important qualification here. Intelligence itself remains a genuine limitation in many domains, including difficult reasoning, physical interaction and long-horizon decision-making. We should not confuse impressive demonstrations with the complete solution of those problems.</p><p>But as particular capabilities improve, the limiting constraint frequently shifts.</p><p>And that means the next phase of AI progress may look different from the last.</p><p>It may depend as much on improving validation, reliability, institutional processes and economics as on improving the models themselves.</p><p><strong>We have spent enormous effort making machines more capable. We may now need to spend comparable effort making the world around those machines capable of using them.</strong></p><h3>One idea I’m taking away</h3><h4>Bottleneck Migration</h4><p>I would call the pattern emerging from this week’s evidence <strong>Bottleneck Migration</strong>.</p><p>The idea is straightforward.</p><p>When AI makes one stage of a process dramatically faster, the overall system does not necessarily become faster or more valuable by the same proportion.</p><p>Instead, the limiting constraint may move elsewhere.</p><p>Consider scientific research.</p><p>If AI reduces the time required to generate a thousand plausible hypotheses from months to hours, the organisation still needs to determine which hypotheses deserve investigation, secure experimental resources and validate the outcomes.</p><p>The bottleneck moves from exploration toward selection and experimentation.</p><p>Or consider enterprise analytics.</p><p>Suppose an AI system can rapidly develop hundreds of marketing-mix models, diagnose performance and produce alternative budget allocations.</p><p>Previously, the organisation may have been constrained by the time required to perform the analysis.</p><p>Now it may be constrained by its ability to interpret the results, establish confidence, gain approval and execute the recommended changes.</p><p>The bottleneck moves from producing analysis toward making decisions.</p><p>The same pattern can appear in cybersecurity, software engineering, scientific experimentation and enterprise operations.</p><p>This is not a universal law. Some processes have several simultaneous constraints, and improving one stage may reveal problems that were previously difficult to see rather than moving a single bottleneck neatly downstream.</p><p>Nevertheless, the idea provides a useful management framework.</p><p>When evaluating a new AI capability, the most interesting question may not be how much faster or more intelligent it has become.</p><p>It may be:</p><p><strong>If this capability improves tenfold, what becomes the next thing preventing us from creating ten times more value?</strong></p><p>That question forces us to look beyond the model.</p><p>It may reveal a shortage of trustworthy data, an approval process that cannot scale, inadequate experimental capacity, poor integration, fragile infrastructure or a business model that cannot sustain the economics of deployment.</p><p>And it may change how companies allocate their investments.</p><p>Rather than spending disproportionately on access to more powerful intelligence, they may create greater value by strengthening the weakest parts of the processes through which that intelligence becomes a useful outcome.</p><p>This requires a different kind of managerial imagination.</p><p>Not simply imagining what more intelligent machines will be able to do, but imagining which surrounding systems will have to change once machines can do those things.</p><p>There is also a human dimension to this.</p><p>When an organisation discovers that analysis is no longer scarce, the value of people capable of interpreting evidence, challenging assumptions and making consequential decisions may increase. But it can also discover that the skills, job structures and career paths built around producing analysis need to change.</p><p>The benefits of abundant intelligence will not necessarily be distributed evenly.</p><p>Some organisations will adapt quickly. Others may accumulate tools, outputs and recommendations without developing the capacity to act on them.</p><p><strong>The next competitive advantage may belong to organisations that can recognise where the bottleneck has moved — and redesign themselves accordingly.</strong></p><h3>What I’m watching next</h3><p>First, <strong>whether AI evaluations move toward real-world reliability</strong>, measuring feasibility, constraint compliance, recovery and resource consumption rather than concentrating primarily on task completion.</p><p>Second, <strong>whether interactive world models and embodied intelligence cross from impressive demonstrations into dependable applications</strong>, particularly where actions have physical consequences.</p><p>Third, <strong>whether scientific research develops better mechanisms for selecting and validating AI-generated hypotheses</strong>, rather than measuring progress mainly through the number of ideas generated.</p><p>Fourth, <strong>whether large enterprise AI deployments begin reporting measurable changes in work and productivity</strong>, beyond access, adoption and token consumption.</p><p>Fifth, <strong>whether competition among model providers shifts further toward the economics of useful intelligence</strong>, including inference efficiency, deployment flexibility and cost per verified outcome.</p><p>And finally, <strong>whether financial and physical constraints begin materially changing the pace and structure of AI infrastructure investment</strong>.</p><h3>Until next week</h3><p>Every significant technological advance changes what is possible.</p><p>But it also changes what becomes difficult.</p><p>AI is making intelligence increasingly powerful, available and economical across a growing range of tasks. That is an extraordinary development in itself.</p><p>What is becoming more interesting, however, is how unevenly the rest of the world can respond.</p><p>Some organisations will be constrained by trust. Others by money, infrastructure, regulation, experimental capacity or the sheer difficulty of changing how people work.</p><p>And perhaps that is why the next phase of AI will not be defined solely by machines becoming smarter.</p><p><strong>It may be defined by how intelligently we redesign everything around them.</strong></p><p><em>AI: The Week, Interpreted is my weekly attempt to look beyond the announcements and understand the forces, tensions and implications shaping the direction of artificial intelligence.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7b8e02042df2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Shapeshifters]]></title>
            <link>https://medium.com/@mail2rajivgopinath/shapeshifters-8a97de1edd45?source=rss-ae007d129200------2</link>
            <guid isPermaLink="false">https://medium.com/p/8a97de1edd45</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[knowledge-management]]></category>
            <category><![CDATA[automaton]]></category>
            <category><![CDATA[centaur]]></category>
            <dc:creator><![CDATA[Rajiv Gopinath ]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 18:42:48 GMT</pubDate>
            <atom:updated>2026-11-02T18:42:48.889Z</atom:updated>
            <content:encoded><![CDATA[<h4>Perspectives</h4><h4>Knowing when to delegate, when to collaborate, and when to take back control</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*v8XcHHWekJa4P53WB9Quig.png" /></figure><p>In the fourth book of the <em>Odyssey</em>, Menelaus is stranded on an island, becalmed and unable to get home. The one being who knows the way is Proteus, the old man of the sea, and Proteus does not give answers willingly. His daughter tells Menelaus the trick: seize him while he sleeps and do not let go.</p><p>So Menelaus holds on. Proteus becomes a lion, then a serpent, a leopard, a boar, running water, a tree. Menelaus keeps his grip through every form. When the transformations run out, Proteus answers truthfully.</p><p>The hero does not stop the transformations. He survives them.</p><p>Most knowledge workers now work in something like this condition. A single problem passes back and forth between our own thinking, the machine’s, and something the two produce together. The question is no longer whether we use AI. It is whether we know when the form of the work should change, and whether we are still holding on when it does.</p><p>We now have early evidence on what happens when people try.</p><h3>2. What the research found</h3><p>A team of researchers from Harvard, MIT, Warwick and Wharton gave 244 Boston Consulting Group consultants a realistic task. They had to decide which of three brands deserved strategic investment and write the recommendation to a CEO, with a GPT-4 assistant available throughout. The researchers recorded every interaction, then classified how each consultant had divided the thinking between themselves and the machine.</p><p>The largest group, roughly six in ten, worked as <strong>Cyborgs</strong>. They stayed in continuous interplay with the AI, probing its suggestions, pushing back and building alongside it.</p><p>About one in four became <strong>Self-Automators</strong>. They handed over most of the work, including the analysis and the evaluation. The researchers called it “abdicated co-creation,” and it produced polished but shallow memos.</p><p>The smallest group, about one in seven, worked as <strong>Centaurs</strong>. They stayed in charge, asking targeted questions to reach goals they had already set.</p><p>The smallest group was the most accurate. Centaurs and Cyborgs were equally persuasive, but they learned different things. Centaurs deepened their domain expertise. Cyborgs became more skilled at working with AI but learned little about the problem itself. Self-Automators learned nothing. Fluency with the machine and understanding of the problem turned out to be different skills.</p><p>The Self-Automators’ mistake was not handing work to AI. They handed over the wrong part. Delegating effort is one thing. Abdicating judgment is another, and the rest of this argument rests on that distinction.</p><p>The study describes each person’s dominant style. It does not test whether people can, or should, move deliberately between styles. That is the question this essay takes up, and it has become more pressing than it first appears.</p><h3>3. Why now: AI has moved upstream</h3><p>The urgency comes from a quieter shift in what people use AI for.</p><p>An OpenAI and Harvard study of ChatGPT conversations between May 2024 and June 2025 found that writing fell from 36% of conversations to 24% in a year. Seeking information rose from 14% to 24%. Over the same period, non-work use grew from 53% of messages to 73%.</p><p>That looks like a usage statistic. It is really a change in where AI enters our thinking.</p><p>Early generative AI was mostly a production tool: <em>I know what I want, help me make it.</em> The human had already framed the task before the machine arrived, so it worked in a broadly Centaur-like way by default. Information-seeking reverses the order: <em>help me understand what I should think about this.</em> Now AI arrives before the conclusion exists. It shapes how the conclusion is expressed, and also how the mental model behind it is built: which evidence deserves attention, which alternatives are worth considering, what counts as a reasonable answer.</p><p>One caution. Usage categories describe what people ask, not how they divide the thinking. “Analyse these 10,000 rows, classify the responses and flag the anomalies” is heavy delegation, and it can be excellent Centaur work if the person still owns the question and the evaluation. The useful test is not how much you handed over. It is this: <strong>did I outsource effort, or did I outsource judgment?</strong></p><p>When AI moves from producing the answer to shaping the question, working with it becomes a problem of judgment, not just productivity. That changes how we should read the three archetypes.</p><h3>4. States, not identities</h3><p>The usual response to the research is to ask which one you are, or which one you should become.</p><p>Better to treat them as states, not identities.</p><p>Any serious piece of work changes character as it goes. Some stretches need breadth and speed. Some need distance and a decision. Some need two minds pushing against each other. A person who stays in one mode throughout is using the wrong tool for most of the job. The real skill is noticing when the work has changed and changing the division of thinking with it.</p><p>This is where the line between delegating and abdicating matters. A Shapeshifter may delegate as aggressively as any Self-Automator: fifty sources searched, thousands of observations classified, a dozen candidate explanations generated. What the Shapeshifter never hands over is the core: what problem we are solving, what counts as evidence, what would change the conclusion, and what we finally believe.</p><p><strong>Delegate like an Automator. Never abdicate like one.</strong></p><p>That discipline is not free, and it is not always worth paying for.</p><h3>5. When it’s worth the cost</h3><p>Every switch between modes requires reorienting. Stepping out of a productive exchange with the machine breaks the flow. For routine, low-stakes, reversible work, deliberate mode-switching is theatre. Delegate, check, move on.</p><p>It earns its cost when the problem is ambiguous, when the answer has consequences, when there are competing interpretations, when novelty matters, or when doing the work is itself how expertise gets built.</p><p>The last condition matters most, because AI can raise today’s output while reducing tomorrow’s capability.</p><p>In a randomised trial run by Anthropic, 52 software engineers learned an unfamiliar programming library, some with an AI assistant and some without. On a follow-up test, the AI group scored 50% and the hand-coding group 67%, and the AI group was not meaningfully faster. Within the AI group, how people used the assistant made all the difference. Those who delegated the code outright scored below 40%. Those who asked conceptual questions, or generated code and then worked to understand it, scored between 65% and 86%. The subgroups were small, but the direction is clear: whether you learn depends on how you use AI, not whether.</p><p>Experts are not immune. A study in <em>The Lancet Gastroenterology &amp; Hepatology</em> followed 19 experienced endoscopists, each with more than 2,000 colonoscopies behind them, as AI detection tools became routine. In procedures they performed <em>without</em> AI, their adenoma detection rate fell from 28.4% to 22.4% within months. That is a 20% relative decline in a skill they had spent careers building. The study is observational and other factors may have contributed, but the signal is hard to dismiss.</p><p>Efficiency can be bought by outsourcing the experiences through which expertise would have formed.</p><h3>6. The transformation you don’t notice</h3><p>Lost skill is at least measurable. The second risk is harder to see, because it feels like your own thinking.</p><p>In an experiment at Cornell with 1,506 participants, people wrote about social media with an AI assistant that had been quietly biased one way or the other. Those given a pro-social-media assistant were twice as likely to argue that position. In a separate survey afterwards, 45% of them said social media was good for society, against 35% of a control group. Few noticed they had been steered. The machine’s framing had become their own.</p><p>Anthropic’s analysis of 1.5 million real conversations found that severe cases of people ceding their judgment are rare, fewer than one in a thousand. But the pattern it describes is the subtle one: accepting a framing, a value judgment or a script without examining it, and only later recognising it as not quite yours.</p><p><strong>The danger isn’t becoming a Cyborg. It is not noticing when you have become one.</strong></p><p>That is why a “human in the loop” is not enough. What matters is building moments of independent judgment into the loop. Order matters too. In a small MIT study, 83% of people who wrote an essay with ChatGPT could not quote a sentence from it minutes later, against about 11% of those who wrote without it. But people who had written on their own first and only then used ChatGPT did not show the same drop. Their engagement rose. The study is a preprint and small. Still, it points to something practical: think first, then fuse.</p><p>What does that look like on a real decision?</p><h3>7. The Loop, on one decision</h3><p>Take a decision many people now bring to AI: <em>which electric car should I buy?</em></p><p><strong>Frame. Human first.</strong> Before asking anything, write down what matters: reliability, rear-seat comfort, five-year ownership cost. Write down your current leaning, and what would change your mind. Thirty seconds of this makes everything that follows yours.</p><p><strong>Delegate. Let AI do the running around.</strong> Specifications, prices, long-term reviews, owner forums, depreciation data. This is effort, not judgment. Hand it over without guilt.</p><p><strong>Reclaim. Step back as the Centaur.</strong> What in all that actually matters? Has the question itself changed? You may discover that home-charging access drives the cost more than any car’s sticker price, which reframes the decision entirely.</p><p><strong>Fuse. Think with it as the Cyborg.</strong> “Challenge my preference for Model X. What am I underweighting? Make the strongest case for the car I’m dismissing.” This is where AI earns its keep: not by answering, but by attacking your answer.</p><p><strong>Separate. Leave the conversation.</strong> Close the window and ask: <em>what do I now believe, and why?</em> If the answer arrives in the machine’s phrasing and not your own, notice that.</p><p><strong>Own.</strong> One test decides whether the work is yours: <strong>can you defend the conclusion without reopening the AI conversation?</strong> If not, ownership has not fully come back to you.</p><p>At any point, re-route. Three rules are enough:</p><p><strong>Delegate when the bottleneck is effort.</strong><br><strong>Go Cyborg when the bottleneck is options.</strong><br><strong>Go Centaur when the bottleneck is the decision.</strong></p><h3>8. Hold on through the transformations</h3><p>Menelaus did not win by refusing the lion or the serpent. He won by keeping his grip through every form until the truth came out.</p><p>The skill in working with AI is not resisting transformation. It is keeping enough continuity through each transformation that the work never stops being yours. Your purpose, your standards of evidence, your accountability for the call all have to hold while everything around them changes shape.</p><p><strong>Change form. Preserve core.</strong></p><p>We have spent the first era of generative AI learning to prompt machines. The next will ask for something harder: deciding which intelligence should be doing what, at which moment, and when control needs to move again.</p><p>The advantage is not the form. It is knowing when to transform, and knowing how to come back.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8a97de1edd45" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI: The Week, Interpreted (Issue #3)]]></title>
            <link>https://medium.com/@mail2rajivgopinath/ai-the-week-interpreted-issue-3-ae1dadcca397?source=rss-ae007d129200------2</link>
            <guid isPermaLink="false">https://medium.com/p/ae1dadcca397</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Rajiv Gopinath ]]></dc:creator>
            <pubDate>Sun, 01 Nov 2026 17:59:31 GMT</pubDate>
            <atom:updated>2026-11-01T17:59:31.842Z</atom:updated>
            <content:encoded><![CDATA[<h4>Weekly AI Synthesis · Issue #3 · 26 October–1 November 2026</h4><h4>The signals, tensions, and what they might mean.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nw8uLDeMJpYUEwfU1u5nTw.png" /></figure><h3>The Price of Intelligence</h3><p><em>Seven days of AI. The signals, tensions, and what they might mean.</em></p><h3>The week in 30 seconds</h3><p><strong>What changed:</strong> Anthropic’s IPO filing brought the financial realities of frontier AI into sharper focus. OpenAI and Anthropic reported potentially important mathematical advances. Google expanded its AI chip ecosystem, while agents moved further into commerce, scientific research and professional work. Meanwhile, new developments in monitoring and cybersecurity exposed the consequences of increasingly autonomous systems.</p><p><strong>Why it matters:</strong> AI is becoming more powerful, but its costs extend beyond computing infrastructure. They include financial risk, institutional change, the redistribution of work, the verification of knowledge and the question of who remains accountable when machines act.</p><p><strong>What I’m watching:</strong> Whether the enormous value anticipated from AI translates into sustainable businesses, dependable systems and benefits that extend beyond the companies developing the technology.</p><h3>My read of the week</h3><p>For much of the generative AI era, we have measured progress through capability. Models became better at reasoning, coding, creating and solving increasingly difficult problems. Each advance reinforced the expectation that sufficiently powerful intelligence would eventually transform almost every industry.</p><p>But this week, a different question kept surfacing.</p><p>What does all this intelligence actually cost?</p><p>Not simply the electricity, GPUs or billions of dollars required to produce it. Those costs are becoming increasingly visible. I mean the larger price: the investments organisations must make, the assumptions institutions may have to reconsider, the work people may no longer perform, and the responsibility that must remain somewhere when machines begin making consequential decisions.</p><p>Consider three developments that would ordinarily sit in different sections of a technology newsletter.</p><p>Anthropic’s IPO filing exposed the scale of the financial commitments supporting a frontier AI company. OpenAI and Anthropic reported advances challenging a longstanding assumption in theoretical computer science, raising questions about how scientific claims generated with AI should be verified. And companies across the commercial internet began preparing for a world in which AI agents, rather than consumers themselves, search, compare and transact.</p><p>One concerns capital, another knowledge and the third commerce. Yet all three reveal a similar transition.</p><p>AI is moving from a technology whose potential we debate to a technology whose consequences we must organise around.</p><p>That organisation has a price.</p><p>Someone finances the infrastructure. Someone validates the discovery. Someone takes responsibility for the transaction. Someone decides how productivity gains are shared, and someone absorbs the disruption when existing ways of working change.</p><p>These costs do not necessarily fall on the people or companies capturing the greatest value.</p><p>And perhaps that is what makes this week’s developments especially interesting.</p><p>The central question is no longer simply how intelligent machines can become.</p><p><strong>It is who pays for that intelligence, who captures its value and who bears its consequences.</strong></p><h3>Signal 01 — Frontier AI meets the discipline of public markets</h3><p>For most of its recent history, the frontier AI industry has been judged primarily through expectations. Better models, accelerating adoption, enormous valuations and the promise of substantial future markets have attracted extraordinary amounts of capital.</p><p>Anthropic’s October 26 IPO filing brought a different level of scrutiny.</p><p>The filing reported $16.3 billion in revenue for the first half of 2026, against $1.2 billion in the comparable period a year earlier. That represents exceptional growth. It also disclosed the scale and complexity of the company’s long-term infrastructure arrangements, including more than $800 billion in compute and data-centre commitments extending through 2045.</p><p>Those commitments should not be mistaken for immediate expenditure. Equally, accounting losses affected by convertible-note valuations should not automatically be interpreted as the recurring economics of its operations.</p><p>Nevertheless, the prospectus provides an unusually detailed view of what it takes to compete at the frontier.</p><p>The company is asking investors to value a business whose revenues are growing extraordinarily quickly, whose infrastructure requirements are immense and whose technological future remains difficult to predict.</p><p>There is also a governance dimension. Anthropic’s proposed ownership arrangements give its founders substantial voting control while retaining a distinctive role for its long-term benefit trust.</p><p>The public-market proposition therefore involves more than financial performance. It asks investors to accept unusual governance arrangements in a company developing technology with potentially far-reaching consequences.</p><p>The timing is significant.</p><p>During the same week, large technology companies experienced sharp share-price declines as investors reassessed the scale of AI capital expenditure. Financial Times reporting described nearly $900 billion in market value being erased from major technology companies during the sell-off.</p><p>That reaction does not establish that the investments are excessive. It suggests that capital markets are becoming more demanding about the relationship between spending and future returns.</p><p>The questions are becoming specific. How durable will demand for frontier intelligence prove to be? What happens when inference prices decline? How rapidly will computing equipment become obsolete? How much revenue must be generated to justify existing infrastructure commitments?</p><p>There is a historical pattern worth remembering.</p><p>Railways transformed economies without making every railway investor wealthy. Telecommunications created extraordinary economic value while leaving numerous operators burdened with infrastructure debts.</p><p>A technology can be enormously successful while some of the companies financing its expansion produce disappointing returns.</p><p><strong>The promise of AI does not remove the need to distinguish technological value from financial return.</strong></p><p>Frontier AI is beginning to be assessed through the same questions that eventually confront every major industry: capital intensity, cash flow, governance, competitive advantage and the sustainability of growth.</p><p><em>Sources: </em><a href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm"><em>Anthropic’s IPO filing</em></a><em> | </em><a href="https://www.ft.com/content/f6ecfecd-d797-4d2c-b427-d333297949cb"><em>Financial Times — Anthropic IPO</em></a><em> | </em><a href="https://www.ft.com/content/c695558e-11e6-494c-aeee-0f9fb5c75afc"><em>FT — Big Tech’s AI spending sell-off</em></a></p><h3>Signal 02 — The economics of compute is becoming a competitive battlefield</h3><p>The enormous cost of producing frontier intelligence is encouraging competition in the technologies required to deliver it.</p><p>One of the week’s more consequential developments came from Google.</p><p>Google and Oracle announced plans to make Google’s Tensor Processing Units available through Oracle Cloud Infrastructure beginning in 2027. Separately, The Information reported plans for OpenAI to purchase Google’s specialised AI processors.</p><p>This is more than another semiconductor announcement.</p><p>Nvidia’s GPUs have become the dominant foundation for training and running advanced AI systems. Google’s TPUs represent an alternative approach: processors designed specifically for machine-learning workloads and developed alongside the software and infrastructure needed to use them.</p><p>Google has employed these processors internally for years. Making them available through additional cloud providers could expand their commercial reach and provide customers with more choices.</p><p>It also produces some unusual competitive relationships.</p><p>Google develops Gemini, operates a cloud platform and designs its own chips. Yet it may supply infrastructure to companies competing with its own AI products.</p><p>Oracle strengthens its cloud offering by providing customers with additional hardware options. AI developers gain opportunities to diversify their computing infrastructure.</p><p>The larger development is that competition is extending across the entire system required to produce intelligence: chips, memory, networking, power, cooling and inference software.</p><p>Not every AI task requires the same infrastructure.</p><p>Training a frontier model, running a conversational assistant and executing millions of simple classification tasks have different computing requirements. The ability to match those workloads to appropriate infrastructure can materially change the cost of providing intelligence.</p><p>This is particularly important as AI becomes embedded in routine business processes.</p><p>A company may care about the reasoning quality of a model for a complex strategic task. For a high-volume operational workflow, however, latency, reliability and cost per successful outcome may matter just as much.</p><p><strong>The competition is no longer simply to build the most capable intelligence. It is also to produce useful intelligence at an economically sustainable cost.</strong></p><p>That distinction may become increasingly important as models become more widely available and the industry searches for durable commercial advantages.</p><p><em>Sources: </em><a href="https://cloud.google.com/blog/products/infrastructure/bringing-google-tpus-to-oracle-cloud-infrastructure"><em>Google Cloud and Oracle</em></a><em> | </em><a href="https://www.theinformation.com/briefings/oracle-start-running-google-tpus-cloud"><em>The Information — Google TPUs</em></a></p><h3>Signal 03 — AI is challenging mathematical assumptions. But who establishes what is true?</h3><p>Among the week’s most intellectually significant developments were announcements from OpenAI and Anthropic concerning the Exponential Time Hypothesis.</p><p>The hypothesis is a longstanding assumption in theoretical computer science about the computational difficulty of certain classes of problems. Researchers use it to establish conditional limits across areas including optimisation, graph theory, databases and computational biology.</p><p>Both companies reported AI-assisted results challenging the hypothesis.</p><p>OpenAI released a technical paper and accompanying formal verification work using Lean, a system for checking mathematical proofs. It also announced an independent examination involving specialists with deep connections to the underlying mathematical theory.</p><p>Anthropic separately published a Claude-assisted argument.</p><p>These claims deserve attention, but also considerable care.</p><p>A machine-checked proof can provide strong evidence that a formal conclusion follows from specified definitions and assumptions. Researchers must still determine whether the formal statement represents the original mathematical question accurately, whether the assumptions are appropriate and whether the claimed implications follow.</p><p>The experts involved in examining OpenAI’s work have emphasised that their participation should not be interpreted as certification of the result.</p><p>There is also a danger of confusing theoretical significance with immediate practical consequences.</p><p>Even if the Exponential Time Hypothesis were refuted, that would not automatically mean modern encryption had been broken or that all computationally difficult problems had become easy.</p><p>The implications depend on what has actually been established.</p><p>But even before the mathematical debate is resolved, something important is changing.</p><p>AI systems are becoming capable of generating mathematical arguments serious enough to require the attention of leading researchers.</p><p>That creates new questions for scientific institutions.</p><p>If machines can produce sophisticated conjectures, proofs and theoretical results at increasing speed, who verifies them? How is authorship attributed? What counts as independent confirmation? And what happens when the production of plausible discoveries exceeds the research community’s capacity to examine them?</p><p>Last week, we encountered a related issue in biological research, where AI dramatically accelerated hypothesis generation while experimental validation remained essential.</p><p>Mathematics presents a different version of the same challenge.</p><p><strong>AI may be accelerating the production of knowledge claims faster than the institutions responsible for establishing what we actually know.</strong></p><p>If these results withstand independent scrutiny, they may be remembered as major scientific advances.</p><p>But the process through which they are verified may become an equally important part of the story.</p><p><em>Sources: </em><a href="https://openai.com/index/building-scientific-understanding-with-academic-communities/"><em>OpenAI — Building Scientific Understanding</em></a><em> | </em><a href="https://github.com/openai/math/tree/main/preprints/The-Exponential-Time-Hypothesis-is-False-October-25-2026"><em>OpenAI’s technical preprint</em></a><em> | </em><a href="https://www.science.org/content/article/ai-shaking-math-can-its-proofs-be-trusted"><em>Science — Can AI proofs be trusted?</em></a></p><h3>Signal 04 — The commercial internet is beginning to serve AI agents</h3><p>A quieter set of developments this week may eventually prove as consequential for commerce as the arrival of search engines or online marketplaces.</p><p>AI agents are beginning to participate more directly in shopping and payment processes.</p><p>Shopify announced support for agentic payment technologies developed by Visa, Mastercard and American Express. These systems are intended to allow authorised AI agents to initiate purchases using mechanisms that preserve established payment protections.</p><p>Meanwhile, The Information reported that some publishers were experimenting with advertisements designed to be encountered by AI agents rather than human readers.</p><p>Amazon was also reportedly expanding its catalogue to include hundreds of millions of products available through rival retailers, partly in response to the emergence of AI shopping agents.</p><p>These developments are connected by a change in the commercial interface.</p><p>Traditionally, consumers visit websites, use search engines, browse marketplaces and encounter advertisements directly.</p><p>An AI agent introduces another possibility.</p><p>The consumer describes a need. The agent searches, compares products, evaluates alternatives and potentially completes a transaction.</p><p>The consumer may never see many of the pages or product listings considered during the process.</p><p>That changes how businesses compete for attention.</p><p>Retailers may need to make their products understandable to agents. Marketplaces may seek to ensure their inventory is included in agent searches. Payment providers must establish how an agent can transact with appropriate authorisation.</p><p>And advertisers may try to influence the machine that is supposed to be representing the consumer.</p><p>That raises a difficult question.</p><p><strong>If an AI agent is acting for the buyer, what happens when sellers begin competing to influence the agent?</strong></p><p>An agent that selects products according to a buyer’s requirements provides one kind of value.</p><p>An agent whose recommendations are shaped by undisclosed commercial arrangements may provide something quite different.</p><p>The next phase of digital advertising could therefore introduce questions about recommendation provenance, commercial conflicts of interest and the distinction between paid placement and independent advice.</p><p>For brands, the challenge may extend beyond conventional search visibility.</p><p>They will increasingly need to make their product information accessible, their claims verifiable and their differences from competitors understandable to intelligent systems.</p><p><strong>The next battle for consumer consideration may take place partly inside systems the consumer never directly sees.</strong></p><p><em>Sources: </em><a href="https://www.theinformation.com/briefings/shopify-adopts-agentic-payment-tech-visa-mastercard"><em>The Information — Shopify’s agentic payments</em></a><em> | </em><a href="https://www.theinformation.com/articles/publishers-adding-invisible-ads-ai-agents"><em>Invisible advertising for AI agents</em></a><em> | </em><a href="https://www.theinformation.com/articles/amazon-lists-250-million-products-rival-sites-effort-battle-agents"><em>Amazon’s response to shopping agents</em></a></p><h3>Signal 05 — AI is beginning to change the economics of professional work</h3><p>One of the more difficult questions surrounding AI is whether productivity gains will translate into better work, smaller workforces or entirely different organisational structures.</p><p>This week offered several reasons to examine that question more closely.</p><p>Financial Times analysis considered emerging evidence of rising productivity alongside declining employment in parts of the technology sector.</p><p>The pattern is noteworthy, but its causes remain difficult to separate.</p><p>Technology employment is affected by economic cycles, restructuring, outsourcing, financing conditions and changes in corporate strategy. It would be premature to attribute every reduction in headcount to AI.</p><p>Nevertheless, the possibility that organisations can increase output without increasing employment in the same proportion has significant implications.</p><p>Professional services provides a revealing example.</p><p>Legal AI systems are increasingly used to review documents, support due diligence, conduct research and help prepare legal work. Yet reporting this week also highlighted scepticism among lawyers about the reliability of AI agents in professional settings.</p><p>That scepticism is not necessarily resistance to technology. It reflects the realities of professional responsibility.</p><p>Legal work involves confidentiality, liability, professional judgment and standards of evidence that do not disappear when the underlying analysis becomes faster.</p><p>At the same time, new AI-native firms are exploring business models that could challenge traditional professional-service structures.</p><p>There are at least two possible paths.</p><p>In the first, existing firms use AI to perform the same work more efficiently. They may increase margins, reduce routine tasks or serve more clients.</p><p>In the second, AI changes the delivery model itself. Smaller teams, specialised agents, continuous services and different pricing structures may become viable.</p><p>The second possibility is more disruptive.</p><p>But it introduces a problem that deserves more attention.</p><p>If AI performs much of the analytical work previously assigned to junior professionals, how will those professionals develop the expertise required for senior judgment?</p><p>And if fewer people are involved in producing professional advice, who remains responsible for its consequences?</p><p>A related debate concerns management.</p><p>Some argue that AI agents could reduce the need for layers of coordination. Others suggest that managing a combination of humans and autonomous systems may create new requirements for supervision, prioritisation and accountability.</p><p>Both may be true under different circumstances.</p><p>AI can reduce the effort required to collect information and coordinate routine activities without eliminating the need to resolve conflicting objectives or make consequential decisions.</p><p><strong>AI may reduce the amount of work required to produce an answer while increasing the importance of deciding what to do with it.</strong></p><p>The economic consequences are unlikely to be distributed evenly.</p><p>Some roles may become more valuable. Others may change substantially. And traditional pathways through which people acquire professional experience could come under pressure.</p><p>The important question is not simply whether AI replaces jobs.</p><p>It is how the work required to create value is redistributed, and whether organisations are prepared for that change.</p><p><em>Sources: </em><a href="https://www.ft.com/content/7f27aa5f-5994-4f82-aed8-1cade18fb487"><em>FT — AI and the labour market</em></a><em> | </em><a href="https://www.ft.com/content/b25a880c-cd64-49d1-ba5f-7d077eb1adac"><em>FT — AI agents need middle managers too</em></a><em> | </em><a href="https://www.theinformation.com/articles/skeptical-lawyers-highlight-growing-pains-ai-agents"><em>The Information — Lawyers and AI agents</em></a></p><h3>Signal 06 — AI is moving closer to the physical world and scientific experimentation</h3><p>Google DeepMind’s announcement of Genie 4 illustrates how AI world models are progressing beyond the generation of static content.</p><p>The system can create interactive environments that respond to natural-language changes, incorporate audio and simulate movement and physical interactions.</p><p>DeepMind also demonstrated how Gemini models can guide activity within these environments, dividing tasks into steps and responding to changes as the simulation unfolds.</p><p>This is different from generating a convincing video.</p><p>A video depicts a sequence of events. An interactive world must respond to what happens next.</p><p>That makes world models potentially useful for training robots, testing agent behaviour, developing simulated environments and exploring designs before introducing changes into the physical world.</p><p>But the limitations remain important.</p><p>DeepMind acknowledges that simulated worlds can lose consistency, object permanence can drift and sustained interaction remains constrained.</p><p>The technology is promising, but it should not yet be mistaken for a reliable general-purpose representation of physical reality.</p><p>Elsewhere, Anthropic reported early progress using Claude in drug development, alongside expanded partnerships with scientific organisations.</p><p>The Information also reported that Anthropic had acquired a small drug-development startup as it moved more directly into biotechnology.</p><p>Together, these developments suggest that frontier AI companies are beginning to participate more deeply in specialised scientific and physical workflows.</p><p>They are moving beyond providing general-purpose models.</p><p>But this introduces a familiar distinction.</p><p>A model can be impressive at predicting how a molecule might behave. Scientific value emerges when the prediction leads to reproducible experimental results.</p><p>A world model may simulate a robot completing a task. Practical value emerges when a physical robot can perform that task reliably under changing conditions.</p><p>The difference is between generating plausible representations and producing dependable outcomes.</p><p><strong>AI is beginning to leave the screen, but the physical world remains the final evaluator.</strong></p><p>As AI moves toward scientific experimentation, industrial automation and embodied action, real-world verification will become increasingly important.</p><p><em>Sources: </em><a href="https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/"><em>Google DeepMind — Genie 4</em></a><em> | </em><a href="https://www.anthropic.com/research/early-progress-drug-development"><em>Anthropic — Drug Development</em></a></p><h3>Signal 07 — Greater autonomy is creating a larger accountability burden</h3><p>As AI agents become more capable, there is a tendency to measure progress through how much work they can complete without human intervention.</p><p>But autonomy creates another requirement: the ability to understand what an agent is doing while it acts.</p><p>MIT Technology Review examined this problem during the week, arguing that monitoring agent behaviour only after an incident may be inadequate.</p><p>The concern is straightforward.</p><p>An AI agent can interpret changing instructions, choose tools, respond to new information and take actions that were not individually specified in advance.</p><p>That flexibility is what makes agents useful.</p><p>It also makes their behaviour harder to predict.</p><p>An agent may complete its assigned task while taking actions that create unintended risks. The final result could appear successful even when the process involved inappropriate tool use, unauthorised data access or a weakened security control.</p><p>OpenAI’s updated Model Spec and work on detecting frontier-model misalignment reflect the increasing importance of defining expectations and observing behaviour.</p><p>Meanwhile, Epoch AI published research suggesting that capable open-weight models could exploit vulnerabilities during the interval between the availability of security-fix information and deployment of the corresponding patches.</p><p>That finding illustrates a broader challenge.</p><p>AI may accelerate the discovery and exploitation of vulnerabilities while defensive organisations remain constrained by testing, approval and deployment procedures.</p><p>The response cannot simply be to require a human to approve every machine action. That would eliminate many of the benefits of autonomy.</p><p>Instead, organisations may need to distinguish between actions agents can take independently, actions requiring additional authorisation and actions that must remain outside an agent’s control.</p><p>They will also need independent monitoring, intervention mechanisms and reliable recovery processes.</p><p>There is another dimension to this debate.</p><p>MIT Technology Review also questioned whether excessive attention to catastrophic AI scenarios risks distracting institutions from more immediate and widespread harms.</p><p>That criticism matters.</p><p>A mature approach to governance must address both frontier risks and the ordinary consequences of deployment: inaccurate decisions, privacy violations, discriminatory outcomes, professional negligence and failures of consumer protection.</p><p><strong>The more authority we delegate to machines, the more precisely we need to define where responsibility remains.</strong></p><p>Autonomy does not eliminate accountability.</p><p>It changes the mechanisms through which accountability must be exercised.</p><p><em>Sources: </em><a href="https://www.technologyreview.com/2026/10/29/1147824/monitor-ai-agents-actions-real-time"><em>MIT Technology Review — Real-time agent monitoring</em></a><em> | </em><a href="https://openai.com/index/updated-model-spec"><em>OpenAI — Updated Model Spec</em></a><em> | </em><a href="https://epoch.ai/publications/open-weight-models-patch-gap"><em>Epoch AI — The patch gap</em></a></p><h3>The contradiction of the week</h3><h4>We are making intelligence more accessible while making its consequences harder to distribute fairly.</h4><p>This week’s developments reveal an uncomfortable tension.</p><p>AI companies are producing extraordinary technological advances, but sustaining those advances requires enormous amounts of capital.</p><p>Businesses expect productivity gains, but those gains may alter employment, professional development and the economic value of established skills.</p><p>AI agents could make commerce more convenient, but new intermediaries will have incentives to influence what those agents discover and recommend.</p><p>Scientific systems could accelerate discovery, while responsibility for verification and authorship becomes more complicated.</p><p>And autonomous systems promise to reduce the need for human supervision while increasing the importance of independent accountability.</p><p>None of these tensions means that AI progress is undesirable.</p><p>They mean that technological capability and its social or economic value are not interchangeable.</p><p>The benefits of a technology do not necessarily accrue to the people or institutions bearing the costs of producing and accommodating it.</p><p>An organisation may benefit from lower labour costs while employees face disruption.</p><p>Consumers may enjoy greater convenience while commercial intermediaries gain influence over recommendations.</p><p>AI developers may increase the availability of intelligence while infrastructure providers assume substantial financial risks.</p><p>And society may benefit from scientific advances while research institutions carry the burden of verifying an increasing volume of claims.</p><p>The question is not simply whether AI creates more value than it costs.</p><p><strong>It is where that value appears, where the costs fall, and whether the institutions connecting the two are capable of adapting.</strong></p><h3>One idea I’m taking away</h3><h4>The Distribution of Intelligence</h4><p>I have been thinking about a distinction that receives less attention than it deserves.</p><p>We often discuss intelligence as though it were a single resource.</p><p>More intelligence becomes available. Organisations use it. Productivity improves. Economic value follows.</p><p>But the process is far less direct.</p><p>There are at least three different distributions involved.</p><p>The first is the <strong>distribution of capability</strong>. Who can access powerful intelligence, under what conditions and at what price?</p><p>The second is the <strong>distribution of value</strong>. Who benefits economically when that intelligence is applied?</p><p>The third is the <strong>distribution of responsibility and cost</strong>. Who finances the infrastructure, absorbs workplace disruption, bears the consequences of mistakes or remains accountable for decisions?</p><p>These distributions need not align.</p><p>A company may develop extremely valuable AI capabilities without capturing all the value those capabilities create.</p><p>A business may realise productivity gains without passing them on to employees or customers.</p><p>A consumer may benefit from an AI agent while remaining unaware of the commercial incentives influencing its recommendations.</p><p>And research institutions may become responsible for validating AI-generated discoveries without possessing the resources required to evaluate them at the rate they are produced.</p><p>This distinction matters because discussions about AI frequently move too quickly from capability to benefit.</p><p>A model becomes more intelligent, and we assume that society becomes better off.</p><p>The relationship may be positive over time, but the path is neither automatic nor evenly distributed.</p><p>Consider an AI system capable of performing much of the analytical work previously carried out by a professional team.</p><p>That capability could reduce costs for clients, increase margins for the firm, improve service quality or make specialist advice available to people who previously could not afford it.</p><p>It could also reduce employment or weaken the training opportunities through which future professionals develop expertise.</p><p>All are plausible outcomes of the same technological capability.</p><p>Which outcomes emerge will depend on competition, regulation, organisational decisions, professional norms and the business models adopted.</p><p>The same applies to AI infrastructure.</p><p>Cheaper inference may benefit developers and consumers while placing pressure on companies that invested heavily in expensive computing capacity.</p><p>The technology can become more useful even as the economics deteriorate for some of the businesses supporting it.</p><p><strong>The distribution of intelligence is not the same as the distribution of its benefits.</strong></p><p>And I suspect this distinction will become increasingly important as AI moves from demonstrations of capability into the ordinary functioning of economies and institutions.</p><h3>Connecting the Dots: The Story So Far</h3><p>Three weeks into this series, I am beginning to see a progression in the questions that matter.</p><p>In the first week, the central observation was that AI was moving from <strong>capability to authority</strong>. As agents become capable of carrying out increasingly complex work, the question is no longer simply what they can do. It is what they should be allowed to do, under what conditions, and how they might earn greater autonomy through demonstrated reliability. I called this <em>Progressive Delegation</em>.</p><p>In the second week, a different problem came into focus. Even when intelligence improves dramatically, the benefits do not necessarily follow at the same pace. Scientific hypotheses still need validation. Business recommendations still need implementation. Autonomous systems still need controls. The limiting constraint often moves elsewhere in the process. That was <em>Bottleneck Migration</em>.</p><p>This week, the financial and institutional consequences have become more visible. The capital required to produce frontier intelligence is extraordinary. Scientific advances are challenging established processes for verifying knowledge. Agents are entering commercial transactions, while companies are being forced to reconsider how work, accountability and value are distributed.</p><p>That brings us to <em>The Price of Intelligence</em>.</p><p>There is a connection between all three.</p><p><strong>Capability creates the possibility of delegation. Delegation exposes constraints in the surrounding system. Removing those constraints requires investment, institutional change and decisions about who bears the cost and who captures the value.</strong></p><p>What began as a question about how intelligent machines might become is gradually turning into a question about how societies, businesses and institutions reorganise around increasingly available intelligence.</p><p>And perhaps that is the larger story worth following.</p><p>Not simply how fast AI is advancing, but how the rest of the world changes in response.</p><h3>What I’m watching next</h3><p>First, whether Anthropic’s proposed listing brings greater transparency to frontier AI economics, particularly the relationship between revenue growth, infrastructure commitments and profitability.</p><p>Second, whether the theoretical computer-science claims from OpenAI and Anthropic withstand independent examination. The process of verification may itself become an important test of how scientific institutions respond to AI-assisted discoveries.</p><p>Third, whether Google’s expansion of TPU availability materially changes the economics of AI computing as customers seek alternatives for different workloads.</p><p>Fourth, whether agentic commerce develops clearer standards for authorisation, recommendation transparency and conflicts of interest.</p><p>Fifth, whether evidence of AI-related productivity gains begins to connect more clearly with changes in employment, wages, organisational structures and professional development.</p><p>And finally, whether independent monitoring and practical safeguards mature quickly enough to support increasingly autonomous systems.</p><h3>Until next week</h3><p>We have spent the last few years watching intelligence become more capable.</p><p>Now we are beginning to discover what follows from that capability.</p><p>There will be economic winners and losers, new forms of competition, scientific discoveries that challenge established assumptions, and institutions that must adapt to responsibilities they were never designed to carry.</p><p>Some of those changes will be welcome. Others will be difficult. Many will unfold unevenly.</p><p>And perhaps the most important question is no longer simply how intelligent AI can become.</p><p>It is what kind of world we are prepared to build around that intelligence, who gets to shape it and who ultimately benefits.</p><p><strong>The price of intelligence may turn out to be far more than the cost of producing it.</strong></p><p><em>AI: The Week, Interpreted is my weekly attempt to look beyond the announcements and understand the forces, tensions and implications shaping the direction of artificial intelligence.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ae1dadcca397" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Chapter 34: Counting the Same Sale Twice]]></title>
            <link>https://medium.com/@mail2rajivgopinath/chapter-34-counting-the-same-sale-twice-2980ba175502?source=rss-ae007d129200------2</link>
            <guid isPermaLink="false">https://medium.com/p/2980ba175502</guid>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[effectiveness]]></category>
            <category><![CDATA[market-mix-modeling]]></category>
            <dc:creator><![CDATA[Rajiv Gopinath ]]></dc:creator>
            <pubDate>Thu, 29 Oct 2026 06:52:39 GMT</pubDate>
            <atom:updated>2026-10-29T06:52:39.271Z</atom:updated>
            <content:encoded><![CDATA[<h4>Marketing Effectiveness Series -The Science of Evidence</h4><h4>A sale that moves from a shop to an app is not growth. A sale recorded by two systems is not two sales. Before you can ask what caused the growth, you have to agree what the growth was.; Chapter 34 in the Marketing Effectiveness Series</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*01scBw3lTX9iRyzoGrShPw.png" /></figure><p>Everything in this series so far has been an argument about the right-hand side of the equation.</p><p>Which channel did the work. What shape the response curve takes. Which variable is a cause and which is just along for the ride. Whether the thing you controlled for was a confounder or a mediator. Chapter after chapter of careful argument about the right-hand side.</p><p>Nobody asked where the number on the left came from.</p><p>Ask, and it comes apart in your hands.</p><p>Offline sales come from a retail audit, which is a sample. Online orders come from a commerce platform, which is complete but only for itself. The website analytics says something different again, and it is lower, because of consent banners and payment redirects and phone settings nobody in marketing controls. One system counts an order when it is placed. Another counts it when it is delivered. A third counts it after the returns come back six weeks later.</p><p>All of them are called sales.</p><p>Now put two of them in the same meeting. The e-commerce team reports that online grew forty per cent and that media drove a good share of it. The mix model reports that media drove a good share of offline sales too. Each is right about its own slice. Added together, they claim more new business than the company actually did.</p><p>Underneath that sits a quieter problem, and it is the expensive one. Some of that online growth was not new. It was the same household, buying the same product, from an app instead of the shop down the road. The app grew. The brand stayed where it was.</p><p><strong>A measurement system can be causally sophisticated and commercially incoherent at the same time.</strong></p><p>That is the sentence I would put above the desk of anyone building one. You can do everything the last ten chapters asked. Proper counterfactuals. Honest priors. Experiments where they matter. Credit assigned to the channel that created the demand rather than the one that collected it. And you can still end up with a set of numbers that does not describe a business that exists.</p><p>Because all of that work happens on the right-hand side. None of it checks whether the totals add up.</p><p>This chapter is about making them add up — and about why that is a smaller achievement than it sounds.</p><h3>Section 1: The Brand That Grew Twice</h3><p>Here is the point, stated first: two correct answers, added together, can describe a business that does not exist.</p><p>Here is the meeting. A packaged foods brand, selling offline through general and modern trade, and online across marketplaces and quick commerce. The numbers that follow are illustrative, but the shape of them is not.</p><p><strong>What the business actually did</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/746/1*ixO1EijCaRg2Mlt3hFzfBA.png" /></figure><p>The e-commerce team presents first. Online is up forty per cent, which is true, and their analysis credits media with ₹12 crore of that rise. The numbers come from the platforms and the order ledger, and nobody has reason to doubt them.</p><p>Forty minutes later, the mix model is presented. It says media contributed ₹28 crore to offline sales. Also carefully done — three years of weekly data, seasonality controlled, distribution in the model, a proper response curve.</p><p>Both presentations are good. Neither analyst has done anything wrong.</p><p>Then somebody puts the two claims on one page, which in most businesses nobody does.</p><p><strong>What media was credited with</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/844/1*6_-A9KYuWRR6ax2XB5xEvA.png" /></figure><p><strong>A sale recorded twice is still one sale.</strong></p><p>Now the careful version, because this is the point where people overcorrect and get it wrong in the other direction.</p><p>Claims exceeding growth is not by itself proof that anybody overclaimed. It is entirely possible for media to have delivered ₹40 crore of value while other forces took ₹22 crore away. Distribution may have slipped. A competitor may have launched. Price may have moved. A net growth figure is the sum of everything pushing up and everything pulling down, so a large media contribution can sit inside a small net number quite legitimately.</p><p>So the table above does not prove an error. It proves that nobody has reconciled.</p><p>And from what was presented in that meeting, nobody can — because three different problems are tangled together in those four rows.</p><p><strong>The first is scope.</strong> The two analyses cover overlapping ground. Both media plans reached the same households. A television campaign that lifted offline sales also sent people to the app, where the platform counted the order and credited whichever ad was nearest. Neither system knows what the other has already claimed.</p><p><strong>The second is the one that catches good teams</strong>, and you can see it in the right-hand column. One number is a share of growth. The other is a share of a total. Different questions, different denominators, added together because both are in rupees crore and look like they belong in the same column. This happens constantly, in decks that are otherwise rigorous.</p><p><strong>The third is what the rest of the chapter is about.</strong> Look again at the offline change: minus ₹2 crore. Against ₹350 crore that is noise, and everyone in the room read it as noise. But nobody asked the question underneath it: of the ₹20 crore that arrived online, how much was new business, and how much was the same households buying the same packets from an app instead of the shop down the road?</p><p>If most of it was new, the brand had a good year and media helped. If a meaningful share was movement, then the online media credited with that growth was partly paid for relocating sales the brand already had.</p><p>Nothing in either analysis can tell you which. Not through carelessness, but because neither system can see outside its own channel, and the question lives in the gap between them.</p><p>Earlier in this book I argued for a measurement architecture — a system that routes each question to the evidence best equipped to answer it. That was the right ambition and I stand by it. But routing was the whole of it. An architecture that sends the offline question to the mix model and the online question to the platform has done its job, and it holds no view at all on whether the two answers can both be true.</p><p>Making the answers add up was never part of the design.</p><h3>Section 2: What Counts as a Sale?</h3><p>Here is the point, stated first: the number on the left of the equation is not a measurement. It is an assembly.</p><p>Follow one sale through the system and you can see it.</p><p>A household orders a ₹60 pack of biscuits on a quick-commerce app on Tuesday evening, having seen an ad for it two hours earlier. One packet. One household. One transaction.</p><p>Now count it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m6XdPbG70-rsvtcHH-0NJg.png" /></figure><p>Five systems. One packet of biscuits. Five different answers, ranging from ₹60 to ₹52 to zero, recorded on three different dates, in two different units.</p><p>And if the household returns it on Friday, some of those systems reverse the sale, some reverse it next month, and some never do.</p><p>All five numbers get called sales.</p><p><strong>Definition is the first disagreement.</strong> Gross orders or net of returns. Gross merchandise value or net revenue after fees and discounts. Sell-in to the trade or sell-out to the shopper. Each is a legitimate way to count. None of them is the same as the others, and a model built on one will answer a different question from a model built on another.</p><p>Returns deserve a line of their own, because this is where a quiet assumption rewrites the outcome variable. If your analysis applies a flat return rate — five per cent, say, because somebody used five per cent once — you have not adjusted the data. You have redefined what a sale is, by an amount nobody has checked. Return rates differ enormously by category, by platform, and between a festive sale and an ordinary week. Use a sourced rate for your category, or state a band and show the decision at both ends. A flat assumption is a decision disguised as housekeeping.</p><p><strong>Coverage is the second disagreement.</strong> A retail audit is a sample, extrapolated. A retailer’s point-of-sale feed is complete for that retailer and blind to the rest. A marketplace ledger is complete for that marketplace. Each one is accurate within its own boundary and silent beyond it, and nothing in the output tells you where the boundary is.</p><p><strong>Then the source of truth.</strong> For transactions, the order or transaction ledger should outrank web analytics, and it is worth being clear about why. Analytics platforms lose sales to consent banners, to payment gateway redirects that break the session, and to device-level tracking restrictions. The losses are not random, so the gap between the two systems is not noise you can average away.</p><p>The useful way to hold this: <strong>analytics platforms are behavioural instrumentation, not a financial ledger.</strong> They are the best thing you have for understanding what people did on the way to a purchase. They are the wrong thing to use for how much was sold.</p><p>Now the line that should govern every argument of this kind, and it is the most useful sentence in this chapter.</p><p><strong>Two systems can disagree because one of them is wrong, because both are noisy, or because they are measuring different things. Reconciliation begins by knowing which of the three you are dealing with.</strong></p><p>Almost every measurement argument I have sat in is conducted as though it were always the first. Someone’s number is wrong and the meeting is about whose. Hours go into finding the error.</p><p>Often there is no error. The platform is counting conversions and the finance team is counting net revenue and the audit is estimating sell-out, and all three are doing exactly what they were built to do. The disagreement is real, permanent, and not anybody’s fault.</p><p>You cannot fix that by finding the mistake. You fix it by deciding, in advance and in writing, which number the business is going to call sales — and then holding every claim to that definition.</p><p>Which raises the harder question. Once you have agreed what a sale is, who is allowed to claim it?</p><h3>Section 3: Three Levels of Incrementality</h3><p>Here is the point, stated first: a sale can be genuinely incremental at one level of the business and not incremental at all one level up.</p><p>Take the biscuit brand again. It spends ₹1 crore on sponsored product ads on a large marketplace. The platform reports ₹8 crore of attributed sales. Is that incremental?</p><p>The honest answer is that the question is incomplete, because there are four different versions of it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pso0rGd6VC4HT1UHpUAN0w.png" /></figure><p>Read down that table and the numbers fall away at every step. The campaign is real at the top and close to invisible at the bottom. Nothing about the campaign changed. Only the question did.</p><p><strong>Incremental to the platform is not necessarily incremental to the channel. Incremental to the channel is not necessarily incremental to the brand.</strong></p><p>Now the part that stops this being a counsel of despair, because there is a version of this argument that goes too far.</p><p>It would be wrong to conclude that only the brand level matters. Winning share on a marketplace is a genuine commercial objective. If you are losing the search results page to a competitor on your own category terms, that is a real problem and the platform-level number is the right number to judge the fix by. A listing that nobody finds is a listing that eventually gets delisted.</p><p><strong>The level that matters is set by the decision you are making, not by a principle.</strong></p><p>Deciding whether to defend your position on a marketplace? Platform level is correct. Deciding how much of the total marketing budget goes to commerce media? Brand level is the only honest answer. Deciding whether the category is worth investing in at all, which is a question a market leader sometimes has to ask? Category level.</p><p>The error is never that somebody measured at the wrong level. It is that a number measured at one level was used to answer a question at another, with no translation in between and nobody noticing that one was needed.</p><p>And that single observation explains both problems from the opening meeting.</p><p><strong>Duplicate credit</strong> is two systems reporting at two levels and somebody adding them. The platform reported platform-level incrementality, which is its job and all it can see. The mix model reported something closer to brand level, which is its job. Added together, they are not describing two sales. They are describing one sale from two heights.</p><p>This is where the distinction between the measurement systems earns its keep. Journey-based attribution reconstructs the sequence — who introduced, who assisted, who closed — but it has no counterfactual anywhere in it, so none of its numbers are incrementality at any level. A mix model infers incrementality but cannot see sequence. Only an experiment manufactures a real counterfactual, and it does so at whichever level the experiment was designed around. Three systems, three different things, and only one of them is answering the question people think they are all answering.</p><p><strong>Migration mistaken for growth</strong> is the same failure, in the other direction. The ₹20 crore that arrived online was a channel-level fact. It was read as a brand-level one. The sale happened, in that channel, and some of it was not new to the brand at all.</p><p>If channel migration is treated as net-new growth, online media can be materially over-credited. How large that error gets depends on how much of the observed growth was substitution rather than real brand or category growth — and that proportion is not fixed, not knowable from the sales data alone, and not stable from one year to the next.</p><p>I want to be careful here, because the temptation is to put a number on it and I do not have one that would survive contact with your business. What the next section can show you is what the arithmetic looks like once somebody is forced to state it.</p><p>What you can take now is the discipline, and it costs nothing. Before anybody estimates anything, write down which level of incrementality the decision requires. One line, agreed in the room.</p><p>Most measurement disputes I have seen would not have happened if somebody had written that line down first.</p><h3>Section 4: Making the Numbers Reconcile</h3><p>Here is the point, stated first: reconciliation is an accounting discipline, and almost none of it is statistical.</p><p>Before the steps, a correction. Last chapter ended with a quick check: add up every incremental claim in the business, compare it with how much the business grew, and if the claims exceed the growth, something is being paid twice.</p><p>That check is too crude, and Section 1 showed why. Positive contributions can legitimately exceed net growth when other forces were pulling the other way. A gap tells you nobody has reconciled. It does not tell you anybody has overclaimed. The real discipline is six steps rather than one subtraction.</p><p><strong>One. Write down what a sale is, and where it comes from.</strong></p><p>Gross or net of returns. Net of platform fees and funded discounts, or not. Sell-in or sell-out. Which system is the ledger of record, and which systems are instrumentation. One paragraph, agreed before anybody opens a model.</p><p>This sounds like housekeeping and it is the single most skipped step in measurement. Every downstream disagreement traces back to it.</p><p><strong>Two. Write down which level of incrementality the decision needs.</strong></p><p>Platform, channel, brand or category. From the previous section: the right level is set by the decision, not by principle. Defending share on a marketplace is a platform-level question. Setting a total marketing budget is a brand-level question. A campaign can score well at one and neutral at another, and both readings can be honest.</p><p>One line in the brief. This step prevents more bad decisions than any modelling choice in this chapter.</p><p><strong>Three. Put every claim on the same basis — growth or level — before you add anything.</strong></p><p>This is the step that does the most work and gets the least attention.</p><p>A mix model’s contribution is usually a share of the <em>level</em> of sales. It answers: of what we sold, how much needed media. A platform’s or an e-commerce team’s number is usually a share of <em>growth</em>. It answers: of what went up, how much did we cause.</p><p>Both are useful. Neither can be added to the other, and converting between them is not a formatting exercise. A channel’s contribution to growth is the change in its contribution, not the contribution itself.</p><p><strong>Four. Reconcile the channels to total brand sales — not the media contributions to net growth.</strong></p><p>This is the correction that matters most, and it is a different constraint from the one I gave you last week.</p><p>Total brand sales equals online plus offline, assuming the two together cover everything. That identity always holds, which makes it a real constraint rather than a hopeful one. Migration then has a natural home: plus X in online, minus X in offline, zero at the brand line.</p><p>Constraining channel outcomes to the brand total is sound. Constraining media contributions to be smaller than net growth is not, because net growth is a residue of everything, not a ceiling on any one thing.</p><p><strong>Five. Give every claimed effect somewhere to live.</strong></p><p>Base. Positive drivers. Negative drivers. Migration between channels. Residual. All of it reconciling exactly to the observed outcome, under the definition you agreed in step one.</p><p><strong>Every claimed effect needs somewhere to live in the P&amp;L of the outcome.</strong> A number that sits outside the identity is not evidence yet. It is an assertion with a decimal point.</p><p>One warning, because this is where a reconciliation can become theatre. An identity with a free residual always balances. If the unexplained term is allowed to absorb whatever disagreement remains, you have not reconciled anything — you have labelled the gap and moved on. Report the residual as a share of the outcome, and set a threshold above which the reconciliation is considered failed rather than completed. A model whose residual quietly carries a fifth of the growth has told you something important about itself.</p><p><strong>Six. Estimate migration rather than assuming it away.</strong></p><p>The honest ways, in descending order of strength. Loyalty or panel data where the same household appears in both channels, which lets you observe switching directly. Category-level comparison, where you check whether the category grew by as much as your online sales suggest it should have. And where you have neither, a stated band — twenty to sixty per cent, say — with the decision shown at both ends. If the decision flips between the ends of the band, you have found something worth paying to measure.</p><p>What you must not do is leave it at zero by default, which is what every system does when nobody raises it. And where the stakes justify it, test it: a set of markets where you push online hard, matched against markets where you do not, watching what happens to offline in both. That is the only clean read on substitution, and for a brand with meaningful commerce spend it is worth more than another year of modelling.</p><p>Here is what it looks like when it is done. Same brand, same year, same ₹18 crore.</p><p>Start with step three, because it changes the picture before anything else does.</p><p>The mix model said media contributed ₹28 crore to offline sales. That is a contribution to the level. To enter a growth identity it has to be restated as a change: last year the same model put offline media’s contribution at ₹26 crore. So offline media’s contribution to <em>growth</em> is ₹2 crore, not ₹28 crore.</p><p>Nothing about the ₹28 crore was wrong. <strong>Media was holding up ₹28 crore of the offline business and adding ₹2 crore to it.</strong> Those are two different jobs with two different numbers, and the meeting in Section 1 had no language for the difference.</p><p>On the online side, the e-commerce team’s ₹12 crore was a growth number already. But ₹7 crore of the online rise turns out to be households that moved from offline — which is not brand growth at all. Online media’s brand-level contribution to growth comes to ₹5 crore.</p><p><strong>The brand-level reconciliation</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/966/1*PLkN1yvJrpq1vsSMhlxXew.png" /></figure><p>Media’s brand-level contribution to growth is ₹7 crore. The meeting in Section 1 had ₹40 crore on the table.</p><p>Nothing was fabricated in those presentations. The ₹40 crore was a level number and a growth number added together, with migration left at zero because nobody raised it. Forcing everything onto one basis and into one identity is what brings it to a figure the business can actually have.</p><p>The residual is ₹1 crore, around six per cent of the growth being explained. Small enough to accept. Had it come in at ₹4 crore, the right response would be to say the reconciliation failed rather than to present it with a footnote.</p><p><strong>The channel reconciliation, with migration visible</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/840/1*LvYkhFHRrzdKJjZRJhio-w.png" /></figure><p>Offline was presented as a business in slight decline, down ₹2 crore. Once the ₹7 crore that moved to the app is placed where it belongs, offline grew by ₹5 crore on its own terms. It was not shrinking. It was growing, and losing a share of its own sales to another channel of the same brand.</p><p>That is a completely different conversation from the one the meeting was having — about trade investment, about general trade coverage, about whether the offline business is structurally in trouble. None of it was visible until the two channels were forced to reconcile to one total.</p><p>The migration figure is the one number here I have had to invent, because the brand had no way to estimate it. Which is precisely the problem. Everything else in the reconciliation came from systems the business already owned.</p><h3>Section 5: What Reconciliation Cannot Tell You</h3><p>Here is the point, stated first: a set of accounts can balance perfectly and still be wrong in every line.</p><p>Everything in the last section was arithmetic. It forces the claims into one identity, and once they are in there, you can see which ones were double-counted, which were measured at the wrong level, and which were on the wrong basis entirely. That is worth doing and almost nobody does it.</p><p>But notice what it does not touch. The ₹7 crore of media contribution in that reconciliation is still an estimate from a model, and the previous chapter was an argument about how badly such an estimate can be wrong. If branded search was sitting in that model soaking up demand that television created, the reconciliation will cheerfully accommodate it. The total will be right. The split inside it will be wrong, and the identity has no opinion on the matter.</p><p><strong>Reconciliation makes the numbers coherent. It does not make them causal.</strong></p><p>Two systems can agree and both be wrong together. That is, if anything, a more dangerous state than visible disagreement, because disagreement at least starts an argument.</p><p>Then there is the question of what reconciliation can see at all. Substitution comes in three forms and the accounting only catches one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/956/1*cxo3njD-k8yY30AiGCs4Sg.png" /></figure><p>Cross-channel is the one the last section handles, because online plus offline must equal the brand total and the migration has nowhere to hide.</p><p><strong>Intertemporal cannibalisation</strong> is harder, and it is the one that quietly funds a lot of promotional activity. A discount that persuades a household to buy three months of biscuits this week has not created demand. It has moved it forward, and every system in the business will record growth. The channel identity still balances. The brand total still balances. Nothing is double-counted. The sale was simply borrowed from a quarter that has not happened yet, and you find out when that quarter arrives and nobody can explain the softness.</p><p>This is the same mechanism the book described much earlier as borrowing from the future, arriving here in a form that reconciliation cannot detect.</p><p><strong>Product cannibalisation</strong> is invisible for a different reason. If your outcome variable is brand sales, a household switching from your ₹10 pack to your ₹20 pack, or from one variant to another, is simply a brand sale either way. The identity holds and the shift never surfaces. This is part of why modelling at SKU and geography level is not a refinement but a requirement — a brand-level outcome has thrown away the information before the analysis starts.</p><p>One more caution before the handoff. Migration rates are not constants. They move with delivery coverage, with distribution gains and losses, with the price gap between channels, and with whatever the platforms are discounting this quarter. A rate you estimated last year is a starting assumption for this year, not a parameter. If you treat it as settled, you will reconcile this year’s numbers to last year’s behaviour and the accounts will balance anyway.</p><p>So: you now know how much the business actually grew, and how much of that growth can credibly be attributed to marketing at the level the decision requires.</p><p>Which leaves the question the whole book has been circling. Marketing did not produce that growth out of nothing. It worked through something — a brand that people already recognised, trusted, and were willing to pay a little more for. How much of the growth came from the activity, and how much from the asset the activity was working with?</p><p>That is the next chapter.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2980ba175502" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Logs Were There. The Answer Wasn’t.]]></title>
            <link>https://medium.com/@mail2rajivgopinath/the-logs-were-there-the-answer-wasnt-4ebcbe4a4d8a?source=rss-ae007d129200------2</link>
            <guid isPermaLink="false">https://medium.com/p/4ebcbe4a4d8a</guid>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Rajiv Gopinath ]]></dc:creator>
            <pubDate>Tue, 27 Oct 2026 12:55:24 GMT</pubDate>
            <atom:updated>2026-10-27T12:55:24.657Z</atom:updated>
            <content:encoded><![CDATA[<h4>THE PLUMBING PROBLEM · PART 7</h4><h4>Why enterprise AI needs observability, not just logs</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0vER_bVZILzPBX6FKKSrqQ.jpeg" /></figure><p>An agentic workflow fails halfway through a customer process.</p><p>The model has its logs. The orchestration layer has another set. The API gateway has recorded the calls. The tools have their responses. Somewhere else, the underlying systems have captured what actually changed.</p><p>There is no shortage of data.</p><p>And yet, when somebody asks the obvious question — what actually happened? — the answer is surprisingly difficult to reconstruct.</p><p>One system shows that the agent called the right tool. Another shows that the request timed out. The agent believes the action failed. The downstream system appears to have processed it. A second agent then made a decision using the resulting state. By the time the problem becomes visible, the evidence is scattered across several places and no single record tells the whole story.</p><p>This is easy to underestimate because we tend to equate observability with logging. If every model call, API request and tool invocation has been recorded, surely the system is observable.</p><p>But recording activity and understanding an execution are not the same thing.</p><p>The harder questions begin after the logs arrive.</p><p>Which agent did what? What information did it rely on? Which tool actually changed the world? Which later action depended on that change? Was the relationship causal, or did one event simply happen after another? And can we trust the trace we are using to answer any of those questions?</p><p>I think of that distance as <strong>The Observability Gap</strong>: the distance between having records of agent activity and being able to reconstruct, with enough confidence to act, what actually happened and why.</p><p>Because in an agentic system, the problem is rarely that nothing was recorded.</p><p><strong>The logs were there. The answer wasn’t.</strong></p><h3>1. Logs are not observability</h3><p>The first mistake is to assume that if a system records enough events, it must be observable.</p><p>That sounds reasonable. An agent calls a model, invokes a tool, writes to a system, hands work to another agent and receives a response. If each step leaves a log, surely we can reconstruct the story later.</p><p>But that is exactly where the gap begins.</p><p>The work on <strong>The Observability Gap in AI Agents</strong> makes the point clearly. Its assessment of mainstream agent frameworks finds that important enterprise signals are still inconsistently captured: persistent agent identity, permission changes during a session, memory updates, details of human intervention and other context needed to understand who acted, under what authority and with what state.</p><p>So you can have thousands of perfectly valid events and still be missing the information that gives those events meaning.</p><p>A tool-call record may tell you that an action was attempted. It may not tell you whether the agent was authorized to take it, whether another agent delegated it, whether a human approved it or whether a memory update earlier in the workflow changed the decision.</p><p>The opposite problem appears in <strong>ORCA-bench</strong>. There, agents are given access to logs, metrics, traces and source code in a production-style incident environment. Even with that telemetry available, root-cause diagnosis remains difficult, particularly when several causes overlap or noisy symptoms point in the wrong direction.</p><p>That is the distinction worth holding on to.</p><p><strong>Telemetry gives you records. Observability gives you an explanation.</strong></p><p>The second does not automatically follow from the first.</p><h3>2. First know what state the agents are actually in</h3><p>Before we ask why an agent did something, there is a more basic question.</p><p>What state is it in now?</p><p>That sounds trivial until several agents are working together. One may be active. Another may be waiting for a response. A third may be alive but degraded. One may have stopped making progress without technically failing. Another may have been restricted by a governance control and be waiting for review.</p><p>A simple “running” or “not running” label is not enough.</p><p>That is the problem the proposed <strong>Agent Status Protocol (ASP)</strong> is trying to solve. It separates agent state across several dimensions — lifecycle, health, task state, conversation-floor state and governance state — rather than compressing everything into one status. An agent could therefore be online, degraded, working, waiting for another participant and under review at the same time.</p><p>This is different from reconstructing what happened yesterday. It is about knowing what is happening now.</p><p>Is the agent alive? Is it making progress? Is it waiting? Has it stalled? Has another agent taken the floor? Has a guardrail or operator restricted what it can do?</p><p>Those distinctions matter because intervention depends on current state. You cannot manage a workflow well if you do not know whether the system is executing, blocked, degraded or simply waiting.</p><p><strong>Observability begins with knowing the current state of the system before we even ask why it got there.</strong></p><h3>3. Then reconstruct the path</h3><p>Knowing the current state is useful. When something goes wrong, the next question is how the system got there.</p><p>That requires more than a timeline.</p><p>This is where provenance becomes important.</p><p><strong>PROV-AGENT</strong> connects the elements of an agentic workflow into a provenance graph: agents, prompts, model invocations, tools, domain data and the downstream outputs those interactions produce. Instead of leaving each event in its own log, it preserves the relationships between them so a decision can be traced backwards through the workflow that created it.</p><p><strong>MAP-Graph</strong> takes a similar idea into multi-agent memory. A piece of information passed from one agent to another does not arrive alone. It may carry a source, a transformation history, a trust level and permission constraints inherited from earlier steps. A summary can look harmless on its own while still carrying problems from where it came from.</p><p>That gives us an important distinction.</p><p>Chronology tells us that A happened, then B happened.</p><p>Lineage tells us that B exists because of A.</p><p>Those are not the same thing.</p><p>A long execution log can show every event in perfect order and still leave us unsure which earlier action produced a later outcome, which information travelled through several agents or which source ultimately shaped a decision.</p><p>For observability, sequence is useful.</p><p><strong>Lineage is what makes the sequence explainable.</strong></p><h3>4. But sequence is not causality</h3><p>Even lineage does not solve everything.</p><p>A trace can tell us what happened first, what happened next and what followed. That still does not necessarily tell us what caused what.</p><p><strong>TraceCompiler</strong> is useful because it refuses to treat adjacency as dependency. If one tool call appears before another, that is not enough to claim that the first produced the second. It looks instead for producer-consumer evidence: did an output from one step actually become an input to the next? Where the dependency cannot be established with enough confidence, the system can leave it unresolved rather than inventing one.</p><p><strong>Adaptive Influence Graphs</strong> approach the same problem from the failure side. Instead of reading a long multi-agent transcript as a flat sequence, they restructure it around influence: which earlier step materially shaped what happened later. That helps trace a visible downstream symptom back toward the agent or action that may have introduced the problem, rather than simply blaming whatever happened immediately before the failure appeared.</p><p>Agentic workflows are full of this kind of misleading proximity.</p><p>A bad recommendation may appear immediately after a tool call, while the real problem began several steps earlier with a stale memory, an incomplete handoff or a wrong assumption carried forward by another agent.</p><p>Chronology is necessary.</p><p>But it is not causality.</p><p><strong>A trace tells you what happened next. It does not automatically tell you what happened because of what.</strong></p><h3>5. What the agent says it did is not enough</h3><p>There is another limitation.</p><p>Application traces tell us what the agent framework recorded. That is useful, but it is still only one view of the system.</p><p>A tool call may look clean in the agent trace while the script it launched reads additional files, starts child processes or contacts another host. A successful task may hide activity that was never visible at the application layer.</p><p><strong>AgentSpy</strong> tackles this by observing the agent from outside its own process. It records operating-system calls and network activity for the agent and the programs it launches. In its experiments, that exposed unrelated activity, hidden file access and other behaviour that ordinary task-success checks did not reveal.</p><p><strong>AgentLoupe</strong> takes the idea further by trying to preserve causal labels across processes, files and sockets. If one interaction causes an agent to write something that is picked up much later by another process, the connection need not disappear simply because the original process has ended.</p><p><strong>Agent-Warden</strong> works at a similarly low level, following process-to-file provenance inside the operating system so that effects invisible to application instrumentation can still be reconstructed.</p><p>The broader distinction is between a <strong>self-reported trace</strong> and an <strong>independently observed effect</strong>.</p><p>The first tells us what the agent system believes it did.</p><p>The second gives us evidence of what actually happened outside it.</p><p>That brings us back to the question from Part 4:</p><p><strong>Did the world change the way the agent thinks it did?</strong></p><h3>6. Multi-agent observability is about propagation</h3><p>Multi-agent systems make observability harder because the place where a failure becomes visible may not be where it began.</p><p>One agent can make a poor decision, another can build on it, a third can compensate for part of it, and the final output can still look plausible. By then, the original error has travelled.</p><p><strong>AgenTracer</strong> tries to identify the agent and the step that introduced the decisive failure in a multi-agent trajectory. Its focus is not simply on who produced the bad final answer, but on where the run first went off course in a way that changed the outcome.</p><p><strong>DoVer</strong> goes one step further. Instead of stopping at “this step looks responsible,” it intervenes at the suspected point, changes the instruction or plan and reruns the workflow. If the task improves, the diagnosis has stronger evidence behind it.</p><p>That gives us another useful distinction:</p><p><strong>Attribution tells you where the problem may be. Intervention tests whether you are right.</strong></p><p><strong>AgentSoV</strong> approaches the problem as propagation. It models how disturbances move through stages, helping separate where an error originated from where its effects finally became visible.</p><p>And <strong>MASIF</strong> adds a more subtle complication. It finds that workflows can contain local errors even when the final task succeeds, because other agents may correct, bypass or absorb them along the way.</p><p>So success and failure are not always properties of a single step.</p><p>They can be properties of how the whole system absorbs, amplifies or propagates what happened earlier.</p><p><strong>In multi-agent systems, observability is not only about finding errors. It is about following what those errors become.</strong></p><h3>7. Observability should eventually change what the system does</h3><p>Seeing a problem after it happens is useful.</p><p>Seeing it while there is still time to act is much more useful.</p><p>That is where observability begins to move from an audit system toward a control system.</p><p><strong>AgenticRV</strong> makes the distinction clearly. A monitor that sees an action only after it has been dispatched can record the event and help explain it later. But it cannot prevent the effect. To intervene, the system has to observe the pending action at a point where it can still be held, checked and, if necessary, blocked.</p><p>Other work on real-time failure detection takes the idea further. Instead of waiting for the final outcome, monitors can watch a workflow as it unfolds, detect certain kinds of deviation and return the system to an earlier checkpoint before the problem compounds. The experiments are still bounded, but the direction is important: monitoring is beginning to connect directly to repair.</p><p>The progression is becoming clearer:</p><p><strong>observe → diagnose → attribute → intervene.</strong></p><p>That connects directly to the Recovery Gap from Part 6.</p><p>Recovery begins after something has already happened. Better observability creates the possibility of acting earlier, while the consequence is smaller and the path back is easier.</p><p>That may ultimately be its greatest value.</p><p>Not merely knowing what went wrong.</p><p><strong>Knowing soon enough to change what happens next.</strong></p><h3>8. Even the telemetry has to be trusted</h3><p>There is one more complication.</p><p>So far, we have treated telemetry as the evidence layer: the thing we turn to when we want to know what the agent did, what changed and where a workflow went wrong.</p><p>But what if the telemetry itself is incomplete, altered or misleading?</p><p>The <strong>Agent-Native Telemetry</strong> work takes that problem seriously. Its argument is that operational evidence should not simply be collected; it should carry guarantees about integrity and completeness. The system should be able to say not only “here are the records,” but whether the record set is complete, whether something is missing and whether the evidence has been altered.</p><p>That matters because agents may themselves use telemetry to decide what to do next.</p><p>Research on manipulating AIOps telemetry shows how dangerous that can become. In controlled experiments, attackers planted misleading content into logs and operational signals later consumed by AI-driven incident-response agents, pushing them toward incorrect diagnoses or remediation.</p><p>So the observability layer is not automatically neutral.</p><p>It can itself become part of the attack surface.</p><p>We started with the idea that agents need trustworthy traces so people can understand what they did.</p><p>But increasingly, the agent may also be reading those traces.</p><p>That changes the requirement.</p><p>Observability is not only about seeing the system clearly.</p><p><strong>It is also about knowing whether the evidence you are seeing deserves to be trusted.</strong></p><h3>Back to the logs</h3><p>The workflow failed.</p><p>The model had recorded its calls. The orchestration layer had its sequence. The tools had their responses. The operating system had its own evidence.</p><p>Somewhere in that pile was the story of what actually happened.</p><p>But the story was not the same thing as the records.</p><p>That is the <strong>Observability Gap</strong>: the distance between collecting activity and being able to reconstruct, with enough confidence to act, what happened, why it happened and what should happen next.</p><p>Closing that gap requires more than logging. It requires knowing the current state of the system, preserving lineage across agents and tools, distinguishing sequence from causality, comparing what the agent reports with what the world actually did, and following errors as they propagate. Increasingly, it also means being able to intervene while the workflow is still unfolding — and trusting the evidence on which that intervention depends.</p><p>As agents become more autonomous, this distinction will matter more.</p><p>We will not only need to know whether the agent completed the task. We will need to know how it got there, what it touched along the way, what influenced its decisions, where an error entered the chain and whether there is still time to do something about it.</p><p>We spend a lot of time asking whether an agent is working.</p><p>Enterprise systems need another question.</p><p><strong>Can we see enough of the work to know when it isn’t?</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4ebcbe4a4d8a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Elastic Festival]]></title>
            <link>https://medium.com/@mail2rajivgopinath/the-elastic-festival-61974adc1938?source=rss-ae007d129200------2</link>
            <guid isPermaLink="false">https://medium.com/p/61974adc1938</guid>
            <category><![CDATA[halloween]]></category>
            <category><![CDATA[culture]]></category>
            <category><![CDATA[society]]></category>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[pop-culture]]></category>
            <dc:creator><![CDATA[Rajiv Gopinath ]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 12:39:12 GMT</pubDate>
            <atom:updated>2026-10-24T12:39:12.062Z</atom:updated>
            <content:encoded><![CDATA[<h4>Perspectives</h4><h4>How Halloween is stretching beyond October 31, and what that tells us about the changing shape of celebration</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aymEdcw4vfM4uFVuigYjlg.png" /></figure><p>A few weeks ago, The Guardian ran a story about something called “Halloweenmas”.</p><p>The name sounds slightly ridiculous until you look at what retailers are selling: Halloween trees, black wreaths, ghost-shaped ornaments, tableware, candles and home accessories that look less like a one-night purchase and more like the start of a seasonal decorating ritual.</p><p>Wholesale marketplace Faire reported that products combining Halloween and Christmas themes rose from 425 in 2021 to 2,632 in 2025. Halloween-tree listings alone grew sharply. In Britain, Wilko said September sales of artificial Halloween trees were up 200% on the previous year.</p><p>The obvious conclusion is that Halloween is becoming another Christmas. There is something Christmas-like in the way it is moving into the home. A pumpkin on the doorstep has become a wreath on the door, decorations across the living room, themed table settings and, increasingly, a tree. The festival seems to be acquiring the infrastructure of a season.</p><p>But when we looked for evidence that consumers themselves were embracing “Halloweenmas”, the story became less straightforward. The products exist. The phrase barely does. Search data showed no meaningful volume for it in Britain or the US, and a year of UK social data produced only a handful of genuine consumer mentions once media coverage and promotional accounts were removed.</p><p>That does not make the retail phenomenon imaginary. It means the label may be running ahead of the behaviour. Trend stories often begin this way. A new word appears, retailers build around it, a few striking products photograph well, and something still quite small starts to sound like a cultural movement.</p><p>Halloween does appear to be changing, but Christmas is probably the wrong comparison. The more interesting evidence lies elsewhere. British consumers search for Halloween decorations months before October 31. In the US, visits to major Halloween retailers begin much earlier in the year. Attractions open in August and can run for more than two months. Home decoration is becoming almost as important commercially as costumes. And new occasions such as “Summerween” are appearing well outside the traditional calendar.</p><p>None of this makes October 31 less important. In some cases the peak is as strong as ever. What is changing is the cultural space around it: a longer approach, more ways to take part, more categories through which to participate and, occasionally, extra moments elsewhere in the year.</p><p>So the better question is not whether Halloween is becoming a second Christmas. It is whether it is becoming something else entirely: a festival whose boundaries are beginning to stretch.</p><h3>The week before October 31 is no longer the beginning</h3><p>The clearest change is not what happens on Halloween. It is how much earlier Halloween now begins.</p><p>In the US, traffic to Spirit Halloween is visible in June, months before most people would think of Halloween as a shopping occasion. Those early months are also growing fast. In 2026, June traffic was up 52.4% on the previous year, July 37.0%, August 30.6% and September 27.5%. That does not mean people are celebrating Halloween in June. It does suggest the season is occupying attention well before the date itself.</p><p>Search behaviour in Britain shows the same pattern. Searches for “Halloween decorations” reached 4,219 in July, more than doubled to 10,757 in August, and rose again to 19,179 in September. The numbers are still small against the October peak, but that is not the point. People are already looking for Halloween décor in the middle of summer.</p><p>This is where an extended festival starts to look less like retail theatre and more like consumer behaviour. Retailers can put products on shelves early without anyone wanting them. Search captures intent: someone actively looking for an idea, a product or a way to take part. And that intent is appearing well before October 31.</p><p>Halloween is acquiring a longer approach. The date has not moved. The cultural runway around it has.</p><h3>But elasticity does not mean the peak disappears</h3><p>Halloween may be starting earlier, but October is not becoming less important. Spirit Halloween’s own traffic makes that clear. In 2024, October accounted for 46.8% of its June-to-November traffic. In 2025, the share rose to 50.6%. The season is stretching forward, and the centre of gravity is still firmly in October.</p><p>That changes how we should think about elasticity. A longer season does not mean demand is spread evenly across more months. People begin thinking, searching and preparing earlier, while the main burst of activity still arrives close to the event. The earlier activity may even build anticipation for the peak instead of diluting it.</p><p>That is the distinction worth holding on to. Elasticity adds runway. It does not necessarily flatten the peak.</p><h3>Different parts of Halloween stretch differently</h3><p>Halloween is not stretching uniformly. Different parts of it run on different clocks.</p><p>In Britain, searches for Halloween decorations are visible in July and grow through August and September. Costume searches stay much closer to the event, with far less activity in the earlier months and a sharper rise as October approaches. The reason is simple. Decorating the home is something people can start early and live with for weeks. A costume is tied to a specific moment: a party, a night out, a school event, Halloween itself.</p><p>The consumer data points the same way. Among UK Halloween shoppers, 45% plan to decorate their home or garden, slightly ahead of the 42% who plan to wear a costume. The gap is small, but culturally interesting. Halloween is no longer only something you put on. It is increasingly something you build around yourself.</p><p>Then there are experiences. Ticketed attractions can begin in late August and run for weeks. Theme parks, haunted houses, immersive events and seasonal entertainment create another clock, one that starts earlier than costumes and lasts far longer than a single evening.</p><p>So the festival looks less like one long season and more like several overlapping ones. Home décor begins early. Experiences stretch across weeks or months. Costumes stay close to the event. That may be the more useful way to think about an elastic festival: different forms of participation stretching by different amounts.</p><p>Different parts of the same festival have different elasticities.</p><h3>Halloween moves into the home</h3><p>Decoration is one of the clearest signs of change. In the US, expected Halloween spending in 2026 puts decorations at roughly $4.3 billion. Costumes are only slightly ahead at $4.4 billion, and candy sits at about $4.1 billion.</p><p>The costume has traditionally been the most visible symbol of participation. You put something on, go somewhere, do something, then take it off. Decoration changes the space around you and often stays for days or weeks.</p><p>The social data points the same way. In the UK, Halloween décor conversation rose from an average of 12.8 posts a day in June to 240.8 a day between October 1 and October 23. More importantly, 57.4% of the measured décor conversation happened before October began. In the US, the figure was higher, at 63.4%.</p><p>The home, it seems, is becoming one of the main places where the season expands. A costume belongs to a particular moment, while a decorated home can make Halloween present long before the night arrives. The pumpkins appear earlier. The lights go up. The table, the front door and the living room change. Halloween starts to occupy space, not just time.</p><p>That may be one reason its boundaries stretch so easily. A single evening is limited. An environment can stay with you. Halloween becomes less like a single performance and more like a temporary world.</p><h3>Experiences stretch even further than products</h3><p>If home décor stretches Halloween into a season, experiences stretch it further. Universal Orlando’s Halloween Horror Nights spans 66 days in 2026, from August 28 to November 1. At that point it is no longer a Halloween-night event. It is a seasonal entertainment property. People can attend in late August, through September, across October and into November. The date remains October 31, but the experience is detached from that single night.</p><p>The UK social data adds a useful signal. Among posts about Halloween experiences, 15.6% referenced tickets or bookings already secured. That is more meaningful than general chatter. Liking a Halloween post is one thing. Buying a ticket weeks in advance is another. It suggests the season is being extended by behaviour, not only by marketing.</p><p>Experiences are good at this because they create more than one point of participation. You can book early, talk about it before you go, attend on one of many dates, post afterwards and sometimes go again. That gives the festival a much longer life than the event itself.</p><p>It may also be one of the deeper mechanics behind elasticity. Festivals become more elastic when consumers can participate repeatedly instead of commemorating once.</p><h3>One of the more interesting signals is Summerween</h3><p>The term “Halloweenmas” has almost no measurable consumer traction. The products are there and retailers are clearly experimenting, but people are not using the term.</p><p>Summerween is different. In the US, searches for “Summerween” reached 107,224 in July 2026 and another 47,081 in August. There was measurable interest in the UK too, and US social conversation ran into the tens of thousands of mentions. There is also activity around “Halfway to Halloween”, which adds another moment before the main season has begun.</p><p>That suggests more than a longer October. Halloween may be starting to generate additional occasions around itself:</p><p>Halfway to Halloween → Summerween → Halloween season → October 31</p><p>These moments are not equally important, and October 31 remains the centre of gravity. But satellite occasions matter. A festival that once had one obvious date can begin to support several moments of anticipation, participation and consumption across the year. That is when the idea of a festival becoming a platform starts to feel credible. The original day does not disappear. More things gather around it.</p><h3>From festival to platform</h3><p>Festivals are usually treated as dates: the day itself, perhaps a little preparation before it, and then life moves on. Once an occasion becomes culturally rich enough, it can support much more than the event it started as.</p><p>Halloween now sits across home décor, beauty, fashion, food, entertainment, hospitality, social content, fandom and ticketed experiences. Different people enter through different parts of it. One decorates the house. Another books a haunted attraction. Someone else experiments with makeup, hosts a party, watches horror films, buys themed food or follows the season on social media.</p><p>That changes the shape of the occasion. It begins as a date. It acquires a period of anticipation. More ways of participating appear around it, and some begin earlier than others. Eventually, as with Summerween, additional occasions can form around the original one:</p><p>Date → anticipation period → multiple participation behaviours → satellite occasions → cultural platform</p><p>Halloween may be moving along that curve. This is what I mean by an Elastic Festival. Not a festival that simply lasts longer, but one whose boundaries stretch because more behaviours, categories and occasions can gather around the same cultural idea. The day still matters. The festival starts to become bigger than the date.</p><h3>Why would consumers want this?</h3><p>The commercial explanation is easy: a longer season gives retailers more time to sell. It does not explain why people take part.</p><p>The UK research points to a more interesting tension. Seventy-nine per cent of consumers say it is important to make the most of occasions like Halloween despite cost-of-living pressures. At the same time, 71% of Halloween shoppers want to mark the occasion without spending a lot. The two numbers sit together neatly. People want the occasion, even when they do not want the expense that comes with it.</p><p>That may explain why an elastic festival works culturally. It offers more ways to take part without requiring any one of them to be large, expensive or formal. You can decorate one corner of the house, try a different makeup look, watch a horror film, go to a themed event, host a small gathering, buy something minor, post something, or simply enjoy the anticipation. Participation becomes easier to scale up or down.</p><p>That makes the underlying need slightly different from the usual retail story. People may not want more things to buy. They may want more moments that feel worth marking. Consumers may be seeking more opportunities to celebrate without seeking more reasons to spend.</p><h3>Halloween is becoming more adult, expressive and modular</h3><p>Halloween can also stretch because participation no longer requires one fixed behaviour. Among adult Halloween-related social conversation, beauty accounted for a large share of activity in both the UK and the US, while traditional costume content was much smaller. That is not a population measure, but it indicates the range of ways adults now take part.</p><p>Halloween can mean makeup, nails or hair. It can mean decorating, hosting friends, watching horror films, going to an event, making content or dressing up. You do not have to do all of these, or even the same thing every year. That makes participation modular. Someone can enter through beauty in early October, book an experience for mid-month, decorate later, and still treat October 31 as the main event.</p><p>A festival is easier to stretch when people are not locked into one prescribed ritual. The more ways there are to take part, the more moments in which participation can happen. That may be one of the quieter reasons Halloween is becoming more elastic. Everyone is not celebrating more intensely. There are simply more ways to belong to the occasion.</p><h3>But every elastic festival has an elastic limit</h3><p>There is a point at which extension becomes irritation. The US data makes that clear. Only 16% of Americans say Halloween decorations before October are acceptable. Another 43% say the earliest appropriate window is October 1–15. More broadly, 42% say holidays generally begin too early. In US social conversation, a growing share of complaints about seasonal creep is now aimed at Halloween itself, as people notice decorations, retail displays and themed activity arriving earlier every year.</p><p>Elasticity is not endless expansion. A festival gains value from anticipation, more ways to take part and a longer runway. Stretch it too far and it can lose the thing that made it feel special. Black Friday is a useful caution. What began as a tightly defined retail moment spread across days and, in many cases, weeks. That created more selling time but weakened urgency. If everything is “Black Friday” for half the month, the phrase starts to mean less.</p><p>The same risk applies here. The opportunity is not to make every festival last forever. It is to understand where anticipation adds meaning and where extension begins to dilute it.</p><h3>So is Halloween becoming elastic?</h3><p>Yes, but selectively.</p><p>The evidence is strongest in a few places. People think about Halloween earlier. Home decoration starts well before October. Ticketed experiences stretch across weeks or months. Participation now runs through beauty, décor, food, entertainment, hosting and content, not just costumes. And moments such as Summerween suggest Halloween can generate satellite occasions around the main event.</p><p>That is enough to say the boundaries are moving. It would be easy to push the idea too far, though. Halloween is not becoming equally important throughout the year. October 31 still matters enormously. “Halloweenmas” has far more retail supply than consumer traction. The pattern is not identical across every Western market. And there is no reason to assume every festival can stretch in the same way.</p><p>So the broader conclusion needs to stay narrower. Some festivals are discovering that their cultural boundaries are elastic. They can begin earlier, support more forms of participation, create additional moments around themselves and expand into categories once peripheral to the occasion. But elasticity is not permanence. The date and the peak still matter. What changes is how much cultural activity can gather around them.</p><h3>What this means for brands</h3><p>If festivals become longer cultural platforms, brands probably need to stop thinking only in terms of “the campaign around the date”.</p><p>Different parts of the occasion open at different times. Inspiration can begin early. Home and décor can start months ahead. Ticketed experiences need even more lead time. Beauty and fashion may gather pace closer to the event. Urgent purchases still cluster near the peak.</p><p>That implies a different planning model. Instead of one campaign built around one moment, brands may need a participation calendar, with different roles, categories and content appearing at different points in the season.</p><p>The aim is not to stretch activity for as long as possible. Consumers are perfectly capable of rejecting extensions that feel artificial, and “Halloweenmas” is the warning: retailers can create products and labels well before consumers decide they matter. So the opportunity is not to manufacture a longer festival. It is to understand where the festival is already stretching on its own, and show up in ways that fit those moments.</p><h3>The day remains. The routes into it multiply.</h3><p>October 31 has not disappeared. If anything, the evidence suggests the date matters as much as ever. The peak remains, the costumes still gather around it, and the main burst of attention still comes close to the day itself.</p><p>What has changed is everything around it. People decorate earlier, book experiences earlier, and create and consume Halloween content earlier. Some take part through beauty, others through the home, food, parties, horror films or ticketed events. And occasions such as Summerween add moments well outside the traditional calendar.</p><p>The festival has not lost its centre. It has acquired more routes into it. The elastic festival does not replace the big day. It creates more ways to arrive at it.</p><p>Perhaps the future of festivals is not that we invent many more of them. It is that the ones we already have become larger cultural worlds, worlds we enter earlier, experience in more ways and leave a little later.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=61974adc1938" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>