<?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 Elliot on Medium]]></title>
        <description><![CDATA[Stories by Elliot on Medium]]></description>
        <link>https://medium.com/@elliotJL?source=rss-79bf89c9d6da------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*wWRm3M2w2vdjGyA0I92gHg.png</url>
            <title>Stories by Elliot on Medium</title>
            <link>https://medium.com/@elliotJL?source=rss-79bf89c9d6da------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 20:38:31 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@elliotJL/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[The Product Engineer and the End of the Handoff]]></title>
            <link>https://medium.com/@elliotJL/the-product-engineer-and-the-end-of-the-handoff-93181f170779?source=rss-79bf89c9d6da------2</link>
            <guid isPermaLink="false">https://medium.com/p/93181f170779</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Elliot]]></dc:creator>
            <pubDate>Fri, 31 Jul 2026 14:23:30 GMT</pubDate>
            <atom:updated>2026-08-11T08:15:14.962Z</atom:updated>
            <content:encoded><![CDATA[<p><em>As the cost of building approaches zero, the cost of deciding becomes the biggest constraint.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*dlgo1kFVHlaz_ggTeSaTmg.jpeg" /><figcaption><a href="https://fineartamerica.com/featured/1-the-boat-builders-by-winslow-homer-1873-winslow-homer.html"><em>The Boat Builders by Winslow Homer, 1873</em></a></figcaption></figure><p>I spent a chunk of the last year building an AI tutor with a small team, and one of the more useful things we learned had nothing to do with the model. We tested the same worked answer against two UK exam boards’ mark schemes. AQA gave it full marks. Edexcel docked a method mark for skipping a step AQA didn’t require. Same maths. Same student. Different rubric, different grade.</p><p>The tech industry is currently running this exact same exam board problem on a new class of builders. When assessing talent, the market seems to default to marking against legacy rubrics. One rubric measures a senior engineer: can you talk convincingly about sharding and the trade-offs a staff engineer has internalised over a decade of production incidents? The other marks against a senior PM: multi-year strategy, board-level bets, and the organisational navigation that comes from running a hundred-person product org.</p><p>Both are real, fair rubrics. Neither measures the job that is actually emerging, because that job doesn’t have a mark scheme yet.</p><h3>The Cost That Changed</h3><p>It’s worth being precise about why “product engineer” isn’t just PM-with-extra-steps or engineer-with-a-roadmap-opinion. It exists because of a specific cost that changed.</p><p>Building used to be the expensive part. A product person had an idea, wrote it up, handed it to engineering, and waited. Splitting problem-definition from problem-building was the correct organisational answer to a world where building took weeks. Today, that seam between product management, engineering, and design is just a speed tax.</p><p>Building doesn’t take weeks any more for a meaningful class of problems. For many product problems, implementation has become the cheap part. The constraint is increasingly judgement: deciding what to build, verifying that it works, and knowing when it’s good enough to ship. That’s a very different optimisation problem from five years ago. Boris Cherny, who built Claude Code, told Lenny Rachitsky in February:</p><blockquote>“the title software engineer is going to start to go away. It’s just going to be replaced by builder.”</blockquote><p>Fiona Fung, who now runs Claude Code and Cowork at Anthropic, said in June that PMs, designers, and non-engineers on her team all check in code. The two profiles that survive are creative builders with product sense and deep systems experts for the parts that genuinely need one.</p><p>That’s remarkably close to what companies like Linear and incident.io have quietly been optimising towards for years: engineers owning customer problems directly, with product judgement distributed into the people building the software rather than sitting behind a handoff.</p><p>When the handoff was the only way to build at all, splitting the job made sense. When the handoff is now the slowest part of the process, whoever can skip it has a real advantage.</p><h3>The Prototype is the Spec</h3><p>I wrote the internal operating document for the tutor team to clarify this exact shift: <strong>one person carries a problem from a prototype straight through to production</strong>. There is no relay race through a second and third person to get there. Working software replaces documents as the unit of alignment, which means the prototype is the actual spec.</p><p>Design still matters enormously, but increasingly as direction rather than prescription. Instead of mockups describing software that will eventually exist, software becomes the thing everyone is reacting to.</p><p>What I’m describing is a different shape of job: you’re accountable for one thing working, end to end. The boundary between “deciding what to build” and “building it” mostly disappears because you’re the one doing both. Badly at first, then less badly.</p><p>It also isn’t a claim that the older jobs are being absorbed. A staff engineer designing a system that has to survive a decade of load is doing work that cannot simply be called “just implementation.” A career product leader running portfolio discovery across a hundred-person org, untangling stakeholder warfare, and defending a three-year bet to a board is practicing a fundamentally different craft built on reps I haven’t had.</p><h3>The Rhythm and the Board</h3><p>The small-scope, high-ambiguity version requires a completely different operating rhythm. To make it work, you have to run forward-deployed engineering: builders sit directly with the end users. Deployment is treated as a discovery surface, not a support function. You don’t need a PM to translate what users hate if you are the one watching them try to use it.</p><p>The rituals that sustain this are minimal. Ownership is carried by the room, the rhythm, and the board. Roadmap ownership is explicit and scales by adding actual owners, not by adding process. We let operating numbers like commit cadence and cycle data shape the bets without bending them into vanity targets.</p><p>The meeting rhythm changes too. Status updates matter less when the person who identified the problem is also shipping the fix. Demos replace reports because working software carries more information than plans ever could.</p><p>This is how our AI tutor went from zero to being awarded a government tender within 2 months, and kept getting faster after that.</p><h3>Evals as the New PRD</h3><p>Right now, I see teams moving toward agentic workflows where autopilot doesn’t require huge token spend, it just requires that you know what you want and have your ticketing and CI/CD in order.</p><p>This is where the actual mechanics of the job change: you stop prompting, and you start orchestrating. The execution is increasingly handled by <strong>agentic loops: systems that find the work, check it, and decide the next step</strong>. My job this week isn’t necessarily writing the code; it is improving the loops that write the code. A shift I wrote about recently in <a href="https://medium.com/@elliotJL"><em>The Loop Was Never the Hard Part</em></a><em>.</em></p><p>Because the loops handle the execution, for this class of product, the bottleneck has moved away from engineering throughput. The scarce skill in this new paradigm is now deciding, precisely enough that a machine can be held to it, <strong>what “correct” means for the thing you’re building, and then checking that the machine actually met it</strong>. Some are calling this error analysis: read real failures before you write a rubric, because you cannot predict the failure modes in advance.</p><p>The interesting shift is that the rubric often emerges after you’ve watched enough failures, not before. You read real transcripts, cluster the mistakes, and only then discover what “correct” actually needed to mean. The evaluation becomes something you uncover rather than invent. Dianne Penn, who shipped the product docs for every Claude model release, calls the practice:</p><blockquote>“evals are the new PRDs.”</blockquote><p>We ran this for real on the tutor. The bare model marked exam answers at 67% accuracy against the real mark scheme. Structured retrieval against the specific exam board’s material got it to 99%. The model hadn’t suddenly become better at maths. It had finally been taught what good looks like. We quickly discovered that the gap was coverage, but only because the person reading a hundred failed transcripts to spot the pattern was also the person building the pipeline to fix it. Every failure improved the system, not just the next answer.</p><p>Neither half alone gets you there. A PM who can read the transcripts but not build the fix waits in a queue. An engineer who can build anything but hasn’t sat with the failures builds the wrong thing well.</p><h3>The Rubric Problem</h3><p>The honest risk in all of this is title inflation. Half of LinkedIn has quietly renamed itself “product engineer” or “builder” in the last year, and a good chunk of that are senior ICs stretching a job title rather than a genuine change in what they do day to day.</p><p>The rush toward “builder” makes sense, but it also hides a distinction. Builder isn’t really a job description. It is a claim about agency: that <strong>someone can carry a problem through the whole loop without waiting for three other functions.</strong> It doesn’t tell you which loop they own. A design engineer carries an interaction into production. A forward-deployed engineer carries a customer’s workflow into production. A GTM engineer turns a commercial bottleneck into a system. Product engineer is one version of the same move: redraw the job around an outcome rather than one stage of making it.</p><p>One reason the naming feels slippery is that the industry is using identities, roles, and methods interchangeably. Builder is an identity. Product engineer is an organisational role. Forward-deployed describes where that role operates. Compound engineering and loop engineering describe how the work gets done. They overlap, but they aren’t synonyms. The title matters only when it records a real transfer of ownership. If the same person still writes the spec and waits for somebody else to build it, nothing changed except LinkedIn.</p><p>The organisational bargain matters too. <strong>End-to-end accountability without the authority, access, or compensation to match is not product engineering. It’s three jobs hidden under one title</strong>. Jack &amp; Jill’s read on the market this year was blunter: the old titles “no longer describe jobs,” and hiring is moving toward traits over titles because the titles stopped being trustworthy first.</p><p>When iteration cycles collapse from weeks to hours, the advantage moves to whoever can carry the loop without stopping at every boundary. The people making the fastest progress are usually the ones who don’t need to stop and hand the problem to somebody else.</p><p>The real test for this emerging class of builders isn’t explaining consistent hashing, and it isn’t defending a three-year roadmap. It’s showing one thing, beginning to end, that you decided needed to exist, then built, then knew was working because you could point at the number that proved it.</p><p>For someone trying to move into this work, the answer isn’t to rename yourself. Pick one problem, get close enough to the users to understand it, build the smallest useful version, instrument it, and stay with it until you can show whether it worked. That evidence is the portfolio. The title can catch up later.</p><p>Right now, the industry is still grading AQA maths with an Edexcel rubric. The rubric is the variable, not the work. Someone’s going to write the new mark scheme eventually. Until then, we have to recognise the seam for what it is, and stop grading builders against the rubrics of a slower era.</p><blockquote><strong><em>EDIT: Postscript, August 2026.</em></strong><em> Two weeks after this went up, Marty Cagan published “</em><a href="https://www.svpg.com/a-fresh-definition-of-the-product-role/"><em>A Fresh Definition of The Product Role.</em></a><em>”</em></blockquote><blockquote><em>In it he says one of the best descriptions of the product role he’s seen in years comes from Benedict Evans: “</em>Writing the code isn’t the hard part, and making the tool isn’t the hard part. The hard part is knowing that it should exist, and knowing how it should exist, and that’s a different person.<em>”.</em></blockquote><blockquote><em>Cagan reads that as the case for a distinct product role, but I still think that in small teams, the person who knows what should exist is increasingly also the person who builds it. Same scarce skill either way. The mark scheme is getting written sooner than I expected.</em></blockquote><p><em>I write about building AI products and what the job actually looks like once you’re doing it. Subscribe if you’re navigating the same shift.</em></p><h3>Reading</h3><ul><li><a href="https://www.lennysnewsletter.com/p/head-of-claude-code-what-happens">The title software engineer is going to start to go away</a> — Boris Cherny on Lenny’s Podcast, Feb 2026</li><li><a href="https://www.lennysnewsletter.com/p/building-the-most-ai-pilled-engineering">What happens after coding is solved</a> — Fiona Fung on Lenny’s Podcast, Jun 2026</li><li><a href="https://www.lennysnewsletter.com/p/why-ai-evals-are-the-hottest-new">Why AI evals are the hottest new skill for product builders</a> — Hamel Husain &amp; Shreya Shankar on Lenny’s Podcast</li><li><a href="https://www.lennysnewsletter.com/p/why-ai-is-going-vertical-again">Why AI is going vertical again</a> — Dianne Penn on Lenny’s Podcast, Jul 2026</li><li><a href="https://www.jackandjill.ai/">Jack &amp; Jill</a> — on titles no longer describing jobs</li><li><a href="https://posthog.com/blog/product-engineer-vs-product-manager">Product engineer vs product manager</a> — PostHog</li><li><a href="https://www.lennysnewsletter.com/p/how-linear-builds-product">How Linear builds product</a> — Nan Yu on Lenny’s Podcast</li><li><a href="https://blog.lawrencejones.dev/2025/">2025 retrospective</a> — Lawrence Jones, incident.io</li><li><a href="https://every.to/chain-of-thought/compound-engineering-how-every-codes-with-agents">Compound engineering: how Every codes with agents</a> — Every</li><li><a href="https://addyosmani.com/blog/loop-engineering/">Loop engineering </a>— Addy Osmani, Jun 2026</li><li><a href="https://newsletter.pragmaticengineer.com/p/forward-deployed-engineers">What are forward deployed engineers?</a> — The Pragmatic Engineer</li><li><a href="https://vercel.com/blog/design-engineering-at-vercel">Design engineering</a> — Vercel</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=93181f170779" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Loop Was Never the Hard Part]]></title>
            <link>https://medium.com/@elliotJL/the-loop-was-never-the-hard-part-5bdd4352acab?source=rss-79bf89c9d6da------2</link>
            <guid isPermaLink="false">https://medium.com/p/5bdd4352acab</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[coding]]></category>
            <dc:creator><![CDATA[Elliot]]></dc:creator>
            <pubDate>Mon, 13 Jul 2026 19:34:37 GMT</pubDate>
            <atom:updated>2026-07-13T19:34:37.495Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Writing a loop takes minutes. Writing down what good looks like, precisely enough that a machine can hold you to it, is the job almost nobody’s been taught.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OnmMA2FH8Xyk-9RYDT2JsA.png" /><figcaption>Infinite Loops (<a href="https://www.behance.net/gallery/94552241/Infinite-Loops">source</a>)</figcaption></figure><p>At six o’clock this morning, a fleet of agents I built found an hour-long interview buried on YouTube, pulled the transcript, worked out which parts of my knowledge base it changed, and filed the updates with reasoning attached. By the time I woke up there was a Telegram message waiting: a debrief on what moved overnight, and a short podcast, generated while I slept, on how it all ties to the job I start this month.</p><p>The system is called <em>argus</em>. Inspired by the Greek <em>Argos</em>, the name of a watchman in Greek mythology who had a hundred eyes. It reads a few hundred sources a day, keeps a corpus of living notes current, and briefs me every morning. It has never missed a day, nor lost the thread that keeps me in the loop.</p><p>I listened to its flawless little production at 8:15 on the Northern line, wedged into a stranger’s armpit, head sore from a 2 AM England game, being dragged at walking pace through a tunnel that still can’t manage air conditioning. The agents have never been slicker, yet the person running them needed a sit down and a Lucozade. Somewhere between Angel and Bank it occurred to me that I hadn’t decided to listen to a podcast that morning; a loop had decided for me, and it had judged, correctly, that I’d want it. I’m not sure which of us is the operator anymore. I’m reasonably sure it’s not the one in the armpit.</p><p>So when Boris Cherny, the man who created Claude Code, <a href="https://thenewstack.io/loop-engineering/">said this in June</a>, I recognised it:</p><blockquote>“I don’t prompt Claude anymore. There are loops running, and it’s the loops that prompt Claude and decide what to do. My job is to write the loops.”</blockquote><p>The numbers behind it: Claude Code wrote 100% of his code contributions over the past 30 days, 259 pull requests. Merges per engineer on his team are up 200%.</p><p>Within days the line had a name, <a href="https://addyosmani.com/blog/loop-engineering/">loop engineering</a>, courtesy of Addy Osmani at Google. Within weeks it had the full hype lifecycle: courses, “the skill that replaces prompting” threads, and a backlash from working engineers calling the whole thing a cron job wearing a hoodie.</p><p>After running these things in production for the past 6 months, I think both camps are missing something. The evangelists are prescribing a new professional identity in a slogan, with no mechanics attached and no mention of what it costs. The sceptics are right that a scheduled prompt isn’t a new invention, and wrong about what actually changed.</p><p>So let’s dig into what a loop actually is, the four places the slogan breaks, and the operating model I’d defend instead.</p><h3>What on Earth is a loop?</h3><p>Not to be confused by Rian Johnson’s criminally underrated movie <a href="https://www.imdb.com/title/tt1276104/">Looper</a>, a loop is an agent repeating cycles of work until a stop condition is met. That’s the whole definition. The interesting part is what you hand over, and there are four levels.</p><p><strong>You hand over the check.</strong> You prompt, the agent works, and instead of eyeballing the output, you’ve encoded what “correct” looks like so the loop can verify its own work. Still one task at a time, but the quality gate is written down instead of living in your head.</p><p><strong>You hand over the stop condition.</strong> You define done, and a second model judges every attempt against it. “Get the page load under two seconds, stop after five tries.” You’ve stopped supervising attempts. You find out when it’s finished, or when it’s failed.</p><p><strong>You hand over the trigger.</strong> The loop runs on a schedule or watches for events, laptop open or not. My 6am brief. A Monday finance audit. They start themselves.</p><p><strong>You hand over the prompt itself.</strong> The loop watches your tools, notices what needs doing, decides, does it, and a second agent reviews before you ever see it. Your only job is the merge. This is where Cherny lives. And as of this month it stopped being one man’s workflow: Anthropic shipped <a href="https://www.anthropic.com/news/introducing-claude-tag">Claude Tag</a>, a persistent agent that lives in Slack and opens work on its own initiative. By <a href="https://www.youtube.com/watch?v=7Ue0yM4J-B8">their own numbers</a>, 65% of their product teams’ pull requests now start with it. The top rung is a thread loving, mention spamming, product now.</p><p>Read it as a ladder. Each rung hands the machine one more piece of what used to be your job: the check, the stop condition, the trigger, the prompt.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*sU-nIggWjkDmMDY4.png" /></figure><p>And nearly everyone with access to these tools is parked at the bottom of the ladder: Anthropic’s own research found <a href="https://www.anthropic.com/research/coding-agents-social-sciences">81% of surveyed researchers had tried AI chatbots while 20% regularly used agents</a>. The capability gap between those two numbers is zilch.</p><h3>The case for Cherny</h3><p>A small team outbuilds funded competitors by refusing to spend humans on anything a loop can hold. The team I mentioned at the top: <strong>six engineers, first commit to a live product in 3 weeks, and it got faster after launch. </strong>Five merged pull requests per engineer per day, the median PR merged in under an hour.</p><p>The loops are what made it possible. Overnight, a fleet of agents named after vegetables (Potato, Turnip, Carrot, Leek, Salad, Radish, Broccoli — long story) triaged the errors, graded the product’s output against a quality spec, watched the costs, and compiled the morning picture. The humans arrived to a brief, not a backlog. A different shape of team, and the only way I know for a handful of people to run something that behaves like it has an ops department.</p><p>The deeper win is consistency. When the checking is encoded, quality stops depending on who was tired that day. When the trigger is handed over, the work happens at 6am whether anyone feels like it or not. A good loop is the first employee who never has an off day.</p><p>And the best small teams have all converged here, mostly without reading each other’s notes. incident.io ships at “70% done” and holds the quality bar with <a href="https://blog.lawrencejones.dev/2025/">eval scores and backtests instead of specs</a>. Lovable’s founder says the bottleneck stopped being engineering and became <a href="https://www.lennysnewsletter.com/p/building-lovable-anton-osika">taste and intuition about users</a>. Every runs whole products on <a href="https://every.to/chain-of-thought/compound-engineering-how-every-codes-with-agents">one person orchestrating agent fleets</a>. Cherny is describing something real.</p><p>The most careful study I found cuts the other way: METR ran a <a href="https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/">randomised trial on experienced open-source developers</a> and found AI assistance made them 19% <em>slower</em>, while they believed they’d been 20% faster. The gains are real and the self-report is worthless. Always measure merged work, never vibes.</p><h3>Where the framing breaks</h3><p><strong>One: the loop is the easy part.</strong> Even the evangelists concede this when pressed. The real work is the stop condition and the context. Think the check that ends it, the budget that caps it, a target that’s actually reachable. Get those wrong and the loop runs all night producing confident garbage.</p><p>Writing a loop takes minutes. Writing down <em>what good looks like</em>, precisely enough that a machine can hold you to it, is the hardest work I’ve done this year. The team that ships Claude Code says it themselves: <a href="https://www.lennysnewsletter.com/p/building-the-most-ai-pilled-engineering">coding is no longer the bottleneck, verification is</a>. The craft didn’t disappear. It moved into the spec of “done”, and almost nobody has been taught to write one.</p><p><strong>Two: loops decay.</strong> We once shipped a loop that read AI chat transcripts and filed tickets. A coping mechanism, built fast, and it worked, briefly. Then attention moved on, nobody owned it, and it rotted into a machine that tagged the same two engineers in every ticket it created until they muted it.</p><p>Another agent, a daily status reporter, spent its early weeks relentlessly cheerful, announcing the greatest day in the product’s history mid-fire. Technically true from the data it saw given we’d launched three days earlier, so every day was the greatest day in the product’s history. Matt Pocock’s <a href="https://www.explainx.ai/blog/loop-engineering-mainstream-ai-skill-june-2026">distrust of self-improvement loops</a> is the same finding from another angle: unattended feedback curdles.</p><p>A loop is a thing you keep, like a junior hire, never a script you finished. An unowned loop doesn’t sit safely idle. It degrades in production, wearing your name, while the meter runs. When you see headlines like Uber <a href="https://simonwillison.net/2026/Jun/3/uber-caps-usage/">exhausting its annual AI budget in four months</a>, it tells us a loop without a stop condition is an outage with a subscription.</p><p>Even the lab furthest ahead treats decay as a product problem. Anthropic’s managed-agents product includes a maintenance loop they call <a href="https://www.youtube.com/watch?v=7Ue0yM4J-B8">dreaming</a>: an agent that audits the other agents’ memory for rot and proposes fixes with the evidence attached. Believe the decay.</p><p><strong>Three: “my job is to write loops” is spoken from the top of the ladder.</strong> Cherny writes the loops. Someone lives downstream of their output. When agents write most of the code, somebody has to review an organisation’s worth of it (review time is already up <a href="https://cjroth.com/blog/2026-02-18-building-an-elite-engineering-culture">91% on agent-heavy teams</a>) and the burden concentrates on the most experienced people, exactly the ones a team can least afford to burn out.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SrSMPwbs2biqxz-w.png" /></figure><p>The mechanics are cruel. Anthropic’s research across <a href="https://www.anthropic.com/research/claude-code-expertise">400,000 Claude Code sessions</a> found the tool compounds the advantage of experienced operators: directing your own agent extends what you already are. Reviewing everyone else’s inverts it. Same expertise, opposite sign, and what flips it is whose work your name is on. There’s already a word for the job: <a href="https://www.glean.com/blog/work-ai-index-productivity-paradox">botsitting</a>.</p><p>It fails quietly, too. Osmani calls the reflex cognitive surrender, and the <a href="https://executiveeducation.wharton.upenn.edu/thought-leadership/wharton-at-work/2026/05/thinking-fast-slow-and-artificially/">Wharton data</a> is grim: when the AI was wrong, nearly three-quarters of people accepted it anyway, and felt more confident than they would have without it. The validator seat burns the senior person out and launders their judgment into a rubber stamp at the same time.</p><p>The engineers calling this a cron job in a hoodie aren’t confused about cron. They’re objecting to being recast as extras in someone else’s job description.</p><p><strong>Four: the reps problem.</strong> The sharpest observation I’ve heard on this came from a working engineer at the end of a larger piece of work: the decisions had been presented to him after the fact, so they never embedded the way they would have if he’d made them. Shipped faster, understood less.</p><p>A <a href="https://www.anthropic.com/research/AI-assistance-coding-skills">randomised trial from Anthropic</a> put a number on that feeling: engineers who worked through AI scored seventeen points lower on a comprehension quiz about their own code — 50% against 67% for those who wrote it themselves. Judgment is built from reps. Loops eat the reps. And the early labour data looks the way you’d fear: <a href="https://www.cnbc.com/2025/08/28/generative-ai-reshapes-us-job-market-stanford-study-shows-entry-level-young-workers.html">employment for developers aged 22 to 25 is down roughly 20% since 2022</a>.</p><p>The industry is spending down a stock of taste it’s no longer replenishing, and nobody selling a loop-engineering course has an answer for where the next generation of people who can write the stop condition comes from. Maybe I’m wrong and judgment can be built from reviewing alone. I’d like that to be true, but I don’t think anyone’s shown it yet.</p><h3>The version I’d defend</h3><p>So for me Cherny is right about the destination and silent about the terrain. What follows is that model. I wrote it down as the operating system for the last team I built with, and it’s the thing I wish someone had handed me a year ago.</p><p>I’m not the only one arriving here. The day I was finishing this, Addy Osmani, who named loop engineering in the first place, published <a href="https://addyo.substack.com/p/own-the-outer-loop">the case for owning the outer loop</a>: agents run the inner loop (investigate, implement, verify) and a human owns the boundary they cross. The verdict, and the accountability contract with your name on it. Read it; it’s the definitive version of the engineering discipline. My argument sits one step to the side. Osmani’s outer loop is about what the accountable human does. Mine is about how you build a team where somebody genuinely owns it.</p><p>And none of it is new, which is the point. Engineers owning the product is the older idea: Gergely Orosz wrote up <a href="https://blog.pragmaticengineer.com/the-product-minded-engineer/">the product-minded engineer</a> in 2019, Basecamp published <a href="https://basecamp.com/shapeup">Shape Up</a> the same year, Amazon ran two-pizza ownership two decades ago. The seams between PM, design and engineering were priced for a world where building was expensive. AI changed the maths. <strong>When building is cheap, the handoff is the expensive part.</strong></p><p><strong>Product engineers own problems, not tickets.</strong> Every person owns a problem space end-to-end: a metric they’re accountable for moving, named users they actually talk to, a queue they re-rank weekly, and the authority to ship against it. Adoption is the score — shipped-but-unused counts as zero. Loops are staff deployed inside that territory. They do the labour; the human’s name stays on the outcome. Pride, it turns out, attaches to ownership, not keystrokes.</p><p>The market has already repriced for this. Recruiting data across hundreds of frontier teams says the old titles (PM, full-stack engineer) <a href="https://www.jackandjill.ai/">“no longer describe jobs”</a>, and the three traits hired for now are cross-functional, leveraged, discerning. Roles are unbundling into archetypes instead: those who prototype, those who build, those who sweep, those who grow, those who maintain.</p><p><strong>Loops are hires, not scripts.</strong> Treat every loop like a junior employee. It gets interviewed before it touches production, which is what evals are. Onboarded, which is a written spec of good. Managed, meaning a named human owner whose reputation rides on its output. Reviewed on a schedule. And if it stops earning its keep, you let it go. Every loop disaster I’ve been near traces back to treating a hire like a script. The fleet works for the humans, never the reverse.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kiX49z-bUXJYlMFF.png" /></figure><p><strong>The verification spec is the craft.</strong> Seniority used to be measured by how well you did the work. Increasingly it’s measured by how precisely you can define done, which is the harder skill. Kent Beck’s version: <a href="https://leaddev.com/ai/nobody-knows-what-programming-will-look-like-in-two-years">when anyone can build anything, knowing what’s worth building becomes the skill</a>. Budget real time for it, and expect it to feel slow. One of the best proposals I heard this year was a no-code fortnight before a major build: humans arguing about architecture until the decisions actually lived in someone’s head, and only then handing the typing to the machines.</p><p><strong>Watch the scoreboard, not the odometer.</strong> Once agents write most of the code, commit and PR counts inflate on their own. They stop being targets and become health signals, watched for regressions, never chased for records. Shipping a lot is the entry requirement; moving the metrics is the score. Throughput flows to wherever a concrete, this-week next thing exists. The project with weekly demoable milestones drew everything; the flagship framed as a big abstract slab starved quietly. If a bet matters, give it a demoable milestone this week, or the loops and the humans will both drift to whatever has one.</p><p><strong>The ladder is climbed by consent.</strong> Each rung renegotiates what somebody’s job is, and a renegotiation you weren’t part of is just a notice. That’s what the 81/20 gap measures. People don’t argue with the tools; they quietly stay at the bottom, because some rung above them asks for the piece of the job they believe makes them valuable, and nobody asked. Mandates don’t fix this: forced rollouts produce compliance that either doesn’t stick or sticks and corrodes. Who climbs, who refuses, and what actually moves them is its own story.</p><p>Cherny compressed his conclusion into one great line, and the industry mistook the compression for the content. My job is to write loops. Fine. But the loop was never the hard part. The hard part is deciding, precisely and in writing, what good looks like, and building teams where every human still owns something when the loops run without them.</p><p>That’s the actual job in July 2026. It’s a heck of a lot more fun than the slogan suggests.</p><p><em>I’m now writeing about building AI products and the teams that ship them over @ Substack: </em><a href="https://substack.com/@elliotjlt">https://substack.com/@elliotjlt</a></p><h3>Reading</h3><ul><li><a href="https://thenewstack.io/loop-engineering/">The Anthropic leader who built Claude Code says he ditched prompting</a> — The New Stack on Cherny, June 2026</li><li><a href="https://www.productmarketfit.tech/p/stop-prompting-ai-and-start-building">Stop Prompting AI and Start Building Loops</a> — the Cherny numbers</li><li>Addy Osmani, <a href="https://addyo.substack.com/p/own-the-outer-loop">Own the Outer Loop</a> — Elevate, Jul 2026 · earlier: <a href="https://addyosmani.com/blog/loop-engineering/">Loop engineering</a>, the coining</li><li><a href="https://www.anthropic.com/news/introducing-claude-tag">Introducing Claude Tag</a> — Anthropic, June 2026 · <a href="https://www.youtube.com/watch?v=7Ue0yM4J-B8">Inside Anthropic: How Claude Tag Is Changing Agentic Work</a> — AI Native Dev, July 2026</li><li><a href="https://www.anthropic.com/research/coding-agents-social-sciences">Coding agents in the social sciences</a> — Anthropic, the 81%/20% gap</li><li><a href="https://www.anthropic.com/research/claude-code-expertise">Claude Code and expertise</a> — Anthropic, 400K sessions</li><li><a href="https://www.anthropic.com/research/AI-assistance-coding-skills">How AI assistance impacts the formation of coding skills</a> — Anthropic, the 17-point comprehension gap</li><li><a href="https://executiveeducation.wharton.upenn.edu/thought-leadership/wharton-at-work/2026/05/thinking-fast-slow-and-artificially/">Thinking Fast, Slow, and Artificially</a> — Wharton, the acceptance data</li><li><a href="https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/">METR: impact of AI on experienced open-source developers</a> (<a href="https://arxiv.org/abs/2507.09089">arXiv</a>)</li><li><a href="https://cjroth.com/blog/2026-02-18-building-an-elite-engineering-culture">Building an elite engineering culture</a> — Chris Roth, the review-time data (<a href="https://www.faros.ai/blog/ai-software-engineering">Faros source</a>)</li><li><a href="https://www.cnbc.com/2025/08/28/generative-ai-reshapes-us-job-market-stanford-study-shows-entry-level-young-workers.html">Stanford Digital Economy Lab on entry-level employment</a></li><li><a href="https://www.sonarsource.com/blog/state-of-code-developer-survey-report-the-current-reality-of-ai-coding/">Sonar, 2026 State of Code</a> — 42% of committed code AI-assisted</li><li><a href="https://www.lennysnewsletter.com/p/building-the-most-ai-pilled-engineering">Fiona Fung on Lenny’s Podcast</a> — verification as the bottleneck</li><li><a href="https://www.glean.com/blog/work-ai-index-productivity-paradox">Glean Work AI Index</a> — “botsitting”</li><li><a href="https://simonwillison.net/2026/Jun/3/uber-caps-usage/">Uber caps Claude Code usage</a> — Simon Willison</li><li><a href="https://www.jackandjill.ai/">Jack &amp; Jill</a> — “Rewriting tech roles”, the titles repricing</li></ul><h3>From the research behind this piece</h3><p>The operating model in “The version I’d defend” is distilled from a longer operating doc I wrote for the last team I built with. The research base, for anyone who wants the full trail:</p><ul><li>Gergely Orosz — <a href="https://blog.pragmaticengineer.com/the-product-minded-engineer/">the product-minded software engineer</a> (2019) · Basecamp, <a href="https://basecamp.com/shapeup">Shape Up</a> (2019)</li><li>PostHog — <a href="https://posthog.com/handbook/engineering/product-engineering">How to do product, as an engineer</a> · <a href="https://posthog.com/blog/product-engineer-vs-product-manager">Product engineer vs product manager</a> · <a href="https://posthog.com/blog/what-is-a-product-engineer">What is a product engineer</a> · <a href="https://posthog.com/founders/product-360">James Hawkins, “I was wrong” on product people</a></li><li>Linear — <a href="https://www.lennysnewsletter.com/p/how-linear-builds-product">How Linear builds product</a> · <a href="https://linear.app/method">Linear Method</a> · <a href="https://linear.app/now/cx-in-linear">triage + CX</a> · <a href="https://www.lennysnewsletter.com/p/linears-secret-to-building-beloved-b2b-products-nan-yu">Nan Yu on scope hammering</a></li><li>incident.io — <a href="https://blog.lawrencejones.dev/2025/">Lawrence Jones, 2025 retrospective</a></li><li>Every — <a href="https://every.to/chain-of-thought/compound-engineering-how-every-codes-with-agents">Compound engineering</a></li><li>Lovable — <a href="https://www.lennysnewsletter.com/p/building-lovable-anton-osika">Anton Osika on Lenny’s Podcast</a></li><li>Figma — <a href="https://www.figma.com/blog/prototypes-are-the-new-prds/">Prototypes are the new PRDs</a></li><li>Anthropic’s operating model — <a href="https://www.latent.space/p/claude-code">Latent Space on the Claude Code team</a> · <a href="https://www.lennysnewsletter.com/p/anthropics-cpo-heres-what-comes-next">Mike Krieger on Lenny’s</a></li><li>Forward-deployed engineering — <a href="https://newsletter.pragmaticengineer.com/p/forward-deployed-engineers">Pragmatic Engineer on FDEs</a> · <a href="https://fde.academy/blog/how-palantir-invented-the-forward-deployed-engineer-model">Palantir’s FDE model</a> · <a href="https://jobsbyculture.com/blog/forward-deployed-engineer-boom-2026">FDE role growth, 2026</a></li><li>Ceremony norms — <a href="https://changelog.com/podcast/587">Raycast on Changelog</a> · <a href="https://posthog.com/handbook/company/sprints">PostHog sprints</a> + <a href="https://posthog.com/handbook/engineering/support-hero">Support Hero</a> · <a href="https://oxide.computer/blog/engineering-culture">Oxide’s demo Friday</a> · <a href="https://blog.invidelabs.com/the-death-of-the-two-week-sprint/">the death of the two-week sprint</a></li><li>Lenny’s Podcast — <a href="https://www.lennysnewsletter.com/podcast">the archive</a>: Julie Zhuo · Scott Wu · Nicole Forsgren · Nan Yu · Todd Jackson · Patrick Campbell · Geoffrey Moore · Judd Antin</li><li>More — <a href="https://newsletter.pragmaticengineer.com/p/ai-impact-on-software-engineers-part-2">Pragmatic Engineer, AI impact on software engineers</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5bdd4352acab" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Claude Keeps Making the Same Mistakes. So I Started Writing Down the Fixes.]]></title>
            <link>https://medium.com/@elliotJL/your-ai-has-infinite-knowledge-and-zero-habits-heres-the-fix-e279215d478d?source=rss-79bf89c9d6da------2</link>
            <guid isPermaLink="false">https://medium.com/p/e279215d478d</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[claude-code]]></category>
            <dc:creator><![CDATA[Elliot]]></dc:creator>
            <pubDate>Wed, 28 Jan 2026 23:03:59 GMT</pubDate>
            <atom:updated>2026-01-28T23:37:03.821Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b6y_PjouYuUW-oCbTErHdw.png" /><figcaption><em>There’s never been a better time to have a problem. Apparently.</em></figcaption></figure><p>There’s a concept in cognitive science that explains everything wrong with AI coding assistants.</p><p>In 1949, philosopher Gilbert Ryle published <a href="https://www.goodreads.com/book/show/695125.The_Concept_of_Mind"><em>The Concept of Mind</em></a> and introduced a distinction that’s haunted epistemology ever since: <strong>knowing that</strong> versus <strong>knowing how</strong>.</p><p><em>Knowing that</em> is declarative knowledge: facts, information, propositions. You know? Like Paris is the capital of France. Water boils at 100°C. Newcastle United aren’t winning the league anytime soon.</p><p><em>Knowing how</em> is procedural knowledge: the ability to actually do things. Riding a bike. Debugging a race condition. Knowing when to stop refining and ship.</p><p>Claude Code is having a<em> real</em> moment right now. No-coders and coders alike are drinking the Kool-Aid of what is at its core <strong>infinite declarative knowledge</strong>.</p><p>Claude knows that you should assess risks before starting a project.</p><p>It knows that good engineers do pre-mortems.</p><p>It knows that perfectionism kills shipping velocity.</p><p>But it doesn’t know how to do these things reliably. Not in the way an experienced engineer does: automatically, consistently, as an ingrained part of their workflow.</p><p><em>“Treat it like a junior dev”</em> is the common advice. But you can supervise a junior dev because you know where they’ll struggle. Claude’s failures are invisible until they’re not. So the default workflow becomes:</p><blockquote><strong><em>Run it, hope, fix what’s broken when you get back</em></strong><em>.</em></blockquote><p>Fine for a hackathon. Less fine for anything that matters.</p><p>Enter <em>Skills</em>. The procedural layer Claude is missing.</p><h3>The Failure Modes</h3><p>I spent a good part of 2025 building with Claude Code. Watching it succeed at mind-blowing speeds and fail just as fast. After a while, I started noticing the failures weren’t random. They followed patterns.</p><p><a href="https://www.linkedin.com/posts/hireelliot_the-best-way-to-learn-how-something-works-ugcPost-7420043304001703938-2mr9?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAABSS5Z8BirDIaInqV1j374hHHf5npKyr2OI">Building to Break: Understanding Software Development with Claude Code | Elliot Little posted on the topic | LinkedIn</a></p><p>Andrej Karpathy , former OpenAI and Tesla, has been documenting the same thing. Which is why his recent notes on coding with LLMs read like a bug report.</p><p>At the core of his thinking, he is talking about the absence of this procedural knowledge. The model has the facts but not the habits.</p><p>After nearly a year coding with Claude, here’s the what kept me awake at night:</p><p><strong>1. No risk assessment.</strong> Claude jumps straight to implementation. It discovers problems mid-way, then patches reactively. Good engineers do the opposite — they imagine failure first, then prevent it.</p><p><strong>2. Overconfidence without calibration.</strong> Claude says “the function returns X” with identical confidence whether it’s certain or guessing. Users can’t tell what to verify.</p><p><strong>3. Hallucination without verification.</strong> The ol’ drunk LLM chestnut. Claude references files, functions, and code that don’t exist. Looks perfect on screen. Fails the moment you run it.</p><p><strong>4. Infinite loops on failed approaches.</strong> Claude fails at something, retries the exact same approach five times, burns 30 minutes. Never tries a different strategy. Never escalates. Just keeps going with the confidence of someone absolutely certain the wall will move this time.</p><p><strong>5. No definition of done.</strong> Perfectionism loops. “We could also add…” “It would be better if…” No clear declaration that something is shippable. Just endless refinement until the heat death of the universe.</p><p><strong>6. Context rot over long sessions.</strong> Instructions from 20 minutes ago get forgotten. The original goal drifts. Quality degrades even within the context limit. Claude’s even forgotten that you swore at it, or has it.</p><p><strong>7. Imposes preferences instead of matching patterns.</strong> Claude suggests modern patterns in legacy PHP codebases. Introduces new libraries when existing ones are already used. Refactors while fixing bugs.</p><p>These aren’t bugs. They’re the predictable consequences of a system with massive declarative knowledge but no procedural layer — no ingrained habits, no automatic workflows, no “muscle memory” for how expert work actually gets done.</p><h3>What Procedural Knowledge Actually Is</h3><p>When you’re learning to drive, everything is conscious: “Check mirrors. Signal. Check blind spot. Turn wheel.” You retrieve each step deliberately.</p><p>After practice, it becomes automatic. You don’t think “check mirrors” — you just check them. The pattern fires on its own: <em>changing lanes → check mirrors</em>. No deliberate retrieval required.</p><p>Cognitive scientists call these <strong>production rules: </strong>if-then patterns that execute without conscious effort. Expert knowledge is mostly this: patterns so ingrained they feel like instinct.</p><p>The same applies to senior engineers. They don’t consciously think “what could go wrong here?” — they just see it. The missing null check. The race condition. The deployment that’s going to fail at 2am. Years of incidents encoded into automatic caution. Claude doesn’t have this layer.</p><p>It has the declarative knowledge — it can <em>tell you</em> about pre-mortems, risk assessment, matching existing codebase patterns. But it doesn’t <em>do</em> these things automatically. The production rules don’t exist.</p><p><strong>Skills are production rules for LLMs.</strong></p><h3>The Anatomy of a Skill</h3><p>A skill has three components that map directly to how procedural knowledge works:</p><p><strong>1. Trigger conditions (the “if”)</strong></p><pre>description: When a user uploads a .csv file or asks to analyse tabular data,<br>immediately run comprehensive data analysis without asking what they want.</pre><p>This is the condition that fires the production rule. Specific, observable, unambiguous.</p><p><strong>2. Procedural instructions (the “then”)</strong></p><pre>## Instructions<br>1. Read the uploaded CSV<br>2. Run data validation (check for missing values, verify column types)<br>3. Calculate summary statistics for all numeric columns<br>4. Generate three visualisations<br>5. Write findings in markdown with key insights highlighted</pre><p>The sequence of actions that execute when triggered. Step-by-step, explicit, no ambiguity.</p><p><strong>3. Behavioural overrides (the constraints)</strong></p><pre>## NEVER<br>- Ask &quot;What would you like me to do with this?&quot;<br>- Wait for user direction before analysing<br>- Offer options instead of taking action</pre><p>This handles the failure modes of Claude’s default personality — suppressing behaviours that interfere with expert execution.</p><p>The three-layer architecture (metadata → instructions → resources) maps to how procedural knowledge loads: pattern recognition first, then retrieval of the procedure, then execution of the steps.</p><h3>The Orchestration Layer: Skill Potions</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*a8tvMP8sLUBLPwHXiaGmuA.png" /></figure><p>Each failure mode I listed has a skill that addresses it. I’ve been building these out in a collection I’m calling <a href="https://github.com/ElliotJLT/Claude-Skill-Potions">Claude Skill Potions</a>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZNw1WvORSRzsaDGHCPsCPw.png" /></figure><p>Each skill encodes one production rule. The interesting part is <strong>combining them</strong>.</p><p>Complex procedural knowledge isn’t a single production rule — it’s a sequence of rules that fire in coordination. Driving isn’t just “check mirrors.” It’s:</p><blockquote>Check mirrors → signal → check blind spot → turn wheel → accelerate through turn → straighten → check mirrors again.</blockquote><p>The same applies to engineering workflows. An engineer doing a significant task doesn’t just “plan.” They:</p><ol><li>Clarify scope (what are we actually building?)</li><li>Assess risks (what could go wrong?)</li><li>Estimate time (how long will this take?)</li><li>Get confirmation (proceed or adjust?)</li><li>Execute with checkpoints</li><li>Declare done</li></ol><p>This is <strong>orchestrated procedural knowledge</strong> — multiple production rules firing in sequence.</p><p>The battle-plan skill in my collection doesn&#39;t contain a massive monolithic procedure. It sequences other skills:</p><ul><li>rubber-duck (clarify scope)</li><li>pre-mortem (assess risks)</li><li>eta (estimate time)</li><li>you-sure (confirm before execution)</li></ul><p>Each component skill does one thing. The orchestrator sequences them. If you improve pre-mortem, battle-plan gets better automatically.</p><p>This is <strong>composable procedural knowledge</strong> — the same pattern that makes human expertise transferable and refinable.</p><p>Karpathy noticed something similar:</p><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;schema=twitter&amp;url=https%3A//x.com/karpathy/status/2015883857489522876%3Fs%3D20&amp;image=" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/b4b056167ae89296775e315a90aedcb8/href">https://medium.com/media/b4b056167ae89296775e315a90aedcb8/href</a></iframe><blockquote>“Don’t tell it what to do, give it success criteria and watch it go. Get it to write tests first and then pass them… Change your approach from imperative to declarative to get the agents looping longer and gain leverage.”</blockquote><p>Skills are the mechanism for encoding those success criteria and making them fire automatically.</p><h3>The Compound Learning Problem</h3><p>There’s one more piece that matters: procedural knowledge improves through practice.</p><p>When you’re learning to drive, every mistake updates your production rules. Misjudged a turn? <strong>The rule refines</strong>. Cut someone off? <strong>The rule updates</strong>. Over time, the rules become robust through accumulated feedback.</p><p>Claude doesn’t have this by default. Each session starts fresh. Mistakes don’t compound into learning.</p><p><strong>But you can build it.</strong></p><p>Every skill in my collection includes a “Failed Attempts” section — documentation of approaches that didn’t work. This gets read more than any other section, because <strong>failure paths save more time than success paths</strong>.</p><p>The self-improving CLAUDE.md pattern takes it further:</p><pre>Reflect on what went wrong<br>    → Abstract the general pattern<br>    → Generalise into a reusable decision framework<br>    → Write it to CLAUDE.md</pre><p>Every correction becomes a permanent lesson. Your procedural knowledge layer gets smarter through use.</p><h3>The Honest Assessment</h3><p>I’ve now spent more time thinking about Skills than I’ve spent using them. That’s either a sign I’m onto something or a sign I should touch grass. Probably both.</p><p>So — is this framework useful, or am I overcomplicating markdown files?</p><p><strong>What’s real:</strong></p><p>The declarative/procedural distinction genuinely explains why Claude can <em>know about</em> best practices but not <em>follow them</em> automatically. <strong>Skills create production rules that fire on specific conditions</strong>. That’s not hype, that’s just what they do.</p><p>The composition pattern works. Individual skills doing one thing, orchestrated into workflows, is more maintainable than monolithic procedures. I’ve tried the monolithic approach. It breaks.</p><p>The compound learning idea is the highest-ROI concept in the whole system. Skills that capture failure patterns and feed them back create genuine improvement over time.</p><p><strong>What’s overhyped:</strong></p><p>Autonomous operation for extended periods. Self-healing workflows. 20-skill orchestration systems. In practice: emergent bugs, mostly. You set up an elaborate autonomous system, go make dinner feeling like a genius, and return to discover it’s been confidently doing the wrong thing for two hours.</p><p>The future is here, and <strong>it’s still your problem</strong>.</p><p><em>The sweet spot in January 2026:</em> I use about five skills regularly. That feels like the right amount — enough to cover the failure modes that actually burn me, not so many that I’m managing a system instead of doing work.</p><p>Start with failure modes you actually hit. Build a skill for each one. Test whether the production rule fires correctly. Iterate based on real failures. Add orchestration only when you need multi-step sequences.</p><p>You’re not building an AI employee. You’re encoding procedural knowledge that compounds over time —<strong> turning implicit expertise into something reusable, shareable, and improvable</strong>.</p><p>That’s the actual value of skills. Not autonomous agents shipping code while you sleep. Not 20-skill operating systems. Just the procedural layer that LLMs are missing — the <em>knowing how</em> that transforms raw capability into reliable expertise.</p><h3>Try It</h3><p>The skills I’ve been building are open source: <a href="https://github.com/ElliotJLT/Claude-Skill-Potions">Claude Skill Potions</a></p><p>Start with pre-mortem if you want Claude to think before it acts. Start with prove-it if hallucination is your main problem. Start with ship-it if you&#39;re stuck in perfectionism loops.</p><p>Each one is a single production rule. Combine them when you need sequences. Add your own failure patterns as you find them.</p><p>The collection is early . I’m iterating based on what actually works in daily use. If you build something useful, open a PR. If something breaks, open an issue. That’s how procedural knowledge improves: through accumulated feedback from real failures.</p><p><em>In the next article, I’ll share the specific skills we’ve built and actually use — the production rules we’ve encoded, what triggered them, and what we learned when they failed.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e279215d478d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Trust Gap: Why Your AI Product Will Fail Without This]]></title>
            <link>https://medium.com/@elliotJL/the-trust-gap-why-your-ai-product-will-fail-without-this-92b8f60391b3?source=rss-79bf89c9d6da------2</link>
            <guid isPermaLink="false">https://medium.com/p/92b8f60391b3</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[product-management]]></category>
            <dc:creator><![CDATA[Elliot]]></dc:creator>
            <pubDate>Sun, 20 Jul 2025 10:09:04 GMT</pubDate>
            <atom:updated>2025-07-20T10:09:04.990Z</atom:updated>
            <content:encoded><![CDATA[<p>OpenAI just dropped ChatGPT Agent for the masses. But if agents work so well, where are all the success stories?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1yv_FXOLILPiwkz6.png" /></figure><p>Source: <a href="https://www.barbaragibson.pl/financialinvestment">https://www.barbaragibson.pl/financialinvestment</a></p><p>AI products are evolving from simple chatbots toward systems that handle real workflows. But they’re all hitting the same wall: trust. Trust isn’t just about whether the AI works. It’s whether people feel confident delegating tasks to it, even when it works imperfectly.</p><p>Research shows that most fail because they’re either <strong>too limited to be useful or too autonomous to be trustworthy</strong>.</p><p>OpenAI’s latest release, ChatGPT agent, represents an interesting case study in this evolution. You can now ask ChatGPT to handle requests like “look at my calendar and brief me on upcoming client meetings” or “analy`e three competitors and create a slide deck.”</p><p>Whether it delivers consistently isn’t the point, it shows where AI products are heading: into <strong>messy, high-stakes workflows where failure erodes trust faster than it builds value</strong>.</p><h3>Why This Evolution Matters (With Caveats)</h3><p>ChatGPT agent represents another attempt to narrow the gap between “AI can help you think about this” and “AI can actually do this for you.”</p><p>It uses its own virtual computer, switching between different tools to complete workflows. The technical approach is interesting: visual browser, text analysis, terminal access, API connections, but the real test will be whether it works reliably in practice.</p><p>OpenAI claims you’re <strong>always in control</strong>, with permission requests before consequential actions and the ability to interrupt at any point. That sounds good on paper, but we’ve heard similar promises before. The challenge with agentic AI isn’t just technical capability but whether people <strong>will trust it enough to actually use it for important tasks</strong>.</p><h3>Trust: The Thing Nobody Talks About Until It’s Gone</h3><p>Here’s the uncomfortable truth about AI products: according to recent global surveys, only <strong>46%</strong> of people trust AI systems, yet <strong>71%</strong> of organisations regularly use generative AI (KPMG, 2025). There’s a massive gap between what companies are building and what people actually trust enough to use.</p><p>The successful AI products aren’t the ones with the fanciest models. They’re the ones that systematically build trust through boring stuff like reliability, transparency, and not making you feel like a mug when they inevitably mess something up.</p><p><strong>Technical reliability isn’t negotiable</strong>. PayPal’s AI fraud detection system handles 430 million accounts with 25% fraud reduction because it works consistently. Amazon’s recommendation engine drives customer loyalty not because it’s clever, but because when it suggests something, it’s usually not completely wrong. Research shows that 66% of users increasingly rely on AI output without verification, which means when your AI gets things wrong, it doesn’t just break workflows, it breaks confidence.</p><p><strong>Transparency builds the kind of trust that actually matters</strong>. The most trusted AI products show their working. GitHub Copilot succeeds because developers understand its capabilities and limitations. Perplexity differentiates itself through source citations that let you verify claims rather than just hoping the AI isn’t having a creative moment.</p><p><strong>Human oversight creates proper safety nets</strong>. Microsoft 365 Copilot maintains user control over final decisions. Notion AI positions itself as a collaborative partner rather than a replacement. There’s a pattern here: training, empowering, and including humans to manage AI tools consistently addresses trust gaps. The trends show that Human-AI collaboration works better than full automation when it comes to user confidence.</p><p>And then there’s the boring-but-essential stuff around data privacy and security. <strong>48%</strong> of businesses are entering non-public information into AI tools, which means robust data governance is now becoming even more of a competitive advantage than before.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/886/0*3lLVLh9sREI0Rux3.png" /><figcaption>@sama on X.com</figcaption></figure><h3>The Four Horsemen of AI Product Failure</h3><p>It sounds like most AI products fail for reasons that have nothing to do with the underlying technology. They seem to follow a pattern so consistent you could set your watch by it.</p><p><strong>First: Building solutions to non-problems.</strong> 34% of AI startups fail due to poor product-market fit — they build something nobody actually wants. The Humane AI Pin ($699 plus $24 monthly subscription) promised to replace smartphones but mostly just overheated while doing things your phone already does better.</p><p><strong>Second: Shipping without proper validation.</strong> McDonald’s AI drive-thru added 260 chicken nuggets to orders and bacon to ice cream before shutting down. A relatively harmless failure (or win if you love nuggets). More concerning: Google Gemini generated historically inaccurate images of diverse Nazi soldiers, and NYC’s MyCity chatbot advised businesses they could fire workers reporting sexual harassment. All shared inadequate testing and insufficient human oversight, but with vastly different stakes.</p><p><strong>Third: Overpromising the user experience.</strong> The Rabbit R1 ($199) seemed smarter by avoiding Humane’s mistakes by using a touch screen instead of a projector, lower price, no subscription. But at launch, it didn’t do most of what its creators promised. It felt as limited as any other GenAI app on your smartphone, just with extra steps and a cute orange case.</p><p><strong>Fourth: Launching without learning mechanisms.</strong> The successful AI products create systems where user interaction makes the AI better over time. The failures launch with fanfare, then wonder why nobody uses them twice. Consider the graveyard of AI note-taking apps like Otter.ai or Granola’s early competitors or the dozens of “meeting summarisers” that promised to capture your discussions but never learned your writing style, tone, or what information you actually found useful. Without continuous improvement based on real usage, even technically impressive AI becomes irrelevant quickly.</p><h3>What Actually Works: Strategic Principles</h3><p>Trends show that the AI products that succeed share specific strategic patterns that have nothing to do with having the fanciest models:</p><p><strong>They solve real problems with AI-native approaches</strong>. Klarna’s AI assistant handles customer service with sub-2-minute resolution times — impossible with traditional methods. It’s not AI for AI’s sake; it’s AI because that’s the only way to solve the actual problem at scale.</p><p><strong>They integrate rather than replace</strong>. Notion AI succeeds because it feels native to the existing product. Microsoft 365 Copilot leverages familiar Office interfaces. The most successful AI products don’t demand workflow changes but make existing workflows better.</p><p><strong>They build connective tissue, not replacement organs</strong>. Anthropic’s Model Context Protocol (MCP) exemplifies this approach. Instead of trying to recreate every tool, it creates standardised connections between AI and existing systems like GitHub, Google Drive, or databases. MCP lets AI agents access the tools you already use rather than forcing you to abandon them. It’s the difference between AI that works with your workflow versus AI that demands you rebuild around it.</p><p><strong>They start narrow and expand gradually</strong>. GitHub Copilot started with code suggestions before expanding to chat and autonomous features. Smart AI companies begin with well-defined problems and expand based on user feedback rather than trying to solve everything on day one.</p><h3>UX That Builds Trust: Applied Design Patterns</h3><p>The successful AI products implement specific interface patterns that systematically build user confidence:</p><p><strong>Transparency patterns make AI decisions understandable</strong>. The most trusted products show confidence levels, cite data sources, and implement “show your work” features. Salesforce’s citations make Einstein generated content readable and clear on its source, giving users the option to validate the source flagged.</p><p><strong>Control patterns maintain human agency</strong>. Easy override mechanisms, granular privacy controls, and customisable AI behaviour. Salesforce implements “Mindful Friction”: system-wide pauses like confirmation steps and delayed actions that create intentional breaks in the user experience to ensure human engagement at critical junctures.</p><p><strong>Reliability patterns create predictable experiences</strong>. Graceful error handling, clear system status indicators, and backup human support options. Successful AI products fail gracefully and recover quickly.</p><p><strong>Personalisation patterns create adaptive experiences</strong>. Context-aware recommendations that improve with use, learning from user corrections, and adaptive interface elements. The most successful AI products get better through user interaction.</p><h3>Where Things Are Heading</h3><p>I don’t think we’re at a point where we can say that AI agent will replace existing software. But they’ll bridge the gaps between systems that were never designed to work together. Companies will use them to connect Enterprise Resource Planning (ERP) systems with customer databases, or automate handoffs between tools that currently require manual data entry.</p><p>The successful AI products will be the connective tissue, not the replacement organs.</p><p>As AI becomes core to business strategy, companies need clear, repeatable methods for tracking its real-world impact. The challenge isn’t proving AI works in demos anymore. <strong>You need to prove it works when real users depend on it</strong>.</p><p>I’m finding that trust-first design isn’t a nice-to-have anymore, the difference between a tool people use once and a product they depend on. Research shows that trusted companies outperform their peers by over <strong>400%</strong>, demonstrating the tangible business benefits of responsible and ethical AI use. Products that prioritise transparency, user control, and explainability will hopefully outcompete technically superior alternatives that users don’t trust.</p><p><strong>Human-AI collaboration beats full automation consistently.</strong> The most successful AI products enhance human capabilities rather than replace them, creating “centaur” experiences where AI handles routine tasks while humans focus on creative and strategic work.</p><h3>The Reality Check We Needed</h3><p>In my current role I’m trying to solve real student problems, and one of the key things I’ve learned so far is how wide the gap still is between what AI can do and what people trust it to do. Most of the promises break down in the messy middle.</p><p>The gap between what AI can do and what people trust it to do is still wide. ChatGPT agent doesn’t feel revolutionary in mid-2025, and maybe that’s the point. The shift isn’t in capability. It’s in how these tools are trying to <strong>earn trust without getting in the way</strong>.</p><p>Whether it works or just moves the problem sideways is still unclear. But it’s a decent test case for what makes an AI product stick, and what makes people come back the next day.</p><h3>References:</h3><ul><li><strong>OpenAI’s official announcement</strong> — “Introducing Operator” (January 2025)</li><li><strong>TechRadar</strong> — “OpenAI’s next ‘giant breakthrough’ tipped to land soon and control your computer”</li><li><strong>BGR</strong> — “ChatGPT Operator AI agent release set for early 2025”</li><li><strong>Substack/Nate’s Newsletter</strong> — “OpenAI Operator: A Complete Hands-On Review with Screenshots”</li><li><strong>Every.to</strong> — “We Tried OpenAI’s New Agent — Here’s What We Found”</li><li><strong>KPMG</strong> — “Trust in artificial intelligence: global insights 2025” (survey of global AI trust levels)</li><li><strong>Harvard Business Review</strong> — “AI’s Trust Problem” (May 2024)</li><li><strong>LangChain Blog</strong> — “The Hidden Metric That Determines AI Product Success” (CAIR framework)</li><li><strong>RAND Corporation</strong> — “The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed”</li><li><strong>Medium/AIProToolkit</strong> — “Top 10 AI Tools Reddit Users Swear By in 2024”</li><li><strong>Creator Economy</strong> — “24 AI Tools Ranked from Essential to Forgettable”</li><li><strong>Buffer</strong> — “8 of the Best AI Productivity Tools to Help You Optimize How You Work”</li><li><strong>Notion AI reviews</strong> from InBound Blogging and All About AI</li><li><strong>CIO</strong> — “12 famous AI disasters”</li><li><strong>Medium</strong> — “13 AI Disasters of 2024” and “99% of AI Startups Will Be Dead by 2026”</li><li><strong>MIT Technology Review</strong> — “The biggest AI flops of 2024”</li><li><strong>Tom’s Hardware</strong> — “Research shows more than 80% of AI projects fail”</li><li><strong>Microsoft Blogs</strong> — “How real-world businesses are transforming with AI — with 261 new stories”</li><li><strong>McKinsey</strong> — “The state of AI: How organizations are rewiring to capture value”</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=92b8f60391b3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What I Learned Shipping AI to Users Who Can’t Afford Bad Advice]]></title>
            <link>https://medium.com/@elliotJL/what-i-learned-shipping-ai-to-users-who-cant-afford-bad-advice-606243049534?source=rss-79bf89c9d6da------2</link>
            <guid isPermaLink="false">https://medium.com/p/606243049534</guid>
            <category><![CDATA[future]]></category>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Elliot]]></dc:creator>
            <pubDate>Mon, 14 Jul 2025 20:24:43 GMT</pubDate>
            <atom:updated>2025-07-14T20:24:43.848Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The EU’s new AI Code of Practice reads like a forensic report on real-world AI failure, &amp; a practical guide on how to avoid repeating them.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/622/0*hWizJnryf7PqOx2f" /><figcaption><a href="https://www.joewebbart.com/store/new-building-blocks-silkscreen">https://www.joewebbart.com/store/new-building-blocks-silkscreen</a></figcaption></figure><h3>The Problem That Keeps Me Up at Night</h3><p>For the past year, my world has involved deploying AI systems to people who can’t afford a hallucination or bad advice from what is becoming a <em>trusted</em> source. I’m talking about the likes of first-generation university applicants, people drowning in visa paperwork, job seekers in brutal markets.</p><p>This isn’t abstract: the wrong suggestion can shatter plans or close doors permanently. Think:</p><p>The legal research chatbot that invented fake court cases, causing a real lawsuit to be dismissed.</p><p>The city-sponsored business chatbot that advised entrepreneurs to break the law by taking a cut of their workers’ tips and discriminating against tenants based on their source of income.</p><p>The medical diagnosis tool that learned to associate COVID-19 risk with whether a patient was lying down in a scan, rather than from the actual medical data.</p><p>The bigger issue isn’t model drift or wallet-destroying pro-plans. It’s a stat I stumbled upon from a 2025 KPMG study, buried deep in the EU’s new AI regulatory documents:<strong> 66% of users don’t check AI outputs</strong>.</p><p>That means two-thirds of the people we’re building for are just <strong>taking the model’s word for it</strong>.</p><p>This month, the European Commission received the final version of the Code of Practice for General-Purpose AI Models. It’s a voluntary framework, but after reviewing it, I realised it’s less a rulebook and more a detailed autopsy report of how AI goes wrong in the real world. And it should change how we build.</p><h3>The Three User Types Your Metrics Are Missing</h3><p>In our user research, people tend to cluster into three patterns of trust behaviour. These aren’t fixed and can change based on the context, but each pays a different cognitive price, a kind of <em>Trust Tax</em>, for engaging with the model.</p><ul><li><strong>The Skeptics (High Trust Tax):</strong> They’ve been burned before. They cross-check everything, demand sources, and live for the system messages. They need verification tools that aren’t just theatre.</li><li><strong>The Practicals (Medium Trust Tax):</strong> They use AI to get unblocked, then spot-check the critical stuff. They get that it’s flawed but useful. These users are your best source of feedback; they find the cracks you didn’t know you had.</li><li><strong>The Default-Trusters (Zero Trust Tax):</strong> They assume software is generally correct. They’re often the most vulnerable and the least likely to question an output. These users represent the greatest risk if guardrails aren’t robust, and are the reason our internal guardrails can’t be an afterthought.</li></ul><p>If your system only works safely for the Practicals, you’ve likely got a trust gap, and it’s the Default-Trusters who’ll fall through it.</p><h3>Breaking Points: Where Your Code Hits a Human Life</h3><p>The EU framework refers to <em>“trigger points”</em> that require extra scrutiny. I think of these instead as <em>Breaking Points</em>:<em> </em>moments where a single AI failure causes harm that isn’t easily reversed, repaired, or forgotten.</p><p>A 2025 <em>California Management Review</em> analysis of 106 AI controversies found that <strong>30%</strong> <strong>stemmed from algorithmic bias</strong>, and <strong>14%</strong> <strong>from decisions no one could explain</strong>. These are proven, repeatable failures and the consequences don’t stay technical.</p><ul><li><strong>Identity Breaking Points</strong>: Your document processor fails on a non-Western name, and a key application is rejected. That’s not a glitch but a denial with real consequences.</li><li><strong>Financial Breaking Points</strong>: AI models used in housing or loan assessments penalise people with non-traditional income. The model sees risk; the human result is being locked out of housing or credit.</li><li><strong>Timeline Breaking Points</strong>: Visa applications run on hard deadlines. A wrong suggestion isn’t a minor mistake but can mean losing a year, or the entire opportunity.</li><li><strong>Trust Breaking Points</strong>: Amazon scrapped its recruitment AI after it penalised resumes that mentioned “women’s.” The system was withdrawn, but the reputational damage, especially with the talent it aimed to attract , was done.</li></ul><p>The regulations talk about building in <em>safety margins</em>. To me, that means assuming your model is less capable in real-world contexts than your internal benchmarks suggest, and designing accordingly.</p><h3>What 106 AI Scandals Tell Us About Saying Sorry</h3><p>The <em>California Management Review</em> study had another killer insight: people judge AI failures through two completely different lenses.</p><ul><li><strong>Capability Failures (“Your AI is incompetent”):</strong> This is usually about bias or explainability. The fix is technical, and you rebuild trust by showing it works consistently over time.</li><li><strong>Character Failures (“Your company has bad values”):</strong> This is almost always about privacy violations. This is more a governance failure than a technical problem. Stakeholders see it as a sign of your company’s malevolent intent.</li></ul><p>When a major tech company’s credit algorithm is found to be biased, people don’t see a simple coding error but <strong>a powerful institution embedding discriminatory values into its products</strong>. Trying to fix a character failure with a technical patch just makes you look like you don’t get it.</p><h3>Implementation That Isn’t Just Security Theatre</h3><p>The EU documents are dense, but they map to things you can actually build.</p><ul><li><strong>Set up a regular risk review: </strong>The EU framework talks about a “Systemic Risk Loop”. In other words, regularly checking your AI for risks: spotting them, understanding how serious they are, deciding if they’re acceptable, and then taking clear steps to reduce or remove them. Doing a simplified version of this monthly for key features can stop mistakes before they affect your users. In practice, this means <strong>making safety part of sprint planning</strong>, not something left to audits or ops.</li><li><strong>Know exactly where your data comes from: </strong>Keep a clear record of the training data your AI uses. EU guidelines suggest simple, practical steps, like respecting the basic robots.txt protocol, a standard that&#39;s been around for decades. If your system can&#39;t handle this basic practice, it&#39;s likely not robust enough to manage more complex data rights issues.</li><li><strong>Test your system from an outsider’s perspective: </strong>The regulations advise having independent experts evaluate your AI just as rigorously as someone trying to break or misuse it would. This might feel challenging or costly, but it’s one of the best ways to uncover critical flaws you’d otherwise miss. This point also came up at the <a href="https://medium.com/@elliotJL/what-i-learned-spending-a-week-with-100-ai-leaders-3517ac3dc16c">AI Summit London</a>.</li><li><strong>When things go wrong, learn openly from mistakes: </strong>If your AI causes significant harm, EU rules say you must report it promptly (within 2 to 15 days). More importantly, you’re expected to carry out a careful investigation into what went wrong and use those lessons to improve your product. The goal here is systematic learning, not assigning blame.</li></ul><h3>Practical Realities of Deploying AI</h3><ul><li><strong>Vulnerable users aren’t beta testers. </strong>Mistakes hit vulnerable users hardest, and the consequences aren’t reversible. Bad AI visa advice or flawed financial guidance doesn’t lead to minor annoyances but ruins lives. If your test strategy puts high-risk users on the front line, you’re creating unacceptable harm.</li><li><strong>Transparency often beats raw accuracy. </strong>A transparent system at 80% accuracy, one that openly shares limitations and clearly communicates risks, is frequently safer than a “black box” system boasting 95% accuracy without explanation. Transparency helps manage user expectations and prevents dangerous overconfidence. <em>But remember: transparency alone won’t fix fundamentally poor-performing models. Accuracy matters most in life-critical decisions.</em></li><li><strong>Your blind spots become your AI’s biases, unless you actively challenge them. </strong>Teams naturally embed hidden biases into products when they rely only on their internal perspective. Regular, structured conversations with users, especially before you build, help surface hidden assumptions and biases. Rather than using vulnerable users to test risky products, you’re engaging them early to ensure you understand their context and avoid causing unintended harm.</li></ul><p>I’m treating these regulatory frameworks like free consulting: an expensive, exhaustive report on how to build AI that doesn’t blow up in your face.</p><p>These frameworks aren’t perfect, but they’re a signal: the burden of proof is shifting. The cost of failure won’t stay technical for long.</p><ul><li>Code of Practice for General-Purpose AI Models — Safety &amp; Security Chapter (EU AI Office, 2025)</li><li>Code of Practice for General-Purpose AI Models — Transparency Chapter (EU AI Office, 2025)</li><li>Code of Practice for General-Purpose AI Models — Copyright Chapter (EU AI Office, 2025)</li><li>KPMG/University of Melbourne Global AI Trust Study (2025)</li><li>“The Reputational Risks of AI” (California Management Review, 2025)</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=606243049534" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Amsterdam Built the ‘Perfect’ Ethical AI System. It Still Failed. Here’s Why.]]></title>
            <link>https://medium.com/@elliotJL/amsterdam-built-the-perfect-ethical-ai-system-it-still-failed-here-s-why-8dc8072beea3?source=rss-79bf89c9d6da------2</link>
            <guid isPermaLink="false">https://medium.com/p/8dc8072beea3</guid>
            <category><![CDATA[politics]]></category>
            <category><![CDATA[ethics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[algorithms]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Elliot]]></dc:creator>
            <pubDate>Tue, 17 Jun 2025 12:55:58 GMT</pubDate>
            <atom:updated>2025-06-17T12:55:58.236Z</atom:updated>
            <content:encoded><![CDATA[<p>One of the most transparent AI projects in history reveals why building fair algorithms might be harder than we ever imagined.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kO-mG9mL5AxG7_RQ1J6Slw.png" /></figure><p>Last week, an investigation by Lighthouse Reports and MIT Technology Review published what might be the most devastating case study in AI ethics history. Amsterdam, a city that essentially wrote the playbook on responsible AI governance, spent five years and €535,000 trying to build a “fair” welfare fraud detection system.</p><p>They did absolutely everything right according to current best practices:</p><ul><li>Transparent algorithms instead of black boxes.</li><li>Extensive bias testing.</li><li>Community consultation.</li><li>Academic oversight.</li></ul><p>They even published their source code on GitHub for the world to scrutinise.</p><p>And it still didn’t work.</p><p>The system they built, called <strong>Smart Check</strong>, was supposed to detect welfare fraud more fairly and effectively than human caseworkers. Instead, it managed to be both biased and rubbish at its actual job by flagging more people while being no more accurate than the humans it was meant to replace.</p><h3>When Transparency Becomes a Weapon</h3><p>What makes this investigation particularly brutal is how transparent Amsterdam was throughout the process. Most companies won’t even admit they use algorithms for important decisions, let alone publish detailed documentation about how they work.</p><p>Amsterdam went the opposite direction. Due to GDPR restrictions, they couldn’t just hand over raw data to investigators. Instead, they created what researchers call a “remote access regime”, running custom analyses on real data and sharing aggregated results. They published multiple versions of the Smart Check algorithm, extensive documentation, bias analysis scripts, and model files.</p><p>It’s probably the most transparent algorithmic audit ever conducted. And that transparency is exactly what makes the failure so damning.</p><p>The system used an Explainable Boosting Machine that evaluated 15 characteristics from welfare applications, assigning risk scores between 0 and 1. Applications scoring above 0.56 got flagged for investigation. Simple, explainable, and theoretically fair.</p><p>Except fairness turned out to be more complicated than anyone anticipated.</p><h3>The Mathematical Impossibility of Fair</h3><p>The first sign that things were off was that the Smart Check model showed strong bias against <strong>migrants and men</strong>.</p><p>So Amsterdam’s team did what every responsible AI guide tells you to do: they re-weighted their training data to eliminate the bias. And it worked! The bias disappeared completely in testing.</p><p>But when they deployed the system in the real world, something fascinating and pretty horrible happened. The bias came back. This time in the opposite direction. Now the system was wrongly flagging <strong>Dutch nationals and wome</strong>n at higher rates. As the investigation notes:</p><blockquote>“Trying to reduce bias according to one definition will invariably result in increased bias according to another definition under certain conditions.”</blockquote><p>There are multiple competing definitions of fairness, and it’s mathematically impossible to optimise for all of them simultaneously.</p><p>Think about it this way: you can’t create a football team that’s simultaneously the fastest, strongest, most technical, and best at headers. Some of these requirements contradict each other. Fairness works the same way. <strong>Different groups have different and sometimes incompatible ideas about what constitutes fair treatment</strong>.</p><p>Amsterdam discovered this the hard way, in front of a global audience.</p><h3>The Netherlands: Ground Zero for AI Gone Wrong</h3><p>The timing of this failure is particularly awkward for the Netherlands, which has become something of a poster child for both AI innovation and AI disasters.</p><p>The country scores first in the Global Index on Responsible AI and has developed extensive governance frameworks that other nations look to for guidance. But their track record with AI in social services is… less impressive.</p><p>In 2021, a childcare benefits scandal saw 35,000 families falsely accused of fraud by an algorithmic system, leading to the government’s resignation. In Rotterdam, an algorithm ranked welfare recipients based on clothing and fluency in Dutch, systematically targeting single mothers with migrant backgrounds.</p><p>Smart Check was supposed to be different. It was going to prove that AI could be made fair with the right processes and oversight. Instead, it’s become another cautionary tale about the dangers of automating social services.</p><h3>The Consultant vs. Reality Divide</h3><p>The investigation reveals a fascinating philosophical split about what these failures actually mean.</p><p>Ethical-AI consultant Jiahao Chen argued: “Why do we hold AI systems to a higher standard than human agents?” He’s got a point — human caseworkers also show bias, but there’s “no attempt to correct it systematically.”</p><p>On the other side, digital rights advocate Hans de Zwart argued that Amsterdam had “used all the tools in the responsible-AI tool kit” but was still pursuing “something that is fundamentally a bad idea.” City council member Elisabeth IJmker worried that the technical process obscured larger questions of “philosophical” and “political values” that hadn’t been addressed.</p><p>To me, this tension reveals something important: <strong>we’re arguing about the wrong things</strong>. The debate shouldn’t be whether AI is more or less biased than humans. It should be whether certain decisions should be algorithmic at all.</p><h3>The Seduction of the Technical Fix</h3><p>As de Zwart puts it: “We are being seduced by technological solutions for the wrong problems. Should we really want this? Why doesn’t the municipality build an algorithm that searches for people who do not apply for social assistance but are entitled to it?”</p><p>This gets to the heart of what’s wrong with most AI ethics discussions. Amsterdam spent €535,000 trying to build a fair fraud detection system. They could have spent that money on better support for welfare recipients, clearer application processes, or addressing the root causes of poverty.</p><p>Instead, they built a system that perpetuates the fundamental assumption that welfare recipients are potential fraudsters who need to be watched and evaluated. Of course, the algorithm didn’t create this worldview, it just encoded it in mathematics and made it seem objective.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hO2AUbauLh-cjgks9qwqRw.png" /></figure><h3>What This Means for AI Governance</h3><p>This investigation comes at a crucial moment for AI regulation. The EU AI Act requires “high risk” systems to be registered. Cities like New York are mandating bias audits. The implicit assumption behind all this regulation is that AI can be made fair with the right processes and oversight.</p><p><strong>Amsterdam’s failure suggests that assumption might be wrong.</strong></p><p>The mathematical impossibility of achieving all fairness definitions simultaneously isn’t a technical problem that can be solved with better algorithms or more data. It’s a fundamental feature of how fairness works in a messy, contradictory world where different groups have different needs and values.</p><p>Recent research shows that organisations implementing responsible AI frameworks struggle with “significant challenges in organisational adaptation, characterised by knowledge gaps, a limited emphasis on stakeholder engagement, and a strong focus on control.” In other words, they’re focusing on technical compliance rather than addressing the underlying social and political questions.</p><h3>The Research That Changes Everything</h3><p>What makes this investigation historically significant is that it represents:</p><ul><li>One of the first comprehensive audits of a system built under responsible AI guidelines with full transparency</li><li>A new model for algorithm accountability through remote access regimes that respect privacy laws</li><li>Evidence of the mathematical impossibility of achieving all fairness definitions simultaneously</li><li>A test case for whether current AI governance frameworks are adequate</li></ul><p>Harvard’s Michael Sandel identifies this as one of three major ethical concerns with AI:</p><blockquote>“Can smart machines outthink us, or are certain elements of human judgment indispensable in deciding some of the most important things in life?”</blockquote><p>Amsterdam’s experiment suggests the answer is clear: some decisions are too important, too complex, and too human to be automated fairly.</p><h3>The Uncomfortable Truth</h3><p>The failure of Amsterdam’s “gold standard” approach raises uncomfortable questions about the entire AI ethics industry. If responsible AI can’t work under ideal conditions: with unlimited transparency, extensive consultation, and mathematical rigor, then maybe the problem isn’t implementation.</p><p>Maybe the problem is that <strong>we’re trying to solve fundamentally political and social problems with technical solutions</strong>.</p><p>This doesn’t mean we should give up on making AI less harmful. But it does mean we need to be more honest about what technology can and can’t do. Some forms of bias can be measured and reduced. Some can’t. It might be that some decisions should never be algorithmic, regardless of how fair we can make the algorithms.</p><p>The most important lesson from Amsterdam isn’t about building better AI systems. It’s about recognising when not to build them at all.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*MKGf24LJi-OOrt3E.png" /></figure><p><strong>Sources:</strong></p><p><strong>Primary Investigation:</strong></p><ol><li>Lighthouse Reports. “The Limits of Ethical AI.” June 11, 2025. <a href="https://www.lighthousereports.com/investigation/the-limits-of-ethical-ai/">https://www.lighthousereports.com/investigation/the-limits-of-ethical-ai/</a></li><li>Lighthouse Reports. “How we investigated Amsterdam’s attempt to build a ‘fair’ fraud detection model.” June 11, 2025. <a href="https://www.lighthousereports.com/methodology/amsterdam-fairness/">https://www.lighthousereports.com/methodology/amsterdam-fairness/</a></li><li>Guo, Eileen, Gabriel Geiger, and Justin-Casimir Braun. “Inside Amsterdam’s high-stakes experiment to create fair welfare AI.” MIT Technology Review, June 11, 2025. <a href="https://www.technologyreview.com/2025/06/11/1118233/amsterdam-fair-welfare-ai-discriminatory-algorithms-failure/">https://www.technologyreview.com/2025/06/11/1118233/amsterdam-fair-welfare-ai-discriminatory-algorithms-failure/</a></li><li>GitHub Repository: Lighthouse-Reports/amsterdam_fairness. <a href="https://github.com/Lighthouse-Reports/amsterdam_fairness/tree/main">https://github.com/Lighthouse-Reports/amsterdam_fairness/tree/main</a></li></ol><p><strong>Government Sources:</strong></p><p>5. Municipality of Amsterdam. “Researchworthiness: Smart check livelihood.” Algorithm Register. <a href="https://algoritmes.overheid.nl/en/algoritme/researchworthiness-smart-check-livelihood-municipality-of-amsterdam/95794697">https://algoritmes.overheid.nl/en/algoritme/researchworthiness-smart-check-livelihood-municipality-of-amsterdam/95794697</a></p><p>6. Municipality of Amsterdam. Algorithm Register. <a href="https://algoritmes.overheid.nl/en/organisatie/gm0363/gemeente-amsterdam">https://algoritmes.overheid.nl/en/organisatie/gm0363/gemeente-amsterdam</a></p><p><strong>Academic and Expert Analysis:</strong></p><p>7. University of Amsterdam. “‘The Dutch childcare benefit scandal shows that we need explainable AI rules’.” February 8, 2023. <a href="https://www.uva.nl/en/shared-content/faculteiten/en/faculteit-der-rechtsgeleerdheid/news/2023/02/childcare-benefit-scandal-transparency.html">https://www.uva.nl/en/shared-content/faculteiten/en/faculteit-der-rechtsgeleerdheid/news/2023/02/childcare-benefit-scandal-transparency.html</a></p><p>8. Rachovitsa, Adamantia, and Niclas Johann. “The Human Rights Implications of the Use of AI in the Digital Welfare State: Lessons Learned from the Dutch SyRI Case.” Human Rights Law Review, Volume 22, Issue 2, June 2022. <a href="https://academic.oup.com/hrlr/article/22/2/ngac010/6568079">https://academic.oup.com/hrlr/article/22/2/ngac010/6568079</a></p><p>9. “Frontrunner model for responsible AI governance in the public sector: the Dutch perspective.” AI and Ethics. <a href="https://link.springer.com/article/10.1007/s43681-024-00596-2">https://link.springer.com/article/10.1007/s43681-024-00596-2</a></p><p>10. Harvard Gazette. “Ethical concerns mount as AI takes bigger decision-making role.” January 3, 2024. <a href="https://news.harvard.edu/gazette/story/2020/10/ethical-concerns-mount-as-ai-takes-bigger-decision-making-role/">https://news.harvard.edu/gazette/story/2020/10/ethical-concerns-mount-as-ai-takes-bigger-decision-making-role/</a></p><p><strong>Policy and Advocacy:</strong></p><p>11. Amnesty International. “Dutch childcare benefit scandal an urgent wake-up call to ban racist algorithms.” November 1, 2021. <a href="https://www.amnesty.org/en/latest/news/2021/10/xenophobic-machines-dutch-child-benefit-scandal/">https://www.amnesty.org/en/latest/news/2021/10/xenophobic-machines-dutch-child-benefit-scandal/</a></p><p>12. WBUR On Point. “What happened when AI went after welfare fraud.” March 13, 2025. <a href="https://www.wbur.org/onpoint/2025/03/13/ai-algorithms-welfare-fraud-benefits">https://www.wbur.org/onpoint/2025/03/13/ai-algorithms-welfare-fraud-benefits</a></p><p><strong>Technical and Regulatory Context:</strong></p><p>13. Project Sherpa. “Ethics of Public Use of AI and Big Data: The Case of Amsterdam’s Crowdedness Project.” <a href="https://www.project-sherpa.eu/885-2/">https://www.project-sherpa.eu/885-2/</a></p><p>14. European Commission. “Ethics guidelines for trustworthy AI.” <a href="https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai">https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai</a></p><p>15. NIST. “AI Risk Management Framework.” <a href="https://www.nist.gov/itl/ai-risk-management-framework">https://www.nist.gov/itl/ai-risk-management-framework</a></p><p><strong>Related Lighthouse Reports Investigations:</strong></p><p>16. Lighthouse Reports. “The Algorithm Addiction.” March 5, 2023. <a href="https://www.lighthousereports.com/investigation/the-algorithm-addiction/">https://www.lighthousereports.com/investigation/the-algorithm-addiction/</a></p><p>17. Lighthouse Reports. “Automating Distrust.” April 26, 2024. <a href="https://www.lighthousereports.com/investigation/automating-distrust/">https://www.lighthousereports.com/investigation/automating-distrust/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8dc8072beea3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What I Learned Spending A Week With 100+ AI Leaders]]></title>
            <link>https://medium.com/@elliotJL/what-i-learned-spending-a-week-with-100-ai-leaders-3517ac3dc16c?source=rss-79bf89c9d6da------2</link>
            <guid isPermaLink="false">https://medium.com/p/3517ac3dc16c</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <dc:creator><![CDATA[Elliot]]></dc:creator>
            <pubDate>Mon, 16 Jun 2025 15:06:20 GMT</pubDate>
            <atom:updated>2025-06-26T09:59:36.961Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mKhTfA1Ez9JKzLGA.png" /></figure><p>When I walked into the 9th <strong>AI Summit London,</strong> it felt like a tech conference that had finally grown up. Gone were the breathless predictions about AI solving world hunger by Tuesday. Instead, we got something more valuable: honest conversations about the messy, complicated reality of actually deploying AI at scale.</p><p>The overarching theme? <strong>AI has officially graduated from shiny tech demo to proper business infrastructure.</strong></p><p>But like any teenager moving out of their parents’ house, it’s discovering that real life is significantly more complicated than anticipated. The summit crystallised around <strong>six</strong> major challenges that anyone serious about AI needs to wrestle with:</p><ul><li>how to reshape work without breaking people</li><li>how to govern something that moves faster than regulation</li><li>why your data is probably not as ready as you think it is</li><li>how to actually scale beyond proof-of-concept hell</li><li>whether AI can genuinely solve real problems (spoiler: it’s complicated)</li><li>how businesses can navigate this without losing their shirts.</li></ul><p>What stuck with me the most was the shift in conversation. A year ago, this event was dominated by “AI will change everything!” evangelism. This year felt refreshingly pragmatic with equal parts optimism and hard-earned skepticism from people who’ve actually tried to make this stuff work in the real world.</p><h3>The Great Job Displacement Myth (And Why It’s More Nuanced Than You Think)</h3><p>Let’s tackle the elephant in the room first: will AI steal your job? The summit’s answer was a collective eye-roll at the simplistic framing, followed by something far more interesting.</p><p>Nick Handley from Impression kicked this topic off by saying that AI isn’t coming for your job — it’s coming for the boring bits of your job. The tedious, soul-crushing tasks that make you question your life choices during particularly grim Monday mornings. His examples from marketing were telling: AI handling routine Google Ads optimisation so humans can focus on actual strategy, or ChatGPT drafting initial client responses so account managers can spend time on relationship building rather than template filling.</p><p>The consensus seemed to be that marketing is heading towards a consultancy-based model. Think less “I’ll run your campaigns” and more “I’ll help you understand what the hell is happening and why.” It’s a pattern that’s likely to ripple across industries: AI handles execution, humans handle meaning-making.</p><p>On the other hand we had Rob Grigg from Holiday Extras, Natalia Konstantinova from BP, and Kris Kang from JetBrains all raising the same concern: workforce displacement is real, even if it’s not the robot apocalypse scenario we’ve been sold. The responsibility falls on organisations to help their people adapt, which in my opinion is a lovely sentiment that ignores how spectacularly most companies have failed at change management in the past.</p><p>Michael Houlihan from Generation UK and Ireland brought some much-needed urgency to the discussion, pointing out that education systems need to integrate AI training now, not in five years when it’s too late for an entire generation. Peter Cheese from CIPD and Minister Feryal Clark MP talked about a “just transition”, borrowing from the energy sector’s playbook to ensure AI adoption doesn’t leave people behind.</p><p>Amanda Slavin from The Future Frequency made perhaps the most important point: <strong>AI should enhance human capabilities</strong>, not replace them. She advocated for cultivating uniquely human skills like creativity, leadership, and self-awareness. It’s a nice counterpoint to the prevailing narrative that humans need to become more machine-like to stay relevant.</p><p>The marketing panel — Matt Garbutt from Brave Bison, Lily McCann from Choreograph WPP, and Kate Ross from eight&amp;four — provided some reassuring data points fo the creative sector. They haven’t seen significant job losses, and there’s still strong demand for creative talent. The key seems to be finding the sweet spot between automation and human oversight.</p><p>But there’s a subtler concern that emerged from multiple discussions: <strong>cognitive displacement</strong>. The worry isn’t just that AI will replace human jobs, but that over-reliance on AI tools could lead to a gradual erosion of foundational skills and critical thinking capabilities.</p><p>When AI handles increasingly complex tasks, humans risk becoming <strong>passive consumers of algorithmic outputs</strong> rather than active participants in problem-solving. It’s the professional equivalent of GPS navigation making us worse at reading maps, brilliant until your phone dies in a cpacital city and you realise it is every human for themselves.</p><h3>The Governance Nightmare (Or: How to Regulate Something That Changes Daily)</h3><p>If the workforce discussion was nuanced, the governance conversation was straight-up chaotic. And not in a fun way.</p><p>The European AI Act dominated much of the regulatory discussion, with Kai Zenner from the European Parliament and Matthias Holweg from Oxford University painting a picture of well-intentioned complexity that’s giving multinational companies proper headaches. The problem isn’t that the EU is trying to regulate AI but that they’re doing it differently from everyone else. The US has NIST, Canada has ADA, and <strong>everyone’s creating their own flavour of AI governance</strong>.</p><p>Lord Chris Holmes of Richmond MBE and Arba Kokalari from the European Parliament called for principles-based, outcomes-focused regulation and international standards through bodies like ISO. The goal is noble: avoid a situation where companies need a small army of compliance lawyers just to deploy a chatbot across borders.</p><p>The human oversight theme kept cropping up, with Ravinder Singh from the Cabinet Office and Dr. Don Kilburg (independent, formerly government service) emphasising that AI should elevate humanity rather than replace it. Which is exactly the sort of thing you’d expect government officials to say, but the underlying concern is valid: we’re building systems that could gradually erode human agency and decision-making capabilities.</p><p>Rob Grigg, Natalia Konstantinova, and Kris Kang made a crucial point about ownership: AI governance isn’t just the tech department’s problem. It’s a business-wide responsibility that requires alignment with actual business goals, not just technical feasibility. For large enterprises, they suggested a federated approach : <strong>essentially distributed ownership rather than centralised control</strong>.</p><p>Sanja Hukovic and Oli Bage from LSEG got into the practical nitty-gritty with their CDMC framework — a unified data control standard developed with major financial institutions and tech companies. It’s exactly the kind of unglamorous but essential infrastructure work that actually makes compliance possible.</p><p>The IEEE CertifAIEd program, discussed by Patrick Murray and Gerlinde Weger, represents an attempt to create standardised ethical benchmarks for AI systems. They walked through certifying an AI-powered HR application, focusing on fairness, accountability, transparency, and privacy. It’s promising, but I can’t help wondering how these frameworks will keep pace with AI development cycles.</p><h3>Data: The Foundation Everyone Thinks They Have (But Probably Don’t)</h3><p>Here’s where the summit got properly technical, and where most organisations probably realised they’re not as ready as they thought.</p><p>Ronny Kerkhofs from IBM tackled the unstructured data problem head-on. Most organisations are sitting on vast amounts of unstructured data (documents, images, videos ) that traditional systems can’t easily process. IBM’s What’s Next.data platform promises to unlock this with 40% better accuracy than vector-only approaches, but the underlying challenge remains: <strong>most companies don’t know what data they have, where it lives, or whether it’s any good</strong>.</p><p>Steve Walker from Computacenter introduced eight steps for data readiness, with a crucial insight: your data doesn’t need to be perfect, it needs to be “<strong>fit for purpose.</strong>” This is refreshingly practical advice in a field often dominated by perfectionist thinking. He emphasised assessing data quality’s impact on specific AI use cases, setting realistic metrics, and fixing issues proactively rather than reactively.</p><p>The data observability discussion felt particularly relevant. Walker highlighted the need to monitor data freshness, volume, and schema changes — basically, knowing <strong>when your data goes sideways before your AI models start producing nonsense</strong>. Combined with a zero-trust approach to security, especially given the risks of downloading pre-trained models that might contain malicious code, it’s clear that data management is becoming a sophisticated discipline in its own right.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qNnyQ-ziGfWT4NGZ.png" /></figure><p>Chris Yearsley from Bell Integration and Debjani Roy from BT introduced the concept of “<strong>data oceans</strong>”: complex integrations of multiple data repositories that often extend beyond organisational boundaries. Managing these requires extreme standardisation and categorisation, with internal users involved in governance framework setup. It’s a far cry from the simple “data lake” metaphor we’ve been using for years.</p><p>Andy McMahon from Barclays made perhaps the most important point: only organisations with solid data and AI foundations can effectively scale generative AI. It’s a sobering reminder that the flashy demos and proof-of-concepts are the easy bit. The hard work is in the infrastructure that nobody sees.</p><h3>Scaling AI: From Demo to Production Hell</h3><p>Speaking of scaling, the transition from proof-of-concept to production emerged as one of the summit’s major themes. And judging by the conversations, it’s where most AI initiatives <strong>go to die</strong>.</p><p>McMahon detailed the evolution from traditional MLOps to LLMOps, emphasising the need to industrialise AI processes and turn them into “organisational muscle memory.” This involves robust frameworks for experimentation, validation, deployment, and monitoring — all the unglamorous operational work that determines whether your AI initiative actually delivers value or just burns budget.</p><p>Konstantinova highlighted a unique challenge with generative AI: how do you evaluate something that’s inherently uncertain and occasionally hallucinates? Traditional quality metrics don’t apply when your output is <strong>creative rather than deterministic</strong>. She discussed specific techniques like Retrieval Augmented Generation (RAG), which is essentially giving AI systems access to verified info sources to <strong>ground their responses</strong> and using Large Language Models as judges to evaluate other LLMs’ outputs. The EU AI Act’s monitoring mandates for high-risk applications add another layer of complexity, requiring integration of technical metrics with regulatory requirements.</p><p>This evaluation challenge extends to the broader operational framework.</p><ul><li>How do you benchmark quality and safety when the system is designed to be unpredictable?</li><li>How do you maintain consistency across multiple AI agents working in parallel?</li></ul><p>These aren’t theoretical questions. They’re the practical headaches keeping AI operations teams awake at night.</p><p>The rise of agentic AI added another layer of complexity. Sama Bali from NVIDIA and Rick Gonsalvez from HP described these systems as capable of autonomous reasoning, planning, and execution, going beyond fixed rules to actually understand context and adapt behaviour. Use cases include financial fraud detection and manufacturing optimisation, but the implementation challenges are significant.</p><p>It’ll come to no surprise that enterprises are struggling with framework selection, software-hardware compatibility, and achieving consistent results from multiple AI agents. The computational demands are also substantial, requiring close collaboration between hardware and software providers to create streamlined development workflows.</p><p>David Colwell from Tricentis brought some interesting realism to the discussion, pointing out that <strong>“hallucinations” aren’t bugs to be fixed but inherent features of how these systems work</strong>. This needs rigorous testing and continuous monitoring, especially in applications where errors are intolerable. Human oversight is necessary but not sufficient, as humans can become complacent when reviewing AI outputs.</p><h3>The Diversity Problem Nobody Wants to Talk About</h3><p>As I’ve said in previous articles: before trying to build AI for social good, we need to address a glaring issue that runs through everything else: who’s actually building these systems.</p><p>Dr. Julia Stamm from She Shapes AI brought some uncomfortable truths to the table about the gender gap in AI. Female-led AI ventures receive disproportionately low funding, and the industry remains overwhelmingly male-dominated. The reality is, this is an issue that directly impacts the systems we’re building and the problems they’re designed to solve.</p><p>Kathy Nicholson from the Australian Institute for Machine Learning championed women’s visibility and representation in AI, highlighting efforts to increase female participation and foster future AI leaders through early education. When the people building AI systems don’t reflect the diversity of the people using them, <strong>we end up with systems that work brilliantly for some and terribly for other</strong>s.</p><p>The implications ripple through everything from healthcare diagnostics to hiring algorithms. If you want AI that works for everyone, you need everyone involved in building it. Simple as that.</p><h3>AI for Good: Moving Beyond the Marketing BS</h3><p>This is where my skepticism usually kicks into overdrive, but the summit’s approach to AI for social good felt refreshingly grounded in reality rather than aspirational PowerPoint slides.</p><p>The chats around healthcare were particularly interesting. Dr. Charit Bhograj from Tricog Health shared an example from rural Kenya where AI-powered ECG accurately identified heart failure, providing life-saving diagnosis without advanced medical infrastructure. Professor David Lowe from the University of Glasgow outlined a four-phase approach for AI adoption in healthcare, emphasising the need for real-world evidence rather than just technical feasibility.</p><p>Nipun Jain from AstraZeneca discussed the ACATLIS network’s role in scaling AI solutions globally, focusing on early screening and diagnosis in underserved areas. Yan Chow from Automation Anywhere highlighted how AI agents can predict critical moments in patient care and reduce administrative burdens. What I loved about these is that they were all<strong> specific applications addressing real problems</strong>.</p><p>The public services discussion felt similarly practical. Ravinder Singh from the Cabinet Office and Dr. Don Kilburg talked about making government services more efficient and user-friendly. The UK government’s Incubator for Artificial Intelligence, led by Victoria Bew, is building cost-effective, secure AI solutions for specific use cases like analysing public consultation responses and converting planning documents into geospatial data.</p><p>But we had another curveball in the, you guessed it, sustainability discussion. These revealed AI’s central paradox in stark detail. Paolo Taticchi from UCL articulated the tension perfectly: AI has vast potential for environmental solutions by optimising supply chains, forecasting weather patterns, accelerating clean tech development, but its energy consumption is potentially catastrophic. André Pienaar from C5 Capital presented some sobering numbers: <strong>AI could consume 1,000 terawatt hours annually by 2030</strong>.</p><p>The solutions being proposed are fascinating and varied. Pienaar championed Small Modular Reactors (SMRs) as a scalable, reliable, and clean solution for powering hyperscale data centres. Sam Livingstone from Ambienta, Nadia Abouayoub, and Scott Wallace from Digital Realty discussed more immediate strategies: optimising energy use through time-of-day trading (running compute when renewable energy is abundant) and location-based computing (moving workloads to regions with cleaner energy grids).</p><p>Interestingly in a separate conversation, Kris Kang from JetBrains made a compelling case for “<strong>focal models</strong>”: smaller, domain-specific AI systems like Melom that are significantly more energy-efficient than large general-purpose models for routine tasks. It’s a reminder that bigger isn’t always better, and sometimes the most sustainable solution is also the most practical one.</p><p>The education transformation discussions were particularly relevant to current work in ed-tech and passion for digital inclusion. Graeme Lipschitz from Duke Corporate Education highlighted AI’s role in addressing the global educational crisis through personalised learning. Case studies like MindSpark in India, Seeing AI for the visually impaired, and Snaplify as Africa’s largest digital library demonstrate real impact.</p><p>Amanda Slavin’s concept that “engagement is the new intelligence” resonated with me pretty strongly. AI tools that track engagement and provide personalised learning strategies could fundamentally reshape how we think about education and skill development.</p><h3>Business Strategy: Beyond the Hype Cycle</h3><p>The business strategy discussions felt way more mature than previous years’ AI summits. Gone were the grandiose claims about disrupting entire industries overnight. Instead, we got practical advice about navigating AI adoption without losing your shirt.</p><p>Marc Schmitt from LOXXESS AG made a crucial point: AI is now a commodity, accessible to anyone with minimal expertise. <strong>The competitive advantage lies not in access to AI tools, but in having a clear strategy for using them to solve real business problems</strong>. Standardisation is crucial for trustworthy results, which requires a management decision to identify genuine problems rather than technology-first solutions.</p><p>The startup advice was alos surprisingly grounded. Albert Aziz-Corson emphasised understanding fundamental problems before rushing to solutions, building defensible core businesses that are profitable from day one with lean capital models. Anil Lakhani from Crane Venture Partners offered a specific framework for fund allocation that felt refreshingly practical:</p><ul><li>60% to customer value creation</li><li>30% to innovation and product development</li><li>10% to culture and team building.</li></ul><p>He stressed using OKRs (Objectives and Key Results) to demonstrate revenue generation for future investment rounds: essentially, <strong>show don’t tell when it comes to AI value</strong>.</p><p>The “AI data flywheel” concept emerged as a crucial differentiator for startups. The TL;DR is:</p><ul><li>Have good data</li><li>Create continuous feedback loops where user interactions improve the AI model</li><li>Which improves the user experience</li><li>Which attracts more users</li><li>Generating more data.</li></ul><p>It’s compound growth for the AI age, and investors are specifically looking for teams that understand how to build and maintain these flywheels.</p><p>Christian Lau from Dynamo AI discussed the complexity of global expansion in an environment where different regions have vastly different AI regulations. Building trusted partnerships with consulting firms, system integrators, and law firms is crucial for adapting AI models to local requirements.</p><p>Jamie Heaslip from Stripe and DC Cahalane from Sure Valley Ventures emphasised that strategic partnerships with large tech companies are essential for AI startups to scale quickly and access global markets. <strong>AI companies are hitting revenue milestones faster than traditional SaaS</strong>, meaning they need clear monetisation and global market entry strategies from day one.</p><p>The investor relations discussion revealed some interesting trends. Barry Downes from Sure Valley Ventures noted that generative AI’s rapid replicability has reduced its business defensibility, shifting investor interest towards <strong>unique datasets and real customer problem-solving</strong>. The “Art of the Pitch” session highlighted what investors seek: teams with both AI developers and domain experts, plus an “AI data flywheel” for continuous improvement.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*WE4HK3U22y83mvo3.png" /></figure><p>Meanwhile, regional AI ecosystems are developing their own characteristics and strategies. Carlo Negri detailed Italy’s remarkable AI market growth, reaching <strong>1.2 billion euros in 2024</strong>, driven primarily by generative AI adoption and advanced solutions. The growth extends beyond just market size: Italy has seen increases in new AI companies, STEM students, and leadership in high-performance computing. The UK-Italy memorandum of understanding and the AI4I institution for industrial AI represent the kind of international collabs that will become essential as AI development becomes more global.</p><p>The general gist was that the regional diversity in AI development approaches, from the EU’s regulation-heavy stance to Asia’s innovation-centric policies, creates both opportunities and challenges for companies looking to scale internationally.</p><h3>The Reality Check We Needed</h3><p>What made this summit feel different was its willingness to acknowledge complexity rather than promise simple solutions. AI is transformative, but transformation is messy, expensive, and often slower than anyone wants to admit.</p><p>I recently completed a course on AI governance, with a key takeaway being that <strong>we’re trying to regulate systems that evolve faster than our ability to understand them</strong>. It’s a sobering reminder that for all our technical sophistication, we’re still figuring out the basics of how to live with the systems we’ve created.</p><p>The summit left me cautiously optimistic about AI’s potential for genuine social good, but realistic about the challenges ahead. The tech is powerful, we all knew that, but its impact will ultimately depend on how thoughtfully we deploy it and how carefully we address its limitations.</p><p>As I left the conference, I couldn’t shake the feeling that we’re at a crucial inflection point. <strong>The proof-of-concept phase is over</strong>. Now comes the hard work of making AI actually work for everyone, not just the privileged few with access to the latest tools and the budget to implement them properly.</p><p>The question isn’t whether AI will transform society, it already is. The question is whether we’ll guide that transformation wisely, or stumble through it and clean up the mess later. Based on the conversations at the summit, I’m cautiously optimistic that enough people are asking the right questions. Whether we’ll act on the answers remains to be seen.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3517ac3dc16c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Learning Anxiety in the Age of Artificial Everything]]></title>
            <link>https://medium.com/@elliotJL/learning-anxiety-in-the-age-of-artificial-everything-0daaf7c6f579?source=rss-79bf89c9d6da------2</link>
            <guid isPermaLink="false">https://medium.com/p/0daaf7c6f579</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[short-story]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[education]]></category>
            <dc:creator><![CDATA[Elliot]]></dc:creator>
            <pubDate>Fri, 23 May 2025 11:28:43 GMT</pubDate>
            <atom:updated>2025-05-23T11:28:43.985Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*mfyJcjz_g-7-JNsF.jpg" /><figcaption><a href="https://www.writermag.com/">https://www.writermag.com/</a></figcaption></figure><p><em>I’ve been off the pace with writing lately (work’s been a bit mental) so here’s a short one to make sure I’m not shadowbanned by the Medium algorithm gods.</em></p><p>There’s this creeping anxiety I’ve been feeling lately, and I suspect I’m not alone. It’s the dread that comes from living in an era where everything seems to be learning faster than we are.</p><p>My phone learns my habits. My car learns my routes. My bloody washing machine apparently learns my laundry preferences (though it still can’t figure out that I don’t want my whites turned pink). Meanwhile, AI models are hoovering up the entirety of human knowledge in weeks, mastering skills that took us millennia to develop.</p><p>And somewhere in the middle of this learning arms race, there’s us. Still trying to remember where we put our keys.</p><h3>The Great Acceleration</h3><p>We’re living through what feels like a weird inversion of natural selection. For most of human history, learning was a luxury. You learned what you needed to survive, maybe picked up a trade, and that was largely that. Now, not learning feels existential. Every week brings another breakthrough that makes last month’s skills feel a bit quaint.</p><p>The pressure to constantly upskill has become so intense that “lifelong learning” has transformed from aspiration to anxiety disorder. LinkedIn is full of people posting about their latest certification like it’s proof they haven’t been left behind by the march of progress. We’re all running on a treadmill that keeps speeding up, desperately trying not to fall off the back.</p><p>But a thought I had recently was what happens when the rug gets pulled out from under this whole system?</p><h3>The Fragility of Always-On Everything</h3><p>We’ve built our learning infrastructure on remarkably shaky foundations. YouTube University. Google as external memory. ChatGPT as a thinking partner. Stack Overflow as a debugging companion. The cloud as our filing cabinet.</p><p>It’s brilliant until it isn’t.</p><p>Imagine waking up tomorrow to a world where the internet just… stopped. Not permanently, but long enough to matter. A massive solar flare, a coordinated cyberattack, or a new Copilot Pro plan which you need to sell your house to access.</p><p>Suddenly, all those people who learned to code through vibes find themselves staring at blank screens with no idea how to actually think through problems. The generation that learned languages through apps discovers they can’t have a basic conversation without their phones. The managers who rely on AI to draft their legal emails realise they’ve forgotten how to actually sign them off (bestest of wishes? .</p><h3>The Outsourcing of Thinking</h3><p>The more I watch AI development, the more I wonder if we’re accidentally training ourselves out of the very skills that make us human. When was the last time you memorised a phone number? Calculated a tip without your phone? Sat with a difficult problem long enough to actually solve it rather than immediately reaching for assistance?</p><p>I’m not being a Luddite here. I use AI tools daily and they’re genuinely brilliant. But there’s something unsettling about how quickly we’ve adapted to having silicon thinking partners for everything. We’re becoming cognitively dependent on systems we don’t understand, controlled by companies whose interests might not align with ours.</p><p>The kids growing up now are digital natives in ways that make millennials look like cave dwellers. But what happens to a generation that never learned to be bored, never developed the patience for slow thinking, never built the mental muscles for wrestling with problems independently?</p><h3>The Paradox of Accelerated Learning</h3><p>Here’s the really twisted bit: the faster AI learns, the more pressure we feel to learn faster too. But humans aren’t designed for that kind of cognitive sprint. <strong>We learn through repetition, failure, reflection, and time</strong>. The messy, inefficient process of human learning isn’t a bug but a feature.</p><p>When we try to compress natural learning into AI-like speeds, we often end up with the educational equivalent of fast fashion: quick, cheap, and falling apart the moment you actually need it to work.</p><p>Real learning, the kind that sticks, that becomes intuition, that helps you navigate genuinely novel situations, takes time to marinate. It requires the kind of deep practice that can’t be optimised away.</p><h3>Maybe We’re Asking the Wrong Question</h3><p>Instead of “How do we learn as fast as machines?” maybe we should be asking “What can humans do that machines fundamentally cannot?”</p><p>Machines excel at pattern recognition and rapid information processing. Humans excel at meaning-making, emotional intelligence, creative problem-solving, and adapting to genuinely unprecedented situations. The trick isn’t competing with AI on its terms but doubling down on our distinctly human capabilities.</p><p>Perhaps the real skill we need to develop isn’t learning faster, but learning better.</p><p>Not consuming more information, but developing better judgment about what’s worth knowing.</p><p>Not optimising ourselves into human-shaped algorithms, but becoming more distinctly, irreplaceably human.</p><h3>Choose Your Dependencies Wisely</h3><p>I’m not suggesting we abandon technology and return to some pastoral fantasy. But maybe we need to be more intentional about what we outsource and what we keep in-house.</p><p>Learn to think without Google occasionally. Practice solving problems without immediately reaching for AI assistance. Develop the patience for slow learning in a fast world. Build redundancy into your cognitive toolkit.</p><p>Because in the end, the real digital divide might not be between those who have access to AI and those who don’t. It might be between those who can still think independently when the machines go quiet, and those who’ve forgotten how.</p><p>And that’s a gap worth bridging before we need it most.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0daaf7c6f579" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Walled Garden: How Zuckerberg Tried To Build His Own Internet]]></title>
            <link>https://medium.com/@elliotJL/the-walled-garden-how-zuckerberg-tried-to-build-his-own-internet-3de0aa0404c2?source=rss-79bf89c9d6da------2</link>
            <guid isPermaLink="false">https://medium.com/p/3de0aa0404c2</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[politics]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[internet]]></category>
            <dc:creator><![CDATA[Elliot]]></dc:creator>
            <pubDate>Tue, 22 Apr 2025 14:57:18 GMT</pubDate>
            <atom:updated>2025-04-24T13:29:10.024Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/900/1*wmcfeLSHk99G51MJe-QOuQ.jpeg" /><figcaption>Source: <a href="https://wallpapersok.com/">https://wallpapersok.com/</a></figcaption></figure><p>I’ve been thinking about connectivity over the Easter weekend. Not the kind that gets measured in bars on your phone screen, but the messy, complicated kind that links us humans together.</p><p>Some of you will remember when getting online meant listening to that screeching dial-up sound. A time when the internet felt like magic. Now it’s essential infrastructure. To the point you or I can type words onto a blank page and publish it for others to see. All for fun.</p><p>If you’re reading this it’s unlikely you’ve ever considered your internet access to be a privilege. I’d be lying if I said it’s something that crosses my own mind on the regular. Yet for billions of people, this free infrastructure simply doesn’t exist or remains tantalisingly out of reach.</p><p>The numbers paint a stark picture: even as of last year, <strong>approximately 2.6 billion people remain offline worldwide</strong>. That’s roughly one-third of humanity with no access to the digital tools and resources many of us take for granted daily.</p><p>The gaps follow predictable patterns of global inequality: while North America and Europe have internet penetration rates above 90%, large parts of Africa hover around 40%. And even these figures mask deeper divides — between urban and rural populations, between men and women, and across economic classes.</p><p>When Facebook launched <strong>Internet.org</strong> in 2013, it seemed like an ambitious but straightforward mission: bring internet access to the billions without it. Who could argue with <em>that</em>? More people online meant more access to education, healthcare information, job opportunities, and yes — social media.</p><p>This wasn’t a solo mission. Facebook partnered with tech heavyweights including Ericsson, MediaTek, Opera, Samsung, Nokia, and Qualcomm in a coordinated attempt to break down the barriers to connectivity. Zuckerberg outlined a three-pillar strategy:</p><ul><li>making access affordable</li><li>using data more efficiently</li><li>helping businesses drive access through new models</li></ul><p>The approach evolved into “<strong>Free Basics</strong>” (rebranded in 2015), a platform offering zero-rated access to selected websites via mobile operators across Africa, Asia, and Latin America. By July 2015, Facebook claimed the service was available to over a billion people across 17 countries. The implementation was rapid, with launches in Zambia (July 2014), Tanzania (October 2014), Kenya (November 2014), Colombia (January 2015), Ghana (January 2015), India (February 2015), and dozens more countries following in quick succession.</p><p>But just like all grand experiments, they rarely unfold as expected.</p><p>What seemed straightforward on a whiteboard in California became considerably more complex when implemented across diverse cultures, economies, and regulatory landscapes. The technical architecture wasn’t the only hurdle; the model itself raised fundamental questions about what <strong>meaningful connectivity</strong> actually means.</p><p>The core tension always centred around a fundamental question: <strong>is limited internet access better than none at all?</strong></p><p>In corner number #1, supporters argued that offering free access to even a subset of the internet would provide immediate benefits to unconnected populations. The idea here is that it would eventually <strong>serve as a gateway to wider adoption</strong>. Facebook presented data claiming the service drove mobile data adoption over 50% faster in markets where it launched, with more than half of Free Basics users starting to pay for data within 30 days.</p><p>Critics worried about creating a <strong>two-tier internet</strong> where the world’s most vulnerable populations received a limited, curated version of the digital world. The platform typically included services like Facebook itself, Wikipedia, AccuWeather, news sites, and some local services related to health, employment, and education — but excluded the vast majority of the open internet. Digital rights organisations argued this violated net neutrality, the principle that <strong>all internet traffic should be treated equally</strong>.</p><p>The limitations weren’t just philosophical. A 2017 study by Global Voices across six countries (Colombia, Ghana, Kenya, Mexico, Pakistan, and Philippines) found significant practical issues: many services offered <strong>weren’t relevant to local needs</strong>, <strong>language support</strong> was often inadequate, and <strong>technical limitations</strong> created a poor user experience. When given options, many users preferred paying for access to the full internet rather than using the limited free service.</p><blockquote>“Internet.org entrenches the digital divide between the haves and have-nots, by delivering a crap version of the internet to two-thirds of the world” — <em>Sarah Wynn-Williams</em></blockquote><p>The debate played out pretty dramatically in India, where digital rights activists, tech entrepreneurs, and citizens mounted a massive campaign under the “<strong>Save The Internet</strong>” banner. They argued the approach violated net neutrality principles and could harm local innovation. Facebook responded with an extensive advertising campaign and prompted users to send support messages to the regulator. The Telecom Regulatory Authority of India (TRAI) ultimately agreed with critics, issuing regulations in February 2016 that effectively banned Free Basics and similar zero-rating services.</p><p>Similar tensions popped up elsewhere, with a variety of outcomes. Egypt suspended the service in December 2015, reportedly after Facebook refused government surveillance demands. Chile’s strong net neutrality laws effectively prevented its launch. Meanwhile, the service continued operating and even expanding in dozens of other countries, particularly across Africa, where it reached nearly 30 nations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3yCy3CnJsSZs-kAY.png" /><figcaption>Source: <a href="https://www.behance.net/">https://www.behance.net/</a></figcaption></figure><p>It’s tempting to frame tech initiatives in simple hero-or-villain narratives. The reality is messier. Connectivity efforts involve complex trade-offs between immediate access, quality of experience, long-term market health, and questions of who controls and shapes digital spaces.</p><p>Corporate connectivity initiatives bring resources, technical expertise, and scale that can make a genuine difference. Free Basics did provide some form of connectivity to previously unconnected populations — Facebook claimed it had connected almost <strong>100 million users by 2018</strong>. The platform evolved over time, opening to third-party developers in 2015 and adding security features like HTTPS support in response to criticism.</p><p>But even well-intentioned initiatives can have troubling implications. Researchers raised concerns about <strong>digital colonialism</strong> — the recreation of historical power imbalances in digital spaces. The platform’s architecture facilitated user data collection regardless of whether someone was logged into Facebook, raising questions about the relationship between connectivity and surveillance. In a pretty major revelation, users in some countries were reportedly <strong>charged hidden fees amounting to nearly $100 million in 2021 alone</strong>, contradicting the “free” premise entirely.</p><p>Facebook itself recognised some of these tensions. By 2020, they began trialing “<strong>Discover,</strong>” an evolved approach offering a daily balance of free data usable on any website, not just pre-approved ones. This directly addressed the walled garden critique while still addressing affordability barriers.</p><p>At the same time, the company steadily shifted resources toward less controversial infrastructure investments, like terrestrial fibre networks in Nigeria and Uganda, edge infrastructure across 44 Sub-Saharan African countries, and the massive 2Africa subsea cable project. These investments aimed to improve the overall internet ecosystem rather than creating preferential access models.</p><p>If there’s one clear lesson from these connectivity experiments, it’s that there’s <strong>no one-size-fits-all solution to bridging digital divides</strong>. Different approaches work in different contexts.</p><p>Community networks have proven super effective in some regions, where local people build and maintain their own infrastructure. Zenzeleni Networks in South Africa’s Eastern Cape, one of the country’s poorest rural areas, evolved from a university collaboration into a licensed cooperative ISP offering data and voice services at a fraction of commercial costs (reportedly 5–20% of standard rates).</p><p>The model emphasises community ownership — “<strong>If it belongs to them, they value it,</strong>” as founder Masibulele Jay Siya puts it. Similar initiatives like Rhizomatica in Mexico have enabled indigenous communities to build their own cellular networks where commercial operators declined to serve.</p><p>Municipal broadband initiatives have transformed connectivity in cities like Chattanooga, Tennessee, where the Electric Power Board built a citywide fibre network. An independent economic study found staggering community benefits totalling<strong> $2.69 billion over ten years</strong> — including <strong>9,500+ jobs created </strong>or retained, business investment, and their HCS EdConnect program providing free high-speed fibre to all economically disadvantaged K-12 students.</p><p>Public-private partnerships, public Wi-Fi initiatives, and mesh networks offer additional models. Johannesburg’s public Wi-Fi project showed significant positive impacts on users’ lives, including improved communication, savings, and income generation opportunities. Solar-powered mesh networks in low-income neighbourhoods have achieved costs 40–50% lower than fibre while maintaining reliability.</p><p>Some of the most promising models focus on <strong>multiple barriers simultaneously</strong> — not just access, but affordability, skills, and relevance. When people see genuine value in getting online, and by that I mean when <strong>content speaks their language, addresses their needs, and respects their context, </strong>adoption follows naturally.</p><p>Meaningfully addressing the <a href="https://medium.com/@elliotJL/when-tech-leaves-people-behind-3b13275a48d7">ASA Gremlin (Access, Skills, Affordability)</a> means looking at holistic approaches that consider local contexts rather than <strong>one-dimensional technical solutions</strong>.</p><p>Everyone talks about the “global village,” but it’s not much of a village when most of the people can’t join the conversation.</p><p>Look beyond the tech hype cycle and you’ll find a sobering reality: the digital revolution that’s reshaping almost every aspect of modern life remains inaccessible to <strong>nearly a third of humanity</strong>. And even among those technically “connected,” millions experience a digital world limited by slow speeds, intermittent access, or crazy costs.</p><p><em>When banking goes digital, what happens to the unbanked?</em> <em>When job applications move online, who gets left in the unemployment line?</em> <em>When healthcare embraces telemedicine, who’s excluded from treatment?</em></p><p>The answers follow predictable patterns of existing disadvantage — the rural poor, the elderly, people with disabilities, linguistic minorities, and women in many regions find themselves disproportionately on the wrong side of the digital threshold.</p><p>But here’s where the story gets interesting.</p><p>The very failures of early connectivity initiatives have sparked more <strong>nuanced, locally-grounded approaches</strong>. Indigenous communities from Oaxaca to South Africa aren’t waiting for Silicon Valley solutions — they’re building their own networks, governed by their own priorities. Cities and towns are treating broadband as essential infrastructure, just like roads or water systems. Policymakers are exploring regulatory frameworks that balance innovation with equity.</p><p>What’s emerging is a <strong>richer understanding that meaningful connectivity isn’t just about technical access</strong>. It requires addressing overlapping barriers simultaneously: physical infrastructure, yes, but also affordability, skills development, relevant content, and supportive policy. It means recognising that a one-size-fits-all approach fails to account for diverse community needs and contexts.</p><p>The most promising initiatives share common elements: they involve affected communities as partners rather than passive recipients, they build local capacity and ownership, they prioritise open access over walled gardens, and they measure success by meaningful usage rather than simplistic connectivity metrics.</p><p>As we look ahead, I’ve come across several concrete approaches that I think need wider adoption:</p><p><strong>1. Overhaul regulatory frameworks:</strong> Craft policies that prioritise underserved populations, especially in the AI era. Update universal service obligations to include broadband, reserve spectrum for community networks (following Mexico’s example), and implement net neutrality protections that prevent discrimination without slowing innovation.</p><p><strong>2. Democratise infrastructure funding:</strong> Create accessible funding pathways for community-led initiatives instead of channeling funds exclusively through traditional telecoms. Redesign Universal Service Funds with participatory governance and simplified applications that grassroots organisations can navigate.</p><p><strong>3. Build digital rights into development:</strong> Integrate data protection, privacy, and algorithmic accountability into every connectivity initiative. The goal should extend beyond technical access to ensure meaningful digital involvement.</p><p><strong>4. Adopt hybrid ownership models:</strong> Combine community ownership with public infrastructure. Models like open access middle-mile networks (government-built core infrastructure used by multiple service providers) balance local control with economic sustainability.</p><p><strong>5. Invest beyond hardware:</strong> Allocate at least 30% of connectivity budgets to digital literacy, local content creation, and skills development. Evidence shows infrastructure alone doesn’t drive adoption without investment in human capacity.</p><p><strong>6. Decentralise decision-making:</strong> Establish formal mechanisms for affected communities to share governance of connectivity programs. Move beyond superficial consultation to shared power over design, implementation, and evaluation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Ly-4xIdQnW6ooH_l.jpg" /><figcaption>Source: <a href="https://www.itcilo.org/">https://www.itcilo.org/</a></figcaption></figure><p>The convo has evolved beyond binary questions of connected versus unconnected. We’re asking richer questions: <em>Connected to what? Controlled by whom? Designed for whose benefit?</em> These questions don’t have single answers, but they’re generating more thoughtful approaches that <strong>recognise connectivity as a means rather than an end</strong> — a tool that communities should be able to shape rather than simply consume.</p><p>The thing about the internet that rightly got everyone excited, it’s <em>revolutionary promise</em> some might say, was a <strong>democratisation of knowledge, creativity, and opportunity</strong>. Achieving that promise means making sure that every person can meaningfully participate in our digital future, not as passive recipients of tech designed elsewhere, but as active creators and shapers of shared digital spaces that reflect humanity’s full diversity.</p><p>The lessons from Internet.org and its evolution are urgent guideposts for how we approach the next phase of global connectivity. As new technologies emerge and digital transformation accelerates, we face a critical choice:</p><p>Will we promote systems that <strong>concentrate power and deepen divides?<br></strong>or <br>Will we build digital ecosystems that <strong>distribute both access and agency</strong>?</p><p>The answer will shape not just who gets online, but what kind of world we create together once we’re there.</p><p>Because an internet that doesn’t work for everyone won’t truly work for anyone. The challenges before us demand solutions as diverse as humanity itself — built from the ground up rather than imposed from above and accountable to communities rather than corporations.</p><p>Anything less would betray the potential of the most powerful communication platform humanity has ever created.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*aLIu7VlR6K39Gcr8.png" /></figure><ol><li>Internet.org Initiative. (2013). “Technology Leaders Launch Partnership to Make Internet Access Available to All.” Press Release.</li></ol><p>2. Zuckerberg, M. (2013). “Is Connectivity a Human Right?” White Paper.</p><p>3. Telecom Regulatory Authority of India. (2016). “Prohibition of Discriminatory Tariffs for Data Services Regulations, 2016.”</p><p>4. Global Voices. (2017). “Free Basics in Real Life.” User Experience Research Across Six Countries.</p><p>5. Joseph Rowntree Foundation. (2023). “Digital Exclusion and Poverty in the UK.” Research Report.</p><p>6. Age UK. (2023). “Digital Inclusion Among Older Adults.” Annual Report.</p><p>7. Ability Net. (2023). “Digital Accessibility and Disability.” Research Findings.</p><p>8. Education Business. (2023). “Digital Poverty in Education.” Impact Assessment.</p><p>9. Facebook. (2015). “One Year In: Internet.org Free Basic Services.” Progress Report.</p><p>10. Analysys Mason. (2023). “The Impact of Facebook’s Connectivity Initiatives in Sub-Saharan Africa.” Economic Analysis.</p><p>11. Chatham House Study. (2024). “The Economic Impact of Digital Divides in Developing Economies.”</p><p>12. Zenzeleni Networks. (2023). “Community-Owned Connectivity in South Africa.” Impact Report.</p><p>13. Rhizomatica. (2022). “Indigenous Community Networks: Outcomes and Challenges.” Project Documentation.</p><p>14. Chattanooga EPB. (2020). “Community Fiber Networks: A Decade of Impact.” Economic Study.</p><p>15. Creative Commons Research. (2024). “Alternative Models for Digital Inclusion.” Comparative Analysis.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3de0aa0404c2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Finding Nemo Taught Me About Tech for Good]]></title>
            <link>https://medium.com/@elliotJL/what-finding-nemo-taught-me-about-tech-for-good-7903fb397b6a?source=rss-79bf89c9d6da------2</link>
            <guid isPermaLink="false">https://medium.com/p/7903fb397b6a</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[tech-for-good]]></category>
            <category><![CDATA[social-impact]]></category>
            <category><![CDATA[film]]></category>
            <category><![CDATA[culture]]></category>
            <dc:creator><![CDATA[Elliot]]></dc:creator>
            <pubDate>Fri, 11 Apr 2025 11:24:36 GMT</pubDate>
            <atom:updated>2025-04-11T11:25:11.318Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*hp3pH0iHuD51uHdk" /><figcaption>Source: <a href="https://characterdesignreferences.com/">https://characterdesignreferences.com/</a></figcaption></figure><p>You know that bit in <em>Finding Nemo</em> where the Tank Gang hatch their elaborate escape plan?</p><p>They’ve thought of everything: the filtration system, the bubbles, the timing. They even account for the one flaw in the plan by having Gill launch himself down the dentist’s sink first to test if fish can survive in the pipes. It’s gold on paper. Every step rigorously tested, every detail considered.</p><p>And then they end up in plastic bags floating in Sydney Harbour with no clue what to do next.</p><p><strong><em>“Now what?”</em></strong> asks Bloat the pufferfish, staring out at the vast ocean.</p><p>That’s how I see Tech for Good in 2025. Precisely engineered solutions trapped in plastic bags, staring at the actual problems they were meant to solve with absolutely no idea how to reach them.</p><p>When <em>Finding Nemo</em> bubbled onto screens in 2003, I was too busy watching it on repeat to realise I was witnessing a masterclass in emotional design. The film won hearts and fundamentally altered how millions of people perceived the ocean. By giving fish human traits and relatable emotions, Pixar accidentally built <strong>one of the most effective pieces of ocean advocacy ever created</strong>.</p><p>It made us laugh, cry, and somehow care deeply about the Great Barrier Reef without delivering a single PowerPoint slide on marine conservation.</p><p>Twenty-two years later, as I navigate the murky waters of the tech-for-good ecosystem, I keep circling back to this question: why can’t our supposedly world-changing tech achieve what a little animated clownfish did so effortlessly?</p><p>The genius of <em>Finding Nemo</em> wasn’t in making us aware of marine conservation issues. Its power was <strong>making us feel something about the ocean</strong>. When Marlin desperately searched for his son, we weren’t thinking about coral reef statistics or ocean acidification metrics — we were emotionally invested in a father’s journey through a suddenly relatable underwater world.</p><p>This hits exactly what design researcher Don Norman pinpointed in his work on emotional design. Good design operates on three levels: <strong>gut reaction, functionality, and personal meaning</strong> [1]. Finding Nemo nailed all three. It was visually stunning, told a coherent story, and left us with meaningful connections to characters who happened to be fish.</p><p>Meanwhile, look at <a href="https://www.ecosia.org/">Ecosia</a>, the tree-planting search engine. Technically brilliant. Genuinely plants trees. Absolutely soulless user experience. Or those carbon footprint calculators that make you feel like shit for existing and offer zero emotional payoff. We’ve built products that can measure dopamine spikes, but barely any that make people cry like a Pixar movie.</p><p>The term “Tech for Good” has spread across industry, policy, and academic circles faster than a sped-up TikTok audio of your favourite song. It sounds like a shift away from pure profit-grabbing toward something nobler [2]. But scratch beneath the surface, and things get murky fast.</p><p><a href="https://techforgoodinstitute.org/">The Tech for Good Institute</a> themselves admit it’s “<strong>catchy but could mean many things</strong>” [2]. In other words, it means nothing. It’s the perfect cover for “impact washing” — slapping virtuous labels on tech that delivers sod-all real change.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*FlbOAASqTPUmC_xH" /><figcaption>Source: <a href="https://characterdesignreferences.com/">https://characterdesignreferences.com/</a></figcaption></figure><p>We’ve seen this movie before with “greenwashing.” Volkswagen faked emissions data while running heartwarming eco-ads. Shell talks up renewable energy while pouring billions into fossil fuels. H&amp;M pretends old polyester shirts become new polyester shirts through fashion reincarnation [6].</p><p>It’s Meta’s “Time Well Spent” campaign while their product team optimises for rage-scrolling. It’s that mindfulness app tracking your engagement metrics to Slack channels while encouraging you to “<strong>be present.</strong>” It’s Silicon Valley word salad that lets companies feel virtuous while changing absolutely nothing about their core business models.</p><p>When <em>Finding Nemo</em> unexpectedly drove interest in clownfish pets, researchers feared a “Nemo effect” would devastate wild populations. Later studies showed something more interesting — there wasn’t a massive spike in clownfish purchases, but there was a huge jump in people seeking information about marine life [7]. People wanted to learn, not just consume. That’s the better outcome.</p><p>Meanwhile, most tech-for-good initiatives measure success through downloads, engagement metrics, or vague “lives impacted” tallies that tell us almost nothing about real change. Sure, the numbers look impressive in investor decks, but do they mean anything substantial? The gap between lofty claims and verifiable results keeps growing, and people aren’t stupid. <strong>They smell the bullshit</strong> [2].</p><p>The “Nemo effect” reveals something crucial: enchantment comes before action. Before we seek information, before we change behaviour, we first need to give a damn. Pixar understood this instinctively by <strong>making us care about the reef before asking us to care about the reef</strong>.</p><p>Research into a newer era of “AI for good” shows a similar pattern. These technologies function as “enchantment tools,” dazzling audiences into believing in their transformative power regardless of whether they actually work [9]. Imagine opening a climate app and feeling the way you did when Dory remembered her parents. <strong>We’d all be driving less tomorrow</strong>.</p><p>But what happens when we trust a spell that doesn’t actually work?</p><p>The core problem with most “Tech for Good” isn’t malicious intent. It’s what critic Evgeny Morozov calls “solutionism” —<strong> the naive belief that complex social problems have neat, tech-shaped solutions</strong> [10]. It’s Google’s balloon internet project (<a href="https://x.company/projects/loon/">Loon</a>) promising connectivity to remote areas but delivering mostly investor presentations. It’s <a href="https://en.wikipedia.org/wiki/One_Laptop_per_Child">One Laptop Per Child</a> failing while PowerPoint slides claimed education revolution. It’s thinking an app can solve homelessness when the problem is housing policy.</p><p>Think about how you find that perfect tech-for-good app. You search. Google Play ranks results based on downloads, reviews, and mysterious ‘prominence’ algorithms. The discovery of “good” technology is itself controlled by systems designed to maximise engagement, not impact. Genuinely useful apps get buried under flashier alternatives with better SEO budgets. The machine is carving up your good intentions before you’ve even tapped “install.”</p><p>What if we built tech with the emotional resonance of <em>Finding Nemo</em>? What if instead of starting with the technical solution, we began with the emotional journey?</p><p>This isn’t about slapping inspirational quotes on your landing page or adding cute mascots to your error messages. It requires fundamentally rethinking how we approach tech-for-good:</p><h3>1. Care about values, not just metrics</h3><p>The <a href="https://www.humanetech.com/">Centre for Humane Technology</a> advocates prioritising human well-being and core values over metrics designed solely to maximise user time and attention [11]. Pixar didn’t optimise for merchandise opportunities or sequel potential; they obsessed over character, emotion, and meaningful narrative.</p><p>Think about when the last time an app actually made you feel deeply understood rather than deeply analysed.</p><h3>2. Bake impact into your business model</h3><p>There’s a massive difference between Toms Shoes’ “buy one, give one” model and Google’s “we’ll donate 0.1% of our ad revenue to something vague.” Real tech-for-good ventures embed their mission into how they make money [2].</p><p>Ecosia plants trees with search revenue. Imperfect Foods sells “ugly” produce that would’ve been wasted. Their profit literally depends on their impact. When your financial incentives fight your mission, guess which one usually wins? (Hint: <a href="https://www.socialmediatoday.com/news/meta-abandons-responsible-innovation-team-as-metaverse-development-contin/631484/">Meta’s integrity team got the axe before their metaverse team did</a>.)</p><p>At <a href="https://www.zerogravity.co.uk/">Zero Gravity</a>, a platform I’m currently building, we connect low-opportunity students to real career pathways. There’s no escaping the fact that <strong>our impact is baked into the product</strong>. If we don’t help people land opportunities, we don’t grow. That’s how it should be. <strong>Tech-for-good without accountability is just PR</strong>.</p><h3>3. Make people feel seen</h3><p>Research shows that users fundamentally want validation of their emotional experiences [14]. People seek confirmation that their feelings, experiences, and beliefs are valid, meaningful, and appropriate.</p><p>Think about how Finding Nemo validates parental anxiety without judgment — yeah I’m going there. Marlin isn’t wrong to be scared because his concerns are legitimate. His journey teaches balance, not elimination of fear.</p><p>Meanwhile, most climate apps treat you like <strong>a morally bankrupt monster</strong> for taking a flight, with zero recognition of the complex emotions involved.</p><p>The apps that last are the ones that make us feel seen, not shamed. Have you noticed Duolingo doesn’t call you lazy for missing Spanish practice? It makes you feel like you’re letting down an owl who believes in you. That’s emotional intelligence.</p><p>Well, it is until Duo starts hunting down your family.</p><h3>4. Hit the identity level</h3><p>As Norman points out, the “reflective level” is where tech becomes part of who we are and how we see ourselves [1]. It asks: “Does this align with my identity? Can I tell a story about why I use this?”</p><p>Spotify’s annual “Wrapped” isn’t just a visualisation of your listening data but a personalised identity statement. “This is who you are, expressed through music.” <strong>People share it because it tells their story</strong>.</p><p>The problem is that most tech-for-good creates negative identity markers: “You’re a carbon offender.” “You’re a water waster.” “You’re not doing enough.” Nobody wants those identities.</p><p>Find a story that makes people heroes, not villains. Bruce the shark wasn’t bad for being a shark; he was brave for fighting his nature. “Fish are friends, not food” works because it’s an identity upgrade, not a scolding.</p><p>Here’s a practical framework for spotting genuine tech-for-good initiatives I’m going to call the ever so catchy <strong>The Do-Good-or-Just-Sound-Good Test</strong>:</p><blockquote><strong><em>🌊 Does it make people actually give a damn?</em></strong> Not just understand intellectually, but feel something about the issue.</blockquote><blockquote><strong><em>🤝 Is impact built into how they make money?</em></strong> If making more profit means creating more positive change (not just funding a CSR program), you’re looking at the real deal. The business model should amplify the mission, not fight against it.</blockquote><blockquote><strong><em>📖 Do users tell their friends about it without sounding like an ad?</em></strong> “I love how <a href="https://www.flashforest.com/">Flash Forest</a> helps people” vs. “My banking app has an ESG score visualisation that made me realise how problematic my investment in… hey, where are you going?”</blockquote><blockquote><strong><em>🧭 Would it survive if the hype died?</em></strong><em> </em>Finding Nemo stayed relevant long after the merchandise disappeared from shelves. Will you?</blockquote><blockquote><strong><em>👥 Does it reflect real human diversity?</em></strong><em> </em>Finding Nemo included characters with disabilities (Nemo’s “lucky fin”) and neurodiversity (Dory’s memory issues) without making them caricatures. Does your tech genuinely serve diverse needs?</blockquote><p>You’re not building a plastic bag in Sydney Harbour. Or at least — you shouldn’t be.</p><p>The tech backlash we’re witnessing isn’t ultimately about tech itself. It’s about the gap between technology’s implied promise and its lived reality. When itpresents itself as being “for good” but fails to demonstrate authentic, verifiable positive impact, <strong>distrust naturally follows</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*mF-EmYJctZHzjBbS" /><figcaption>Source: <a href="https://characterdesignreferences.com/">https://characterdesignreferences.com/</a></figcaption></figure><p>I’ll admit it…I’ve been suckered. I once downloaded an app that claimed it would plant trees every time I logged on. Three months of logging in later, I discovered their “tree planting algorithm” was still “being finalised” and exactly zero saplings had been planted. The app had raised millions in venture funding. Their metrics were perfect. Their impact was a mirage.</p><p><em>Finding Nemo</em> didn’t save the ocean, but it moved people in a way that decades of conservation messaging couldn’t. That emotional movement matters deeply. It created a generation that saw the ocean differently — not as an abstract concept but as a world filled with personalities and stories worth caring about.</p><p>Tech for good can’t rely on functionality and “doing good” alone. We’re building tools for humans, and humans are fundamentally emotional creatures. We’re narrative-weavers, meaning-makers, emotional validators seeking confirmation that <strong>our feelings and experiences matter</strong>.</p><p>The most impactful technology won’t just solve problems — it will make us care about them in the first place. It will validate our emotional connection to the issues and communities it serves. It will tap into our need for meaning, identity, and emotional resonance.</p><p>In a world of <a href="https://medium.com/@elliotJL/the-algorithmic-gentrification-of-your-sunday-roast-cd916c79a668">algorithmic gentrification</a> and impact washing, the real revolution isn’t building better metrics. It’s building tech that makes people <strong>feel something authentic</strong> that ripples through them long after the screen goes dark.</p><p>Like that moment when Nemo finally found his way back home, and we all pretended not to cry.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/0*LxJ75GLeBT4biHrH.png" /></figure><h3>References</h3><ol><li>Norman, D. A. (2004). Emotional Design: Why we love (or hate) everyday things. Basic Books.</li><li>Tech for Good Institute. (2023). Tech for good: What it means and how we can deliver on it. World Economic Forum. <a href="https://www.weforum.org/stories/2023/03/tech-for-good-what-does-it-mean-and-how-can-we-deliver-on-it/">https://www.weforum.org/stories/2023/03/tech-for-good-what-does-it-mean-and-how-can-we-deliver-on-it/</a></li><li>Powell, A. et al. (2022). What’s not good with ‘Tech For Good’? Critical Educational Technology. <a href="https://criticaledtech.com/2022/09/16/whats-not-good-with-tech-for-good-notes-on-radhakrishnan-powell-et-al/">https://criticaledtech.com/2022/09/16/whats-not-good-with-tech-for-good-notes-on-radhakrishnan-powell-et-al/</a></li><li>Good Tech Lab. (2020). Defining Impact Tech. Extract from Chapter 2: The Rise of Impact Tech. <a href="https://medium.com/good-tech-lab/defining-impact-tech-819922112408">https://medium.com/good-tech-lab/defining-impact-tech-819922112408</a></li><li>Bethnal Green Ventures. (2021). BGV Impact and Learning Report. <a href="https://uploads-ssl.webflow.com/625a6a50e620b457b31b660f/627d5d9b5f85fa6df5c1f395_BGV%20Impact%20and%20Learning%20Report%202021.pdf">https://uploads-ssl.webflow.com/625a6a50e620b457b31b660f/627d5d9b5f85fa6df5c1f395_BGV%20Impact%20and%20Learning%20Report%202021.pdf</a></li><li>Provenance. (2025). Retailer Greenwashing Examples 2025: Major Cases and Key Lessons for Retailers. <a href="https://www.provenance.org/news-insights/greenwashing-examples-2025-major-cases-and-key-lessons-for-retailers">https://www.provenance.org/news-insights/greenwashing-examples-2025-major-cases-and-key-lessons-for-retailers</a></li><li>Buckley, K. A., et al. (2017). Finding Nemo’s genes: A chromosome-scale reference assembly of the genome of the orange clownfish Amphiprion percula. Molecular Ecology Resources, 17(6).</li><li>Gell, A. (1992). The Technology of Enchantment and the Enchantment of Technology. In J. Coote &amp; A. Shelton (Eds.), Anthropology, Art and Aesthetics (pp. 40–63). Oxford University Press.</li><li>Nonhuman humanitarianism: when ‘AI for good’ can be harmful. (2021). Information, Communication &amp; Society. <a href="https://www.tandfonline.com/doi/full/10.1080/1369118X.2021.1909100">https://www.tandfonline.com/doi/full/10.1080/1369118X.2021.1909100</a></li><li>Morozov, E. (2013). To Save Everything, Click Here: The Folly of Technological Solutionism. PublicAffairs.</li><li>Centre for Humane Technology. (2023). Ethics &amp; Technology: The New Frontier of Responsible Board Governance. <a href="https://www.iodglobal.com/blog/details/ethics-technology-the-new-frontier-of-responsible-board-governance">https://www.iodglobal.com/blog/details/ethics-technology-the-new-frontier-of-responsible-board-governance</a></li><li>Nesta Impact Investments. (2019). Impact Investments Annual Report. <a href="https://nestainvestments.org.uk/wp-content/uploads/2019/07/impact_investments_annual_report_v4.pdf">https://nestainvestments.org.uk/wp-content/uploads/2019/07/impact_investments_annual_report_v4.pdf</a></li><li>The Impact of Emotional Design on User Engagement. (2023). Designlab. <a href="https://designlab.com/blog/impact-of-emotional-design-user-engagement">https://designlab.com/blog/impact-of-emotional-design-user-engagement</a></li><li>Let’s get emotional: emotion research in human computer interaction. (2022). ResearchGate. <a href="https://www.researchgate.net/publication/221518006_Let&#39;s_get_emotional_emotion_research_in_human_computer_interaction">https://www.researchgate.net/publication/221518006_Let&#39;s_get_emotional_emotion_research_in_human_computer_interaction</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7903fb397b6a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>