<?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 Victor Moreno on Medium]]></title>
        <description><![CDATA[Stories by Victor Moreno on Medium]]></description>
        <link>https://medium.com/@thecodingteacher_52591?source=rss-6c2ef86448d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Fuax9tzJNmuru4P0PvbQBg.jpeg</url>
            <title>Stories by Victor Moreno on Medium</title>
            <link>https://medium.com/@thecodingteacher_52591?source=rss-6c2ef86448d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 15:14:31 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@thecodingteacher_52591/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[Why Your Team Isn’t Delivering More with AI: Theory of Constraints]]></title>
            <link>https://medium.com/@thecodingteacher_52591/why-your-team-isnt-delivering-more-with-ai-theory-of-constraints-427cbeab37c5?source=rss-6c2ef86448d------2</link>
            <guid isPermaLink="false">https://medium.com/p/427cbeab37c5</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Victor Moreno]]></dc:creator>
            <pubDate>Sun, 28 Jun 2026 19:37:13 GMT</pubDate>
            <atom:updated>2026-06-28T21:06:44.057Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jF31JfhexMEoDO6bG760qQ.png" /><figcaption>“generate a cool abstract image representing theory of constraints”</figcaption></figure><p>I’m writing this article to explain how software delivery actually works, applying constraint theory and simulations to the delivery process in a way that I’ve never seen done before. But the article wouldn’t be complete without exploring the question of why so much of the online discourse is about fake 100x productivity gains nobody who works on real systems seems to be able to emulate.</p><p>Many “gurus” online brag about pushing code into GitHub in the order of 50k lines per day. To those of us who are actually building real things used by real people, the 100x productivity claims are laughable. And yet, these claims take up a very large proportion of the online discourse on AI. “Peter Steinberger says he just writes loops! omg!”</p><p>We need to explore why this is. Justin Welsh explained this phenomenon extremely well in <a href="https://justinwelsh.substack.com/p/certainty">his blog yesterday</a>. Long story short: these people are marketing themselves and their companies. The screenshot of 55 github PR’s in one day is marketing collateral. The post telling you that if you can’t ship 50k loc per day it’s a skill issue, and the arguments in the comment section… Those are the goal, not an accidental side effect. These types of claims are what goes viral.</p><p>This is a problem because of second-order effects. These things dominate the AI discourse so much that otherwise reasonable people start believing them. The following are all concerning things I’ve seen or heard about in the real world, from smart people who are good at business and decision-making in general:</p><ul><li>“As soon as they start delivering 3x with AI, I’ll give everyone a raise.”</li><li>“We’re going to break up our team into pods of 4 so we can deliver faster with AI”</li><li>If online reporting is to be believed, meta has been in full ai psychosis mode for a while and that’s why they shipped an account takeover bug that allowed anybody to claim an account so long as they were within the expected geofence associated with the account.</li></ul><p>In this article, I aim to do a few things:</p><ol><li>Give you an intuitive heuristic how to quantify the actual impact of AI code generation on feature release output.</li><li>Explain theory of constraints by summarizing the explanation given in the excellent book <a href="https://amzn.to/4uRJhAq">The Goal</a> (affiliate link).</li><li>Explain how people <em>think</em> software delivery works.</li><li>Explain why they’re wrong and how the theory of constraints applies to software delivery.</li><li>Explain how accelerating the typing of code cannot possibly be expected to speed up software delivery.</li><li>Explain what to do about it to get actual speed gains.</li></ol><p>But I can’t do any of that without you first taking a step back and understanding that the vast majority of the 10–100x productivity stuff is just marketing. Now that we did that, let’s get into it.</p><h3>A Thought Experiment Showing Why Very Fast Code Doesn’t Make Much of A Difference</h3><p>Even before AI, there was almost no feature that a highly competent dev couldn’t prototype in a day or two. By “prototype”, I mean a white plain html page with no styles, connected to a nodejs backend with no security, which demonstrates the functionality. For the purposes of this article, let’s call this a <em>white-page prototype</em>.</p><p>Here’s the thing: if you have a decently mature team, the amount of code required for a white page prototype is not too much higher than a <em>maximal tech debt production implementation</em>. A mature team already has a security middleware that takes care of security, and you just embed your code into a route that’s protected by that middleware. A mature team already has a component library that takes care of UI, so you just need to use components from this library instead of plain divs. A mature team already has a throttling/rate control middleware, an authorization middleware, etc, so you just need to add a few config flags to these.</p><p>The difference between a white-page prototype and the minimal, most unmaintainable version of the production implementation is perhaps 5–10% more code.</p><p><em>This has always been the case</em>.</p><p>Here’s your intuitive heuristic to estimate the impact of AI code generation on your feature release timelines: the next time your team gives you an estimate of 3 months to deliver something, ask them: “how long would it take you to deliver a white-page prototype by hand, without using AI assistance in coding?” If the answer is 3 days, then the impact of AI code generation over the 3 month timeline is 3 days.</p><h3>Theory of Constraints</h3><p>In the book <em>The Goal</em>, the protagonist Alex leads his son’s Boy Scout troop on a long hike and treats it like a scheduling problem: the “goal” is getting the entire troop to the destination together, not just the fastest kids. As they walk in a line, he notices large gaps opening up between the faster boys and the slower ones, with the slowest boy, Herbie, consistently falling behind and effectively determining the pace of the whole group.</p><p>This situation reveals two key phenomena: <strong>dependent events</strong> (each boy’s position depends on the one ahead) and <strong>statistical fluctuations</strong> (each boy’s walking speed varies over time). Alex realizes that trying to push the fast kids harder doesn’t help the troop reach the goal faster because everyone ultimately waits on Herbie, the bottleneck.</p><p>Alex is from a manufacturing background, so from his frame of reference, <em>efficiency</em> matters a lot. If you were to use the boy scout walk as an analogy to a factory, then the distance between kids is equivalent to unfinished parts on hand. That means that keeping the troop tight is an important goal - in the context of a factory, being able to keep the group tight means fewer unfinished parts on hand, which in turn means less working capital demands, which in turn means less cash flow strain.</p><p>To improve efficiency, Alex moves Herbie to the front of the line. This works very well to increase efficiency, but the line is still moving too slow. In order to address that, Alex looks at Herbie’s backpack and realizes he’s carrying something like 50lb of stuff. Alex then spread Herbie’s unreasonable load throughout the rest of the troop. This increased Herbie’s walk speed significantly.</p><p>Once the slowest scout became the pace-setter and got extra support, the gaps shrunk (efficiency increased), and the troop’s throughput (distance covered together per unit time) increased. This becomes Alex’s vivid, physical model of the Theory of Constraints: the system’s performance is governed by its constraint, and improvement comes from managing that constraint, not from optimizing non-constraints.</p><p>The book is really interesting because the book <em>also</em> shows the phenomenon whereby you can learn a simple insight from a simplified experience and then struggle to apply it in a more complex context.</p><p>In the book, it takes much brainstorming, trial and error, and even consulting with the book’s Yoda character to figure out how to increase the throughput of the factory.</p><p>In practice, there are no recipes for accomplishing these things. Alex and the other factory managers needed to get really creative. They did things like rescheduling lunch breaks, redistributing work, reassigning workers, and even engaging in “bad practices” like letting certain machines sit idle if running them would just create excess parts inventory that the bottleneck couldn’t process.</p><p>This is much the same as it is with every complex discipline. True elite performance never comes from a 10-step checklist. It comes from being focused on results and relentlessly pursuing the right results.</p><p>In the next sections, I talk about what the right result is for software development.</p><h3>How People think Software Delivery Works</h3><p>This is how most people think software delivery works, and they’re wrong:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9B4Ifkh49JrvFdQhYgQ3AQ.png" /></figure><p>This is wrong for two main reasons:</p><ol><li>It considers a <em>single change</em>, rather than considering <em>the throughput of the whole team</em>.</li><li>It ignores defect rates. It’s assumed that once code is merged, everything is sunshine and rainbows from that point. This is wrong. Your production defect rate is <em>the most significant constraint on the system</em>. This is the constraint to every <em>real*</em> tech team, even the lowest sophistication teams with zero monitoring or observability, because the second you have lots of real customers, they let you know when the software is not working.</li></ol><p>Non-functional defects are of particularly large importance. In AWS, my job is probably 50% dealing with non-functional defects. Database schema doesn’t support new access pattern. Data backfills. Some operation is too slow. Some input crashes the api. This is the stuff that AI is particularly bad at.</p><h3>Theory of Constraints applied to Software Delivery</h3><p>To correctly understand software delivery, the central thing you need to look at is <em>the software release pipeline</em>. The software release pipeline is the logical reflection of your software factory.</p><p>I’ve created a <a href="https://swolebrain.github.io/ddb-simulator/team-throughput/">simulator</a> to help you visualize the software delivery pipeline. The default values reflect what I would consider a very unhealthy (but pretty common) team. Here’s a picture of the simulator, after running through 254 code changes with that team.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XVRss3tfRS4rXE6lu_CZlQ.png" /><figcaption>Picture of simulator, hosted at <a href="https://swolebrain.github.io/ddb-simulator/team-throughput/">https://swolebrain.github.io/ddb-simulator/team-throughput/</a>. This simulation is based on merging changes directly to main and never working in feature branches. Feature branches are stupid and make all these problems worse. There are some simplifying assumptions I’ve made as well. For instance, feature flags can be used to significantly lower the risk of a production deployment, such that it wouldn’t force you to do a rollback. However, wide use of feature flags can introduce its own failure modes (for example, interactions between many feature flags can create a combinatorial explosion of possible code paths — modality is usually bad and hard to test).</figcaption></figure><p>I encourage you to play around with the simulator. It will reveal interesting insights and help you really grok everything I say below.</p><p>The basic insight is understanding the release management dynamic: <strong>the more changes you put into a deployment, the more likely that deployment contains at least one bad change</strong>.</p><p>Even if each individual change has a low defect probability, the deployment risk compounds as the batch gets larger. A deployment with 2 changes is much safer than a deployment with 50 changes, because there are 50 chances for something to break.</p><p>For software teams, the implication is that if your team keeps producing code faster than your release process can safely validate and ship it, unreleased work piles up. Then each deployment gets bigger. Bigger deployments are more likely to have a defect. Production deployments with defects create fixes, retesting, coordination overhead, context switching (a dev who had moved on to the next task now needs to fix a defect), and thus more slowness.</p><p>This behaves as a negative feedback loop:</p><ol><li>Developers produce more changes.</li><li>More changes accumulate before release.</li><li>Deployments get larger.</li><li>Larger deployments fail more often.</li><li>Failures create fixes, retesting, and pauses.</li><li>The release queue grows even more.</li><li>Eventually production releases become rare or stop entirely.</li></ol><p>This is why teams with heavy manual QA often end up doing big-bang releases with code freezes. It is not because that is inherently the best way to ship software. It is because once the validation process cannot keep up with the rate of change, the organization has to stop new work just to stabilize the release candidate.</p><p><em>The counterintuitive lesson is that merging more code into main does not necessarily mean deploying more code.</em></p><p>The bottleneck is not developer productivity. The bottleneck is the team’s ability to safely validate and release small changes continuously.</p><p>The healthier pattern is smaller batches, faster feedback, and fewer changes per deployment. That reduces the probability that any one deployment is bad, makes failures easier to diagnose, and keeps production releases moving.</p><p>Now of course there are steps you can take here. You can break up your system into smaller systems that deploy independently. You can speed up your staging validations. You can use property-based testing in addition to unit tests to discover edge cases. You can use mutation testing to discover more tests you should write. You can have a suite of benchmark tests. None of these things can be done well by AI without a human designing them. This takes me to my next point.</p><h3>How to Use AI Effectively</h3><p>The key to using AI effectively is to <em>work on different things</em>. Stop hoping that AI will make your regular feature roadmap faster without you changing your systems. It won’t. It can’t.</p><p>AI shines at all the important-but-not-urgent stuff that constantly takes a back seat. You should use AI to build the things that will make you faster in the long term.</p><p>AI also shines at lowering <em>developer friction</em>. The main reason some higher level engineers and managers stopped coding is because they have too much scope to know the details of anything. They’re also often not up-to-date with the tools — a manager or PE might have left off at browserify and gulp, and the team has migrated from gulp to webpack to rollup to vite since. With AI, none of this matters. AI can figure out how to work your build tools in 5 minutes, without you having to wrestle with it for 2 days.</p><p>AI is to cognitive workers what the mechanical plow was to farming. <strong>The biggest impact of AI is to change the calculus of what’s worth building, and how good/robust it should be.</strong> Below are a bunch of examples of how AI can actually improve the quality of what you build. Many of them are based on things my team has done to improve our systems.</p><p><em>Before AI</em>: We’re spending 100k a month on RDS Postgres because we need to support very high traffic for this one db entity, but we don’t have the bandwidth to migrate.<br><em>After AI</em>: Let’s ship dual-writes to a dynamodb table, migrate that write path off postgres and save $50k a month.</p><p><em>Before AI</em>: We keep having to maintain this 6k loc wrapper around library X because we need to overcome its limitations.<br><em>After AI</em>: We can understand library X in 30 minutes now, let’s add the feature and get rid of our workaround.</p><p><em>Before AI</em>: let’s release this and hope it actually improves performance.<br><em>After AI</em>: let’s write a simulation to see how this performs for each of our largest customers.</p><p><em>Before AI</em>: we’re moving too fast and there’s too much product ambiguity to have comprehensive test coverage.<br><em>After AI</em>: AI can write your tests and update them in 2 minutes if the requirements actually merit updating the hundreds of tests.</p><p><em>Before AI</em>: “man, I <em>know</em> this is a 1-line change but i don’t even know how to run a typescript project.”<br><em>After AI</em>: kiro-cli chat --trust-all-tools &quot;add account id XXXXX to the allowlist to use the feature YYY. Send a pull request and tag Brent and Ali when done. The codebase is at ssh://.../MyTeamServiceAllowlist.git&quot;</p><p><em>Before AI</em>: This process consumes a whole lot of time, but it’s too hard to automate.<br><em>After AI</em>: let’s vibe code a cli tool and write agent SOP that uses the cli tool to execute the process autonomously, with human review at the end.</p><p><em>Before AI</em>: we need some performance benchmarks, but nobody has time to do it and only 2 people in the team have experience with the criterion library.<br><em>After AI</em>: your manager can totally add these.</p><p><em>Before AI</em>: let’s make some horrible balsamiq mocks and hope they actually communicate the intent.<br><em>After AI</em>: PM can vibe-code the whole thing against a simplified local backend, and you can iterate on functionality together without having to build the real thing.</p><h3>Conclusion</h3><p>I am a huge AI advocate. I write about 90% of my code with AI, and use AI in some way or another to assist me in almost everything I do. It pains me to see such a transformative technology being misapplied.</p><p>There are way too many leaders who see that AI can make an app that looks pretty nice and is functionally equivalent to a white-page prototype, and they mistakenly assume from this that the whole software delivery machine can be accelerated in this way. Leaders and IC’s alike need to learn to separate “this feels so fast” from “my team is delivering value faster.”</p><p>It doesn’t help that there’s so many snake oil salesmen selling bullshit like “claude code clean room python prompt in 2 hours” (that thing didn’t even run, and it had 2300 LOC). Or making up bullshit terms like “loop engineering”. These people are not motivated by you learning. <em>They’re just marketing themselves and their companies</em>.</p><p>Don’t let the nonsense from these marketers convince you that software delivery has drastically changed. Coding the white-page prototype has always been a <em>minute</em> part of delivering software. Teams that ignore the fundamentals of software delivery and get pressured into misusing AI will get <em>worse</em>, not better.</p><p>Use AI to manage cognitive load and you will get speed and stability. Use it for speed, and you will get slowness and throw away your company.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=427cbeab37c5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rationality always wins in the end.]]></title>
            <link>https://medium.com/@thecodingteacher_52591/rationality-always-wins-in-the-end-051cf43bf747?source=rss-6c2ef86448d------2</link>
            <guid isPermaLink="false">https://medium.com/p/051cf43bf747</guid>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Victor Moreno]]></dc:creator>
            <pubDate>Sat, 23 May 2026 11:42:22 GMT</pubDate>
            <atom:updated>2026-05-23T11:42:22.728Z</atom:updated>
            <content:encoded><![CDATA[<h3>Rationality always wins in the end. Junior SWE job roles will come back, but they’re not gonna be the same thing.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kWdVlLctEiFEjbfCuR0Kaw.png" /><figcaption>pictured: AWS SWE in 2019</figcaption></figure><p>When I first interviewed at AWS in 2019, they flew me out to Seattle for a 3 day mini vacation for my interview. After I got an offer, they gave my family a free paid vacation to visit Seattle before my start date. They provided corporate housing in a luxurious apartment. They paid for my relocation (all the way to moving my cars and my massive dog who couldn’t fit in airplanes). They even paid for us to have a local friend. Seriously, there was a lady whose job it was to meet with us, show us places, give us the scoop on the various neighborhoods and schools, etc.</p><p>This level of pampering extended into the workplace too. Even though Amazon is known as supposedly the least pampering of the big tech firms, here’s what I found:</p><p>- I got to work at 8am sharp on my start date only to notice that everyone else started strolling in at 10–1030. Some of them were getting there at those times because of the brutal seattle commutes (traffic and distances) but many still had bed head when they walked to the office.</p><p>- I take my work seriously and work intensely, and then I look a few desks over and there’s a guy who “works” with anime playing on his second monitor. All day. Nobody bats an eye. This guy still works for AWS today.</p><p>- Although the “extra amenities” are sparse (at least compared to what i hear from other places), there’s hardly a day that the ping-pong table and the console room don’t see at least 2–3 hours of use.</p><p>Fast forward to today, none of this is happening. People get to the office at 9, stay late. New buildings don’t have “extra amenities” on every floor, if they do at all. This is the unspoken thing that leaders accomplish with mass firings; it nudges back towards rationality.</p><p>But it’s imperfect. A lot of our best people were the ones to get canned. There’s many possible explanations as to why so many of the wrong people got fired. Some say downsizings need to be business based in order to not risk legal action or whatever. My take: leaders can’t actually tell who is a true value-add and who is fluff.</p><p>The reality is the exuberant, irrational conditions of big tech resulted in attracting a certain percentage of employees who simply decided to game the interview. They’re not intellectually curious. They don’t read books or papers, in their field or any other. They’re just here for the pay, and they gamed the interview to seem like they were intellectually curious without being so. And now the company is still stuck with most of them.</p><h3>How it Should Have Been Shows What the Future Might Hold</h3><p>There’s an interesting funny phenomenon with smart people. Their first inclination is always to reinvent everything. They never think to look at what others have done. Big tech hiring was no different. Bunch of smart people who never bothered to look at how other industries worked (or if they did, they thought they knew better).</p><p>Top tech tried to figure out how to create an elite workforce from scratch. In the process, they painted themselves into a corner that will take years to get out of. Natural attrition will do its work but it will take many years.</p><p>If top tech wanted to learn how to attract and retain top talent, they needed to just look at how top law firms, top accounting firms, and top management consulting firms do it.</p><p>Management consulting firms still do puzzles (kinda the business version of leetcode), but the reputation is totally different. Management consulting firms figure: if I’m gonna hire a kid right out of school who doesn’t know sh** about f**, that kid better be willing to go through a meat grinder.</p><p>And so it is. Top management consulting, accounting, and law firms have the reputation that if you join as a junior associate you better clench them cheeks. You’re in for a rough ride. If you join one of this firms, this is where you <em>really</em> cut your teeth. This is where you <em>really</em> build a career.</p><p>This is how software development in big tech should be, and how I predict it will be. You simply have way too much to learn as a new grad. There’s no reason for the company to spend 1–2 years financing you becoming not-useless at your own pace.</p><h3>The Future Junior</h3><p>All the exciting change going on with AI will definitely impact the nature of <em>what</em> juniors do. I personally think that the best post-AI junior is one who actually uses the type of app you work on. They’re the ones that will come with the product skills you need, and they can get the engineering mentorship from inside.</p><p>If I’m AWS, I’d rather hire some solution architect who can code as a SWE than a kid fresh out of school. If I’m peloton, I want someone who uses peloton. Etc.</p><p>But I have near-zero percent confidence in that prediction, there’s too many variables there.</p><p>However when predicting <em>how</em> juniors will work, we have a framework. We can look at these other industries and see how they evolved.</p><p>There’s no question about it: juniors at top tech will have to <em>work hard</em>.</p><p>Juniors outside top tech will probably get paid abysmal (by comparison to 2014–2021) wages. Perhaps they won’t exist. Nobody outside companies like Bain hires “junior management consultants,” after all. Junior accountants and junior lawyers exist outside top firms, but not management consultants. Accounting and law are geographically bound, so that’s a difference. Why hire a junior in the US when I can get a senior from South America for the same price? I don’t know. But the days of this being an easy, pampering profession are done.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=051cf43bf747" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Common Ways to Suck at A/B Testing]]></title>
            <link>https://medium.com/@thecodingteacher_52591/common-ways-to-suck-at-a-b-testing-2ef3a59e2dc5?source=rss-6c2ef86448d------2</link>
            <guid isPermaLink="false">https://medium.com/p/2ef3a59e2dc5</guid>
            <category><![CDATA[data-driven-decisions]]></category>
            <category><![CDATA[data-driven-marketing]]></category>
            <category><![CDATA[ab-testing]]></category>
            <dc:creator><![CDATA[Victor Moreno]]></dc:creator>
            <pubDate>Fri, 01 May 2026 13:48:50 GMT</pubDate>
            <atom:updated>2026-05-01T13:48:50.974Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fHuoyj3Y5gU1qytS" /><figcaption>Typical IYI A/B testing setup</figcaption></figure><p>A lot of people talk about being “data driven” but suck at it. Many of the people who think they’re good at being data driven, because they know what an A/B test is, still suck. They aren’t really data driven, they’re acting on whims, but with extra steps.</p><p>By far the most common way to suck at being data driven is by using data gathering methodologies that can never provide real quantitative insights, only qualitative. There’s a lot of people out there running “cohort analysis” with “cohorts” of 13 people, or running usability studies with 8 people and then reporting results like “12.5% of users preferred x”. That’s just one user, Johnny, chill. You’re literally passing off one user’s opinion as “data” by dressing it up as a percentage.</p><p>But like I said, even in the A/B testing crowd, it’s often not much better.</p><p>A/B testing has this reputation as a clean, scientific way to make product decisions. In reality, most teams are running something closer to las vegas slot machine statistics — pressing buttons, watching dashboards, and convincing themselves they’ve improved business outcomes.</p><p>This article is about the most common ways teams quietly sabotage their own experiments. Most things here apply whether you’re testing ads or product pages or in-app experiences.</p><h3>Noise: the silent killer</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Nsp33EMDLJbNMhzb" /><figcaption>The same exact software over time has different usage patterns.</figcaption></figure><p>The biggest threat to any experiment isn’t bias — it’s noise.</p><p>Temporal noise is especially dangerous. Your product metrics naturally fluctuate over time due to factors completely unrelated to your experiment: day of week, seasonality, marketing campaigns, even infrastructure quirks.</p><p>I’ve seen this firsthand in several different contexts. At one point, while researching task completion rates for my team in AWS, I noticed that re:Invent week not only featured vastly higher number of users trying out products, but also massively high task completion rates. If your product’s onboarding wizard usually has 40% completion, during re:Invent it might have 80–90%.</p><p>Another example is even more surprising. If you have a system using dynamodb and you look at dynamodb latencies, there are <em>clear</em> temporal noise patterns. Even when running the same traffic again and again (automated test traffic) there will sporadically be windows of time when latencies (which are otherwise pretty teady at around 10ms) spike up by 50 or even 100% (to 15–20ms).</p><p>The skinny is this: If you run an experiment over too short a window, or if your experiment variations don’t run simultaneously, you’re not measuring your change — you’re measuring whatever slice of time you happened to capture.</p><p>A good experiment ensures both variants experience the same temporal conditions — and runs long enough to smooth out the noise. You’d be surprised at how many people just ship a change to 100% of users, then compare metrics before and after, and call that an A/B test.</p><h3>Statistical significance vs. MDE (what people get backwards)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*0Z18DAJsmWBXOm1I" /><figcaption>Statistical significance pertains to this sample only, MDE is a feature of the experiment’s design.</figcaption></figure><p>Most teams treat statistical significance as the goal. It’s not.</p><p>Statistical significance is a property of the sample you already collected. In the extreme, you can achieve it with a single sample if the observed effect is large enough. All it tells you is: “given this data, the difference is unlikely to be zero.”</p><p>What it does <em>not</em> tell you is whether your experiment had a real chance of detecting an effect in the first place.</p><p>That’s where minimum detectable effect (MDE) comes in.</p><p>MDE is about controlling Type II error — the probability of missing a real effect. When you define an MDE, you’re saying: “if an effect of at least this size exists, I want a high probability of actually detecting it.”</p><p>And that requirement forces your hand on sample size.</p><p>To reliably detect an effect, your sample needs to be large enough that:</p><ul><li>Real differences don’t get drowned out by noise.</li><li>You have enough power to surface the signal if it exists.</li></ul><p>This is why MDE-driven sample sizes are so much larger. You’re not just asking “is this sample surprising?” — you’re designing the experiment so that it <em>would have found something meaningful if it were there</em>.</p><p>A useful mental model:</p><ul><li>Statistical significance is backward-looking: given this sample, does it look non-random?</li><li>MDE is forward-looking: if a meaningful effect exists, will this experiment actually catch it?</li></ul><p>Skip MDE, and you get the worst of both worlds: underpowered experiments that miss real improvements, but still occasionally produce “significant” results which are actually just noise.</p><h3>P-hacking your way to nonsense</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-9YvYNW6l9hV7K0D" /><figcaption>If you run this many meaningless experiment, you’re bound to find <em>some</em> random wins.</figcaption></figure><p>If you run enough experiments, some of them will “work” purely by chance.</p><p>P-hacking refers to trying lots of variations, slicing data endlessly, stopping early when results look good, or just running enough tests until something crosses the threshold.</p><p>In practice this might look like running dozens and dozens of experiments testing the hero image of the landing page.</p><p>The result is a steady stream of false positives.</p><p>Combine this with weak or nonexistent MDE discipline, and you start shipping changes that:</p><ul><li>Don’t actually improve anything.</li><li>Add complexity and maintenance cost.</li><li>Crowd out better ideas.</li></ul><p>Over time, your product fills up with “wins” that never really worked.</p><h3>Experiment interaction: when tests collide</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*daWrIhi_wJn9vgm7" /><figcaption>Plesae forgive for the terrible image. Hopefully it gets the point across about what experiment stacking means and how experiment interaction can be awful.</figcaption></figure><p>Most experimentation platforms allow multiple experiments to run at once. That sounds efficient — but introduces a subtle failure mode.</p><p>Experiments can interact.</p><p>Example:</p><ul><li>Experiment A changes a button’s text color to red.</li><li>Experiment B changes the same button’s background to yellow.</li></ul><p>Individually, both might be fine. Together, the ui becomes really jarring to the eye.</p><p>This isn’t edge-case weirdness — it’s a structural flaw in how most systems assign experiments independently.</p><p>When I worked at Noom, we used a model based on “experimental surfaces.”</p><p>Each surface (like a landing page) could have zero or more experiments, but each user was assigned at most one experiment per surface. If two experiments targeted the same surface, they were never shown together.</p><p>Instead, traffic was split across:</p><ul><li>Experiment 1 (control + variant)</li><li>Experiment 2 (control + variant)</li></ul><p>Yes, each experiment takes longer to reach its MDE. That’s a feature:</p><ul><li>No interaction effects.</li><li>Less temporal noise.</li><li>More trustworthy results.</li></ul><p>Most systems optimize for speed. Better ones optimize for validity <em>and</em> speed.</p><h3>Not running ablation tests</h3><p>Teams love shipping incremental improvements. What they rarely do is verify that those improvements actually compounded into something meaningful.</p><p>That’s what ablation tests are for.</p><p>An ablation test compares your current product against a previous baseline — say, the version from 10 months ago.</p><p>You’re asking: did all these “wins” actually add up?</p><p>It’s common to discover that:</p><ul><li>Improvements canceled each other out.</li><li>Some changes degraded other parts of the experience.</li><li>The net effect is negligible — or negative.</li></ul><p>Without ablation tests, you’re accumulating changes without auditing their combined impact. You’re paying for a growth department without knowing whether you’re even getting anything out of it at all.</p><h3>UX flickering: when your tooling sabotages you</h3><p>Frontend-driven tools like Optimizely often run their A/B tests by injecting scripts that run after the page loads and then mutate the DOM.</p><p>This creates flicker:</p><ul><li>The user sees the original version.</li><li>Then elements shift into the experiment layout.</li></ul><p>Even if the experiment is good, this degrades the experience because layout shifts are jarring. More importantly, in most cases it is only your experimental version that has a weird layout shift/ui flicker. The control group then gets a smoother experience. This is a hurdle for the experiment to overcome; you can never know if the experimental arm lost because of the degraded UX brought about by the layout shift or because the idea you had actually didn’t work.</p><p>You end up rejecting good ideas because your experimentation layer introduced noise.</p><p>The fix is architectural: apply experiments before render — server-side or at build time.</p><h3>Ignoring the backend</h3><p>Frontend experiments are easy, so teams over-index on them.</p><p>But the biggest wins are often in the backend.</p><p>One of the most impactful experiments I’ve run was a pricing test. We tried several price points and chose the second-highest — it maximized revenue.</p><p>That single experiment produced an annualized $50M lift. It took about a week to implement.</p><p>You don’t get outcomes like that by tweaking button colors or signup forms. You need an experimentation framework that reaches:</p><ul><li>Pricing logic</li><li>Eligibility rules</li><li>Billing flows (eg renewal cycles, installments, etc)</li></ul><p>If you’re only testing UI, you’re playing in the shallow end.</p><h3>Getting stuck in local maxima</h3><p>A/B testing is great at small, local improvements. It’s terrible at big jumps.</p><p>At Noom, about 90% of users were women in a specific age range. Any experiment that improved conversion for men would get statistically drowned out.</p><p>Even if it worked, you wouldn’t see it. That’s a local maximum.</p><p>More broadly, a lot of A/B testing tends to ossify your product structure and prevent innovation. For example, let’s say you run an ecommerce shop and you’ve hyper-optimized everything. But now, agent-first shopping experiences are a big deal. Your first crack at trying to convert your website to be agent-first is just that… your first crack. It doesn’t have the benefits of years of A/B testing optimization to back it. So it’s bad. Conversion numbers and AOV’s take a hit. You scramble to go back to the comfort of the known.</p><p>Escaping local maxima requires boldness. Sometimes you have to accept worse short-term metrics to find a better long-term position.</p><p>A/B testing won’t do that for you automatically. That’s where your strategic mind comes in.</p><h3>Having no principles</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G_RW0pB_MJpWVLknlrAVzQ.png" /><figcaption>Sexualized females in game ads is a rampant unethical technique for achieving winning results in A/B tests</figcaption></figure><p>Not every “winning” experiment is worth shipping.</p><p>There’s a well-known pattern in gaming ads which luckily has been stifled by regulation: creatives featuring sexualized imagery dramatically outperform others — even when they misrepresent the product.</p><p>They work. They also erode trust, and expose your company to risk of having its funnel snapped shut by regulatory changes.</p><p>Similarly, fake urgency tactics like countdown timers boost conversions while quietly damaging your relationship with users.</p><p>There are always tactics that drive clicks:</p><ul><li>Sex</li><li>Outrage</li><li>Fear</li></ul><p>But optimizing for short-term conversion at the expense of trust is a losing strategy.</p><h3>Ignoring code quality</h3><p>Most A/B testing setups rely on feature flags scattered throughout the codebase.</p><p>Over time, this creates modal behavior, which in turn creates a combinatorial explosion of possible states for users. This is very bad because it ends up putting you in a position where at any given time there’s a bunch of bugs in your app which you can only see if you have a user enrolled in the right experiment combination.</p><p>At Noom, experiments were deployed as separate static bundles. Instead of adding conditionals, you replaced the baseline implementation entirely. The backend decided which version to serve.</p><p>Experiment code would be hacky and scrappy but you just throw it away when the experiment is done. It didn’t have to live alongside your main codebase because it was deployed from a separate branch. Branches with a particular prefix in their name were automatically deployed to their own s3 location, and the backend could serve them from there.</p><p>This approach keeps code deterministic, eliminates flag debt, makes experiments easier to reason about, and most importantly, avoids cleanup. Your real codebase doesn’t need to be changed until the idea has proven its merit from an A/B test, and at that point it makes sense to come up with a good implementation that is well thought through and considerate of long term tech debt.</p><h3>Conclusion</h3><p>There’s a deeper write-up about the Noom experimental system (also written by me) in here: <a href="https://medium.com/noom-engineering/how-noom-uses-technology-to-herd-the-experimental-cats-256d3713c761">https://medium.com/noom-engineering/how-noom-uses-technology-to-herd-the-experimental-cats-256d3713c761</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2ef3a59e2dc5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why “vibe coding” Can Never Work: 2 Case Studies]]></title>
            <link>https://medium.com/@thecodingteacher_52591/why-vibe-coding-can-never-work-2-case-studies-22bdfb903041?source=rss-6c2ef86448d------2</link>
            <guid isPermaLink="false">https://medium.com/p/22bdfb903041</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Victor Moreno]]></dc:creator>
            <pubDate>Fri, 24 Apr 2026 20:24:57 GMT</pubDate>
            <atom:updated>2026-04-27T14:03:07.516Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*rlFOPgnWGePP5cIU" /></figure><h3>Introduction: Why Code Quality Matters</h3><p><strong>Let me preface this by stating I hand-type less than 10% of my code. I write mostly with AI.</strong></p><p>The fact that code quality still matters is pretty much IYKYK, but I’m gonna try anyway. Whether you’re technical or non-technical, if you have had <em>any</em> level of success in tech, that means you’ve moved up the food chain from worse teams/companies to better ones. If this is the case, you <em>probably</em> have noticed that some teams/people are <em>vastly</em> more effective at making software than others. Not only do they ship faster, they ship significantly fewer bugs.</p><p>You have probably also noticed that <em>everyone</em> can ship v1 of a greenfield app very quickly. Similarly, launching a brand new feature that doesn’t integrate with existing code is usually very fast. Even devs who kinda suck can add net-new code pretty quickly.</p><p>What you probably don’t know is why. If you’re non-technical, you probably think the difference is raw intelligence. If you’re technical, then you’ve noticed about 500 things that good teams do differently from bad ones, and so you probably think the difference between good and bad teams are a combination of those 500 things.</p><p>Neither of those answers are correct. The reason why good teams are better than bad teams is that they have developed good software design skills, they have the mental discipline to apply them, and the business has the discipline/trust to let them.</p><p>Here’s the reality: a 150 IQ dev with sloppy habits will <em>still</em> become very slow to deliver over time, and ship lots of bugs, regardless of their intelligence. A less gifted dev with good habits, who actually expends effort in software design, will end up being much faster in the long run than the more gifted counter-part.</p><p>Similarly, developers who obsess about “best practices” but don’t apply them thoughtfully also develop garbage slop systems. <em>Good teams are good because of how they position their software — and therefore the business — for the future. </em>Everything else is just specifics. Continuous integration, automated deployments, automated testing… name any good practice. If it’s actualy helping the business it’s because it was thoughtfully implemented, with an eye towards the future.</p><p>Let’s go through a very intuitive exercise. Look at the two server rooms below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tk_XQoFJJ8ax4J5l22MB1g.png" /><figcaption>Bad cable management vs good cable management</figcaption></figure><p>Now consider a project like “We need to retire all our 2023 servers and replace them with new ones,” which in software would be the equivalent of something like “we need to replace Paypal with Stripe.” Which server room do you think can pull this off more easily?</p><p>Now, add one more constraint to the thought experiment. The employee who has to move the servers is brand new, with no context and no memory of past decisions. This is the situation you’re in with AI. At best, it comes into each project with a very coarse summary of the system. And it doesn’t reason. It can’t count the r’s in strawberry without receiving specific examples to do it.</p><p>Why would AI be any different than the low quality devs you’ve worked with? Given that it doesn’t reason, why would you expect that it can keep shipping slop on top of slop without ever getting to the point that shipping one feature creates two bugs? It’s still writing the same kind of code that sloppy devs — it’s just doing it faster. Why do you expect different long term results?</p><p>I think there can be valid arguments to make here — for example you use formal methods, your domain is easy to test and you can run millions of simulation tests, etc. But to just accept the premise that AI is 10x speed boost forever is foolish, especially for most companies that only have their code, 10 automated tests, and mostly rely on metrics and customer feedback to find bugs.</p><p>Sure, it seems like the fact that you can cook up a full stack app in an afternoon means everything is fundamentally different. It is not. It’s really no different from how the quality coming out of sloppy devs degrades more and more over time. The only difference is that ai types <em>really fast</em>, so the number of calendar months before you reach the unsustainable situation is much lower than with human coding.</p><p>If you’re technically inclined, I have some case studies to show you.</p><h3>Case Study 1: Contributing Claude yolo mode to Deckard</h3><p>Depending on the type of task I will be engaging in, I often launch agents in “yolo mode” (skipping permissions). Whenever I’m just using the agent for code assist, that’s how I like it. In my opinion, coding agents are useless if you have to click “allow” for every single thing they do. The tradeoff in quality is not worth it if you can’t just plan the task and go do something else.</p><p>But when I’m not just doing code assist but im having the agent help me with infrastructure or server maintenance, I <em>definitely</em> want the reins. I don’t want an agent to just delete my db or some nonsense.</p><p>So, I thought I would make a simple contribution to Deckard, the tabbed terminal emulator I’ve recently been using to manage my claude code sessions in my weekend coding.</p><p>Here’s my PR: <a href="https://github.com/gi11es/deckard/pull/86">https://github.com/gi11es/deckard/pull/86</a></p><p>I don’t know swift so I didn’t even look at the code. The code package’s creator, who actually understands the code, made three observations. The second and third one are very important.</p><p>The second comment is the type of thing that creates a TON of surface area for future bugs. The way AI chose to do this is to create two different places for handling arguments passed to claude code. This is a sure source of bugs at some point in the future. I didn’t know because I’ve never read the code.</p><p>The third comment is even more important because it means the feature AI built doesn’t even work properly.</p><p>And this is just the thing, AI is really good at <em>appearing</em> like the software it makes is working.</p><h3>Case Study 2: Adding Formatting Support to the Cedar Monaco Editor</h3><p>Here, I have the prompt-by-prompt play still saved in my terminal. I think it’s instructive to analyze it play by play.</p><p>In this session, what I wanted to do was add code formatting support to <a href="https://github.com/cedar-policy/cedar-monaco-editor">@cedar-policy/cedar-monaco-editor</a>. This is a react component to edit cedar policies. Under the hood, it uses microsoft’s “monaco” package which is derived from the vscode source.</p><p>Some editors have optional toolbars, and you can configure the toolbar to either show or not. I thought maybe monaco had a toolbar but it was hidden by default.</p><p>In the sequence below I’ll blank out curse words. Anyways I started off with this prompt:</p><blockquote><strong>Prompt</strong>: is there a way to configure monaco to have a button to auto-format code?</blockquote><p>Interesting caveat here: I made this prompt with a fresh session and the agent was running within a folder of a project that consumes the cedar-monaco-editor package. Obvious in hindsight, but this lead the AI down a rabbit hole of stupidity. After several minutes of analysis and thinking, it decided to create a brand new component that adds a button and uses a ton of hacks to format the policies within the cedar-monaco-editor.</p><p>The approach is absolute slop and not portable at all. Also visually incompatible with the design of the page. It’s just completely wrong. So then, when I noticed what it was doing, I interrupted it and prompted:</p><blockquote><strong>Prompt:</strong> hold the ____ up man what are you ____ing doing adding an external format button. I didn’t say to do that _______ ____ing ____. I deleted that EditorFrame component. I refuse to believe that monaco doesn’t have an internal way to make this work. I own @cedar-policy/cedar-monaco-editor too so i can enable it there. Now let’s explore and give me some options that arent ________</blockquote><p>Then it spent literally like 15 minutes thinking. It provided several options, all of which were terrible except the first one, which it recommended. This was the first option:</p><blockquote><strong>AI</strong>: Recommendation: do Option 1 in `@cedar-policy/cedar-monaco-editor`. ~15–20 lines total. Want me to write the patch for that repo? It would add:</blockquote><blockquote>1. A useFormatProvider hook in components/ that wraps the policy worker’s existing format message into a DocumentFormattingEditProvider<br>2. Register it from registerCedarLanguages(monaco) for the cedar language<br>3. Optionally enable formatOnPaste: true as a default in the options merged by CedarPolicyEditor<br>4. Bump the package version to 0.2.0 and publish</blockquote><blockquote>Consumer site needs zero changes after that — users press Shift+Alt+F (or right-click → Format Document) and it works.</blockquote><blockquote>Which direction do you want to take?</blockquote><p>I told it to go ahead and implement option 1. In the final implementation I noticed there was something off. I never read monaco docs but it still felt wrong that it was calling so much functionality from the react component itself. I asked an exploratory question:</p><blockquote><strong>Prompt</strong>: why do we need both ensureCedarFormatProviderRegistered and bindFormatterToModel?</blockquote><p>It gave me a detailed and thorough explanation of it. It turns out ensureCedarFormatProviderRegistered needs to be only called once per page load, while bindFormatterToModel needs to be called for every editor instance.</p><p>The codebase already has a global method to initialize and do all the stuff that onle needs to happen once per page load. This initialization method is called configureCedarEditors. So then I asked the AI:</p><blockquote><strong>Prompt</strong>: wouldn’t configureCedarEditors be a better place to call ensureCedarFormatProviderRegistered than from within the component?</blockquote><p>Its answer:</p><blockquote><strong>AI</strong>: Yes, you’re right. configureCedarEditors is the app-startup hook — it already calls registerCedarLanguages(c.monaco) for exactly the same reason (register language-level stuff once at config time, not every time an editor mounts). The format provider belongs right there.</blockquote><h3>Closing remarks</h3><p><strong>AI doesn’t have taste or common sense</strong>. The need to have a global initialization function is a pattern that any developer who’s ever written a decently complicated client-side application will know. AI just imitates, it doesn’t have common sense like this. Typically it will not care about doing things like this that are good for your business and result in code that it will be better able to maintain later. Humans are needed for this.</p><p><strong>AI is most impactful in tasks where you have conceptual strengths but don’t know the particulars of the tool you’re using</strong>. I outsourced a lot of thinking to the ai for this very-small task. If I was familiar with monaco internals, this was a 15 minute task. It ended up taking 2 hours. However, this is still a net gain because I didn’t have to read the monaco docs and become familiar with it. AI saved me the 10+ hours it would have taken to achieve this.</p><p><strong>The more thinking you outsource to AI, the more you have to verify the result</strong>. When I began this task, I didn’t even look up how to access monaco’s formatting capabilities (it’s a context menu and a keyboard shortcut, not a toolbar like I expected). I didn’t spend time thinking about the code to guide AI on its implementation. A prompt that <em>may</em> have succeeded at this would have said something like:</p><blockquote>Add auto-formatting support to Cedar policy editors on the website. I want to use Monaco’s built-in formatting UX (Shift+Alt+F, right-click → Format Document, command palette) — **not** an external button. No consumer-side component wrappers; this belongs in @cedar-policy/cedar-monaco-editor (which I own, at /Users/me/cedar-monaco-editor). The package already has a formatworker handler calling cedarWasm.formatPolicies; the missing piece is that nothing registers a DocumentFormattingEditProviderwith Monaco for the cedarlanguage. Create a new file called src/path/to/filename.ts and add that registration code there. Then invoke the registration code from configureCedarEditors(same place registerCedarLanguages is called — it’s the once-per-monaco-instance setup hook). Handle multiple &lt;CedarPolicyEditor/&gt; instances on the same page correctly (each owns its own worker). Scope: policy editor only; schema and JSON are out of scope.</blockquote><p><strong>The best results will happen when you are able to reach a level of clarity where you are ready to <em>only</em> outsource the tedium of typing out the code</strong>. In order to write the prompt above, I would have had to do enough research to know exactly how I’d have implemented the thing by hand. Of course, AI can also help with the research. But in order to maintain code quality with AI, you need to reach the step immediaetly above code itself in terms of clarity. Your problem definition should mention file names, classes/functions, define function signatures, etc. This takes time. AI doesn’t make you 10x faster at this. The speed-up is maybe 20%.</p><p><strong>There IS a time and place to just let go and let the agent slop it out</strong>. Don’t waste time reviewing code quality for some little tool you build to simplify customer data intake, or whatever. In my team in AWS we have a full slop tool to add throttling rule overrides (When someone wants to call one of our APIs faster than the default max). The code quality of the tool doesn’t matter. The tool generates the rules, we review the rules. Who cares if the tool’s own code is slop.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=22bdfb903041" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Comparison, achievement, happiness, and a life well-lived]]></title>
            <link>https://medium.com/@thecodingteacher_52591/comparison-achievement-happiness-and-a-life-well-lived-0cd18291ed1c?source=rss-6c2ef86448d------2</link>
            <guid isPermaLink="false">https://medium.com/p/0cd18291ed1c</guid>
            <category><![CDATA[happiness]]></category>
            <category><![CDATA[mental-health]]></category>
            <dc:creator><![CDATA[Victor Moreno]]></dc:creator>
            <pubDate>Fri, 27 Mar 2026 20:00:51 GMT</pubDate>
            <atom:updated>2026-03-27T20:00:51.346Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e6QCmfzH4PPMK_9AhU2M8Q.png" /><figcaption>What’s the real enlightenment?</figcaption></figure><p>Recently, I’ve been noticing a tremendous amount of material on social media that promotes what I think are unhealthy comparisons.</p><p>“I got a $300k job at 24, and you can too! Just sign up to my newsletter.”</p><p>“I made principal engineer at AWS at 30 years old, sign up for my career coaching service.”</p><p>It seems like everyone who’s made some decent money wants to peddle you their career building advice. I don’t think there’s anything wrong with this in and of itself. Hell, I give career advice all the time.</p><p>But I do think you can give career advice without obnoxiously flashing your compensation numbers. I think that this extreme focus on compensation numbers and timelines to achieve them is a breeding ground for psychological malaise and unhealthy comparisons.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nnZhgf0d27gwxRFdAnJM6w.jpeg" /><figcaption>The average faang bro just wants to show you how big his compensation is so you can buy his course. Nevermind the fact he didn’t even like the job enough to stay at it.</figcaption></figure><h3>Comparison is the thief of joy, but the fuel of achievement</h3><p>Unhealthy comparisons are awful for your mental health, but you need <em>some</em> degree of goal-setting, and comparisons are inevitable when setting goals. If you <em>only</em> compare yourself to yourself, you can get stuck making vanishingly slow progress, miles away from your potential, and never even realize it.</p><p>The place where this is by far the most evident is the gym. There are very few people who go to the gym for years on end and don’t have some kind of fitness goal. The vast majority of men (I’m a man so I feel more qualified to speak to the psyche of men) who frequent the gym want to be getting leaner, stronger, more muscular, or a combination of those.</p><p>Yet, a significant majority of people who go to the gym don’t progress at all. This is because they’re phoning it in. In spite of the fact that they’re showing up, they’re not showing up with the intensity required to make good gains.</p><p>One could say that this is because they simply don’t care enough. Fitness is something they sort-of want, but not something that their heart and soul desires with any level of real passion.</p><p>A huge number of people who go to the gym don’t even bench 200lbs.</p><p>The other day, I took a cardio kickboxing class. All the people in that time slot were friends — it’s one of those social workout joints designed so you make friends with the people you work out with. You know, so you get addicted to the social aspect and keep going.</p><p>Anyways all the folks in that class have been doing it for a long enough time to be friends, <em>yet literally none of them could do real pushups. They all were doing pushups from the knees. Even the men</em>.</p><p>When the class finished I was practically swimming in a pool of my own sweat. I averaged 168 heart rate for the whole hour (as a 40 year old man). Only two other people in the class had broken a sweat, and neither of them sweat enough to wet the floor.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/414/1*4U6DmX6lYO_eCD_U7E-zYQ.png" /><figcaption>What it looks like when you’re there to push yourself.</figcaption></figure><p>The key here is that what you can and should compare is inputs. Do not compare outputs. Comparing your career trajectory to someone else’s is a recipe for being miserable. But you can and should compare your level of effort and input. Compare it to yourself last year but also to other people. <strong>If you outwork everyone, you will outperform a lot of people</strong>, <strong>but you won’t outperform everyone</strong>. There will always be someone who gets better results than you with less effort. That is okay.</p><p><strong><em>So that’s lesson 1: never compare results. Only compare inputs.</em></strong></p><h3>On discipline</h3><blockquote>“Discipline is doing what you hate to do, but doing it like you love it” — Mike Tyson</blockquote><p>Motivational speakers and self-help gurus often paint discipline as the golden key to unlocking potential.</p><p>I have nothing against self-help gurus or motivation guys (not in general, although some of them are very cringe). In fact, one of the things I hope you take away after finishing this article is that sometimes you DO have to white-knuckle it in life. Sometimes you do have to go <a href="https://www.youtube.com/watch?v=ry2Cj08OY3A">Goggins</a>.</p><p>I asked claude to summarize the way motivational gurus think about self discipline and amongst the things it said was this:</p><blockquote>Picture this: you’re on a boat, and your dreams are on a distant island. Discipline is your motor — without it, you’re just floating, hoping the wind does you a solid. These gurus hammer in the idea that discipline isn’t about restriction, it’s about steering your life where you want it to go, full throttle.</blockquote><p>I think that’s really accurate. That’s exactly the perspective of people who push discipline as the answer to everything. You’re always at sea at risk of dying, and you need discipline to make it to some distant island, which is your goal.</p><p>The problem with this view is twofold:</p><ol><li>You are not in fact at sea at risk of dying.</li><li>When you live your life this way, you never make it to the island. It always moves away from you, or disappears once you reach it and then you have to paddle to a new island.</li></ol><h3>On Passion and Your Zone of Genius</h3><p>So, as a response to the discipline-bros, there’s another movement of people who say that you need to just follow your passion. If you follow your passion, they say, you won’t need discipline, because your work won’t actually feel like work.</p><p>Here’s the problem with that: most people don’t have a single burning passion that neatly aligns with something valuable in the market. Even if they do, turning that passion into a consistent source of income requires grinding through lots of tasks that don’t feel “passionate” at all — logistics, marketing, client management, tweaking the copy of your landing pages, Facebook ads A/B tests, etc. In a nutshell: boredom.</p><p>The myth of passion persists because it’s emotionally comforting. It promises that if you “just find your calling,” joy and success will magically intersect. Reality’s messier — your <strong>zone of genius</strong> is usually built, not discovered. You develop skill in a field through consistent effort, feedback, iterative improvement, and most importantly <em>real-world rewards</em>. That positive spiral of development transforms into competence that <em>creates</em> passion over time.</p><h3>On Happiness and Contentment</h3><p>It’s so interesting the journey many of us go on when learning about happiness. Many of us realize that the pursuit of accomplishment is a sort of treadmill. If you focus on outcomes, you never achieve anything because the second you attain an outcome you’re already focused on the next.</p><p>So that creates the space to realize that happiness is a choice, or a skill as my friend <a href="https://www.youtube.com/watch?v=nqk7n4h4WV0">Metin puts it</a>. But then, even though it should have been obvious, it’s surprising to later learn that the better you get at the skill of happiness, the less impetus you have to strive towards accomplishment.</p><p>The more I’ve developed my happiness skill, the less Ifind myself working brutal hours and investing Herculean effort into my goals.</p><p>Basically what I’m saying here is: you better spend a bunch of time under effort, outworking everyone around you, before you find your soulmate and have kids. Because relationships are ultimately what makes for a fulfilling life, and by the time you build a fulfilling life, you will not <em>want</em> to outwork everyone around you. So by then, you better have built a tremendously valuable set of skills that can get you paid.</p><h3>On a Life Well-Lived</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/480/1*O2YnWJsLYEHtrr3Bir9pwg.gif" /><figcaption>“Conan, what is best in life? To crush your enemies, see them driven before you, and hear the lamentation of their women”</figcaption></figure><p>So, then, what is best in life? What makes for a life well-lived? How can you know that, when you’re on your deathbed, you’ll have no regrets?</p><p>Because you didn’t waste your time on fruitless endeavors, nor relinquished happiness by wishing for outcomes you can’t control.</p><p>Because you were the man/woman you wanted to be, in the ways you can control — strength of mind, body, and spirit.</p><p>Because you were father/husband (or mother/wife) you’d wish for your kids.</p><p>Because whatever you did, you did with passion and fire, with enough effort to actually improve. And that goes for everything — from your profession to being a parent and a spouse.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0cd18291ed1c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Intelligence, Cain, 3pstein, Demons and Satan]]></title>
            <link>https://medium.com/@thecodingteacher_52591/intelligence-cain-3pstein-demons-and-satan-09bd13732733?source=rss-6c2ef86448d------2</link>
            <guid isPermaLink="false">https://medium.com/p/09bd13732733</guid>
            <category><![CDATA[morality]]></category>
            <category><![CDATA[religion]]></category>
            <category><![CDATA[ethics]]></category>
            <dc:creator><![CDATA[Victor Moreno]]></dc:creator>
            <pubDate>Thu, 05 Mar 2026 02:46:17 GMT</pubDate>
            <atom:updated>2026-03-05T03:15:53.689Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pTltDYd2HoPpnpJ8cnDINQ.png" /></figure><p>The extent of the horrible crimes of human trafficking revealed by the files is atrocious. But a lot has been said about that, so I’m gonna talk about something else. Something nobody is talking about. I will describe the phenomenon whereby someone achieves that level of evil. Let’s start with some observations.</p><h3>A Weird Obsession With Intelligence</h3><p>It’s kind of a shame that the guy loved to talk about smart cars so much. If you search for the word “smart” in the files you get 7,797 hits, but most of them are about smart cars. This makes it tedious to get a real sense of the extent of jeevacation’s obsession with intelligence.</p><p>But I perused 1% of the hits. Let’s see what we find.</p><p>In this one, he’s apparently looking for an assistant. Seems like the only criterion is that they be smart:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jLat1SSYht1-4CHUukdRFA.png" /></figure><p>In this next one one, he refers someone who’s “very smart” to someone else. The other guy wasn’t impressed. Note how jeevacation didn’t even really know what the supposed smart guy did for a living (we’ll come back to that, it matters).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wwoyeFyfFWSjgDi0utzSAQ.png" /></figure><p>Discussing intelligence with yet another dude, some Robert Schank fellow who produces PhD’s but can’t produce a signature line with his own name capitalized:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jRS_EQDxgFHaKe3Gofetwg.png" /></figure><p>Talking to good ole’ Roger again about where to find smart people:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Xa6ygVqPttl44Wa8xInNIA.png" /></figure><p>Whenever jeevacation looks for a doctor, it’s gotta be a smart doctor:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*C0VZJcs8AUcfTGAU5ynEvw.png" /></figure><p>An exhilarating recap of an AI summit where I’m betting nobody had a CS background and nobody could even create an empty neural net in sklearn if their life depended on it:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HcPKFFFUJoFsOnEjcz_qqQ.png" /></figure><p>And finally the god old one everybody has seen:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MPKgzoIfkgW-fb31jH1ezA.png" /></figure><p>Does this strike you as odd? Jeevacation was a college dropout who only ever had two real jobs — uncredentialed math teacher and unqualified stock trader. He failed at both. He could barely spell, had very poor grammar, and he couldn’t even use a computer very well (randomly adding extra spaces and using commas instead of periods for an ellipsis). How does someone like that find himself obsessing about intelligence? How do they find themselves in a spot where they think that they have some sort of sixth sense to know intelligence, so much so that he thinks he can gauge it without even knowing much about the person’s background? And why does this barely-literate sycophant in the image above propose jewness to jeevacation as a measure of intelligence?</p><p>Stick around if you want to know my hypothesis. But first, some background.</p><h3>Cain &amp; Abel</h3><p>Two brothers bring offerings to God: Abel gives what is first and best, while Cain brings an offering that is somehow lesser — an outward act that doesn’t fully contain the inward sincerity and self-correction real sacrifice requires. When Abel is “accepted” and Cain is not, Cain meets a moment every person recognizes: the sting of being judged, the humiliation of comparison, and the suspicion that reality itself has singled you out.</p><p>God’s response is not condemnation but a warning and an invitation: if you give your sacrifice with sincerity, you can be accepted, and if you don’t, sin waits like a predator at the door. Cain can’t tolerate that counsel; instead of imitating what’s good, he isolates his brother and kills him — eliminating the living reminder of what Cain could have been. Afterward he lies (“Am I my brother’s keeper?”), becomes alienated and restless, and then God curses him forever.</p><p>Cain’s central error isn’t mere anger; it’s the progression from disappointment to <strong>resentment</strong>, then to the decision to destroy the ideal (Abel) rather than correct himself and offer something better.</p><p><strong>If the outcomes of your efforts don’t meet your desires, you must either increase your effort or reduce your desires.</strong></p><p>That’s the way a reasonable, good person things, anyway. In reality, there are two paths one might take. The path towards resentment and nihilism, or the path towards virtue (where you raise your effort or lower desire).</p><p>Can you see it yet? Can you see the parallel between jeevacation and Cain? Can you see jeevacation progressing from failing at college, to cheating his way into a position as a math teacher, to failing as a math teacher, to cheating his way into wall street, to failing at wall street, while his ego and sense of self-satisfaction grew the whole time?</p><h3>Measurement and Ego</h3><p>The thing about real life is that, unlike written stories, it has a long time arc. In reality, when you embark on Cain’s road of resentment, it doesn’t take a few pages of reading until you’re cursed and your life is ruined. It takes a long time. Throughout that time, it seems like things are working. After all he did move from math teacher to wall street to financial advisor to Leslie Wexler. It was easy for him to rationalize and to believe he was crushing it.</p><p>He never accomplished anything that was cognitively hard. He never proved a math theorem. Hell, he couldn’t even graduate college. But why was he so obsessed with intelligence?</p><p>Because you can’t escape yourself. Deep down these people know that they’ve never achieved anything cognitively hard. Most of us are trying to win our respective metaphorical chess matches by getting better at chess, while people like jeevacation just slip poison into their opponent’s drink. That’s how they go through life. While some of us graduate as mathematicians through toil and hard work, they get a position teaching math through a religious friend. While some of us grow from the point of being perplexed by discrete math to knowing the chinese remainder theorem by heart, they talk about percentage jewness as a measurement of intelligence. While some of us spend time accumulating money and investing in a rational and conservative way, they seduce a billionnaire, film them abusing underage girls, and then blackmail a fortune out of them.</p><p>Being able to win at chess by hitting your opponent over the head with a baseball bat doesn’t make you a good chess player. Those of us who have virtue and have actually achieved cognitively challenging things know this. Deep down they know this too, and what’s worse, they know we know. Hence the circle of sycophants and all the weirdness about intelligence.</p><h3>Demons and Sin</h3><p>Everything I’ve thus far discussed can be tied back to one demon — pride.</p><p>By demon, I don’t mean a little red goblin. A demon is a negative force that attracts you towards whats bad for you. “<a href="https://en.wikipedia.org/wiki/Resistance_(creativity)">Resistance</a>,” as defined in “The War of Art,” is a demon. The feeling of sloth that seduces you into doom scrolling for 90 minutes straight is a demon. A compulsion towards lust and fapping is a demon. Any compulsion towards any vice is a demon.</p><p>Most of us, most of the time, can keep these demons at bay. Most men can fap once a week without spiraling into complete desensitization.</p><p>But most of us also have that <em>one</em> demon that really gets us. If you’re over 300lbs yours is most likely gluttony (as mine was when Iwas 330lbs).</p><p>For jeevacation, his demon was pride.</p><p>There is an interesting phenomenon with demons and sin though. If your personal demon defeats you — if it convinces you that its demonic way is the righteous way, it will bring other demons with it.</p><p>This is how someone like Chris Farley might go from gluttony to drugs to prostitutes every night to death.</p><p>And this is how, I think jeevacation went from the wounded pride of a failed Math student to traffickign young girls.</p><p>This is the main lesson for you to learn too. You know yourself. You know the demon who’s your nemesis. At any point in time you’re either slipping towards defeat or climbing towards victory. Every time you slip towards defeat you make it more likely that another battlefront with another demon will open up.</p><p>And if you keep letting demons pull you down the road towards sin, it won’t be too long before you’re completely seduced by Satan, unrecognizable to the parents who raised you and knew you as a lovely little kid.</p><p>And that’s just what I think Satan is. Communing with Satan is living in a space where you’re committing acts of great evil that would be unrecognizable to yourself as an innocent child, and yet not even realizing you’re doing evil.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=09bd13732733" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Insight is not Enough]]></title>
            <link>https://medium.com/@thecodingteacher_52591/insight-is-not-enough-4100a56f9ea8?source=rss-6c2ef86448d------2</link>
            <guid isPermaLink="false">https://medium.com/p/4100a56f9ea8</guid>
            <dc:creator><![CDATA[Victor Moreno]]></dc:creator>
            <pubDate>Sun, 25 May 2025 02:22:37 GMT</pubDate>
            <atom:updated>2025-05-26T16:04:01.700Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JmqNxuhf7Q1ABRY29yivnA.jpeg" /><figcaption>The hydra, a symbol representing a very important concept: anti-fragility.</figcaption></figure><p>I’ve been toying with this idea for a long time. When it comes to doing hard things, (such as entrepreneurship, learning something difficult, quitting an addiction, etc.) I have noticed that there’s a lot of people with tremendous amount of insight who fail. They’ve read all the books and they say all the right things when talking at a high level, yet when it comes to taking action, they totally fuck it up. They fail so spectacularly that observing them is like observing someone drive off a bridge in real time.</p><p>Conversely, I’ve noticed people who are barely able to spell, who don’t read, don’t know statistics, don’t know math, and yet have a clarity of thought that makes them succeed in business. They cut through bullshit like a light sabre through a piñata.</p><p>I am not one of those people.</p><p>I am profusely imbued with <a href="https://nassimtaleb.org/2016/09/intellectual-yet-idiot/">Intellectual-Yet-Idiot (IYI)</a> DNA, as if my mom had grabbed me by the ankle when I was a baby and dipped me in a fountain of over-intellectualization and naïve scientism to protect me from prophesized enemies. Over the past 15 or so years I have managed to make a lot of progress away from these tendencies, though. In this article, I want to share the latest perspective on this.</p><p>The idea is this: <strong>academic style insights are neither necessary nor sufficient to accomplish things in real life </strong>(in the realms of business, weight loss, love, etc). When it comes to the most random areas of life, conceptual insights are neither necessary nor sufficient to achieve high performance, <strong>and high performance can be achieved without conceptual insights.</strong></p><p>I must acknowledge I stole the phrase “Insight is not enough” from psychologist Orion Taraban. I am not linking the video where I got it from because the video contains a lot of red pill bullshit that paints men and women as caricatures, and I don’t wanna promote it. However, you can often learn from people whose philosophies and ideas you don’t agree with and this is one of those cases. Dr. Taraban gave one example that really drove the point home for me so I’m repeating it here:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RLlrpq2gRHC9xpjITXtDuA.png" /><figcaption>Dr. Taraban describing how different people approach addiction.</figcaption></figure><p>Dr. Taraban says that when he worked with alcoholics, he noticed that there was a group of what you might call “frequent flyers” who were often in and out of rehab. They knew the psychological jargon about addiction better than he himself did, and yet they couldn’t manage to kick the habit. Meanwhile, other people with very little insight just decided “hey, this is ruining my life, I should stop” and then stopped.</p><h3>A Made-Up Parable About Implicit Wisdom</h3><p>Imagine you are given a coin to flip. You’re free to flip it however you want.</p><p>You flip it 100 times, in many different ways — using a cup to shake it, shaking it in the hollow between both of your hands, flicking it with your thumb, even bouncing it off your own calf muscles. All 100 flips land on heads. You are offered the opportunity to gamble on the 101st coin flip. What do you do?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/675/1*fNtp_XIvLAl3hANRsMTSWQ.jpeg" /><figcaption>The difference between the super-idiot and the Jedi is in the details. Even though their actions might fit the same description, they are not the same.</figcaption></figure><p>The idiot will bet on tails because he expects that the streak has to end. He thinks that the “streak” of tails must be about to end. He doesn’t know these events are independent. He has never read a book on probabilities and is just a plain idiot.</p><p>The IYI knows that coin flips are independent events. He’s studied statistics quite significantly in fact, and he scoffs at the attempt to trick him into betraying his knowledge.</p><p>The Jedi realizes the coin is rigged. He guesstimates that the probability of getting heads again in the 101st toss is very high, close to 100%. He places a bet that makes him slightly uncomfortable, perhaps a thousand or two thousand dollars, depending on how wealthy he is. He wins.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/675/1*hFwupvhKw0_XVZAJM5jfGQ.jpeg" /><figcaption>The thought process and first principles that drive wisdom often don’t require deep book knowledge. Rather, they rely on careful observation and analysis.</figcaption></figure><p>Note that the Jedi in this case probably doesn’t know any statistics. He might have taken it in college or whatever, and forgot all of it. Conceptual understanding of randomness is neither necessary nor sufficient to recognize the true nature of this situation and thus profit from it.</p><h3>The Curse of Knowledge</h3><p>It just so happens that those of us imbued with Intellectual-yet-idiot DNA love accumulating knowledge, and this can be a trap. The more you know, the bigger the toolbox you have to over-analyze situations and focus on the wrong aspects of each decision in front of you. <em>Over-reliance on formal knowledge can obscure simple truths</em>.</p><p>But this can go a step further. Sometimes the “knowledge” you have is wrong. Some examples:</p><p><strong><em>Being “data driven” in business</em></strong></p><p>In business, it is a meme that one should be “data driven.” To this end, just about every software company installs telemetry tools and event tracking.</p><p>The overwhelming majority of the time, the way these data are used is completely wrong.</p><p>I have included faulty data in pitch decks (with sample sizes of &lt;100 and no randomization) to help companies raise funds, and VCs don’t bat an eye. They think these metrics I fed them actually constitute data.</p><p>I have seen in my day job at AWS how faulty data is used as evidence and nobody bats an eye either. It is beyond the scope of this article to explain how to conduct proper, statistically sound analysis of metrics, but almost everyone does it completely wrong — inadequate sample sizes, not measuring statistical significance, not controlling for noise, not computing minimum detectable effect, etc. And then they make dumbass decisions based on emotions masquerading as data.</p><p><strong><em>Semmelweis and Hospital Mortality</em></strong></p><p><a href="https://www.perplexity.ai/page/semmelweis-and-hospital-mortal-rEd3t2VhTayl2PlI1ie.Xg">Ignaz Semmelweis</a>, a Hungarian physician, noticed that doctors’ unwashed hands spread fatal infections in maternity wards. Despite data showing handwashing reduced mortality, the medical establishment rejected his idea. Why? Doctors were steeped in the miasma theory (disease from “bad air”) and refused to believe they could be the causing harm.</p><p>Like the Jedi questioning the coin’s fairness, Semmelweis prioritized observation over dogma. His peers, however, were cursed by their education — they dismissed a lifesaving fix because it conflicted with their “knowledge.”</p><p><strong><em>Kodak’s Digital Camera Blind Spot</em></strong></p><p>Kodak engineer Steven Sasson invented the first digital camera in 1975, but the company shelved it to protect its film business. Executives, experts in chemical photography, couldn’t fathom a world where digital tech would dominate. Their deep knowledge of film became a liability, blinding them to disruption.</p><p>Kodak’s experts behaved like the Intellectual-Yet-Idiot — overanalyzing risks of cannibalizing film sales while missing the obvious: <em>they owned the rigged coin</em>. The future was digital, they owned the future, and yet they spent their time and energy on spreadsheets and forecasts for their film products.</p><p>This takes us to what makes IYI’s so bad at decision-making: <em>the plethora of cognitive tools they have available (whether they’re actually using the cognitive tools correctly or not) distract them from thinking about the things that actually matter.</em></p><h3>IYI’s Learn the Wrong Lessons</h3><p>We’ve seen that IYI’s tend to focus on the wrong things to study and to make decisions. But there’s one more thing they do which makes them even worse decision-makers. IYI’s tend to have a habit of compounding on their own wrongness. They learn the wrong lessons and the older they get the further they steer away from sound thinking and good decision-making.</p><p>As an example, let’s go back to the way some orgs in AWS analyze metrics. Let’s say there’s some wizard to create IAM instance profiles and they want to improve the completion rate of this wizard. So what they’ll do is ship a change to that particular wizard on, say, a Tuesday. The change takes 5 days to deploy through all regions. Then they look at a metric like <em>task completion rate</em> during the one week prior to the change, and the week when the change happened. They ignore the fact that the change deployed in phases through regions. They ignore temporal noise (the assumption that the population of users on week 1 and week 2 are the same is faulty). They don’t perform any sample size/minimum detectable effect computations. They don’t compute statistical significance (not that statsig would even tell you anything with small samples). And then, after all this, they will generalize the results of the experiment into something like “actually, conventional wisdom is wrong, increasing the number of steps in a wizard does not hurt completion rates.”</p><p>This is an <em>insane</em> conclusion to draw. Not only because the statistical methodology is more precarious than a soap bubble’s survival rate in a cactus farm, but because establishing causation in that way is nowhere near being <em>something that statistics can even do</em>.</p><p>It’s important to know that we are all prone to these mistakes. Just because you make a mistake like this, it doesn’t make you a full-time IYI. What makes you an IYI is defaulting to this mode of thinking, and compounding on these mistakes over time. Some more quick hits on poor reasoning and drawing terrible conclusions:</p><ul><li>The viral <a href="https://www.youtube.com/watch?v=tVz_hf4Jbe0">37 year old broke loser</a> who thinks software engineering consulting doesn’t work even though there’s multi-million dollar consulting firms.</li><li>Anybody who believes in “gentle parenting.”</li><li>People who try to emulate Pietr Levels.</li><li>People who staunchly believe in design patterns, clean code, communism, and other ideological myths.</li><li>The average founder goes through a “trough of disillusionment” where they feel lost and try to act on all the random, stupid, often-conflicting advice they get from VC fund reps who’ve never started a business themselves.</li></ul><p>But the fact that IYI’s make stupid decisions doesn’t explain why <em>some </em>people who appear so unpolished end up making such great decisions. To explore that, we need to take a little detour into the nature of knowledge.</p><h3>Concepts Exist Whether you Know Their Name or Not</h3><p>In <em>Anti-Fragile</em>, Nassim Taleb belabors the point that there are many concepts that exist and don’t have words associated with them. The ancient Greeks, for example, <a href="https://www.reddit.com/r/worldbuilding/comments/2vwu2r/ive_been_reading_about_how_the_ancient_greeks/">only had words for four colors</a>. Yes, this is mind-boggling. They all had access to eyes and walking under the sky. Thus, they had the concept of blue-ness. Yet, they didn’t have a word for it.</p><p>Similarly, the concept that there are things that gain strength when you try to harm them has existed for thousands of years. We know this concept exists because it’s what the hydra is. Many things in life are hydra-like. For example, as they say, “no publicity is bad publicity.” Often, CEOs and public personalities actually end up benefitting from scandals. News outlets and rumor mongers try to harm their reputation and end up strengthening their aura instead.</p><p>Attempts to suppress an industry often strengthen it. Look at the firearms industry in America, for example. It keeps growing, and many people say it’s because of the forbidden fruit syndrome.</p><p>Attempts to ban a book, movie, or song often make it more popular and sought after (e.g., “banned in Boston” was once a selling point).</p><p>Nassim Taleb calls these things “anti-fragile,” and he notes that we don’t have a word to describe this concept.</p><p>But I digress. The point I’m making is not about anti-fragility but about knowledge itself.</p><p><strong>Some people understand certain concepts without knowing the name of the concept, and without having studied it formally.</strong></p><p>This should be self evident if you think about how you yourself have a facial recognition algorithm in your head, and you don’t know how to describe it. You can recognize faces without any conceptual framework for doing so.</p><p>The ancient Greeks could recognize colors without having words for them.</p><p>Some people recognize when they’re fighting a hydra without having a word for it, and without having read <a href="https://www.perplexity.ai/page/heracles-and-the-hydra-U9beVWIXRPuOBEcD6Mdqlg">the story where Hercules fights the hydra</a>.</p><p>And some people sometimes have good intuitive understandings of life, the world, and business, without having studied the academic explanations for the phenomena that belie their understanding.</p><p><em>I challenge you to think about this right now. I challenge you to identify people in your life who seem to make good choices without much theoretical backing.</em> I know many examples in my personal life. I know you have some too.</p><p>One of my friends recently purchased a gym. He doesn’t work out and doesn’t know anything about working out. He never went to business school and doesn’t know anything about gyms. He just saw a distressed asset and bought it at a steep discount, then brought it to break-even in 2 months. How? By using <em>printed flyers</em>. He noticed that there’s a lot of businesses in the area near the gym and created a flyer targeting office workers to come to the gym after getting off work and before going home. “Don’t commute during rush hour,” or something like that.</p><p>Meanwhile, people like me wouldn’t buy a business without reading 5 books on business acquisition first, and “learning” all these “lessons” such as “you shouldn’t buy a business in an industry you don’t know or care about.” I’d have spent 100 hours over 2 months working on local SEO instead of just sending out some flyers.</p><h3>How to Defeat your IYI Tendencies</h3><p>I cannot tell you how to develop the kind of “implicit wisdom” my friend who bought the gym has. Like I said, I don’t have it. I have IYI DNA instead.</p><p>What I can tell you is what I’m trying to do. I’m trying to change my identity. I’m trying to stop identifying as an erudite. I’m trying to stop connecting everything to abstract math. I’m trying to become a <a href="https://grugbrain.dev/">grug</a> instead. I think all my readers are engineers but in the off-chance that you’re not and you don’t understand that grug-brained article, it’s saying the same thing as this one:</p><p>Stop overthinking. Stop overanalyzing. Stop defaulting to solving every real world problem with some mental model from a book. Develop the habit of looking for rigged-coin situations instead of an incremental 2% gain from applying a more clever equation.</p><p>Only resort to the big brain solution once you are <em>extremely convinced</em> that there are no rigged coins to look out for. And make sure when you do resort to a big brain solution, you use <strong>rigor</strong>. Do not use charts to explain your feelings.</p><p>If you don’t know enough statistics to <em>correctly</em> run an A/B test, do not run it.</p><p>If you don’t know enough data analysis to <em>correctly</em> make business forecasts, do not create that ETL pipeline and do not create the business forecasts.</p><p>If you haven’t actually spoken to a single customer yet, don’t do <em>that</em> stupid “Customer personas” exercise every consultant does.</p><p>This is not an attack on education, academic knowledge or concepts learned from books. Robust mental models, <em>when correctly applied</em>, always help you. Instead, this is an attack on the IYI personality trait that blinds us into <em>reducing every problem to book concepts without even looking at the situation holistically first</em>. This is always a bad thing to do, but it’s especially bad when paired up with the other vice of IYIs: thinking our grasp of the book concepts is much better than it actually is.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4100a56f9ea8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Contrarian Use of AI — Making Humans Smarter Instead of Dumber]]></title>
            <link>https://medium.com/@thecodingteacher_52591/contrarian-use-of-ai-making-humans-smarter-instead-of-dumber-c01b56b91934?source=rss-6c2ef86448d------2</link>
            <guid isPermaLink="false">https://medium.com/p/c01b56b91934</guid>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[ec2]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Victor Moreno]]></dc:creator>
            <pubDate>Fri, 18 Apr 2025 10:48:18 GMT</pubDate>
            <atom:updated>2025-04-18T10:48:18.633Z</atom:updated>
            <content:encoded><![CDATA[<h3>Contrarian Use of AI — Making Humans Smarter Instead of Dumber</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*_CdUNJhcoD8SP4S3wn6CNQ.png" /><figcaption>Perplexity’s idea of me dancing with an AI. I’m fatter but also more jacked than that guy.</figcaption></figure><p>While the intellectually challenged masses scramble to use AI as a crutch for their pitiful cognitive abilities and intellectual laziness, a few of us are weaponizing it to become superhuman. The average person posting about AI is a drooling simpleton who dreams of outsourcing their work so they can spend their entire life daydreaming. They so desperately want their “idea guy” badge to be vindicated.</p><p>But that’s not gonna happen. Those of us with functioning neurons realize AI is the ultimate mental steroids. Below, I will give an example of how I use AI to deepen my expertise. I do this basically daily, sometimes multiple times a day.</p><p>You won’t understand any of the conversation I had with AI if you don’t know anything about AWS though, so stick around this preamble if you wanna learn a bit about AWS.</p><p>In order to understand the conversation I had with AI you have to understand what the ec2 instance metadata service does. So here’s an explanation about the ec2 instance metadata service (skip it if you already know):</p><blockquote>Think of the Instance Metadata Service as a secure internal directory or help desk specifically for each Amazon EC2 server (instance) you’re running. Just like how employees can look up company information on an internal portal, your EC2 instances can query this service to get important information about themselves.</blockquote><blockquote>Key points that make it valuable:</blockquote><blockquote>* It provides essential information like:|<br>* The instance’s own ID and network configuration<br>* The AWS region it’s running in<br>* Startup scripts it should run<br>* <strong>By far the most important thing:</strong> Security credentials it needs to access other AWS services</blockquote><p>Basically, each ec2 instance launches with a little internal http server you can query to get all those bits of info above. In particular, you can get the credentials that the ec2 instance itself has in order to access other AWS services. With those credentials, you can perform whatever AWS API calls the instance can perform (API calls to push logs, retrieve secrets, etc.).</p><p>There’s a second version of the instance metadata service that’s supposed to be more secure. Security vulnerability scanners make it a big deal if you’re not exclusively using instance metadata v2.</p><p>Now that the background is done, here is my conversation with AI:</p><p><strong>Me:</strong> I never understood the big deal about ec2 instance metadata v2. I understand that imdv2 protects against some weird attack vectors with 3p firewalls or vectors that exist when you have the ec2 host act as a reverse proxy and you misconfigure it. But it seems to me that the primary attack vector against an ec2 instances is someone gaining access to the host itself. If someone has shell access to the host, or RCE access to the host, then they can call the instance metadata service and fetch all the secrets from secrets manager and escalate privilege from there. Is there a way to protect against that?</p><p><strong>Claude</strong>:</p><p>You raise a good point. IMDSv2 primarily protects against SSRF (Server-Side Request Forgery) attacks, but you’re correct that if an attacker gains shell access or RCE, they can still access the instance metadata service.</p><p>However, there are several strategies to protect against this:</p><p><strong>IAM Role Session Duration</strong></p><pre># Set short session duration for credentials<br>aws iam update-role --role-name YourRole --max-session-duration 3600</pre><p><strong>Block IMDS Access at Network Level</strong></p><pre># Disable IMDS completely<br>aws ec2 modify-instance-metadata-options \<br>    --instance-id i-1234567898765432 \<br>    --http-endpoint disabled</pre><p><strong>Use AWS Systems Manager Session Manager</strong> instead of direct SSH access</p><p><strong>Implement IP Tables Rules</strong></p><pre># Block all access to IMDS except from specific processes<br>iptables -A OUTPUT -m owner --uid-owner specific-user -d 169.254.169.254 -j ACCEPT<br>iptables -A OUTPUT -d 169.254.169.254 -j DROP</pre><p><strong>Use ECS/EKS Task Roles</strong> instead of instance roles where possible</p><p><strong>AWS Nitro Enclaves</strong> for sensitive workloads</p><pre># Enclave can&#39;t access IMDS directly; must go through parent instance<br>aws ec2 run-instances --enclave-options &#39;Enabled=true&#39;</pre><p>The most secure approach is usually a combination of these methods along with proper application security practices.</p><p>Would you like me to elaborate on any of these approaches?</p><p><strong>Me:</strong> yeah tell me more about the iptables rule approach</p><p><strong>Claude:</strong></p><p>The iptables approach lets you control IMDS access at the network level. Here’s a more detailed breakdown:</p><pre># Allow specific user/process to access IMDS<br>iptables -A OUTPUT -m owner --uid-owner application-user -d 169.254.169.254 -j ACCEPT<br><br># Allow specific group to access IMDS<br>iptables -A OUTPUT -m owner --gid-owner application-group -d 169.254.169.254 -j ACCEPT<br><br># Block all other access to IMDS<br>iptables -A OUTPUT -d 169.254.169.254 -j DROP<br><br># You can also be more specific with ports<br>iptables -A OUTPUT -p tcp --dport 80 -d 169.254.169.254 -j DROP</pre><p>Some important considerations:</p><p>Make these rules persistent:</p><pre># Save rules<br>iptables-save &gt; /etc/iptables/rules.v4<br># Restore on boot (Ubuntu/Debian)<br>apt-get install iptables-persistent</pre><p>You can also use process names:</p><pre># Allow only specific process<br>iptables -A OUTPUT -m owner --cmd-owner your-application -d 169.254.169.254 -j ACCEPT</pre><p>Monitor blocked attempts:</p><pre># Log dropped packets<br>iptables -A OUTPUT -d 169.254.169.254 -j LOG --log-prefix &quot;IMDS ACCESS DENIED: &quot;</pre><p>However, be cautious because:</p><ul><li>Some AWS services might need IMDS access</li><li>System updates might need IMDS access</li><li>Boot processes might need IMDS access</li></ul><p>A more complete example might look like:</p><pre># Allow system processes<br>iptables -A OUTPUT -m owner --uid-owner root -d 169.254.169.254 -j ACCEPT<br><br># Allow specific application<br>iptables -A OUTPUT -m owner --uid-owner app-user -d 169.254.169.254 -j ACCEPT<br><br># Log attempts<br>iptables -A OUTPUT -d 169.254.169.254 -j LOG --log-prefix &quot;IMDS BLOCKED: &quot;<br><br># Block everything else<br>iptables -A OUTPUT -d 169.254.169.254 -j DROP</pre><p><strong>Me:</strong> Ilove the idea of restricting by process name! What are some aws services that might need imds access? Probably codedeploy, ssm agent, cloudwatch agent?</p><p><strong>Claude:</strong></p><p>Yes, exactly! Here’s a common list of AWS services/agents that typically need IMDS access:</p><p>AWS Core Services:</p><pre># SSM Agent<br>iptables -A OUTPUT -m owner --cmd-owner amazon-ssm-agent -d 169.254.169.254 -j ACCEPT<br><br># CloudWatch Agent<br>iptables -A OUTPUT -m owner --cmd-owner amazon-cloudwatch-agent -d 169.254.169.254 -j ACCEPT<br><br># CodeDeploy Agent<br>iptables -A OUTPUT -m owner --cmd-owner codedeploy-agent -d 169.254.169.254 -j ACCEPT</pre><p>Additional Common Services:</p><pre># AWS CLI (for system tasks)<br>iptables -A OUTPUT -m owner --cmd-owner aws -d 169.254.169.254 -j ACCEPT<br><br># ECS Agent (if running ECS)<br>iptables -A OUTPUT -m owner --cmd-owner amazon-ecs-agent -d 169.254.169.254 -j ACCEPT<br><br># Inspector Agent<br>iptables -A OUTPUT -m owner --cmd-owner awsagent -d 169.254.169.254 -j ACCEPT<br><br># AWS X-Ray Daemon<br>iptables -A OUTPUT -m owner --cmd-owner xray-daemon -d 169.254.169.254 -j ACCEPT</pre><p>Package Management (might need access during updates):</p><pre># Package managers<br>iptables -A OUTPUT -m owner --cmd-owner yum -d 169.254.169.254 -j ACCEPT<br>iptables -A OUTPUT -m owner --cmd-owner apt-get -d 169.254.169.254 -j ACCEPT</pre><p>A practical approach might be to:</p><ol><li>Monitor IMDS access first</li><li>Build your allowlist based on observed legitimate access</li><li>Then implement restrictions</li></ol><p>You can monitor access with:</p><pre># Log all IMDS access before implementing restrictions<br>iptables -A OUTPUT -d 169.254.169.254 -j LOG --log-prefix &quot;IMDS ACCESS: &quot;</pre><p>Then check your logs:</p><pre>grep &quot;IMDS ACCESS&quot; /var/log/syslog</pre><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c01b56b91934" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Modern Agile is Stupid: on why all great ideas are killed by simple-minded idiots, why Jira kills…]]></title>
            <link>https://medium.com/@thecodingteacher_52591/modern-agile-is-stupid-on-why-all-great-ideas-are-killed-by-simple-minded-idiots-why-jira-kills-02080b937de2?source=rss-6c2ef86448d------2</link>
            <guid isPermaLink="false">https://medium.com/p/02080b937de2</guid>
            <category><![CDATA[agile]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Victor Moreno]]></dc:creator>
            <pubDate>Sun, 16 Mar 2025 01:59:54 GMT</pubDate>
            <atom:updated>2025-04-30T01:38:34.534Z</atom:updated>
            <content:encoded><![CDATA[<h3>Modern Agile is Stupid: on why all great ideas are killed by simple-minded idiots, why Jira kills software, and how to be a good manager</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_3Xg0_fdp-j910BBxFvpTw.png" /></figure><p>An idea in your mind is 𝗻𝗼𝘁 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝘁𝗵𝗶𝗻𝗴 as the words you use to convey it.</p><p>Similarly, a process your software development team follows is 𝗻𝗼𝘁 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝘁𝗵𝗶𝗻𝗴 as the document outlining it.</p><p>Language is low resolution, and by definition it will leave out a lot of the nuance and details.</p><p>Written-down processes are like the string that keeps a kite from flying off and crashing. They give the team some basic parameters to play within. They’re like a rulebook to a sport.</p><p>But the execution of the actual process is the kite itself, flying, adjusting to the wind, moving omnidirectionally.</p><p>Many managers and “Agile practitioners” cannot tell the difference between the two.</p><p>They spend their day writing docs and adding processes like ticket templates, planning pokers, definitions of done, etc.</p><p>What’s worse, they actually think these things help.</p><p>Such people are one of the biggest wastes of air and space in software development teams. They are actively harmful to your team’s output.</p><h3>Why Managing by Jira is a Terrible Idea</h3><p>When I say “managing by Jira” I mean the phenomenon whereby some managers want to add more and more pieces of templatized processes. They want to add lengthy issue templates, blockers so you can’t merge code that isn’t linked to an issue, worthless rituals like planning poker, etc. This is often referred to as “meta work.” Excessive meta-work is terrible for three main reasons:</p><ol><li><strong>The sheer waste of time it is</strong>. In any given year, a team’s total working hours are finite. Out of 2000-2500 hours an engineer might put in on a given year, <em>if you make them spend an hour more than absolutely necessary on meta-work you’re a bad manager</em>. You’re actively divesting their time away from feature and operational work that is beneficial to customers, and investing it into bullshit. The amount of time that’s necessary to be spent on meta work, even in a BigTechCo like AWS, is somewhere in the vicinity of 10% (around 1% in a startup with a team of 2–3). If you make your team spend 30% of their time on meta work you’re holding the gross productive output back by <em>at least</em> 20% (in reality much more, as we shall see).</li><li><strong>It creates a self-perpetuating vicious cycle of bureaucracy</strong>. An environment that places undue emphasis on meta-work invites a particular type of employee: the sandbagger. These individuals excel at creating the appearance of productivity through meticulous documentation and process adherence, while their actual contribution to the product is minimal. In Jira, their work may appear impressive, but a glance at their Git commits tells a different story. This discrepancy is particularly problematic in large tech companies, where there are ample opportunities for paper-pushing. These sandbaggers not only fail to contribute meaningfully to the product but also set a dangerous precedent that values form over function. Then they get promoted and as they do, will typically worsen the situation with more meta-work. After all, they were rewarded for it. Note that not all sandbaggers are actually sandbagging in a sense of putting in low amounts of effort. Some of them work <em>hard</em>. It’s just that they’re sandbagging the stuff that actually matters, the productive work that benefits customers.</li><li><strong>It kills actual productivity</strong>. Perhaps most importantly, meta-work interrupts the flow state that engineers need to produce their best work. I’ve lost track of the number of times when I’ve taken immediate action and fixed significant issues in a shorter amount of time than what it would have taken me to fill out the ticket template to the standards of some of the most Jira-worshipping managers out there. The death of inspiration and flow happens in teams where your calendar is filled with useless rituals like planning poker, multi-hour sprint planning, velocity analyses, or where you have to do things like over-segmentation of tasks into sub-daily lengths and story pointing.</li></ol><h3>Some Meta-Work is Necessary</h3><p>I’ve been name-calling Jira a lot, hopefully it’s obvious I don’t mean Jira specifically but any task management software (Asana, Clickup, or if you also work in Amazon our <em>beloved</em> “Simple Issue Manager” and its younger sibling Taskei).</p><p>I also want to be clear that there is nothing wrong with the most basic scrum/Agile processes per se (the “advanced” ones like burndown charts are all bullshit). There are common needs that arise in software delivery, and scrum rituals offer some tools to address them:</p><ol><li>The need for communication and avoiding duplicate work. Scrum standups are a fine mechanism to address this.</li><li>The need to learn from mistakes. Retrospectives are a fine mechanism to facilitate this.</li><li>The need to work on the highest impact tasks at hand. Breaking work into sprints and re-evaluating every 2 weeks can facilitate this as well.</li></ol><p>What I am saying is that just because you meet every two weeks and have a “retro” that doesn’t actually mean you’re learning from your mistakes. It is possible — indeed likely — to have retros with no true self reflection.</p><p>Whether you use the tools in scrum or not, solving those problems of communication, learning from mistakes, and prioritizing is is the difference between teams that move erratically like flies and teams that fly towards objectives like a peregrine falcon attacking its prey. Moving fast in no particular direction is actually <em>worse</em> than moving slow due to wasting time on Jira. Erratic movement tends to create much more cruft and tech debt.</p><p>Being a peregrine falcon is obviously better, and it requires <em>some</em> meta-work.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FMqGo3G9LL2o%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DMqGo3G9LL2o&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FMqGo3G9LL2o%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/583d2cc1ce85c0820554cdf24e65cb5b/href">https://medium.com/media/583d2cc1ce85c0820554cdf24e65cb5b/href</a></iframe><p><strong>I want to also be very clear that I am not hating on checklists or step-by-step processes in a generic sense</strong>. It is a well-known fact that preoperative checklists and pre-flight checklists have a tremendous positive effect in avoiding accidents. I myself have written dozens if not hundreds of runbooks, step by step processes, and checklists for specific tasks in my career. Boiling down atomic processes to known checklists (or better yet, automating them) is usually good. <strong>The damage is in thinking that the entire product delivery process can be boiled down to a checklist</strong>.</p><h3>Some “Actual Work” Looks like “Meta-Work”</h3><p>I also wanna point out that some of your work output is actually best placed in Jira (or in my case and other Amazonians, Taskei).</p><p>Just 2 days ago I worked on an issue. The ticket was automatically cut because of a latency alarm — our service’s latency in a particular AWS region was higher than expected. This was a second occurrence of this ticket so I looked at the work that two previous on-call operators had done. They done log diving and ascertained the following:</p><ol><li>Identified the top calling customer (about 5000 queries per second into our system, mostly sustained during business hours)</li><li>Identified a different customer that seemed to have some kind of syncing workflow triggering every 3 hours, where they would hit us at 10,000 queries per second for a few minutes and then send zero calls until the next workflow</li><li>Chased down a couple of leads that didn’t pan out.</li></ol><p>They posted screenshots and log outputs of all these things into the ticket. This is not meta work. It is real work.</p><p>From there I went on to investigate our I/O latencies with that customer’s call path and saw that most of their latency was coming from a thing that should have been cached. So then I went on to investigate caching metrics for that customer, saw that our cache was not working as well as it should, and immediately introduced a small TTL increase to the cache.</p><p>My teammates saved me re-doing multiple hours of work by putting the work into the ticket. But of every entry that went into that ticket, 0% of it was meta work. 100% of it was real work.</p><p>So please understand, interpreting what im saying as a call to ban Jira or something like that would be utterly stupid. Google docs or quip are much worse places for this type of work.</p><h3>Why Idiots Think This Way (Prescriptions)</h3><p>Business methodologies, business frameworks, and other forms of watered down step-by-step guides are catnip for idiots. It’s simple: weak minds crave simple solutions. They want a paint-by-numbers approach to success, and Agile promises just that.</p><p>These halfwits see a world of complexity and instead of rising to the challenge, they retreat into a comfort zone of checklists and buzzwords. Agile, with its neat little ceremonies and easy-to-digest principles, offers them a security blanket. It’s the corporate equivalent of thinking a green smoothie will solve all your health problems.</p><p>The truth is, software development — hell, any complex endeavor — doesn’t work that way. It’s messy, it’s unpredictable, and it requires actual thinking. But that’s too much for these simpletons. They’d rather play scrum master and feel important moving sticky notes around a board.</p><p>It’s the same mentality that has CEOs swearing by their morning ice baths or meditation routines. As if freezing your ass off at 5 AM is what separates you from Jeff Bezos. It’s magical thinking for adults who should know better.</p><p>Real success in software development comes from deep understanding, creative problem-solving, and the ability to adapt to changing circumstances. It comes from experience, intuition, and sometimes, just rolling up your sleeves and grinding it out.</p><p>But that’s HARD. It’s difficult to resist the siren song of Agile authors promising you better results without you having to stay technically sharp.</p><h3>Jira as the Maker of Terrible Managers</h3><p>Probably the worst kind of manager is the toxic asshole manager, who chooses fear over charisma, threats over understanding, appeals to authority over data, yelling over persuasion, self over team, etc.</p><p>Another terrible kind of manager is the micromanaging manager, and there’s many other ways in which managers can be bad. But those other ways don’t concern us today, as the topic of this article is meta work.</p><p>Past a certain size, a team cannot operate well without a good manager. But managers obsessed with meta-work are, at their very best, like this guy:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fes0nPjLNsi0%3Fstart%3D29%26feature%3Doembed%26start%3D29&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Des0nPjLNsi0&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fes0nPjLNsi0%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/e56d7bf4ac22d8246d545968a605a180/href">https://medium.com/media/e56d7bf4ac22d8246d545968a605a180/href</a></iframe><p>The reason why these managers want to much to go into Jira is so they can have a sense that they understand wtf is going on with their team. This is usually because of laziness, incompetence (lack of technical ability), or both.</p><p>As a result, their concept of management is to encourage the team to adhere to these processes. These terrible managers obsessed with meta-work stay on the surface of things. An example:</p><p>In AWS our CI/CD system has a feature where it shows a little box at the top right with a warning if they notice you’ve taken too many manual actions. CI/CD should be fully automatic. If you’re having to retry integration tests, or having to override deployment failures in order to promote code with broken tests to production, or take any manual actions of the sort, that can be a sign of a team that’s in terrible operational health.</p><p>A bad manager will see this metric and bring it up in standup in a manner like this one:</p><blockquote>Team, I noticed our pipeline is giving us a warning for too many manual interventions over the last 30 days. Let’s make sure that we’re not manually intervening the pipeline unless absolutely necessary.</blockquote><p>This is the corporate equivalent of a shitpost. It’s the kid above pretending he’s actually helping the train start and stop. It’s a complete waste of air.</p><p>A good manager, upon seeing the warning on the pipeline tool, would actually go into the log of manual actions and look over the specific manual interventions that were taken (blocking pipelines to run load tests is a very, very different thing than promoting code that has broken tests). Then they might talk to some of the people who took manual actions in order to to better understand the situation. Then they would make a decision as to whether the “too many manual interventions” metric represents an actual problem or not. If it does pose a problem, they’d ask one of the senior engineers <em>how</em> to solve it and <em>roughly</em> how long it would take. Then they would prioritize solving the problem by deprioritizing something else.</p><h3>How to be a Good Manager</h3><p>And that’s the crux of the issue: a manager should be technically competent enough to comprehend the value of their team’s work by reviewing pull request descriptions, design docs, and engaging in meaningful discussions about the operation and architecture of the service. A manager should be technically competent enough to make decisions and go through investigations like the one I outlined above.</p><p>A manager doesn’t need to be able to write the code themselves. It’s typically better if they can, but this is way too rare and not terribly important — Sorry Elon fans, he’s wrong.</p><p>This hands-on approach is akin to a coach on the sidelines, observing and guiding their team in real-time. The coach doesn’t have to be able to rope 3-pointers himself. He doesn’t have to be able to dunk or cross over 22-year-olds. But he needs to understand the game.</p><p>Imagine a coach who never attends games and instead “manages” by looking at a spreadsheet of player statistics twice a week. When he talks to a player, all he can say is “keep it up” or “your ppg is down, you need to score more.” Such an approach would be universally recognized as ineffective in sports, and yet it’s sold as a panacea of software engineering productivity in the world of Agile. It’s snake oil. It’s garbage.</p><h3>Conclusion</h3><p>Let’s refer back to the original agile manifesto:</p><blockquote>Individuals and interactions over processes and tools<br>Working software over comprehensive documentation<br>Customer collaboration over contract negotiation<br>Responding to change over following a plan</blockquote><p>This is very clearly an anti-prescription stance. And it’s a good stance.</p><p>It’s a sad reality of life that small minds will always try to find a paint by numbers approach to achieving difficult things. You just have to make sure you go places where the density of these people is very low, and leave their teams if you find yourself in that unlucky situation.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=02080b937de2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Learn Things Worth Learning, With Examples]]></title>
            <link>https://medium.com/@thecodingteacher_52591/how-to-learn-things-worth-learning-with-examples-41f3af99f1c5?source=rss-6c2ef86448d------2</link>
            <guid isPermaLink="false">https://medium.com/p/41f3af99f1c5</guid>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[learning-and-development]]></category>
            <category><![CDATA[learning-to-code]]></category>
            <dc:creator><![CDATA[Victor Moreno]]></dc:creator>
            <pubDate>Tue, 07 Jan 2025 02:00:38 GMT</pubDate>
            <atom:updated>2025-01-07T02:08:35.489Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1019/1*fblc1aDlKhXrbqehvNeoLA.jpeg" /><figcaption>Enjoy the irony of an AI image in an article about meat-based learning</figcaption></figure><p>There’s a question I get rather frequently, and that question is related to learning. The question takes many forms:</p><ul><li>What are the most important things to learn in software engineering?</li><li>How do I learn X? (Could be a programming language, digital marketing, DB performance, meta ads, etc).</li><li>Do you have any books to learn [thing that cannot be learned from books, like sales, marketing, product thinking, high scale distributed systems, etc]?</li></ul><p>I find myself giving a lukewarm half-answer most of the time. That’s because in the middle of a conversation, it would be grossly pedantic of me to start a lecture about how the person I’m talking to is asking the wrong question.</p><p>This article is the expanded answer at what I think are the real questions behind those questions. These are:</p><ul><li>When you say you want to learn X, what do you hope to get?</li><li>What is the role of passion?</li><li>Is X a thing I should bother learning?</li><li>How much academic study should I engage in, and how do I pick resources to study?</li><li>What if I have no passions and no aptitudes?</li></ul><p>I intend for this article to evolve with my thinking over the years and become a resource I can share with anyone wanting to learn a difficult skill.</p><h3>When you say you want to learn X, what do you hope to get?</h3><p>Before embarking on any kind of learning journey, I think it’s important to be really clear on what it is one hopes to get out of it.</p><p>Too many are the entrepreneurs who start a business because “they wanna work for themselves” only to find themselves working for 8 bosses instead of 1.</p><p>I myself have made this mistake, <em>twice</em>. The first time, it was when I decided to start a consulting practice. I landed several clients and quickly found myself in a position where I just had several shitty bosses. I actually made <em>less</em> money than a full time employee of my skill level could. There’s a whole book on this topic called <a href="https://a.co/d/984qzrQ">The E-myth Revisited</a> which belabors this point thoroughly: the skillset to run and scale a business is a very different skillset from that of executing the technical work that the business sells. In other words, being a good developer and decent closer will not make you a successful owner of a software development consulting firm.</p><p>The second time was when I wanted to be an indie hacker following the <a href="https://medium.com/@yumaueno/who-is-pieter-levels-an-entrepreneur-who-earns-300-million-a-year-by-developing-various-products-4f2cde322aee">Pietr Levels</a> model. I started several businesses: a print-on-demand store, a book to teach people how to pass software engineering interviews, an AI transcription assistant, and a wellness app (for which I had co-founders).</p><p>In both cases, the main mistake I made was that I had a completely wrong, cartoonish idea of what the actual day-to-day life is in those businesses. In the case of becoming a freelance consultant, I expected to spend <em>more</em> time doing technical work and less time dealing with people. The complete opposite was true.</p><p>In the case of the indie hacker bootstrapped projects, I expected to be able to live a lifestyle of total creative freedom, just coding my little heart out, solving great technical problems. The complete opposite was the truth. For every 100 hours I devoted to these projects, about 99 of them were spent tweaking things like the CSS styles of the pages, google ads phrasing, meta tag descriptions, privacy policies, experimenting with free shipping, SEO, improving copy, getting backlinks, etc.</p><p>For someone with my interests, both of these businesses sucked. I don’t like doing any of that crap. Especially when I already make a <a href="https://www.levels.fyi/?compare=Amazon&amp;track=Software%20Engineer">Senior FAANG salary</a> which means my regular day job values my work at hundreds of dollars per hour. With these businesses, I invested 100s of hours doing stuff that was 10x more boring and dreadful than the most boring Amazon meeting, and in the end I generated very little money. Less than a dollar per hour in most cases. Even with the help of AI accelerating everything, it’s still nowhere near worth it to me.</p><p>So here is some concrete advice: if you want to learn digital marketing, or amazon FBA, or software engineering, or whatever else, first decide why exactly you want to learn this. For example, say you want to learn Amazon FBA so that you can be a trend-spotting re-seller of up-and-coming goods. You identify trends in China/Korea and then bring those trends to the US. Armed with this clarity, talk to several people who already do this for a living and get a dispassionate view of the detail of their day-to-day. This is crucial information to decide whether you pivot or persist.</p><h3>What is the role of passion?</h3><p>Passion is what makes the 100s of hours of sub-$1 work worthwhile. If I was passionate about copywriting, marketing, and click funnels, I would have eventually succeeded in one of my ventures.</p><p>When you choose to work on something you are passionate about, something you feverishly believe should exist in the world, that’s when you will be able to stick with it for the long haul in spite of no rewards.</p><p>I love writing and teaching. I have made zero dollars from it. Yet I continue to do it, and will likely continue forever. Writing clarifies thought. Writing in a way that is consumable and useful to others goes beyond clarification of thought, into the realm of discovering paradigms and crystallizing underlying philosophies. Occasionally, I get a tremendous amount of <a href="https://www.linkedin.com/posts/thecodingteacher_an-api-endpoint-should-never-perform-an-unbounded-activity-7268647549333299200-NxP_?utm_source=share&amp;utm_medium=member_desktop">validation </a>from people finding value in the knowledge I share, and that motivates me a lot more, but the truth is a lot of what I write I don’t even publish. I love writing.</p><p>If you are not passionate about click funnels and marketing, you probably do not want to start a regular SaaS app. If you are not passionate about experimenting with messaging to try and hack the human mind, you probably do not want to learn digital marketing. If you are not passionate about feverishly competing for the best prices/niches and the best store listing, you probably do not want to start an amazon FBA. These are all over-simplifications, but hopefully they get the point across. Make sure that the day-to-day of the thing you want to learn is something you’re passionate about.</p><p>Make sure you’re doing it for the journey and not because of some idealized, romanticized outcome.</p><h3>Is X a thing I should bother learning?</h3><p>Now we get into the realm of <em>relative aptitude</em>. Is X a thing you <em>could</em> be good enough at?</p><p>Allow me to illustrate by way of extreme examples. Imagine you’re considering a career in basketball. There are only about a handful of professional leagues in the whole world where you can make a good living. There’s about 500 players in the NBA, so let’s say overall there’s 3000 professional players in the world. The probability of you being in the top 3000 people playing the sport is, for the average person, vanishingly small. This is not a good career move.</p><p>Similarly, let’s say you want to make a living as a YouTube influencer making silly entertainment videos. The probability of success in that is even smaller and perhaps more random (you’re somewhat at the mercy of the algorithm’s whims).</p><p>Or let’s say you want to make a living doing Amazon FBA. If you count the number of people who took an FBA course as the pool of people who tried, the success percentage is also vanishingly small.</p><p>But all of this changes with aptitude and passion. If you’re 6&#39;9 with a 44 inch vertical, then your chances of success at basketball skyrocket. If you’re already a practiced comedian and you routinely kill it at the comedy club, your chances of succeeding on YouTube skyrocket. If you’ve already been running an import/export business for years, your chances of being able to expand it with Amazon FBA are high.</p><p>When evaluating relative aptitude, it is important to have a realistic picture of the skill requirement. People often get this wrong due to the sheer amount of internet grifters that make false promises of making quick money with Amazon FBA or Facebook ads.</p><p>Here’s a thing that I have found to apply more or less universally: the lower the barrier to entry to a field, the higher the competition will be. With high competition comes an <strong>invisible skill barrier</strong> that is hidden from beginners.</p><p>Consider Amazon FBA or posting YouTube videos. Literally anybody can open an Amazon FBA account. Anyone can film a video with their smartphone and upload it to YouTube. That means that in order to stand out, in order to actually succeed at a thing that anyone can try, you have to develop a complex set of skills and strategies that are not immediately apparent to beginners.</p><p>For Amazon FBA sellers, the invisible skill barrier includes:</p><ol><li>Product Research: Identifying profitable niches with manageable competition levels. <em>And doing this faster than most.</em></li><li>Inventory Management: Balancing stock levels to avoid excessive storage fees while maintaining availability.</li><li>Pricing Strategy: Utilizing dynamic pricing tools to stay competitive without sacrificing profitability.</li><li>Listing Optimization: Creating compelling product listings that stand out in search results.</li><li>Competition Analysis: Understanding and outmaneuvering competitors in a crowded marketplace.</li></ol><p>For YouTube creators, the invisible skill barrier includes:</p><ol><li>Content Strategy: Developing a unique voice and niche in a saturated platform. But if it is unique, then by definition there’s nobody to copy, and therefore risky. It’s time consuming to find a tone of voice you can take that actually appeals to people, if it even exists for you.</li><li>Video Production Skills: Creating high-quality content that engages viewers.</li><li>SEO and Keyword Optimization: Ensuring videos are discoverable in YouTube’s search algorithm.</li><li>Audience Engagement: Building a loyal community around your content.</li><li>Consistency and Persistence: Regularly producing content despite initial low viewership.</li></ol><p>Additionally, all fields with a low barrier to entry suffer from the following general difficulties:</p><ol><li>Since the barrier to entry is so low, new ideas and new approaches pop up all the time. Your journey of keeping your skills sharp is lifelong.</li><li>Since you’re failing over and over in the beginning, improving is actually quite difficult. Going from failure to failure often leaves you lost and makes it difficult to learn anything. When I started my print-on-demand t-shirt store, I could drive traffic to it but the traffic never converted. I revamped the website twice and never got a conversion. Going from zero sales, to a new website look, and still having zero sales doesn’t teach you anything. I don’t even know if my hypothesis (that my hastily-made website didn’t inspire trust) was correct. Maybe my t-shirt designs were only good enough for people to click on the ads but not to actually buy.</li></ol><p>All of this is to say that if you hope to turn some low-barrier-to-entry field into your profession, you need to make sure that there’s some aspect of the day-to-day of performing this skill that you actually enjoy. Make sure there’s some aspect you would do even if there was no promise of economic gain. This is important because none of the courses or resources out there will get you all the way to your goal. There are no recipes. There are no instruction manuals. You can only get the very basics from courses and books, and it won’t even take you 5% of the way to your destination. You need to be ready for much trial and error.</p><p>In contrast, fields with a high barrier of entry in terms of technical skills or regulatory requirements (think accounting, law, medicine, etc.) do tend to have a more pre-defined “recipe” to get into them. Be the right age, get into college, get your degree, pass your bar, get the job. Of course, getting a <em>good</em> job and progressing within your career are by no means formulaic or guaranteed. But at least gaining entry into the field is more formulaic. I think this is a path that a tremendous amount of people ignore. A lot of people paying for facebook ads courses should really be spending their time trying to become a licensed electrician, mechanic, plumber, dental assistant, etc. The <a href="https://en.wikipedia.org/wiki/Expected_return">expected return</a> on their time would be much higher if they did.</p><p>So here’s my rule of thumb on this subject: if you’re opting for a path with no recipe (for example content creation, social media marketing, SEO, FBA, etc), <em>you should assume it will take two years of skill-building before you’re able to see real results</em>. Yes, I know such-and-such did it after two months. Statistically, you can assume you’re not him. It’s irrational to go into an endeavor only having considered the best-case scenario.</p><h3>How much academic study should I engage in, and how do I pick resources to study?</h3><p>Ok, so at this point you’ve decided you still want to learn SEO, facebook ads, Amazon FBA, content creation, or some other skill with a super low barrier to entry. Here are my guidelines on how to proceed:</p><h4>Pay nothing for courses.</h4><p>Because people don’t understand that lower barrier to entry means more competition, these fields like content creation or Amazon FBA are often perceived as a quick or easy way to make money. This is, as we have seen, not true, but it doesn’t stop grifters from selling bullshit courses.</p><p><strong>It is of utmost importance that you never ever pay for any of these courses, especially ones worth in excess of $1k.</strong></p><p>NONE of these courses contain information that will directly cause success. Paying $25k for a course is like paying $25k for basketball coaching in the hopes of becoming Lebron James. <strong>You only become successful through creating your own recipe, never from imitating someone else’s.</strong></p><p>Here’s the good news: these fields move a lot of money. Because of this, there’s an abundance of companies selling tools to better compete in these fields. The providers of these tools, as well as the platforms themselves, are incentivized to give you all the generic knowledge you need in order to get started on your learning journey. Some examples:</p><ul><li>To become a YouTube creator, one of the skills you’ll need to master is YouTube keywords/SEO. The tool VidIQ greatly helps with this. Naturally, since they make more money off of content creators actually succeeding, they <em>want</em> you to succeed. They know that nobody will keep paying for VidIQ after the 18th month of their channel sitting on zero views. That’s why the <a href="https://academy.vidiq.com/collection">VidIQ academy </a>exists, and they teach you the basics of YouTube SEO for free. VidIQ also has a pod</li><li>To learn Amazon FBA, Amazon themselves teaches you a bunch of good stuff. Check out the <a href="https://sell.amazon.com/learn">Amazon sellers academy</a>. Of course, with FBA there are also tool providers that help you make more money. One such provider is Helium 10. Helium 10 helps with product research, keyword research, listing optimization, competitor research, etc. They offer a lot of learning resources, for example the <a href="https://academy.helium10.com/">Helium 10 academy</a> and their <a href="https://open.spotify.com/show/31CTCHQGSghLavsf0otYQq">Serious Sellers podcast</a>.</li><li>To learn SEO, you can look at Google’s own marketing courses as well as Ahrefs academy and moz.com.</li><li>And so on and so on.</li></ul><h4>Spend no more than 10 hours taking courses without action.</h4><p>I’m fucking serious here. The main thing that holds people back (especially the kinds of analytical people who ask the kinds of questions that prompted writing this article) is never starting; never taking action.</p><p>You <em>must</em> make a pledge to yourself that you’re not going to spend months consuming content about your dream without taking action. Make that promise to yourself <em>right now</em>.</p><h4>Create the circumstances that force some degree of permanent immersion</h4><p>You need to be ready for the long haul here. You’re going to need to make some adjustments to your life to fit in habits that keep you engaged in your new pursuit. When I was rebuilding my engineering career, I accidentally managed this by taking a teaching job where I had to teach data structures and algorithms. That meant I basically got paid to practice the kind of problems elite employers ask. By the time I interviewed with Amazon the first time, I smoked the interview so bad I had to use psychological tricks like pretending to be nervous in order to not come off cocky.</p><p>So in your quest to learn this new thing, I recommend you find a way to force yourself to bring it into your life. Here are some strategies I have seen that paid off very well:</p><ol><li>Find a way to get paid to learn. This is in effect what I was doing when I took that teaching job. A friend of mine, who wanted to learn to code, offered to make his employer a website, for free. Eventually the website was so good, his employer promoted him to web developer.</li><li>Find a community of people doing the thing you wanna be doing. Find whatever works for you, maybe it’s an IG group, or a linkedin group, or a facebook group, or a discord server, or a web forum, or a subreddit. Find one that isn’t a chore to log into and that actually appeals to you. Stay engaged in it.</li><li>Maintain a <em>slow burn</em> of content consumption. Get through books and courses <em>slowly</em>. Most analytical people (including me) have this stupid, wrong mental habit that finishing courses or books actually counts for something. It fucking doesn’t. If you read ONE chapter in one book and put ONE lesson into practice, that’s worth more than finishing 10 books without taking action on anything.</li></ol><h3>What if I have no passions and no aptitudes?</h3><p>I included this only to specifically state that there is no easy answer. I think people who feel this way probably have some degree of screen addiction or dopamine insensitivity problem which they need to fix first before the path to a good life can reveal itself to them. But take that with a grain of salt.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=41f3af99f1c5" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>