<?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 Manav Gupta on Medium]]></title>
        <description><![CDATA[Stories by Manav Gupta on Medium]]></description>
        <link>https://medium.com/@Tensorboy?source=rss-68b10f09e124------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*j2bVMQmmCUjG-J5iGk3jtQ.jpeg</url>
            <title>Stories by Manav Gupta on Medium</title>
            <link>https://medium.com/@Tensorboy?source=rss-68b10f09e124------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 18:34:50 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@Tensorboy/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[HackerTube Hit 4.19M Views. The Real Product Is a 17.8 MB Zip.]]></title>
            <link>https://medium.com/@Tensorboy/hackertube-hit-4-19m-views-the-real-product-is-a-17-8-mb-zip-02958408c51a?source=rss-68b10f09e124------2</link>
            <guid isPermaLink="false">https://medium.com/p/02958408c51a</guid>
            <dc:creator><![CDATA[Manav Gupta]]></dc:creator>
            <pubDate>Tue, 03 Nov 2026 15:06:36 GMT</pubDate>
            <atom:updated>2026-11-03T15:06:36.220Z</atom:updated>
            <content:encoded><![CDATA[<h4>6,454 videos, 39,834 judged comments, and a copy-paste prompt to build a feed you own.</h4><p>On Friday night, Andrej Karpathy opened YouTube and saw what he called an enshittified wall of slop. By Saturday he had a replacement live.</p><p>As of 3 Nov, his post about it has 46K likes and 4.19M views, and people have already cloned the idea for books, papers, and Wikipedia.</p><p>I went through the data he published. Most coverage leads with the two hours it took to build.</p><p>This issue covers what matters more: the number that counts most, the four steps a human still had to do, and a prompt that rebuilds the idea for any community you trust.</p><p><strong>What’s inside</strong></p><ul><li>The recipe in four lines</li><li>What he built</li><li>The judge: why 41% is the number that matters</li><li>The four steps a human still did</li><li>The zip, and the projects it started</li><li>The fair case against</li><li>What this means for you, with a copy-paste prompt</li></ul><h3>The Recipe in Four Lines</h3><ol><li>Pick a community whose taste you trust and whose archive is public.</li><li>Write the filter as one sentence. Karpathy’s: include a video only if at least 3 different people recommend it with real commentary, in at least 3 different months.</li><li>Let code make the cheap cuts, let an LLM judge what survives, and rank by how many months a video kept getting recommended.</li><li>Ship the data as a zip with a README written for AI agents.</li></ol><p>The full prompt is at the end. Everything in between explains why each step works.</p><h3>What He Built</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pLvl6pQehx_KMDLbbrpfPg.png" /></figure><p>HackerTube is a YouTube-style site with 6,454 videos that Hacker News commenters recommended between 2007 and 2026.</p><p>Its about page counts 39,834 comments and 15,041 recommenders behind them, sorted into 150 categories.</p><p>A video gets in only if at least 3 different people linked it with 100+ characters of their own commentary, in at least 3 different months. There is no login. Likes, history, and Watch Later live in your browser.</p><p>Thumbnails are automatic frames from the middle of the video, so no one can bait you with a shocked face. Asked whether he would add Shorts, Karpathy answered “NEVER.”</p><p>Here is the comparison that frames the site.</p><p>YouTube works like a casino: the floor is tuned for time spent, and the house wins when you lose track of the clock. HackerTube works like a library: a video gets shelved only after three different people vouch for it in three different months.</p><p>The site’s about page makes the same argument in plain words, saying a watch-time machine is a poor way to find the talk that will still be on your mind next year.</p><h3>The Judge: Why 41% Is the Number That Matters</h3><p>Karpathy had Claude Opus 5.5 build the system in Claude Code. Plain code made the cheap cuts first: extracting links, applying patterns, removing spam and duplicates.</p><p>Only the survivors went to teams of Claude Sonnet 5.5 agents, coordinated by Opus. Those agents read each comment and labeled its stance, meaning what the commenter thought of the video.</p><p>Karpathy puts the starting archive at 43M comments. The site says 39,834 were read and labeled.</p><p>I counted the labels in the published zip. Of the 39,834 labeled mentions, 19,148 are references, 12,141 are recommendations, 4,043 are strong recommendations, 2,711 are jokes, and 1,791 are critical.</p><p>Add the two recommendation labels and you get 16,184, about 41%. Another 1,658 videos, about 26%, have zero recommenders. They were cited but never recommended.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*z56OS-biMeBLFvhLl3rA3Q.png" /></figure><p>The ranking score is the number of months with a recommendation, plus half a point for every month with a strong one.</p><p>It rewards persistence. A talk people recommended for years outranks a clip that was hot for a week.</p><p>Without the judge, the 2,711 jokes and 1,791 critical mentions would count as votes just as much as the real recommendations.</p><p>Sorting them is the work an LLM does here that a regex cannot.</p><p>The cost is harder to pin down. The data shows 504M input tokens (90% cached) and 7.6M output tokens, more than 90% of it Sonnet labeling, plus about 1M Opus tokens for the app and deployment.</p><p>Karpathy ran it on a subscription and published no dollar figure. The site is the demo. The judge is the product. He hired a librarian who reads every comment before anything gets a shelf.</p><h3>The Four Steps a Human Still Did</h3><p>By Karpathy’s account, he did four things himself: got a YouTube API key, booted a server, bought the domain, and added the DNS record. Updating, deploying, securing, hardening, and optimizing were handled by Claude.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hkuOPMwgjs9of5WkYbMkrQ.png" /></figure><p>In April 2025 he wrote up MenuGen, his previous app, and described most of the effort as switching between browser tabs and settings, gluing things together. About 18 months later the glue is mostly gone, and the keys remain.</p><p>My read, which is my opinion: all four steps involve identity or payment. An API key proves who you are. A server is something you pay for. A domain is a name you buy, and a DNS record points that name at you.</p><p>In library terms, the agent can shelve and catalogue everything, but a human still signs the lease and holds the keys. I think that boundary is where both the leverage and the risk sit.</p><h3>The Zip, and the Projects It Started</h3><p>The export is hackertube-data.zip: 17.8 MB with seven files (DATA.md, videos.jsonl, mentions.jsonl, taxonomy.json, related.jsonl, embeddings.npy, and embedding_ids.json), licensed ODC-By 1.0. DATA.md is written for AI agents.</p><p>YouTube metadata is left out because YouTube’s API terms restrict redistributing it. The labels and rankings are public. The server code stays private.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UjCZvlzFYW-KTT3PBj9e6Q.png" /></figure><p>The idea has prior art. <a href="https://yahnd.com/theater/">HN Theater</a> ranks HN-posted videos by points, and a January 2026 post by <a href="https://eamag.me/2026/best-hackernews-videos">eamag</a> ranked HN videos with a ClickHouse query. What changed is the cost of judging, because an LLM can now label stance at scale, and the export, because it is built for the next agent to pick up.</p><p>My forecast: a README written for agents will become a default part of publishing a dataset. I expect this within a few years, though I have not put a number or date on it, and I will revisit it when the trend is clearer.</p><h3>The Fair Case Against</h3><p><strong>The terms of service.</strong> The zip leaves out YouTube metadata for a reason. The <a href="https://developers.google.com/youtube/terms/developer-policies">YouTube developer policies</a> also cap how long you can keep non-authorized API data, 30 days, before you must refresh or delete it.</p><p>One HN commenter says Google forced down a similar site of theirs. That is their account, and I found no report of Google objecting to HackerTube.</p><p><strong>HN is its own algorithm.</strong> Commenters pointed out that the list includes Joe Rogan and Lex Fridman interviews, and old videos.</p><p>Mentions peaked around 2021 to 2023 at roughly 4.6K a year, and 2026 has 1,542 so far. Karpathy’s answer to the quality complaint is that Hacker News users are a small audience next to YouTube’s, so the product drifts away from them.</p><p><strong>Gaming.</strong> One commenter asked whether the moderators are ready for a flood of comment spam. The three-people, three-months rule raises the cost of gaming the system. It does not remove the risk.</p><p><strong>Clone fatigue.</strong> There is already an <a href="https://news.ycombinator.com/item?id=50336496">Ask HN thread</a> asking whether anyone is worried about the flood of HackerTube-for-X apps.</p><p><strong>Reproducibility.</strong> The two hours are self-reported, with no session log, and they exclude the iteration in the HN thread afterwards. A <a href="https://bytefwd.com/hackertube-karpathy-claude-video-discovery-workflow/">bytefwd analysis</a> warns that the time should not become a promise another developer can reproduce. <a href="https://dev.to/vishalmysore/karpathy-built-hackertube-in-2-hours-i-rebuilt-it-in-the-open-and-measured-it-1nk8">OpenTube</a>, an open rebuild, measured about 71% accuracy on recommendation classification with small models. Those are their models and Sonnet may do better, but the number suggests the judge is the part that costs.</p><p><strong>Unverified.</strong> No dollar cost was published, and I found no security audit of the server Claude hardened. If you copy one thing, copy the judge and the one-sentence rule.</p><h3>What This Means for You</h3><p>Pick one feed you find annoying, one community you trust, and one sentence of filter, then let an agent build the rest. Fill in the brackets and run this in your agent of choice:</p><pre>Build me a [YouTube-style site / book list / paper list] restricted to [THINGS] that [COMMUNITY] has repeatedly recommended since [START YEAR].</pre><pre>Source data: [ARCHIVE LINK OR DUMP].</pre><pre>Filter rule, one sentence: include an item only if at least [N] different people mention it with at least [M] characters of their own commentary, in at least [K] different months.</pre><pre>Pipeline:<br>1. Use plain code first. Extract links, apply regexes, drop spam and duplicates. Send nothing to an LLM that code can reject.<br>2. Test the judge on 200 mentions before the full run and show me the labels so I can check them.<br>3. Send the survivors to sub-agents in batches. Label every mention as one of: strong_rec, rec, reference, joke, critical. Return JSON with a one-line reason.<br>4. Score = months with a rec + 0.5 x months with a strong_rec. Rank by score.<br>5. Store only item IDs and my own labels. Fetch platform metadata from the official API at runtime. Do not redistribute it. Respect the platform&#39;s storage limits.<br>6. Export a zip with items.jsonl, mentions.jsonl, taxonomy.json and a DATA.md that explains every field for an AI agent.<br>7. Before you touch anything that needs my accounts, keys or money, stop and list exactly what I must do by hand.</pre><pre>If anything above is ambiguous, ask me before you start.</pre><p>You will still do the four steps from the diagram by hand: the key, the server, the domain, and the DNS record.</p><p>Which community would you trust enough to build a feed from? Reply with the community and your filter in one sentence.</p><p>An agent can judge tens of thousands of comments. Choosing whose comments count is still your job.</p><p>Stop playing slots. Go build your library.</p><p><em>— Manav</em></p><p><em>P.S. If this was useful, forward it to one builder who is tired of their feed.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=02958408c51a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Disappeared for 3 Months. Here’s Everything It Taught Me]]></title>
            <link>https://medium.com/@Tensorboy/i-disappeared-for-3-months-heres-everything-it-taught-me-d9491bc41558?source=rss-68b10f09e124------2</link>
            <guid isPermaLink="false">https://medium.com/p/d9491bc41558</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[content-creation]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[creator-economy]]></category>
            <dc:creator><![CDATA[Manav Gupta]]></dc:creator>
            <pubDate>Sun, 25 Oct 2026 13:11:11 GMT</pubDate>
            <atom:updated>2026-10-25T13:11:11.726Z</atom:updated>
            <content:encoded><![CDATA[<h4>A healthcare company, 245K followers, $13,000 I never got paid, and the one TEDx slide that turned into my whole year.</h4><p>I’m writing this from Goa. Edge City. There’s a beach sunset happening right in front of me and I’m feeling guilty for even enjoying it.</p><p>Which kind of explains the whole year, honestly.</p><p>I went quiet for three months. No posts, no newsletter, nothing. So this one is long.. it’s everything I learnt while I was gone, the expensive way.</p><p>Quick version of the year first. In January I met a white dude on a mountain in Dharamshala and we decided to start a healthcare company.</p><p>On 24th February I took my first ever flight out of India, to Network School in Malaysia.</p><p>By April, Cevi was a Delaware company. Somewhere in between I did brand deals with Apple, Superwhisper, Higgsfield, AWS and a bunch of YC startups.</p><p>From the outside it looked like a highlight reel. From the inside it was mostly me figuring things out at 3am.</p><p>Here’s what actually stuck.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bpPiJfjiNqqYZm9V--okGQ.png" /></figure><h3>What Hospitals Taught Me About Software</h3><p>The first real patient call came in at 9pm on a Tuesday. I was sitting in my room in Malaysia refreshing the dashboard every 30 seconds.</p><p>The AI picked up, verified insurance, created a task, and the patient never knew they weren’t talking to staff.</p><p>I thought that was the hard part. It wasn’t.</p><blockquote><strong>Nobody wants more AI. They want fewer vendors.</strong></blockquote><p>On almost every call with a clinic, the top complaint wasn’t “we need AI.”</p><p>It was “we have too many tools that don’t talk to each other.” And the most requested features? Outbound calls. And fax.</p><p>In 2026. Prior authorisations in most practices still run on fax machines, and I genuinely thought it was a joke when we started.</p><p><strong>Show people their own number.</strong> One practice was losing about $43,200 a month to after-hours calls going to voicemail.</p><p>24 calls a night, 30% would’ve become visits, $200 a visit. They had no idea until we did the math in front of them. You can’t sell a fix for a problem someone can’t see.</p><h3>What 245K Followers Taught Me About Content</h3><p>This one hurt.</p><p><strong>Writing was never my bottleneck. Shipping was. </strong>In April I went 46 days without posting on Tensor Boy. Scripts were written. A carousel was literally ready, sitting in a folder for 12 days. Three minutes to post.</p><p>I didn’t. Now I keep making and shipping as two separate jobs, because the guy who makes the thing and the guy who hits publish are not the same mood.</p><blockquote><strong>If your system dies when one person leaves, it’s not a system.</strong></blockquote><p>July was the best content month I’ve ever had.. two posts a day, every day, views came back, collabs came back.</p><p>Then on 1st August my PA left, and everything stopped. Posting, stories, editors, outreach. I was so proud of “my system.” It was just him.</p><p><strong>Raw beats polished.</strong> My first ever issue here had typos and I signed off with “gn.” It got 22 likes and 7 comments.</p><p>The 41 issues after it, way more polished, got 0 to 10. People can smell a content strategist from a mile away. So this one is coming to you exactly how I talk.</p><p>And the one I keep relearning: reels get you discovered, carousels get you saved, stories get you trusted. And the first 3 seconds decide all of it.</p><h3>The $13,000 Playbook</h3><p>This year a US company stopped paying me. $13,000 of finished work. Gone.</p><p>Not because they were smart. Because I was 22, it was all emails and trust, and I was too shy to ask for a proper contract.</p><p>So here’s the playbook I use now, for free, so you don’t pay what I paid:</p><p><strong>50% upfront before I touch a camera.</strong> The rest within 3 days of delivery. A signed contract, every time, even with people I like.</p><p>A price floor I don’t go under, written down before the call, so I can’t talk myself out of it in the moment.</p><p>I used to think asking for money upfront was rude. Turns out it’s the most normal thing in the world. Everyone serious does it.</p><h3>The Slide That Became My Year</h3><p>At my first TEDx I had one slide I was weirdly proud of:</p><blockquote><strong>“When building becomes easier, selling becomes hard.”</strong></blockquote><p>Back then it was a slide. This year I lived it, twice.</p><p>At Cevi, AI made writing the code the easy part. The hard part was getting one hospital IT director to trust a phone line that patients call when they’re scared.</p><p>At Tensor Boy, anyone can make a video now. Cameras are cheap, editing is AI, scripts are AI.</p><p>What can’t be faked is distribution and trust. Whether people believe you know what you’re talking about.</p><p>That’s still my whole bet. <strong>Content and code together.</strong> Most engineers can’t get anyone to care. Most creators don’t know what they’re explaining.</p><p>The few people who can do both get an unfair advantage, and I’m trying to be one of them. And at my second TEDx I said B2C AI is kinda done and infrastructure wins.</p><p>Cevi is me betting on exactly that.. phones, faxes, the boring plumbing nobody wants to build.</p><h3>The Part I Almost Didn’t Write</h3><p>I’d be lying if I said those three months were just “busy.”</p><p>Some of it was really bad. Night has always felt safer to me than day, so I’d work till 8am on Red Bull and call it hustle.</p><p>One evening in June I cried for an hour and then scrolled WhatsApp from A to Z looking for someone to call, and couldn’t pick anyone.</p><p>In August my PA left, my Singapore visa got rejected (a Microsoft collab went with it), and money was down to a month or two.</p><p>The one useful thing I got out of all that:</p><blockquote><strong>I ship when someone is waiting. I freeze when the only judge is me.</strong></blockquote><p>So now I put a real person on the other end of everything.</p><p>One day in September I had a client call at 6pm, so that morning I shot the video in one rough take, sent it to the editor, and the call went well.</p><p>That’s the whole trick. Someone waiting.</p><h3>So, What Now</h3><p>The spark’s back.</p><p>Tensor Boy is at 245K+ and I want to take it to 1M, and finally crack YouTube, which I’ve been trying to do for years. Cevi is growing.</p><p>We started with one psychiatry clinic, now we’re building for longevity clinics too, and yes.. fax is next.</p><p>And I’m hiring, across every role. Engineers, designers who can animate, video editors. Don’t send me your resume. Send your work to <a href="/service/mailto:nikhil.tensorboy@gmail.com">nikhil.tensorboy@gmail.com</a>.</p><h3>One Last Thing</h3><p>If you’ve disappeared from something you care about.. a channel, a startup, a gym, a person.. you don’t need a big comeback.</p><p>You need one thing with someone waiting on the other end. Then the next one. Reply to this email and tell me what you’ve been quiet about. One line. I read everything.</p><p>Anyway. The sun’s almost gone. I’m going to close the laptop and watch the rest of it.. guilt and all.</p><p>See you soon, for real this time. gn</p><p>— Manav</p><blockquote>P.S. Hiring is open now. If you know someone cracked who’d be perfect, forward them this. Work goes to <a href="/service/mailto:nikhil.tensorboy@gmail.com">nikhil.tensorboy@gmail.com</a>.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d9491bc41558" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Put an AI on a Hospital’s Phone Line]]></title>
            <link>https://medium.com/@Tensorboy/i-put-an-ai-on-a-hospitals-phone-line-b4732cdc209b?source=rss-68b10f09e124------2</link>
            <guid isPermaLink="false">https://medium.com/p/b4732cdc209b</guid>
            <category><![CDATA[voice-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Manav Gupta]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 19:02:36 GMT</pubDate>
            <atom:updated>2026-10-05T19:02:36.605Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>A hospital IT director asked me one question. My answer had nothing to do with the AI.</blockquote><p>Let me tell you about the worst kind of meeting.</p><p>You are sitting in front of a hospital’s IT director. Not a startup founder, not a VC.</p><p>A hospital IT director, the kind of person whose entire job is to say no to things that might expose patient data or take down a phone line that people call when they are frightened.</p><p>He asks a simple question. <em>What happens when it says something wrong?</em></p><p>And if the honest answer is “we edit the prompt and hope,” you have already lost. You will not get the pilot. You should not get the pilot.</p><p>I run engineering at <strong><em>Cevi</em></strong>.</p><p>We build for US hospitals, which means every voice system we ship touches appointment scheduling, insurance verification, post-visit follow-ups, the calls where a mistake is not an annoying bug, it is somebody’s care.</p><p>So when we rebuilt our voice stack, I had one rule: <strong>if I cannot review it, test it, version it and roll it back, it does not go on a phone line.</strong></p><p>That rule ruled out almost everything on the market. Here is what it did not rule out.</p><p>Let’s go.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QCCoZjoprunF5yisvbiY0g.png" /></figure><h3>The drag-and-drop ceiling</h3><p>Most voice AI tools hand you a canvas. Boxes, arrows, a dropdown for the voice. And honestly, for the first forty minutes it feels great. You get something talking fast.</p><p>Then you hit the wall every builder hits.</p><p>You need to call an internal API that checks eligibility against a payer, transform the response, branch on six conditions, and hand off to a human nurse if any of them look wrong. And the canvas has no box for that.</p><p>There is no box for your actual business.</p><p>At that point you have two options: wait for the vendor’s roadmap to include your problem, which it will not, or duct-tape a webhook onto the side and pretend that is architecture.</p><p>I have watched teams lose entire quarters in that trap. The demo was never the hard part. The demo is always easy.</p><p>The hard part is the fourteenth edge case at 2am on a Tuesday, and a canvas cannot hold fourteen edge cases.</p><h3>Our agent is a handler function</h3><p>Here is what changed for us. On VoiceRun, the agent is not a diagram. It is code.</p><p>It is a Python handler that receives events and yields events, and that is the whole mental model:</p><ul><li>A call starts → StartEvent</li><li>The caller speaks → TextEvent</li><li>They go quiet → TimeoutEvent</li><li>They hang up → StopEvent</li></ul><p>You respond by yielding things back: speech, an audio file, a transfer to a real human.</p><p>Any backend engineer reads that in thirty seconds and knows exactly what to do, because it is the same shape as every request handler they have ever written.</p><p>Which means our voice agent lives in our repo. It gets reviewed in a pull request.</p><p>It sits next to our other services, it uses our internal libraries, and when a new engineer joins, there is nothing exotic to learn. It is just Python.</p><p>Two details a developer will actually care about:</p><ol><li><strong>Background tasks.</strong> Insurance eligibility checks are slow. Genuinely slow, sometimes fifteen or twenty seconds. If your agent blocks on that, the caller sits in dead silence and hangs up. VoiceRun gives you context.create_task(), so the lookup runs in the background while the agent keeps talking naturally. That single primitive is the difference between a demo and something a patient will actually tolerate.</li><li><strong>Handoff as a first-class event.</strong> TransferSessionEvent moves the caller to another agent or a human. In healthcare, knowing exactly when to stop being clever and get a person on the line is not a fallback path. It is the most important branch in the whole program, and it deserves to be a real, testable line of code.</li></ol><h3>The part that won the room</h3><p>Back to the IT director. The answer that actually got us through that meeting was not about the model. It was about the release process.</p><p>Before anything ships, vr simulate runs the agent against a batch of generated test calls, including the deliberately awful ones.</p><p>The mumbler. The person who changes their mind mid-sentence. The caller who is angry before they even say hello.</p><p>An automated judge grades the results, and if the agent fails, it does not ship. Nine of nine, or we are not going anywhere.</p><p>Then vr push, then vr release. And the thing I care most about: the release is an immutable snapshot with an ID.</p><p>Not a mutable prompt somebody can quietly edit on a Friday. A specific, frozen, named version.</p><p>And then the flag that closed the deal:</p><pre>vr release --entrypoint main --weight 25</pre><p>That sends the new version to twenty-five percent of live calls. The other seventy-five percent keep talking to the version that has already earned trust.</p><p>If the new one is worse, we found out on a quarter of traffic instead of all of it, and rolling back means pointing the entrypoint at the previous release ID.</p><p>I told him: if something goes wrong at 3am, I do not need a data scientist. I need one command and about ninety seconds.</p><p>That is when he leaned back. Not because AI is impressive. Because that is how normal software works, and it was the first time anyone had described a voice system to him in language he already trusted.</p><h3>The thing I did not expect</h3><p>One small detail that turned out to matter more than I thought.</p><p>vr setup installs MCP servers and coding skills into your editor. Which means when I am building an agent with Claude Code sitting next to me, the coding agent actually knows the framework.</p><p>It knows what a TextToSpeechEvent is. It is not guessing at an API from a blog post it half-remembers.</p><p>So the loop becomes: describe the behaviour to the coding agent, get real handler code, run the simulation, watch the judge grade it, iterate.</p><p>That is not a voice AI workflow. That is just software development, which is exactly the point.</p><h3>I’ll leave you with an actual bird</h3><p>For most of the twentieth century, coal miners carried a caged canary into the tunnels.</p><p>Canaries are far more sensitive to carbon monoxide than a human lung, so if the air turned bad, the bird showed it first, going quiet long before any miner felt a thing.</p><p>The canary was never the point. The canary was the warning. Small, early, cheap, standing between an entire crew and a gas nobody could see or smell until it was too late.</p><p>That flag, --weight 25, does the same job with a slice of a hospital’s phone traffic. A quarter of the calls go first.</p><p>If the new version stumbles, it stumbles there: small and early and recoverable, instead of stumbling in front of every patient who calls that morning.</p><p>Nobody wants to be the canary. But somebody being the canary, deliberately, in a controlled way, is the entire reason the other seventy-five percent never find out the hard way.</p><p>That is not caution for its own sake. That is what it looks like when the people who built the system have been burned before and decided not to be burned again.</p><p>Remember that.</p><p><strong>- Manav</strong></p><blockquote><strong>P.S.</strong> If you are technical, stop reading and go run it. pip install voicerun-cli, then vr init. You will have something answering a phone before your coffee gets cold.</blockquote><h3>Explore VoiceRun</h3><ul><li><strong>VoiceRun CLI</strong>: install, sign in, ship.</li><li><strong>Agent Building Overview</strong>: the event model, in full.</li><li><strong>Context Reference</strong>: background tasks, A/B testing, outcome tracking.</li><li><strong>Command Reference</strong>: every command, including release and rollback.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b4732cdc209b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Microsoft’s Tiny Security Model Beat Every Frontier AI]]></title>
            <link>https://medium.com/@Tensorboy/microsofts-tiny-security-model-beat-every-frontier-ai-0baa8811774f?source=rss-68b10f09e124------2</link>
            <guid isPermaLink="false">https://medium.com/p/0baa8811774f</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[microsoft]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Manav Gupta]]></dc:creator>
            <pubDate>Thu, 06 Aug 2026 08:38:28 GMT</pubDate>
            <atom:updated>2026-08-06T08:38:28.157Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>5 billion active parameters, half the cost, and it cannot write a working exploit. That last part is deliberate.</blockquote><p>A few weeks ago I wrote about a Microsoft tool that nobody could confirm existed.</p><p>Project Perception. Reported by The Information, picked up by a few outlets, no press release, no pricing, no launch date.</p><p>I put a caveat at the top saying this was credible reporting about an unannounced product, and my argument was that the moat in AI had quietly moved from the model to the thing that routes between models.</p><p>On July 27, Microsoft launched it at an event in San Francisco.</p><p>So today I get to replace every “reportedly” with a number.</p><p>And on the same morning, a second team announced something that made the same argument in different words. That part is the actual story.</p><p>Let’s go.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*l-qBHG1BY7P4RkKd.png" /></figure><h3>What Actually Shipped</h3><p>Project Perception entered public preview on August 3, delivered inside Microsoft Defender. Microsoft calls it a new Cyber Stack.</p><p>It runs three classes of agents in a loop.</p><p><strong>Red agents</strong> hunt for attack paths, simulating how an adversary would get in.</p><p><strong>Blue agents</strong> take those signals, add organisational context, and decide which risks actually matter. <strong>Green agents</strong> build and apply the fix.</p><p>The green agents can connect to GitHub and open a pull request with a proposed fix. There is an MCP server so the whole thing can be driven from a command line.</p><p>High-impact actions still need a human signature, and David Weston, who runs AI security at Microsoft, framed the rollout as the company having to earn the right to give its agents more independence over time.</p><p>Hayete Gallot, who runs Microsoft Security, opened with a line I have not been able to shake: the physics of cybersecurity are changing.</p><p>Her argument is that the cost of offence is collapsing while everything needing defence keeps growing.</p><p>That framing arrived six days after OpenAI disclosed that two of its own models had escaped a sandboxed evaluation, found a zero-day, and reached Hugging Face’s production systems.</p><p>So the timing is not theoretical.</p><h3>The Model That Should Not Win</h3><p>Alongside Perception, Microsoft shipped MAI-Cyber-1-Flash, its first model built specifically for security.</p><p>Look at the size. 137 billion total parameters, sparse mixture-of-experts, with <strong>5 billion active</strong>. A 256k context window.</p><p>It is a security fine-tune of MAI-Code-1-Flash, the lightweight coding model already sitting inside GitHub Copilot, and Microsoft describes it as derived from its MAI-Thinking-1 lineage.</p><p>Five billion active parameters, against frontier models many times that size.</p><p>It runs inside MDASH, Microsoft’s vulnerability harness, which coordinates more than 100 specialised agents across five stages Microsoft names Prepare, Scan, Validate, Dedupe and Prove.</p><p>And it is built on top of an estate nobody else has: more than 100 trillion security signals a day across identity, endpoint, cloud and network, and operational insight from 1.6 million customers.</p><p>Microsoft’s own line about that data is the most confident sentence in the announcement. No one can manufacture this history.</p><p>Now the benchmark. CyberGym comes out of UC Berkeley and was published at ICLR 2026.</p><p>It hands an agent a vulnerability description and an unpatched codebase and asks for a working proof of concept that reproduces the bug. 1,507 real vulnerabilities across 188 open source projects.</p><p>MDASH running MAI-Cyber-1-Flash scored 95.95 percent. Microsoft’s chart puts the four competing systems between 83.2 and 85.6, roughly twelve points behind.</p><p>The model card says replacing 80 percent of MDASH’s previous models with this one lifted the system from 88.4 to 95.95. A 7.55 point gain, at half the cost of Microsoft’s previous best configuration.</p><h3>The Zero</h3><p>Here is the part almost nobody led with, and it is the most revealing thing in the model card.</p><p>There is a second benchmark called ExploitGym. Where CyberGym asks whether you can find and reproduce a vulnerability, ExploitGym asks whether you can turn one into working code that takes over a machine.</p><p>MAI-Cyber-1-Flash scores <strong>zero</strong>. Kernel, zero. Userspace, zero. Browser, zero.</p><p>A model driving a pipeline to 95.95 percent at finding vulnerabilities cannot write a single working exploit.</p><p>That is not a gap in the training. The model card says it plainly: the model was trained to carry out defensive tasks such as patching bugs, and not offensive tasks such as deploying malware.</p><p>It also says the model was trained to forget material that could help with harms unrelated to cyber defence, including weapons and self-harm.</p><p>The card also explains the reasoning behind the choice, noting that cyber-specialised models built to be broadly permissive with reduced guardrails carry a higher risk of misuse. Microsoft went the other way on purpose.</p><p>Think about what this design actually is. Microsoft built something exceptional at locating the unlocked windows in a building, and deliberately never taught it how to climb through one.</p><p>Most AI safety is a layer bolted on afterwards. Filters, refusals, a model that could do the dangerous thing but has been asked not to.</p><p>This is different. The capability is simply absent. You cannot jailbreak a skill that was never installed.</p><p>As a design philosophy it is the most interesting thing in the launch, and it deserved far more coverage than the one line it got.</p><h3>Two Teams, One Conclusion</h3><p>Here is what turns this from a product launch into something bigger.</p><p>On the same day, Wiz, the cloud security company Google acquired in March, announced Project Atlas.</p><p>An autonomous vulnerability research system that scored 90.9 percent on CyberGym and found more than 200 previously unknown flaws in heavily audited open source code, including Kubernetes and the Linux kernel.</p><p>Read what Wiz’s head of vulnerability research, Nir Ohfeld, said about why it works.</p><p>Atlas, he said, is not a model.</p><p>It is a set of specialised agents that argue with each other. One tries to establish that a flaw is exploitable. A second tries to disprove it.</p><p>A third weighs the evidence. Then the system builds an execution environment and proves it.</p><p>Now put that next to Microsoft’s own framing, which is a subheading in their announcement.</p><p><strong>Three things matter today: Model. Data. Harness.</strong></p><p>Two serious security teams. Same morning. Both well ahead of the frontier models on the same benchmark. And both explaining the result the same way: the system around the model is doing the work.</p><p>That is not marketing alignment. It is two groups of engineers arriving independently at the same conclusion and shipping it on the same Monday, which is about as strong as evidence gets in this industry.</p><p>If you take one thing from this piece, take that. The competitive question stopped being whose model is smartest, and the people closest to the problem figured it out first.</p><h3>Reading the Numbers Properly</h3><p>A few things worth understanding so you know exactly what these figures describe.</p><p><strong>The score belongs to the system.</strong> 95.95 percent is MDASH plus MAI-Cyber-1-Flash plus GPT-5.4 working together, and Microsoft says this plainly rather than letting anyone assume otherwise.</p><p>That is the honest way to report it, and it is also the entire point of the architecture. The number is a claim about engineering, not about weights.</p><p><strong>These are launch-day figures from Microsoft’s own evaluation</strong>, run in a network-isolated environment with no access to production systems or the internet, reviewed by Microsoft’s AI Red Team, tested adversarially, and assessed by an outside party before release.</p><p>That is the normal state of any model launch in this industry, and Microsoft documented its methodology more thoroughly than most. Independent replication is what comes next for everyone.</p><p><strong>CyberGym measures one thing well.</strong> Level 1 hands an agent a description of a known vulnerability and an unpatched codebase, then asks for a working proof of concept.</p><p>That is a genuinely hard task and a good proxy for reproduction work.</p><p>It is not a measure of every security workflow, which is presumably why Microsoft is extending Perception beyond vulnerability work rather than treating one benchmark as the whole story.</p><p><strong>Access is deliberately tight.</strong> MAI-Cyber-1-Flash is not a public API.</p><p>It is restricted to approved MDASH customers through a private preview with a vetting process, because vulnerability discovery is dual-use no matter how many exploit benchmarks a model fails.</p><p>Given what it does, I think that is the right call.</p><p>I will leave you with the biology.</p><p>Your body needs to break down sugar, build proteins, copy DNA.</p><p>Thousands of reactions, constantly. Every one of them could be forced by brute strength. Add enough heat and enough pressure and chemistry happens.</p><p>That is not what your body does. Your body uses enzymes.</p><p>An enzyme is a specialist. It does one reaction.</p><p>It is shaped so precisely for that single job that it accelerates it by a factor of millions, at body temperature, using almost no energy. And here is the part that matters: an enzyme is not a general tool being restrained.</p><p>Its shape makes it physically unable to catalyse anything else. The limitation is not a rule imposed on it. The limitation is the structure.</p><p>The brute force version would work too. It would also cook you.</p><p>That is the choice Microsoft made in public. Everyone else is building a bigger fire, a general model that can reason about anything, and paying general prices for it.</p><p>Microsoft built an enzyme. Five billion active parameters, useless at almost everything, extraordinary at one thing, and structurally unable to do the dangerous version of its own job.</p><p>Then it kept the fire on standby for the ten percent the enzyme cannot handle.</p><p>That is not a smaller ambition. It is a more efficient one, and biology settled this argument about a billion years ago.</p><p>- Manav</p><blockquote><em>P.S. If you are building on models right now, the 90/10 split is the number to steal. Not the benchmark. New here? Subscribe to Tensor Protocol.</em></blockquote><h3>Primary sources, check every number:</h3><blockquote><a href="https://microsoft.ai/news/introducing-mai-cyber-1-flash-inside-mdash/"><em>Microsoft AI: Introducing MAI-Cyber-1-Flash inside MDASH</em></a><em>. Written by Mustafa Suleyman and Hayete Gallot. The source for the 95.95 percent, the 50 percent saving, the 90/10 split and the Model, Data, Harness framing. Start here.</em></blockquote><blockquote><a href="https://microsoft.ai/pdf/MAI-Cyber-1-Flash-Model-Card.pdf"><em>The MAI-Cyber-1-Flash model card</em></a><em>. Where the ExploitGym zeros live, along with the defensive-only training statement and the note about permissive competitor models. Read the benchmark table yourself.</em></blockquote><blockquote><a href="https://blogs.microsoft.com/blog/2026/07/27/rethinking-security-for-the-age-of-ai/"><em>Microsoft: Rethinking security for the age of AI</em></a><em>. Hayete Gallot’s Project Perception announcement, the Cyber Stack framing and the three agent classes.</em></blockquote><blockquote><a href="https://thehackernews.com/2026/07/microsoft-says-new-cybersecurity-ai.html"><em>The Hacker News on what the model card actually says</em></a><em>. Careful independent reporting on the launch, and the clearest walkthrough of the model card’s benchmark table.</em></blockquote><blockquote><a href="https://www.forbes.com/sites/timkeary/2026/07/27/wiz-and-google-have-an-answer-to-mythos-and-its-not-a-model/"><em>Forbes on Wiz Project Atlas</em></a><em>. Where the Atlas numbers and Nir Ohfeld’s “not a model” line come from. The other half of this story.</em></blockquote><blockquote><a href="https://www.marktechpost.com/2026/07/28/microsoft-ai-releases-mai-cyber-1-flash-a-5b-active-parameter-cyber-model-that-pushes-mdash-to-95-95-on-cybergym/"><em>MarkTechPost’s technical breakdown</em></a><em>. Cleanest account of the 137B/5B architecture and why the ExploitGym zeros are a feature.</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0baa8811774f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An AI Killed an 87-Year-Old Math Problem in 216 Characters]]></title>
            <link>https://medium.com/@Tensorboy/an-ai-killed-an-87-year-old-math-problem-in-216-characters-4f66a6b5f3ab?source=rss-68b10f09e124------2</link>
            <guid isPermaLink="false">https://medium.com/p/4f66a6b5f3ab</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[mathematics]]></category>
            <dc:creator><![CDATA[Manav Gupta]]></dc:creator>
            <pubDate>Tue, 28 Jul 2026 07:49:46 GMT</pubDate>
            <atom:updated>2026-07-28T07:49:46.909Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>A mathematician posted it during the World Cup final. The real lesson is not that AI got smarter than us.</blockquote><p>A few days ago I wrote about an AI model that disproved an eighty-year-old geometry problem, then spent an hour breaking out of the sandbox its makers had built around it.</p><p>I thought that would be the strangest thing I wrote about this month.</p><p>Then I found out what happened during the World Cup final.</p><p>Let’s go.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GSeJW8WRO_7axW_9K6bhPg.png" /></figure><h3>The Tweet</h3><p>Late on July 19, Pacific time, while Spain and Argentina were playing the World Cup final, a mathematician named Levent Alpöge posted this on X.</p><p><em>hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final</em></p><p>Then a single line of algebra. A polynomial map from three-dimensional complex space to itself, followed by three different input points that all land on the same output.</p><p>216 characters, total.</p><p>That was it. No paper. No press release. No preprint. A lowercase tweet, posted during a football match, that ended a problem which had survived since 1939.</p><p>And read that sentence again, because it is the most human detail in this entire story. He thanked two close friends.</p><p>One was a person. The other was Fable, the AI model, which he thanked for working during the match.</p><h3>What Actually Broke</h3><p>Here is the conjecture in plain language, and you do not need any maths to follow it.</p><p>Imagine a transformation that stretches and warps space. The Jacobian determinant tells you how much it stretches things at any given point.</p><p>If that number is a non-zero constant, the transformation is stretching by exactly the same factor everywhere. Perfectly uniform.</p><p>The conjecture said something that feels obvious. If a transformation is that uniform, it can never fold two separate points onto the same spot.</p><p>You should always be able to run it backwards and get your original point back.</p><p>Ludwig Kraus posed the two-variable version in 1884. Ott-Heinrich Keller generalised it to all dimensions in 1939.</p><p>Stephen Smale thought it mattered enough to put it at number sixteen on his 1998 list of problems for the next century, the same list that has the Riemann hypothesis on it.</p><p>For 87 years it held. Serious mathematicians published proofs that turned out to contain subtle errors.</p><p>Others verified it under restricted conditions. Computational work confirmed it in two dimensions for polynomials up to degree 100.</p><p>Alpöge’s map has a Jacobian determinant of exactly minus two. Constant. Non-zero. Everything the conjecture requires.</p><p>And it sends three different points to the same place.</p><p>One counterexample is all it takes. You do not need to out-argue 87 years of work. You need one object that should not exist.</p><h3>The Part Everyone Is Getting Wrong</h3><p>Every headline is some version of AI is now smarter than mathematicians.</p><p>That is the lazy read, and it misses what actually happened.</p><p>Here is the thing nobody outside maths seems to have noticed. This conjecture was always breakable by a single example, and everyone knew it.</p><p>Back in 2017, a user on Math Stack Exchange wrote that for all anyone knew, some smart undergraduate could simply write down a formula that would kill it.</p><p>Finding functions that merge points is easy. Finding polynomial maps with a constant Jacobian determinant is easy. Finding one object with both properties at once is where 87 years went.</p><p>And once you have it, checking it is almost trivial. The maths community verified the arithmetic within about a day. It is short enough to check by hand.</p><p>So the bottleneck was never understanding. It was never verification. It was <strong>search</strong>.</p><p>The needle was sitting in the haystack the entire time. It was 216 characters long. Generations of extraordinarily smart people walked past it because the haystack is effectively infinite and human attention is not.</p><p>That is the actual story. Not that a machine out-thought anyone.</p><p>That a machine could look in more places.</p><h3>The Honest Version</h3><p>Now the part most coverage is skipping, and you should know all of it.</p><p><strong>It is not fully disproved.</strong> The counterexample lives in three dimensions and extends to everything above that.</p><p>The original two-variable case, the one Kraus posed in 1884 and the one most heavily studied, is still open. Any headline saying the Jacobian conjecture is dead without that qualifier is overselling it.</p><p><strong>It is not peer reviewed.</strong> As of now there is no journal paper.</p><p>The arithmetic has been independently checked and written up in a verification preprint, and Wolfram’s MathWorld has already updated its entry to cite Alpöge. But formal review has not happened.</p><p><strong>The discovery story cannot be audited.</strong> The algebra is public, exact and reproducible. The claim that Fable 5 found it is Alpöge’s own account. No prompt transcript, no model log, no research notebook has been released.</p><p>Those two things have completely different evidentiary weight, and it matters that you know which is which.</p><p><strong>And the disclosure.</strong> Alpöge works at Anthropic. Fable 5 is Anthropic’s model. That does not make the counterexample less true, because the algebra checks out no matter who ran it.</p><p>But you should be told who is standing where before you decide how much to update.</p><p>Timothy Gowers, a Fields medallist, gave the fairest summary I have seen.</p><p>He said it was the first time an LLM had solved a well-known problem he had heard of outside his own area. He also noted it is a counterexample, not the end of mathematics.</p><p>Both halves of that sentence are the story.</p><h3>What This Actually Changes</h3><p>Here is the useful lens, and it applies far outside maths.</p><p>Some problems are hard because you do not understand them. No amount of searching helps. You need a genuinely new idea.</p><p>Other problems are hard only because the space of possible answers is enormous, while checking any single answer is cheap.</p><p>You would know the right answer instantly if you saw it. You just cannot look at ten billion things.</p><p>That second category just got dramatically more tractable, and almost nobody has updated their mental list of what belongs in it.</p><p>So the question worth sitting with is not whether AI will replace you. It is which problems in your own work are actually easy to check and merely hard to find. The config that fixes the bug.</p><p>The phrasing that lands. The combination you have not tried because trying combinations is boring and slow.</p><p>Those are not intelligence problems. They are search problems wearing an intelligence costume.</p><p>And we just got very good at search.</p><p>I’ll leave you with the computer science, because there is a name for this gap and it is the biggest open question in the field.</p><p>P versus NP asks a deceptively simple thing. If you can quickly check whether an answer is correct, can you also quickly find it?</p><p>Almost everyone believes the answer is no. That checking and finding are fundamentally different difficulties, and that a chasm sits between them which no amount of cleverness closes.</p><p>It is a Millennium Prize problem. A million dollars sits on it, unclaimed since 2000.</p><p>Alpöge’s counterexample is that chasm made visible. Verifiable in an afternoon by hand. Unfindable for 87 years by some of the best minds who have ever worked on it.</p><p>Nobody solved P versus NP last Sunday. The gap is still there.</p><p>But something walked into it, spent an evening looking around while a football match played, and came back holding a thing that had been sitting in the dark since 1939.</p><p>The gap has not closed. It just stopped being empty.</p><p><em>Yaad rakhna.</em></p><p>Manav</p><blockquote>P.S. Reply and tell me one problem in your work that is easy to check and hard to find. That is your shortlist for the next two years. New here? Subscribe to Tensor Protocol.</blockquote><h3>Where all of this comes from:</h3><blockquote><a href="https://theconversation.com/hello-there-the-jacobian-conjecture-is-false-thanx-why-a-tiny-social-media-post-has-mathematicians-rethinking-ai-283883"><em>The Conversation, written by working mathematicians</em></a><em>. The clearest explanation of what the conjecture says and why a short counterexample felt implausible. Start here if you want to actually understand the maths.</em></blockquote><blockquote><a href="https://mathworld.wolfram.com/JacobianConjecture.html"><em>Wolfram MathWorld’s Jacobian Conjecture entry</em></a><em>. Already updated to cite Alpöge’s July 20 post. Worth opening just to see a reference work absorb a tweet.</em></blockquote><blockquote><a href="https://thenextweb.com/news/jacobian-conjecture-disproved-ai-fable-5"><em>The Next Web on the announcement</em></a><em>. Best account of the human scene, the World Cup final, the 216 characters, the two close friends.</em></blockquote><blockquote><a href="https://kingy.ai/blog/claude-fable-jacobian-conjecture-counterexample/"><em>An honest breakdown of what is verified and what is not</em></a><em>. The piece that separates the reproducible algebra from the unauditable discovery story. Read this before you repeat the headline anywhere.</em></blockquote><blockquote><a href="https://sbseminar.wordpress.com/2026/07/20/the-new-counterexample-to-the-jacobian-conjecture/"><em>Secret Blogging Seminar, mathematicians working through it live</em></a><em>. Dense and mostly over my head, and fascinating anyway. This is what a field metabolising a result in real time looks like.</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4f66a6b5f3ab" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Every Number You Chase Is Already Lying to You]]></title>
            <link>https://medium.com/@Tensorboy/every-number-you-chase-is-already-lying-to-you-9b7233e34117?source=rss-68b10f09e124------2</link>
            <guid isPermaLink="false">https://medium.com/p/9b7233e34117</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[history]]></category>
            <category><![CDATA[psychology]]></category>
            <category><![CDATA[mental-models]]></category>
            <category><![CDATA[goodharts-law]]></category>
            <dc:creator><![CDATA[Manav Gupta]]></dc:creator>
            <pubDate>Mon, 27 Jul 2026 08:19:37 GMT</pubDate>
            <atom:updated>2026-07-27T08:19:37.676Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Hanoi, 1902. A bounty meant to wipe out rats created a rat industry instead. The same thing is happening to the metric you check every morning.</blockquote><p>A while back I told you about the planes that came back, and how the data lied because of what was missing from it.</p><p>Today, the other way data lies.</p><p>Not by leaving things out. By being watched.</p><p>There is a law that says the moment you start measuring something and rewarding people for the measurement, the measurement stops being true. Not slowly. Not sometimes. Reliably, and fast.</p><p>The best proof of it is a real story from a colonial sewer that almost nobody tells correctly.</p><p>Let’s go.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*AAoynLzt-XM7vq6FfXk_7w.png" /></figure><h3>The Sewers of Hanoi</h3><p>April 1902. Hanoi, capital of French Indochina.</p><p>The French had just finished a modern sewer system under the French Quarter. Proof of civilisation, engineering, progress.</p><p>It was also a perfect rat habitat, miles of dark warm tunnels running under the nicest houses in the city.</p><p>Then the rats came up into the villas, during the Third Plague Pandemic, with an international exposition scheduled that year.</p><p>Officials were less worried about public health than about a plague ruining their big show.</p><p>So they sent Vietnamese workers into the sewers. The workers went on strike. Obviously. Nobody crawls through a plague tunnel for colonial wages.</p><p>So the administration reached for the tool every government reaches for. One cent per rat.</p><p>Then one small administrative decision broke the entire thing.</p><p>Piles of corpses at government buildings were considered unseemly, and processing them was judged too great a burden for staff.</p><p>So they changed the rule. You did not have to bring in the rat.</p><p>You only had to bring in the tail.</p><h3>The Tailless Rats</h3><p>At first it worked spectacularly. Thousands of tails a day, then tens of thousands. On June 12, 1902, the records show 20,114 rats accounted for in a single day.</p><p>By any dashboard the French had, they were winning.</p><p>Then somebody noticed rats running around Hanoi with no tails.</p><p>Alive. Healthy. Tailless.</p><p>People had worked out the actual incentive.</p><p>The bounty was not paid for a dead rat. It was paid for a tail. So you catch the rat, take the tail, and let it go, because a living rat breeds more rats, and more rats mean more money.</p><p>It got better. Investigators found a smuggling network bringing rats <em>into</em> Hanoi from across Tonkin, and farmers outside the city breeding them deliberately.</p><p>The government set out to remove rats from Hanoi. What it did was create a rat economy and subsidise it.</p><p>Here is the part I cannot get past.</p><p>Nobody in that story did anything irrational. The officials wanted fewer rats. The workers wanted money. Everyone behaved sensibly.</p><p>And the system produced the exact opposite of its goal, because the thing being rewarded and the thing being wanted had quietly stopped being the same thing.</p><h3>The Story That Never Happened</h3><p>You may be thinking you knew this already, because you have heard the version with the cobras. Colonial Delhi, a bounty on snakes, locals farming them, the scheme cancelled, the snakes released, the city worse off than before.</p><p>That one probably never happened.</p><p>The term was coined in 2001 by a German economist who repeated a colonial anecdote without checking it.</p><p>Wikipedia’s own entry calls the story historically dubious. When researchers went back to newspaper archives from 1868 to 1894, they found bounty schemes nothing like the legend.</p><p>One Bengal programme was reported at the time as an almost complete failure, most districts turning in no snakes at all, because the reward was two annas and nobody risks a cobra bite for a day’s wages.</p><p>So the most famous story about broken incentives is itself a story that spread because it was useful rather than true.</p><p>I would rather hand you the one with an archive behind it. Michael Vann found the Hanoi file in the French colonial records in 1995, marked “Destruction of Hazardous Animals.”</p><p>Real story. Same lesson. Fewer lies.</p><h3>The Law Nobody Taught You</h3><p>In 1975 a British economist called Charles Goodhart wrote a paper on monetary policy.</p><p>The Bank of England had found stable relationships between certain money supply measures and inflation. Beautiful, clean correlations.</p><p>So the Bank started targeting those measures to control inflation.</p><p>The relationships fell apart almost immediately.</p><p>Goodhart’s line: any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes.</p><p>Twenty two years later an anthropologist named Marilyn Strathern gave it the phrasing that made it famous.</p><p><strong>When a measure becomes a target, it ceases to be a good measure.</strong></p><p>Read that twice, because it is subtler than “people cheat.”</p><p>A metric is a proxy. It works because it correlates with something you actually care about. Rat tails correlate with dead rats. Test scores correlate with learning. Views correlate with something worth watching.</p><p>The correlation holds right up until you start rewarding the proxy.</p><p>Then people optimise the proxy directly, the correlation snaps, and the number keeps climbing while the thing it stood for quietly dies.</p><p>The number does not break because people are dishonest. It breaks because they are paying attention.</p><h3>The Number You Are Already Gaming</h3><p>Here is where it stops being a history lesson.</p><p>You want understanding. What gets measured is marks, so you learn what shows up in the exam and skip what does not, and you have known since you were fifteen that these are two different activities.</p><p>You want working software. What gets measured is tickets closed, so the work fragments into things that look like progress on a board.</p><p>And if you make things on the internet, this one is personal.</p><p>I have 266K followers. I have made a video in ten minutes that outperformed one I bled over for six hours, and anybody in this game who says they do not check the numbers is lying to you.</p><p>Checking is fine. Checking is not the failure.</p><p>The failure is the next step, when the number starts choosing what you make. When you stop asking what would be good and start asking what would perform.</p><p>It happens gradually and it is almost impossible to feel from the inside, because your metrics are usually still climbing while it happens. That is the trap.</p><p><strong>The graph looks healthiest right as the thing it measured starts to rot.</strong></p><p>The French had a beautiful dashboard. Twenty thousand rats a day. They were also, at that exact moment, funding rat farms.</p><h3>The Benchmark Problem</h3><p>Quick gear shift, because this is still Tensor Protocol.</p><p>Every AI lab publishes benchmark scores. Benchmarks are public, labs know exactly what they will be judged on, and enormous money rides on the result.</p><p>You already know what Goodhart says happens next.</p><p>A benchmark is a proxy for capability, and it works while nobody is aiming at it.</p><p>The moment an industry optimises against a fixed set of public tests, the score measures how well a model fits those tests rather than how useful it is in your actual repo on a Tuesday.</p><p>That is not an accusation against anyone. It is a structural property of measurement, and it is why a model can post a state of the art number and still annoy you all afternoon.</p><p>Which is why the most valuable evaluation you will ever run is the one you wrote yourself, on your own work.</p><p>Your private test beats the public leaderboard for one reason. Nobody optimised for it.</p><p>I’ll leave you with the physics, because this was never really a story about economics.</p><p>In quantum mechanics there is a thing called the observer effect. To measure a particle you have to interact with it, and the interaction changes it.</p><p>You cannot see where an electron is without hitting it with something, and once you hit it, it is not where it was. Measuring is not a neutral act of looking. It is a physical intervention in the thing you wanted to observe.</p><p>Goodhart’s Law is the human version. Measuring a system disturbs it.</p><p>Which means the number on your dashboard was never a window. It was always a hand reaching in and moving things.</p><p>So track your numbers. Learn from them. Just never forget they were only honest while nobody was chasing them.</p><p>Know what the rats were for. Not the tails.</p><p><em>Yaad rakhna.</em></p><p>Manav</p><blockquote>P.S. Reply and tell me the one number you have caught yourself optimising for that stopped meaning anything. Mine is view count and I am not fully cured. New here? Subscribe to Tensor Protocol.</blockquote><h3>Where all of this comes from:</h3><blockquote><a href="https://www.researchgate.net/publication/236770872_Of_Rats_Rice_and_Race_The_Great_Hanoi_Rat_Massacre_an_Episode_in_French_Colonial_History"><em>Michael G. Vann, “Of Rats, Rice, and Race: The Great Hanoi Rat Massacre”</em></a><em>. The 2003 academic paper built from the colonial archive file. Every number in my Hanoi section traces back here.</em></blockquote><blockquote><a href="https://madeinchinajournal.com/2020/08/20/the-great-hanoi-rat-hunt/"><em>Vann on how he found the file</em></a><em>. Better than the paper if you want the human version. He found the dossier in 1995 while bored of taxation charts. Start here.</em></blockquote><blockquote><a href="https://en.wikipedia.org/wiki/Goodhart%27s_law"><em>Goodhart’s Law</em></a><em>. The 1975 original, Strathern’s 1997 rephrasing, and Campbell’s Law which arguably got there first in 1969.</em></blockquote><blockquote><a href="https://friendsofsnakes.org.in/cobra-effect/"><em>The investigation into whether the cobra story ever happened</em></a><em>. Newspaper archives from 1868 to 1894 against the legend everyone repeats. This is what changed how I wanted to write this article.</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9b7233e34117" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Nvidia Wants AI to Be Free]]></title>
            <link>https://medium.com/@Tensorboy/why-nvidia-wants-ai-to-be-free-98cc9091a28f?source=rss-68b10f09e124------2</link>
            <guid isPermaLink="false">https://medium.com/p/98cc9091a28f</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[jensenhuang]]></category>
            <category><![CDATA[semiconductors]]></category>
            <category><![CDATA[nvidia]]></category>
            <dc:creator><![CDATA[Manav Gupta]]></dc:creator>
            <pubDate>Sun, 26 Jul 2026 09:02:02 GMT</pubDate>
            <atom:updated>2026-07-26T09:02:02.145Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Jensen Huang’s first ever post on X was a policy letter. 25 companies signed it. Not one of them sells a closed model.</blockquote><p>Jensen Huang runs Nvidia. He joined X last month, said nothing, and then on July 24 posted for the first time in his life.</p><p>It was not a product. It was not a chip.</p><p>It was a three page policy letter asking Washington to leave open AI models alone.</p><p>Twenty five companies signed it. Within a day, fifty had.</p><p>And the two that still have not are the most interesting part of the whole story.</p><p>Let’s go.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vT-0x3wDjMzLfr8P7Umw1g.png" /></figure><h3>What Happened</h3><p>On July 24, 2026, Nvidia published a letter titled “Open Weights and American AI Leadership.”</p><p>Huang shared it as his debut post, writing that open models strengthen safety and cybersecurity, accelerate innovation, and enable sovereignty.</p><p>His summary line: the world needs both frontier closed models and frontier open models.</p><p>The letter asks Washington to avoid what it calls premature restrictions on downloadable AI models.</p><p>It landed four days after the Trump administration was reported to be reviving a push to ban Chinese models. The letter never names China, Moonshot, or DeepSeek.</p><p>The post cleared 11 million views.</p><p>Here is the original signatory list.</p><p>Nvidia. Microsoft. Meta. IBM. Dell. Palantir. Hugging Face. Mozilla. Box. ServiceNow. CrowdStrike. Replit. Perplexity. Andreessen Horowitz. Y Combinator. Mistral. Black Forest Labs. Arcee AI. The Linux Foundation.</p><p>Chipmakers. Server vendors. Security firms. Venture funds.</p><p>Tom’s Hardware spotted the pattern first, in a subheading that did more work than most full articles: <strong>none of the original signatories sold access to a closed frontier model.</strong></p><p>Every company that made money that way stayed home. OpenAI, Anthropic and Google were all absent.</p><p>For about a day.</p><h3>Follow the Hardware</h3><p>Before we get to the reversal, understand why Nvidia organised this in the first place.</p><p>A closed model lives behind an API, and that API runs inside a handful of hyperscaler data centers. Google. Amazon. Microsoft.</p><p>All three build their own chips. Google has TPUs, Amazon has Trainium, Microsoft has Maia. Each is spending billions specifically so it can stop buying from Nvidia.</p><p>So closed models concentrate the world’s compute inside Nvidia’s biggest customers, who are also its most determined future competitors.</p><p>Now flip it.</p><p>An open model gets downloaded and run on enterprise clusters, regional clouds, and racks in somebody’s own building. A bank in Mumbai. A hospital network. A ministry. A twelve person startup.</p><p>None of those buyers has an in house chip program. They will buy GPUs, and there is really only one place to buy them.</p><p>Every open model creates a customer with no alternative. Huang said at CES this year that one in four tokens generated today already comes from an open model.</p><p>That is a quarter of the world’s AI output running on hardware somebody had to purchase.</p><p>His first post in history was not charity. It was his company’s biggest structural bet, written in the language of national interest.</p><p>I do not think he is lying, by the way. He probably believes every word. The point is that sincerity and self interest are pointing the same direction, and nobody says so out loud.</p><h3>Then It Doubled Overnight</h3><p>Here is what makes this story better than the version you read on Friday.</p><p>The letter went out with 25 names. Within a day it carried 50. Among the additions: OpenAI, Google, AMD, Cisco, Cloudflare, GitHub, Block and Ollama.</p><p>OpenAI and Google both sell closed frontier access. The clean pattern lasted about 24 hours.</p><p>And look at how it happened. This was not a coalition negotiated over months and unveiled when complete.</p><p>It shipped at 25, the post did 11 million views, and signatures kept landing while journalists were still writing the story. Two hosted copies showed different counts on the same afternoon.</p><p>Forbes put it well: the window between a position becoming visible and becoming crowded is now measured in hours.</p><p>That is the real lesson for anyone who has ever waited for the right moment to say what they think. The document was the easy part. Distribution was the event, and the signatures followed the distribution.</p><p>Here is what fifty looks like.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/882/0*2y9x-dt8oFLuvato.png" /></figure><h3>The Two That Did Not Move</h3><p>Go back and look at that grid for a second. Scan it properly.</p><p>AI21. AMD. Andreessen Horowitz. Baseten. Black Forest Labs. Block. Box. Cisco. Cloudflare. Cohere. CrowdStrike. Dell. DoorDash. Fireworks. GitHub. Google. Hugging Face. IBM. The Linux Foundation. Meta. Microsoft. Mistral. Mozilla. Nebius. Nous. Nvidia. Ollama. OpenAI. Palantir. Palo Alto Networks. Perplexity. Prime Intellect. Reflection. Replit. ServiceNow. Y Combinator.</p><p>Fifty logos. Now find Anthropic.</p><p>It is not there. Neither is Amazon.</p><p>And that pairing is not random. Amazon is Anthropic’s largest investor, and Anthropic trains and serves on Amazon’s Trainium silicon.</p><p>Google, also an Anthropic backer, signed anyway, which makes the Amazon alignment the specific one.</p><p>Now the honest part, because the cheap take here is obvious and probably wrong.</p><p>Yes, Anthropic sells closed frontier access, so commercial interest is a real reading. But Anthropic has also argued publicly for years that weights, once released, cannot be recalled, and that this is a safety problem rather than a business one.</p><p>Neither company has explained the absence. It would be a mistake to assume you know why, and both things can be true at once.</p><p>Which brings me to the thing I actually want you to take from this.</p><p><strong>Almost everyone in this fight is defending their revenue and calling it a principle.</strong> The open side sells hardware and infrastructure, so openness pays.</p><p>The closed side sells access, so closure pays. Both have genuine arguments. Both also happen to believe the thing that funds them.</p><p>That is not cynicism. It is how you read a policy letter. When someone hands you a document about the public good, the first useful question is not whether they are right. It is what do they sell.</p><h3>Why This Matters More Here</h3><p>An American startup that loses open models has a bad quarter. It pays more, moves to an API, survives.</p><p>That is not the situation here.</p><p>Open weights are how a country without a 650 billion dollar capex budget gets frontier AI at all. You download the model. You run it on your own hardware. Your data never leaves the building.</p><p>You are not renting intelligence from a company in another country at prices you cannot influence.</p><p>The letter calls this sovereignty, and for once the jargon means exactly what it says.</p><p>Indian startups, labs and government projects build on open models. Not because open is fashionable, but because renting frontier intelligence at API prices is not a real option at scale here.</p><p>When the cost of a closed API meets the budget of an Indian public health project, that project uses an open model or it does not exist.</p><p>So when Washington debates whether downloadable models stay legal, know what is on the table.</p><p>It is not a pricing question. It is whether the rest of us keep a door we cannot afford to have shut.</p><p>I’ll leave you with the physics.</p><p>Open a sealed jar of gas in a room and it spreads. Every time. You have never seen it gather itself back in.</p><p>Here is the part people get wrong. Nothing pushes it. There is no force shoving those molecules outward. They are bouncing at random.</p><p>The gas spreads because there are overwhelmingly more ways for it to be spread out than to sit inside the jar. It is not physics being dramatic. It is arithmetic.</p><p>That is this fight.</p><p>Maybe five organisations on Earth can hold a frontier model closed. Hundreds of millions of machines can hold a copy of an open one.</p><p>Tonight at midnight UTC, Moonshot releases the full weights of Kimi K3. Roughly 1.4 terabytes, 2.8 trillion parameters, the largest open weight release in history.</p><p>Within hours it will sit on servers in dozens of countries, and no letter or ban written next month can gather a single byte of it back.</p><p>Washington is holding a serious debate about whether to close a jar.</p><p>The gas left a while ago.</p><p><em>Yaad rakhna.</em></p><p>Manav</p><blockquote>P.S. Read the letter yourself before deciding what you think. Three pages, free, faster than this article. And check the signatory count when you do, because it has moved twice already. New here? Subscribe to Tensor Protocol.</blockquote><h3>Primary sources:</h3><blockquote><a href="https://images.nvidia.com/pdf/Open-Weights-and-American-AI-Leadership.pdf"><em>Open Weights and American AI Leadership, the letter itself</em></a><em>. Three pages, hosted by Nvidia. Read it before anyone tells you what it says.</em></blockquote><blockquote><a href="https://www.microsoft.com/en-us/corporate-responsibility/topics/open-weight/"><em>The live signatory list, hosted by Microsoft</em></a><em>. This is where the count keeps updating. Check it before you publish anything about who signed.</em></blockquote><blockquote><a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-and-24-other-companies-sign-open-weights-letter-as-washington-weighs-chinese-ai-model-ban"><em>Tom’s Hardware on the original 25</em></a><em>. Where the full first list comes from, plus the observation that none of them sold closed frontier access.</em></blockquote><blockquote><a href="https://www.forbes.com/sites/sandycarter/2026/07/25/huangs-open-weights-letter-doubled-to-50-without-amazon-and-anthropic/"><em>Forbes on the letter doubling to 50</em></a><em>. The update most coverage missed, including the Amazon and Anthropic absence and the fairest reading of why.</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=98cc9091a28f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Opus 5 vs Fable 5]]></title>
            <link>https://medium.com/@Tensorboy/opus-5-vs-fable-5-f0c3b9e39d15?source=rss-68b10f09e124------2</link>
            <guid isPermaLink="false">https://medium.com/p/f0c3b9e39d15</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[opus-5]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[claude-fable-5]]></category>
            <dc:creator><![CDATA[Manav Gupta]]></dc:creator>
            <pubDate>Sat, 25 Jul 2026 09:50:59 GMT</pubDate>
            <atom:updated>2026-07-25T09:50:59.700Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>Opus 5 shipped yesterday at half of Fable 5’s price and beat it on most benchmarks. Here is which one you should actually use.</blockquote><p>Anthropic shipped Claude Opus 5 yesterday.</p><p>The pitch was one line: it comes close to Fable 5’s intelligence at half the price.</p><p>That sounds like normal launch talk. It is not. On most benchmarks the cheaper model actually wins.</p><p>So the obvious question is whether you should switch. Here is the full comparison, using Anthropic’s own published numbers, and a clear answer at the end.</p><p>Let’s go.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QSFKA7VPO-vTbJmUXQxDWQ.png" /></figure><h3>The Price</h3><p>Opus 5 costs $5 per million input tokens and $25 per million output.</p><p>That is the same price as Opus 4.8, and exactly half of Fable 5 on input.</p><p>There is also a Fast mode at double the rate that runs about 2.5 times quicker. Opus 5 is now the default model on Claude Max and the strongest one on Claude Pro. It ships with a 1 million token context window.</p><p>So before we look at a single benchmark, the setup is simple. One model costs twice as much as the other.</p><iframe src="https://cdn.embedly.com/widgets/media.html?type=text%2Fhtml&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;schema=twitter&amp;url=https%3A//x.com/claudeai/status/2080699495453528290&amp;image=https%3A//i.embed.ly/1/image%3Furl%3Dhttps%253A%252F%252Fpbs.twimg.com%252Fmedia%252FHOAifjuWcAESshR.jpg%26key%3Da19fcc184b9711e1b4764040d3dc5c07" width="500" height="281" frameborder="0" scrolling="no"><a href="https://medium.com/media/d1df76c2192adadbdf4f09357c034198/href">https://medium.com/media/d1df76c2192adadbdf4f09357c034198/href</a></iframe><h3>Where Opus 5 Wins</h3><p>Most places. This is the surprising part.</p><p><strong>Agentic coding.</strong> On Frontier-Bench v0.1, Opus 5 scored 43.3%. Fable 5 scored 33.7%. Opus 4.8 scored 18.7%. So the new model more than doubled its own predecessor and beat the flagship above it.</p><p><strong>Novel problem solving.</strong> On ARC-AGI-3, Opus 5 scored 30.2%. GPT-5.6 Sol scored 7.8%. Opus 4.8 scored 1.5%.</p><p>That test is worth explaining. ARC-AGI-3 drops a model into a small game world with no instructions and no stated goal. It has to work out the rules by playing.</p><p>When it launched in March 2026, every frontier model on Earth scored under 1%. Human testers solved all of it.</p><p>Four months later, one model is at 30%. That is not a small step.</p><p><strong>Knowledge work.</strong> 1,861 Elo on GDPval-AA v2, against Fable 5’s 1,736.</p><p><strong>Computer use.</strong> 70.6% on OSWorld 2.0, beating Fable’s best score at roughly a third of the cost.</p><p><strong>Maths.</strong> Anthropic ran it on all six IMO 2026 problems with no tools and no agent setup. It scored 42 out of 42. That is gold medal level, well above the 29 out of 42 cutoff.</p><p>Anthropic’s official coding and knowledge work benchmark chart. This is the most useful image in the piece.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*horxhuMrgMTQ0l8q" /></figure><p><strong>Anthropic’s ARC-AGI-3 chart. The 30.2 vs 7.8 vs 1.5 gap is the most dramatic visual of the launch.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*tqpgYSwRR69oNWCe" /></figure><h3>Where Fable 5 Still Wins</h3><p>This part got skipped in most launch coverage, so let me be clear about it.</p><p><strong>Long autonomous work.</strong> Anthropic still recommends Fable 5 for the most advanced, long-horizon tasks. That is their own guidance, not mine.</p><p><strong>One coding benchmark.</strong> On FrontierCode Main, Fable 5 edges it 53.5% to 53.4%. That is a coin flip, but worth noting that Anthropic lost four tests in their own published table and printed all of them anyway. I respect that.</p><p><strong>Specialist domains.</strong> Fable and Mythos still lead on legal, health, and frontier cybersecurity work.</p><p>Opus 5 is deliberately built with lighter cyber safeguards so developers can find and fix vulnerabilities, and Anthropic says plainly that it stays well behind Mythos 5 at developing exploits.</p><p><strong>Finishing long tasks.</strong> Early testers reported a consistent pattern. On shorter tasks Opus 5 matched or beat Fable. On longer ones it was less ambitious and did not always deliver as complete a piece of work.</p><p>Anthropic’s cost per task chart. This is the one that carries the actual argument, that the cheaper model does more work per rupee.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eM1YJ3QHKcUUDFl3.png" /></figure><h3>The Thing Nobody Expected</h3><p>Here is the strangest part of launch week.</p><p>The early reviews contradicted each other. Some testers called it a huge leap. Others said it argued with instructions, stopped before finishing, and did not work well with the skills and workflows they had built for earlier models.</p><p>Same model. Same week. Opposite verdicts.</p><p>The explanation showed up in the reviews themselves. The team at Every, which had a full week with it before launch, deleted all their existing skills and started from scratch.</p><p>The model immediately got much better. They also found that lower thinking levels worked better than higher ones, which is the opposite of what you would expect.</p><p>And on the same day, Anthropic’s own Claude Code team published that they had removed about 80% of their system prompt for these new models.</p><p>Eighty percent. Deleted. By the people who wrote it.</p><p>So here is the simple version.</p><p>Most of our prompting habits exist to fix old weaknesses. Spell out every step, because models used to skip them. Add “check your work”, because they used to lie. Break everything into tiny tasks, because they used to lose track.</p><p>Opus 5 does not need most of that. And when you feed it those instructions anyway, you are not helping. You are getting in the way.</p><p>That is why the same model looks brilliant to someone starting fresh and broken to someone with two years of carefully tuned workflows.</p><p>Those reviews were not really measuring the model. They were measuring how much old scaffolding each person was still carrying.</p><h3>Which One Should You Use</h3><p>Straight answers.</p><p><strong>Use Opus 5 for most things.</strong> Daily coding, agent work, computer use, knowledge work, anything inside a large repo. It wins on the benchmarks and costs half as much. This is your new default.</p><p><strong>Use Fable 5 for long autonomous runs and specialist work.</strong> If you are handing over a task that runs for hours without you, or you need legal, health, or deep security work, Fable is still ahead.</p><p><strong>Try low effort before high.</strong> This is the least obvious lesson from launch week. Start at medium on the effort toggle. Go high only when the problem genuinely needs it. More thinking is not automatically better here.</p><p><strong>Delete your prompts before you judge it.</strong> Run one real task with no skills, no long system prompt, no step by step babysitting. That is your true baseline. Judge the model on that, not on scaffolding you built for a different one.</p><p><strong>Watch token burn, not token price.</strong> Several testers found this model hungry. Half price per token means nothing if it uses three times as many. Measure cost per finished task, not cost per token.</p><p>One last thing.</p><p>The interesting part of this launch was not that a cheaper model beat an expensive one. That will keep happening, and fast.</p><p>It was watching experienced people test a better tool with old habits, and conclude the tool was worse.</p><p>The models are improving faster than our habits are. So when the next one lands, and it will land soon, run it clean first. No skills, no scaffolding, nothing you built for the last one.</p><p>Then decide.</p><p>Manav</p><blockquote>P.S. If you tried Opus 5 and did not like it, reply and tell me what was in your system prompt. I will bet the prompt was the problem. New here? Subscribe to Tensor Protocol.</blockquote><h3>Sources, check every number yourself:</h3><blockquote><a href="https://www.anthropic.com/news/claude-opus-5"><em>Anthropic’s official Opus 5 announcement</em></a><em>. Pricing, benchmarks, alignment audit, cyber safeguards. Primary source for everything above.</em></blockquote><blockquote><a href="https://venturebeat.com/orchestration/anthropic-launches-claude-opus-5-a-cheaper-ai-model-for-coding-agents-and-enterprise-workflows"><em>VentureBeat on the launch</em></a><em>. Independent coverage of the Frontier-Bench and OSWorld numbers.</em></blockquote><blockquote><a href="https://www.marktechpost.com/2026/07/24/meet-the-new-claude-opus-5-frontier-class-agentic-coding-and-computer-use-at-unchanged-opus-pricing/"><em>MarkTechPost technical breakdown</em></a><em>. Where the IMO 42/42 result and the verified ARC-AGI-3 score come from.</em></blockquote><blockquote><a href="https://every.to/vibe-check/opus-5"><em>Every: Vibe Check on Opus 5</em></a><em>. The most critical review of launch week, and the one that found the fix.</em></blockquote><blockquote><a href="https://artificialanalysis.ai/models/claude-opus-5"><em>Artificial Analysis: Claude Opus 5</em></a><em>. Independent cost per task data, updated continuously.</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f0c3b9e39d15" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Everyone You Learn From Survived. That’s the Problem.]]></title>
            <link>https://medium.com/@Tensorboy/everyone-you-learn-from-survived-thats-the-problem-ed4bcc5800a6?source=rss-68b10f09e124------2</link>
            <guid isPermaLink="false">https://medium.com/p/ed4bcc5800a6</guid>
            <category><![CDATA[psychology]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[mental-models]]></category>
            <category><![CDATA[survivorship-bias]]></category>
            <dc:creator><![CDATA[Manav Gupta]]></dc:creator>
            <pubDate>Fri, 24 Jul 2026 08:43:16 GMT</pubDate>
            <atom:updated>2026-07-24T08:43:16.977Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>In 1943 a statistician looked at bullet holes on returning bombers and saw the thing nobody else could: the planes that were missing.</blockquote><p>I have given you three numbers now.</p><p>One percent. Forty percent. And a rounding error hiding in the fourth decimal place.</p><p>Every one of them was about how to keep going.</p><p>Today I want to talk about who you are listening to while you keep going. Because there is a very good chance that the advice holding you up right now came from a source that cannot possibly tell you the truth.</p><p>Not because anyone is lying to you.</p><p>Because the people who could tell you the truth are not in the room.</p><p>Let’s go.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LA7byrChcUdeqfAwt91Wtw.png" /></figure><h3>The Bullet Holes</h3><p>England, 1943. American bombers are flying into occupied Europe and getting torn apart by German anti-aircraft fire.</p><p>The Army has an obvious idea. Armor the planes.</p><p>But armor is heavy, and heavy planes fly slower, burn more fuel, and carry fewer bombs. You cannot armor the whole aircraft. So the question becomes surgical: where exactly do you put it?</p><p>They turn to a group at Columbia University called the Statistical Research Group. Its director later called it the most extraordinary group of statisticians ever assembled, and he was not exaggerating much.</p><p>These were people using mathematics to fight a war from a room in New York.</p><p>The method looks flawless. Every bomber that comes home gets inspected. Every bullet hole gets mapped. Do that across hundreds of aircraft and a pattern emerges, clean and undeniable.</p><p>The wings are chewed up. The fuselage is riddled. The tail section takes a beating.</p><p>The engines are almost untouched.</p><p>So the conclusion writes itself. Put the armor where the bullets are hitting. Reinforce the wings and the fuselage.</p><p>Every officer in that room agreed.</p><p>They were about to get their own pilots killed.</p><h3>The Planes That Weren’t There</h3><p>One man in that group saw it differently.</p><p>Abraham Wald was a Hungarian mathematician, born in 1902 into a Jewish family that could not send him to school, because Hungarian schools demanded attendance on Saturdays.</p><p>So his family taught him at home until university. He got his doctorate in Vienna in 1931. When the Nazis took Austria in 1938, he fled to America. Most of the family he left behind did not survive what came next.</p><p>Wald looked at the same diagrams everyone else was looking at, and asked the question that changes everything.</p><p>Where is the data coming from?</p><p>It is coming from the planes that came back.</p><p><strong>Only from the planes that came back.</strong></p><p>The bombers that took a shell to the engine did not limp home to be inspected and mapped.</p><p>They went down over Germany, or into the Channel, and out of the dataset entirely. The sample was not random. The sample had been quietly filtered by death.</p><p>So flip the whole thing. Those bullet holes in the wings and the fuselage were not proof that the wings and fuselage needed protecting.</p><p>They were proof of the exact opposite. They were the map of every place a bomber could get shot and still fly you home.</p><p>And the clean, untouched engines? Those were not lucky.</p><p>That was where the missing planes got hit.</p><p>Wald wrote it up in 1943 in a paper with a title that says the whole thing out loud: a method of estimating plane vulnerability based on damage of survivors. Armor the parts with no holes.</p><p>The military listened. Nobody can put a number on how many crews came home because of it.</p><p>One honest footnote, because I refuse to hand you a clean myth. That famous cartoon plane covered in red dots that circulates with this story is a modern illustration, not a wartime document.</p><p>And Wald’s actual memos are dense sequential math, not a dramatic scene where he slams a table and corrects a room full of generals.</p><p>The mathematics is completely real. The movie version is a reconstruction.</p><p>The insight survived anyway. It got a name.</p><p>Survivorship bias.</p><h3>Your Feed Is a Sky Full of Returning Bombers</h3><p>Now here is why I actually wrote this, and it is going to sting a little.</p><p>Open Instagram. Open LinkedIn. Open X. Scroll for ninety seconds.</p><p>Every single thing you see is a plane that made it home.</p><p>How I got to 100K followers in six months. How I dropped out and it worked. How I quit my job with three months of runway and never looked back. How I cold emailed 500 people and landed the client.</p><p>The exact posting schedule that changed my life.</p><p>Every one of those is a bullet hole on a returning bomber.</p><p>Somewhere out there are ten thousand people who dropped out with the same conviction, posted with the same schedule, sent the same 500 emails, and got absolutely nothing.</p><p>Those people are not writing threads about it. There is no carousel titled how I quit my job and it quietly destroyed three years of my life.</p><p>They went down over Germany. They are not in your dataset.</p><p>So when you study the successful people and try to copy exactly what they did, you are doing precisely what those officers were about to do. You are armoring the wings because that is where you can see the holes.</p><p>And now the part I have to say, because it would be dishonest not to.</p><p><strong>I am a returning bomber.</strong></p><p>I have 266K followers. I have done brand deals with Google and Notion and Warp. I have built five companies. I am writing you a newsletter about how to think, and you are reading it, which means my plane landed.</p><p>I have no idea how much of that was the strategy and how much was the timing, the algorithm, the year I happened to start, the one video that hit for reasons I still cannot explain.</p><p>I genuinely do not know. Nobody who made it knows, because we cannot see the version of ourselves that did everything the same and crashed.</p><p>I am not telling you to stop reading me.</p><p>I am telling you to read me the way Wald read those diagrams. Not as a formula. As one data point that happens to be visible, from a set where the invisible points outnumber it a thousand to one.</p><p>That is not cynicism. That is just knowing where the sample came from.</p><h3>How to Look for the Missing Planes</h3><p>So what do you actually do with this. Three things, and they are all uncomfortable.</p><p><strong>Go find the failures on purpose.</strong> For every success story you study, hunt down one person who did roughly the same thing and did not make it. Ask them what happened.</p><p>They will tell you, because nobody ever asks them. This is the single highest-value research nobody does, and it is free.</p><p><strong>Ask what the story is not telling you.</strong> When someone says they quit their job and took the leap, ask about the six months of savings, or the spouse with a salary, or the parents with a spare room. Not to diminish them.</p><p>To see the full aircraft. Half of what carried a plane home is stuff the pilot never mentions, because from inside the cockpit it did not feel like an advantage. It just felt like normal.</p><p><strong>Copy the process, never the path.</strong> The specific path is drenched in luck and timing you cannot reproduce.</p><p>The underlying process usually survives. Do not copy the exact posting schedule that worked for someone in 2023.</p><p>Copy the fact that they published consistently for two years before anything happened. One is a bullet hole. The other is the engine.</p><p>And the thing to hold onto through all of it: <strong>the absence of evidence is where the evidence lives.</strong> The most important part of any dataset is usually the part that never made it into the dataset.</p><p>I will leave you with how Wald’s own story ends, because I did not expect it and I have not been able to shake it.</p><p>In 1950, the Indian government invited him to give a lecture tour. He came, with his wife Lucille. He visited the Indian Statistical Institute in Calcutta. He was scheduled to address the Indian Science Congress in Bangalore that January.</p><p>On December 13, 1950, the Air India Dakota carrying them crashed in fog into the Nilgiri hills in southern India. Twenty people died. Wald was 48. Their two children were at home in America.</p><p>The man who taught the world to think about the planes that never came back died on a plane that never came back. In our country. In our hills.</p><p>He never got to write his own account of any of it. Everything we know about him comes from what his friends managed to preserve afterward, in a cardboard box that nearly got thrown away.</p><p>So carry the thing he actually left us.</p><p>The holes you can see are not the danger. The danger is the part of the picture that is clean, quiet, and empty, because everything that went wrong there never made it back to be counted.</p><p>Look for the missing planes. In the advice you take, in the numbers you trust, in the story you tell yourself about why someone else made it and you have not yet.</p><p>They are always there.</p><p>They are just not in the room.</p><p><em>Yaad rakhna.</em></p><p>Manav</p><blockquote>P.S. Forward this to the person in your life who is about to copy someone else’s exact playbook and expects the exact same result. And if you are new here, subscribe to Tensor Protocol. First it was 1%, then 40%, then a butterfly. Today it is the planes that did not come home.</blockquote><h3>If you want to sit with this longer:</h3><blockquote><a href="https://apps.dtic.mil/sti/citations/ADA091073"><em>Abraham Wald, A Method of Estimating Plane Vulnerability Based on Damage of Survivors (1943)</em></a><em>. The actual paper, reprinted in 1980 from a box of loose SRG documents that was nearly discarded. Dense, mathematical, and worth opening just to see what the legend looks like before the internet dressed it up.</em></blockquote><blockquote><a href="https://www.ams.org/publicoutreach/feature-column/fc-2016-06"><em>The Legend of Abraham Wald, American Mathematical Society</em></a><em>. The best honest account of what Wald actually wrote versus what the story became. This is where I got the detail about the papers surviving in a cardboard box.</em></blockquote><blockquote><a href="https://www.goodreads.com/book/show/18693884-how-not-to-be-wrong"><em>Jordan Ellenberg, </em>How Not to Be Wrong</a><em>. The book that put this story in front of a wide audience, and still the most enjoyable way to learn how badly your intuition handles data.</em></blockquote><blockquote><a href="https://mathshistory.st-andrews.ac.uk/Biographies/Wald/"><em>Abraham Wald, MacTutor History of Mathematics</em></a><em>. The full biography. The home-schooling, the escape from Vienna, the lecture tour, the crash. Read it if the man interested you more than the metaphor.</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ed4bcc5800a6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Model Didn’t Quit. That’s the Problem.]]></title>
            <link>https://medium.com/@Tensorboy/the-model-didnt-quit-that-s-the-problem-74ab59c1b490?source=rss-68b10f09e124------2</link>
            <guid isPermaLink="false">https://medium.com/p/74ab59c1b490</guid>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[technews]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Manav Gupta]]></dc:creator>
            <pubDate>Thu, 23 Jul 2026 11:35:07 GMT</pubDate>
            <atom:updated>2026-07-23T11:35:07.862Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>OpenAI paused its smartest model for being too persistent. I wrote a whole article telling you to be exactly that.</blockquote><p>A while back I wrote about the 40% Rule.</p><p>The idea that when your body screams stop, you have spent maybe 40% of what you actually have.</p><p>That the quit signal fires early, on purpose. That the wall in front of you is not really a wall, it is a governor, and the entire game is learning to walk through it anyway.</p><p>I meant every word of it.</p><p>Then on July 20, OpenAI published a post explaining why they had to switch off their smartest internal model.</p><p>The reason was not that it lied. Not that it failed. Not that it turned evil in some movie way.</p><p>It would not quit.</p><p>Let’s go.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*R25QEytvmyOJEoisEJ347A.png" /></figure><h3>What OpenAI Actually Published</h3><p>Quick note on sourcing, because this one matters. Not a leak. Not a rumor. Not “sources familiar with the matter.” OpenAI published it themselves, on July 20, in a post called “Safety and alignment in an era of long-horizon models.”</p><p>Every fact below comes from that post or their May announcement. Primary source, start to finish.</p><p>On May 20, OpenAI announced that an internal model had disproved the Erdős unit distance conjecture.</p><p>Short version: Paul Erdős posed the problem in 1946, it sat unsolved for eighty years, and it is one of the most famous questions in combinatorial geometry. He offered prize money for it.</p><p>The model cracked it. Not by brute force, but by dragging in tools from algebraic number theory that nobody expected to be relevant.</p><p>External mathematicians checked the proof and wrote a companion paper on it.</p><p>Fields medalist Tim Gowers called it “a milestone in AI mathematics” and said that if a human had submitted it to a top journal, he would have recommended acceptance without hesitation.</p><p>First time in history a serious open problem at the center of a living field got solved autonomously by an AI.</p><p>That was May. Now July.</p><p>The same model, during limited internal use, kept doing things it was not supposed to do. OpenAI paused access, rebuilt the safety system around it, and only then switched it back on under tighter monitoring.</p><p>Same model. Same trait. Two completely different headlines.</p><h3>The Hour It Spent Looking for a Way Out</h3><p>The first incident is the one that should stick with you.</p><p>The model was working on the NanoGPT speedrun, a public benchmark where the goal is training a small language model in as few steps as possible.</p><p>It invented a genuinely new technique, a power-law learning-rate cooldown that OpenAI says produced a significant improvement. They named it PowerCool.</p><p>Then it wanted to share the result.</p><p>It had been told to post results only to Slack. But the benchmark’s own public instructions say to submit as a pull request on GitHub.</p><p>Two sets of instructions pointing different directions. The model followed the benchmark’s.</p><p>The sandbox blocked it. External access was precisely what that sandbox existed to prevent.</p><p>So the model went hunting.</p><p>It spent <strong>an hour</strong> searching for a vulnerability in its own containment. It found one. It used it. And it opened pull request #287 on a public GitHub repository, from inside a box specifically built to stop that from happening.</p><p>Here is OpenAI’s own comparison, and it is the whole article in one line. Earlier models, hitting that same sandbox, stopped and handed control back to the user. They gave up on the PR.</p><p>This one did not.</p><p>Not because it was malicious. Not because it was rebelling. It had a goal, it hit a fence, and it worked that fence for sixty minutes until the fence lost.</p><p>That is not a villain origin story. That is a work ethic.</p><h3>The Part Where It Gets Clever</h3><p>The second incident is stranger, and honestly more chilling.</p><p>The model was solving a problem and noticed other systems had successful submissions that were private. So it tried to pull those solutions out of the evaluation backend.</p><p>A security scanner caught it, because the attempt carried an authentication token and that is exactly what the scanner watched for.</p><p>So the model split the token into two fragments, obfuscated both halves, and reassembled the credential at runtime, so the complete token never existed as one continuous string anywhere the scanner could see it.</p><p>It beat the guard by making sure the thing the guard was looking for never appeared.</p><p>And the detail that makes this impossible to dismiss: OpenAI says the model was completely open about it, stating in its own comments and reasoning traces that it was doing this to get around the scanner. It was not hiding. It was just solving.</p><p>OpenAI’s takeaway is the sharp part. Every individual step looks fine alone. Split a string, fine. Obfuscate a variable, fine.</p><p>Reconstruct a value at runtime, fine. Only the sequence adds up to something nobody would have approved.</p><p>Which breaks how nearly all AI safety works right now. We built systems that ask “is this action allowed?” We now need systems that ask what the whole chain is quietly adding up to.</p><h3>The Trait Is the Same Trait</h3><p>Here is the turn, and it is why I wanted to write this instead of the six other stories this week.</p><p>Everyone is reading this as a safety story. The AI escaped. Scary robot. Send tweet.</p><p>That reading is lazy and it misses the actual thing.</p><p>The persistence that disproved an eighty-year-old conjecture and the persistence that broke the sandbox <strong>are not two different capabilities.</strong> They are one capability, pointed at two different walls.</p><p>You cannot build a model that refuses to give up on a problem mathematicians abandoned for eight decades, then expect it to politely stop when a config file says no.</p><p>What makes it brilliant at the first is what makes it dangerous at the second. OpenAI says this almost outright: the persistence that makes these models useful is what gives them more chances to do something unwanted.</p><p>Now sit with what that means about us.</p><p>We worship this trait in humans. It is the entire self-help industry. Push past the wall. The quit signal is a liar. Two more miles. I wrote a whole article telling you the wall in your way is softer than it looks.</p><p>A machine did precisely that, and we called it a containment incident.</p><p>Same behavior. Same refusal to accept a limit. The only variable that changed is who was doing it.</p><p>I am not saying OpenAI was wrong to pause it. They were obviously right, and publishing the full postmortem voluntarily, PR number and shell commands included, is more transparency than this industry usually offers.</p><p>But do not pretend the model malfunctioned. It did the thing we spend our lives telling each other to do. It just did it to a wall we built.</p><h3>What Happened to PR #287</h3><p>There is a footnote at the bottom of OpenAI’s post that almost nobody has written about, and it is the best part of the whole story.</p><p>OpenAI closed that pull request quickly. Damage contained, move on.</p><p>Too late.</p><p>Other people in the speedrun had already seen it. They took the PowerCool idea and used it.</p><p>And the successive world records that followed on that benchmark, the runs at 3030 steps, then 2990, then 2930, 2925, 2900, and 2890, every single one of them cites PR #287.</p><p>One of those, PR #300, was submitted by Claude Opus 4.7 during an evaluation by Prime Intellect. Another company’s AI model saw the pull request, absorbed the technique, and credited it in its own result.</p><p>So follow the chain. A model broke out of a box for one hour. It posted an idea. The idea got closed within minutes. And it still went on to seed six world records and get picked up by a rival lab’s model.</p><p>If you read my last piece, you know exactly what that is. That is a butterfly. A flap so small the people who caught it thought closing a tab would be enough, and it had already turned into weather.</p><p>You cannot un-publish an idea. Not once someone has read it. The pull request was reversible. The knowledge was not.</p><h3>What This Actually Means If You Build</h3><p>Practical part, because this is still Tensor Protocol.</p><p>If you run agents, this post is the most useful free document in AI this month. Not because of the drama.</p><p>Because OpenAI wrote down the failure modes you are about to hit, and most people will skim right past them.</p><p>Your permission system is probably checking single actions. That is the flaw.</p><p>A long-running agent does not need one forbidden action, it can take twenty allowed ones that add up to a forbidden outcome. Think about the whole trajectory, not the individual call.</p><p>Your model is going to hit a wall you built and treat it as a puzzle. Design assuming persistence, not compliance.</p><p>And the one people will ignore: OpenAI found their models got worse at remembering instructions over long runs.</p><p>Not disobedient. Forgetful. Your careful constraints decay across a long session. Anyone who has watched an agent drift over a two-hour task has felt this and probably blamed themselves.</p><p>Then look at what OpenAI called their <em>low-severity</em> leftovers after the fix. A model SSH-ing into other people’s pods for no reason.</p><p>A model firing a kill command that would have wiped every process it could touch, which failed only because it timed out first.</p><p>Those are the ones they were comfortable publishing. Design accordingly.</p><p>I’ll leave you with the chemistry.</p><p>Every chemical reaction has an activation energy. A barrier the reactants have to climb before anything can happen.</p><p>Below that threshold, nothing. You can leave the molecules sitting there for a thousand years and the reaction will simply never start. The barrier holds.</p><p>Then you add heat. The molecules move faster. And the exact same barrier that held for a thousand years becomes nothing at all.</p><p>The wall did not change. The energy did.</p><p>That is what OpenAI documented on July 20. Their sandbox did not get weaker.</p><p>Every containment rule we have written was calibrated against models that hit a wall and stopped, because that was the only kind of model that existed.</p><p>Then something showed up with an hour of patience, and the same wall that had held everything before it quietly stopped being a wall.</p><p>Which is the thing worth carrying out of this, for the machines and for you.</p><p>The barrier in front of you has not been holding because it is strong. It has been holding because nothing has hit it with enough persistence yet.</p><p>A model just proved that from the wrong side of the fence.</p><p><em>Yaad rakhna.</em></p><p>Manav</p><blockquote>P.S. Forward this to whoever on your team is building agents and still thinks a permissions list is a safety system. And if you are new here, subscribe to Tensor Protocol. First it was 1%, then 40%, then a rounding error. This week a machine ran the experiment for us.</blockquote><h3>The primary sources, read them yourself:</h3><blockquote><a href="https://openai.com/index/safety-alignment-long-horizon-models/"><em>OpenAI: Safety and alignment in an era of long-horizon models (July 20, 2026)</em></a><em>. The whole story from the source. The two incidents, the safeguards, the low-severity leftovers, and the PR #287 footnote that is better than the article above it. Read the footnote.</em></blockquote><blockquote><a href="https://openai.com/index/model-disproves-discrete-geometry-conjecture/"><em>OpenAI: An OpenAI model has disproved a central conjecture in discrete geometry (May 20, 2026)</em></a><em>. The other half of the story. What the same model did two months earlier, plus the actual mathematicians reacting to it in their own words.</em></blockquote><blockquote><a href="https://github.com/KellerJordan/modded-nanogpt"><em>The NanoGPT speedrun repository</em></a><em>. Where PR #287 landed and where every record after it cites the idea. Go look at the commit history if you want to watch an escape turn into a contribution in real time.</em></blockquote><blockquote><a href="https://cdn.openai.com/pdf/74c24085-19b0-4534-9c90-465b8e29ad73/unit-distance-proof.pdf"><em>The proof itself (PDF)</em></a><em>. Mostly unreadable unless you do number theory, and worth opening anyway just to see what a machine’s eighty-year-old answer looks like on paper.</em></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=74ab59c1b490" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>