<?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 Eugène Rastignac on Medium]]></title>
        <description><![CDATA[Stories by Eugène Rastignac on Medium]]></description>
        <link>https://medium.com/@gene.rastignac8?source=rss-5e601784a999------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*PvlU0kyc_cKG1Yf2ITQKgg.png</url>
            <title>Stories by Eugène Rastignac on Medium</title>
            <link>https://medium.com/@gene.rastignac8?source=rss-5e601784a999------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 15:18:03 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@gene.rastignac8/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[What the DeepSeek-v4.1-Flash-1017 Update Changes for Engineers Building Agents]]></title>
            <link>https://medium.com/@gene.rastignac8/what-the-deepseek-v4-1-flash-1017-update-changes-for-engineers-building-agents-94e7be1de0b7?source=rss-5e601784a999------2</link>
            <guid isPermaLink="false">https://medium.com/p/94e7be1de0b7</guid>
            <category><![CDATA[deepseek]]></category>
            <category><![CDATA[china]]></category>
            <category><![CDATA[post-training]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Eugène Rastignac]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 12:54:05 GMT</pubDate>
            <atom:updated>2026-10-19T12:54:05.885Z</atom:updated>
            <content:encoded><![CDATA[<h4>The same 552B base model got a post-training update. Three parts of the announcement say more than its slogan</h4><figure><img alt="Two models, 1017 and Opus 5, each reach the same green Pass check, but 1017 sits on a single coin while Opus 5 sits on a tall stack of coins labelled cost per task." src="https://cdn-images-1.medium.com/max/1024/1*VbbMgxu5nBf5jVoGlIIP7A.png" /></figure><p>V4.1-Flash-1017 is a post-training snapshot on the same 552B MoE base, and its real news is cost per task rather than “wish-driven programming.” On DeepSWE v1.1 it scores 73% to 77% at about 1 to just over 2 CNY per task, while Claude Opus 5’s top tier reaches about 75% at 50 to over 100 CNY. It still trails clearly on HLE, and scores depend on the harness.</p><h3>TL;DR</h3><ul><li><strong>1017 is a post-training snapshot of the same model.</strong> The API name and price are unchanged, and the base is still the 552B MoE with 1M context and an FP4 KV cache.</li><li><strong>All eleven benchmarks went up, most on long-horizon tasks.</strong> Terminal-Bench 4.0 gained 11.1 points and ProgramBench 8.8, though the footnotes tie scores to the harness and HLE still trails Opus 5 by 13.7 points.</li><li><strong>Cost per task is where 1017 leads.</strong> On DeepSWE v1.1 it reaches 73% to 77% for about 1 to just over 2 CNY per task, against about 75% at 50 to over 100 CNY for Claude Opus 5.</li><li><strong>The demos count because they took very few interaction turns.</strong> Few turns mean the model got most of the project right on the first try, and self-feedback RL probably explains why this showed up first on games and 3D scenes.</li><li><strong>Search is now a paid product.</strong> DeepSeek’s in-house Search and Fetch APIs cost 5 CNY per 1,000 calls, and server-side search in DeepSeek Harness, Claude Code and Codex is free until billing starts on November 1.</li></ul><p>The phrase the official announcement pushes hardest is “wish-driven programming,” where you describe what you want and get a project. Read the announcement from start to finish, though, and “wish-driven programming” turns out to be only the entry point. The real information sits in three other places: a benchmark table with eleven items, a scatter plot that puts “cost per task” on the horizontal axis, and an item most people skimmed past, the newly opened in-house search API. The sections below take these three in turn.</p><p>V4.1 Flash was released on September 10, and the October 17 update arrived 37 days later. The name keeps the dated snapshot format (V4.1-Flash-1017), where the suffix marks the release date. The update keeps the existing model rather than shipping a new one: the API name and price are unchanged, and the web client, the app and the API all switched over at the same time.</p><p>The announcement credits the gain in project-level building to a training technique it calls “reinforcement learning from self-feedback.” That makes V4.1-Flash-1017 a post-training iteration with the architecture left as it was, so the base is still the approach from last month’s technical report: a 552B MoE with 1M context and an FP4 KV cache.</p><p>Once you know what kind of release this is, the rest of the data is easy to read: the base model is the same, and the better behavior comes from the training method and the harness, the agent framework that runs the model, working together.</p><h3>Gains on All Eleven Benchmarks, Biggest on Long-Horizon Tasks</h3><p>Here is the official comparison table. Alongside the previous V4.1 Flash, it includes four models for comparison: Kimi K3, GLM 5.3, GPT 5.6-Sol and Claude Opus 5.</p><figure><img alt="Benchmark table: DeepSeek V4.1-Flash-1017 vs V4.1 Flash, Kimi K3, GLM 5.3, GPT 5.6-Sol, Claude Opus 5 on eleven benchmarks" src="https://cdn-images-1.medium.com/max/1024/1*4hs6Jzzr6jDfcBdW7o00gQ.png" /><figcaption>DeepSeek V4.1-Flash-1017 beats the previous V4.1 Flash on all eleven benchmarks.</figcaption></figure><p>(Image: <a href="https://api-docs.deepseek.com/zh-cn/news/news261017/">DeepSeek’s official announcement</a>)</p><p>Against the previous version, all eleven items went up, by about 5 points on average. The two biggest gains:</p><ul><li>Terminal-Bench 4.0: from 31.2 to 42.3, an absolute gain of 11.1 points, or about 35.6% in relative terms.</li><li>ProgramBench: from 20.3 to 29.1, an absolute gain of 8.8 points, or about 43.3% in relative terms.</li></ul><p>Terminal-Bench 4.0 and ProgramBench are both long-horizon execution benchmarks. The model has to keep working in an environment until the job is done, rather than answer one question at a time, which is the same capability as “project-level building.”</p><p>Three of the other gains were also substantial: HLE (Humanity’s Last Exam) rose from 36.8 to 42.6, Apex Bench v1.1 from 64.6 to 72.1 and Automation-Bench from 54.8 to 60.5. DeepSWE (74.2 to 77.0) and NL2Repo-Bench (65.4 to 68.2) moved less, under 3 points each. But both were already near the top in the previous version, and the closer a model gets to the top, the harder each extra point is to win.</p><p>Across the six models and eleven items, here is where V4.1-Flash-1017 landed:</p><ul><li><strong>It took four firsts.</strong> DeepSWE v1.1 (77.0, ahead of Claude Opus 5’s 74.0), Automation-Bench V1.0.6 (60.5, with Claude Opus 5 second at 50.3), Agents’ Last Exam (35.4) and Apex Bench v1.1 (72.1, against 65.8 for Claude Opus 5).</li><li><strong>It came second on five more, close behind the leader.</strong> Terminal-Bench 4.0 (42.3, behind only Opus 5’s 51.8), ProgramBench (29.1, behind only Opus 5’s 37.0), NL2Repo-Bench (68.2 against 75.3), Chartography (83.0 against Opus 5’s 84.0) and BabyVision (92.5 against 94.1).</li><li><strong>It placed third on ZeroBench-main.</strong> Its 51.0 sits behind GPT 5.6-Sol’s 53.0 and Opus 5’s 52.0.</li></ul><p>HLE is where it falls clearly behind: 42.6 against Opus 5’s 56.3, a gap of 13.7 points, with GPT 5.6-Sol (44.5) and Kimi K3 (43.5) also ahead of it. HLE tests expert-level questions across academic fields, a different capability from “writing code and building projects.” The benchmark table makes DeepSeek’s position clear: it can’t keep up with the top tier on general-knowledge Q&amp;A, but on hands-on benchmarks it either ranks first or sits right behind the leader.</p><p>Two lines of footnotes under the table are easy to miss, but they matter. First, Terminal-Bench 4.0 was evaluated with the Claude Code harness and DeepSWE with the Mini-SWE-V2 harness. In other words, much of what the model can do only shows up through an external agent harness, so the same model scores differently when the harness changes. Writing the harness straight into the footnote amounts to admitting that results are tightly coupled to the runtime environment. When you read a benchmark score, you can’t talk about the model apart from its harness.</p><p>Second, ZeroBench-main uses a Pass@5 scoring convention, and the starred HLE numbers cover only the text subset. These two differences in scoring convention are easy to overlook when you compare across tables.</p><h3>The Performance-Cost Chart: Cost per Task Completes the Story</h3><p>The performance-cost chart says more than the benchmark table does. On DeepSWE v1.1, DeepSeek plotted each model’s pass rate against “the average cost of completing one task,” with the horizontal axis showing task cost on a log scale:</p><figure><img alt="Scatter plot of DeepSWE v1.1 Pass@1 versus cost per task on a log axis; DeepSeek V41 Flash-1017 reaches 77.0%." src="https://cdn-images-1.medium.com/max/1024/1*LEjvmI5-VQX9u_8cXjwvng.png" /><figcaption>DeepSWE v1.1 pass rate (Pass@1) versus average cost per task.</figcaption></figure><p>(Image: <a href="https://api-docs.deepseek.com/zh-cn/news/news261017/">DeepSeek’s official announcement</a>. Public-model data come from the DeepSWE v1.1 leaderboard with the mini-swe-agent harness, as of September 22, 2026; dollars are converted at 1 USD = 6.6862 CNY.)</p><p>V4.1-Flash-1017’s three data points sit between 73% and 77%. The cheapest tier reaches about 73% for roughly 1 CNY, and the top point, 77.0%, costs just over 2 CNY per task. For comparison:</p><ul><li><strong>Claude Opus 5:</strong> about 75% at its top tier, at a cost that runs from 50 CNY to over 100. That is tens of times what 1017 costs, for a score just two points lower.</li><li><strong>GPT-6.1 Sol:</strong> about 71% at its top tier, at around 10 CNY.</li><li><strong>Kimi K3:</strong> about 69% at max, at somewhere between 10 and 20 CNY.</li><li><strong>GLM-5.3:</strong> about 67% at max, at close to 20 CNY.</li></ul><p>So 1017 gained 3.5 points at a lower price, and that small shift up the chart is what the announcement’s claim of “shorter task run time and output length” looks like on the books.</p><p>Put the performance-cost chart next to the benchmark table and you get a more meaningful conclusion than another round of “so-and-so is sweeping the leaderboards.” On agent tasks, “score” and “cost per task” can be weighed in the same coordinate system, and the upper-left corner of the Pareto frontier, the curve of the best score at each cost, now belongs to 1017.</p><p>For an individual user, the question is whether they can afford it. If your team has put agents into its production pipeline, the question is whether the bill hurts when it comes due. For the past two years, talk about evaluations has been dominated by scores, and the very act of putting cost on the horizontal axis says more about where the market is heading than any single high score.</p><h3>“Wish-Driven Programming”: Four Demos and One Premise Tucked Out of Sight</h3><p>Beyond scores and cost, the most widely shared part of the announcement was its four demos: a 3D model of the Forbidden City (Palace Museum), a browser-based racing game, a remake of “Portal 2” in the Godot game engine, and Valley of Ages, a 3D scene that runs from prehistory to the modern era. Here are GIFs of the first two demos:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/480/1*_nu0iXIBii4ofz-W-psBgQ.gif" /><figcaption>One of the demo cases: a 3D model of the Forbidden City generated from a prompt.</figcaption></figure><figure><img alt="Web racing game demo: a teal convertible on a neon city street with NEON VELOCITY title and a stats panel." src="https://cdn-images-1.medium.com/max/480/1*axp8-ZfCImmHKC8rbfwfIA.gif" /><figcaption>The web racing demo ships a full UI: title screen, car stats panel and start control.</figcaption></figure><p>(Image: official announcement)</p><p>The level of finish you can make out in the GIFs deserves a mention. In the racing game, the cockpit view, the speedometer and gear display, the track minimap and the lap timer are all there, which takes it past “a demo that runs” and closer to a complete small project with a UI and a state system. The “Portal 2” demo even reproduces the dialogue text of GLaDOS (the Portal 2 AI character) and the level structure of the test chambers.</p><p>One sentence in the announcement, though, pins down exactly how much the “wish” is worth: the four demos were built “with relatively few prompts and reference assets, inside the DeepSeek Harness framework, in very few interaction turns.” That makes three conditions: few prompts, inside the harness, few turns. The first two are easy to understand, but the real key metric is “very few interaction turns.”</p><p>In project-level generation, the biggest worry is not failing to write code but writing it wrong and never getting it back on track. Few turns mean the model got most of the project structure, the file layout and the engine API usage right on the first try, instead of grinding it out over dozens of rounds of trial and error. The very short turn count and ProgramBench’s 43% gain in the benchmark table point to the same thing.</p><p>On “reinforcement learning from self-feedback,” the announcement gives only a one-sentence characterization with no implementation details, and nobody should fill in the specifics on DeepSeek’s behalf. In the public literature, self-feedback RL generally has the model produce its own evaluation signals (self-scoring, self-correction, feeding execution results back) and then uses those signals as rewards to update the policy.</p><p>Games and 3D modeling have a natural advantage here. Whether the output runs, whether it renders and whether what renders is a sensible scene can all be checked automatically, so the reward signal needs no human labeling. Self-feedback RL is probably why the ability to build a complete project from scratch showed up first on games and 3D modeling, since games and 3D scenes happen to be tasks that come with their own verifiers.</p><h3>The Easy-to-Miss News: The Search API Opens Up</h3><p>Everything covered so far is the most prominent part of the announcement. Its second half also announced that DeepSeek’s in-house Search and Fetch APIs are now open, priced at 5 CNY per 1,000 calls.</p><p>DeepSeek also described the machinery underneath. Its in-house search engine handles hundreds of millions of search requests from around the world every day and indexes on the order of a hundred billion web pages. The company made a point of stressing that it filters out SEO spam, GEO (generative engine optimization) manipulation and low-quality AI-generated content. The same announcement also tucks in a date. Server-side search calls from DeepSeek Harness, Claude Code and Codex are free for now, and billing starts at 00:00 Beijing time (UTC+8) on November 1.</p><p>Put the pricing and the billing date together and the intent is fairly plain. An agent has to look things up on its own before it can get work done. That retrieval step used to rely either on third-party search APIs or on each harness wiring up its own. Now the model vendor is building that infrastructure straight into its stack and selling it per call. People who build RAG already know how 5 CNY per 1,000 calls compares with similar search APIs.</p><p>As for billing from November 1, two readings both make sense: the free period builds usage habits, and it also gets the accounting worked out up front. It shows, too, that DeepSeek means to run “server-side tool calls” as a real product. This isn’t charity.</p><h3>The One Thing to Remember</h3><p>In one sentence, V4.1-Flash-1017 means this: on agent tasks, DeepSeek has moved from “the top tier of open source” to “right behind the strongest closed models, and more than one order of magnitude cheaper,” while also filling in the retrieval piece.</p><p>Real-world performance isn’t the same as leaderboard numbers. Scores wobble when the model changes harness or workload, and the 13-point gap on HLE is still there. But one post-training snapshot in 37 days, 3.5 more points at a lower cost and “cost per task” put on the table make the direction clear: DeepSeek is betting on cost per task instead of an arms race over individual scores. For people building applications, cost per task is the number that can actually go into a business model.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=94e7be1de0b7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Rules vs Examples for In-Context Learning Across 5 Task Types]]></title>
            <link>https://medium.com/@gene.rastignac8/rules-vs-examples-for-in-context-learning-across-5-task-types-0fa8db747095?source=rss-5e601784a999------2</link>
            <guid isPermaLink="false">https://medium.com/p/0fa8db747095</guid>
            <category><![CDATA[writing-prompts]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Eugène Rastignac]]></dc:creator>
            <pubDate>Mon, 19 Oct 2026 01:54:14 GMT</pubDate>
            <atom:updated>2026-10-19T01:54:14.359Z</atom:updated>
            <content:encoded><![CDATA[<h4>You can state the rule or show worked examples. One study measured which one the model follows more accurately.</h4><figure><img alt="A stated rule and a set of input-output examples each feed the same language model, and on the same test the rule path returns all correct answers while the examples path returns a mix of correct and wrong ones." src="https://cdn-images-1.medium.com/max/1024/1*DaiNNYkqadxFCbptld12Hw.png" /></figure><p>Giving an LLM the rule for a new computational task beat showing it examples by an estimated 17.1 percentage points across games, arithmetic and language judgments. The gap was widest on tasks that compose operations or relate attributes. Adding examples to a clearly written rule gained 1.2 percentage points, short of statistical significance. The advice to try rules first holds only where the rule can be written faithfully and clearly.</p><h3>TL;DR</h3><ul><li><strong>Rules win clearly on three tasks.</strong> The operator function (40.1 percentage points), the Set card game (31.9) and noun-class agreement (20.8) all reach p&lt;0.001, while Tapatan and the two lexical-category judgments show no significant rule advantage.</li><li><strong>Across all tasks, rules beat examples by an estimated 17.1 percentage points.</strong> That is the linear mixed model’s pooled estimate for the rules-only condition over the examples-only condition.</li><li><strong>Adding examples to a clear rule showed no consistent gain.</strong> Rules plus examples beat rules only by 1.2 percentage points, short of statistical significance, and no task-by-task comparison found a significant gain.</li><li><strong>Try explicit rules first when you can fully specify the task.</strong> The advice holds only where the rule can be written down faithfully and clearly, so whether to add examples still has to be tested on your task.</li></ul><p>To get a large language model to perform a new task, you can state the rule directly in the prompt. Or you can give a few inputs with their correct answers and let the model infer the rule from those examples. An experiment spanning games, arithmetic and language judgments compared these two prompting approaches. Explicit rules performed better overall, and the gap was wider on tasks that require composing operations or judging relationships between attributes. The advantage was not statistically significant on every task, though.</p><p>The operator-function task makes this concrete. It defines a new function f(x,y,z)=(x×y)+z and asks the model for f(-19,-14,-7). The rule prompt says directly to multiply first and then add. The example prompt shows only a few input-output pairs, so the model still has to work out how the operations relate. Both conditions change only the context: neither updates the model weights, and both use the same test questions.</p><blockquote><strong><em>[Figure 1: Visualization of the experimental design]</em></strong><em> The left side shows the five task types, with lexical-category inference split into an AND variant and an OR variant. The right side uses the same arithmetic problem to compare three prompting conditions: rules only, examples only, and rules plus examples.</em></blockquote><figure><img alt="Experiment design: five tasks on the left; rule, example and combined prompt modes on the right." src="https://cdn-images-1.medium.com/max/995/1*PnkCIkiZQCs-0ikwf9pxng.png" /><figcaption>The same operator-function problem is posed under rule-only, example-only and rules-plus-examples prompts.</figcaption></figure><h3>Rules Win Mainly on Operations, Card Relationships and Word-Form Matching</h3><p>The operator-function task shows the largest rule advantage. The study’s linear mixed model controls for instruction tuning, model size and task difficulty, and it also accounts for differences between model families. It puts the “rules minus examples” accuracy difference at 0.401, or 40.1 percentage points. That figure is a difference estimated by a statistical model. It does not mean accuracy reached 40.1%, and it does not mean every model improved by that much.</p><p>The Set card game asks the model to pick, out of nine cards, the three whose attributes satisfy the required relationship. The rule advantage there is estimated at 31.9 percentage points. The noun-class agreement task asks whether word forms in an artificial language match, with an advantage of 20.8 percentage points. Counting the operator-function task, all three of these differences reach p&lt;0.001. By contrast, no significant rule advantage was detected for win/loss judgments on the Tapatan board game or for the two lexical-category judgments. [1]</p><blockquote><strong><em>[Figure 4: Linear mixed model contrasts by task for rules minus examples (left) and rules plus examples minus rules only (right)]</em></strong><em> On the left, three task types show a significant rule advantage, while on the right none of the differences reaches significance. Error bars are 95% confidence intervals. The two panels use different scales, so you cannot directly compare line lengths between the left and right panels.</em></blockquote><figure><img alt="Forest plots: rules beat examples on three tasks; adding examples to rules shows no significant gain." src="https://cdn-images-1.medium.com/max/1024/1*CW40DSD1runY5kK5a-8q0A.png" /><figcaption>Rules help significantly on three tasks, but adding examples to stated rules yields no reliable gain.</figcaption></figure><h3>No Consistent Gain Detected from Adding Examples to a Clear Rule</h3><p>With all tasks analyzed together, the rules-only condition beat the examples-only condition by an estimated 17.1 percentage points in accuracy. Rules plus examples beat rules only by 1.2 percentage points, which falls short of statistical significance, and task-by-task comparisons detected no significant gain either. The experiment shows there is not enough evidence to confirm a gain in these settings. It does not establish that examples contribute exactly zero.</p><p>These estimates come from the base and instruction-tuned versions of Gemma 3, Qwen 2.5, Qwen 3 and OLMo 3/3.1, at 7B–32B parameters. GPT-5.4 served as a separate reference and was left out of the statistical fit. The examples were not picked at random either. Each set tries to cover the task’s basic components, its representative cases and the applicable output labels, and the number of examples varies with the task and its difficulty. A supplementary experiment that added more examples likewise showed no consistent gain.</p><blockquote><strong><em>[Figure 3: Main linear mixed model effects of learning condition, instruction tuning, model size and difficulty on accuracy]</em></strong><em> The first two rows give 0.171 for rules versus examples and 0.012 for rules plus examples versus rules only. Both use an accuracy scale from 0 to 1, so multiply by 100 to convert them to percentage points.</em></blockquote><figure><img alt="Forest plot of LMM effects on accuracy: rules +0.171, instruct tuning +0.122, large size +0.067, hard tasks -0.150." src="https://cdn-images-1.medium.com/max/645/1*M3QF2iPtar_WmOZsABd8gQ.png" /><figcaption>Rules beat examples by 0.171 in accuracy, while adding examples to rules changes little (0.012).</figcaption></figure><h3>For a New Computational Task You Can Fully Specify, Try Explicit Rules First</h3><p>Some tasks have an order of operations or attribute constraints that can be clearly defined. For those, this experiment supports trying explicit rules first rather than asking the model to infer the rule from examples alone. The gaps were smaller on the board-game and lexical-category tasks. The board game lays out the spatial arrangement directly, and the lexical judgments can draw on the model’s existing semantic knowledge. These task features may help models make use of examples, but the experiment did not show that they are what narrowed the gap.</p><h4>The Honest Take</h4><p>This recommendation holds only where the rule can be written down faithfully and clearly. The rule prompt supplies both the task description and the valid answer format, while the example prompt leaves the model to infer them. So the two conditions are not fully symmetric in the information they present. Real-world rules may be vague, incomplete or noisy, and the experiment does not show that examples should be cut from few-shot prompting across the board. Whether to add examples still has to be tested on your specific task.</p><p>📄 Original title: LLMs Learn Better In-Context from Rules than from Examples</p><p>🔗 Original link: <a href="https://arxiv.org/abs/2609.03213">https://arxiv.org/abs/2609.03213</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0fa8db747095" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How the OpenCode Breach Drained Claude and GPT Credits Overnight]]></title>
            <link>https://medium.com/@gene.rastignac8/how-the-opencode-breach-drained-claude-and-gpt-credits-overnight-07900f54e45e?source=rss-5e601784a999------2</link>
            <guid isPermaLink="false">https://medium.com/p/07900f54e45e</guid>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[open-code]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[security]]></category>
            <dc:creator><![CDATA[Eugène Rastignac]]></dc:creator>
            <pubDate>Sun, 18 Oct 2026 16:14:33 GMT</pubDate>
            <atom:updated>2026-10-18T16:14:33.694Z</atom:updated>
            <content:encoded><![CDATA[<h4>Stolen provider API keys ran up bills while developers slept. Only one fix actually shuts a leaked key down</h4><figure><img alt="A stolen API key slips out of a cracked legacy system and flows into a provider API whose credits drain, with a revoke-here marker on the provider side showing where the key must be shut off." src="https://cdn-images-1.medium.com/max/1024/1*lgsaqEKuW76Z-2HYo1_8Mw.png" /></figure><p>Attackers broke into a legacy OpenCode system that was being decommissioned and took the provider API keys 1.4% of OpenCode Console users had stored in their workspaces. OpenCode linked the breach to a large-scale distillation attack: within hours, the attackers turned those credits into high-quality tokens from top models to train their own. Only revoking each key at the provider that issued it stops the charges.</p><h3>TL;DR</h3><ul><li><strong>A forgotten legacy system was the way in.</strong> Attackers raided a system OpenCode was decommissioning but had not yet shut down, exposing the provider API keys that 1.4% of OpenCode Console users had saved in their workspaces.</li><li><strong>The bills landed within hours.</strong> Victims posted charges of $400 on Anthropic, more than $1,000 on Google Cloud and more than $1,700 in the worst case.</li><li><strong>The attackers wanted training data.</strong> Stolen keys gave them top-model outputs at the victims’ expense, sidestepping both the cost of that data and the providers’ bans on training with it.</li><li><strong>Revoke each key where it was issued.</strong> A new password or a deleted OpenCode config leaves a leaked key valid, so revoke it in each provider’s console, then check your usage and bills.</li></ul><p>Developers woke up to a nasty surprise. Today, a lot of OpenCode developers around the world got “robbed.” In just a few hours, nearly all the tokens in their accounts were burned through, and the whole AI community was in an uproar.</p><figure><img alt="Tweet showing an OpenCode dashboard with monthly usage at 100% used, weekly at 86% and rolling at 5%." src="https://cdn-images-1.medium.com/max/1024/1*lt0dVwgTn3ifSxnvQvbrTA.png" /><figcaption>Figure: One affected user’s OpenCode dashboard shows the monthly quota fully used. — Source: 新智元</figcaption></figure><p>OpenCode, the open-source AI coding tool, rushed to investigate and then acknowledged a security breach. <strong>This time, about 1.4% of OpenCode Console users had the bad luck to be caught in the crossfire.</strong> The API keys they had stored in their workspaces for “the big three” (OpenAI, Claude, Gemini) were left completely exposed.</p><figure><img alt="OpenCode’s post announcing a security leak tied to a distillation attack that affected 1.4% of Console users." src="https://cdn-images-1.medium.com/max/1024/1*_NKRfWAIFyd91pzksa4uvA.png" /><figcaption>OpenCode confirms a security leak linked to a distillation attack that affected 1.4% of its Console users.</figcaption></figure><p>Once the hackers had the keys, they showed no restraint and hammered the top models with request after request. One victim posted their bills: more than $400 gone from their OpenAI and Anthropic keys in the blink of an eye. One developer had it even worse: <strong>more than $1,700 was run up on his account</strong>.</p><figure><img alt="X thread: users report stolen OpenCode API keys racking up $400 and $1,700 in fraudulent charges overnight." src="https://cdn-images-1.medium.com/max/1024/1*XgocbmOqIcwNgoGJrY2BZA.png" /><figcaption>Users of OpenCode reported hundreds to thousands of dollars in fraudulent charges overnight after their API keys were stolen.</figcaption></figure><p>OpenCode emailed developers overnight with a warning and said it was working with Anthropic, OpenAI and Google “to remediate.” Anthropic’s Lydia Hallie was quick to respond as well: “We’ve been working with OpenCode on yesterday’s breach and have revoked Claude API keys we believe were affected.”</p><figure><img alt="Post from Lydia Hallie: Claude API keys affected by the OpenCode vulnerability were revoked; create a new key to be safe." src="https://cdn-images-1.medium.com/max/1024/1*GZZWDSeXMUntSuUI_wCKGg.png" /><figcaption>Anthropic’s Lydia Hallie confirms the affected Claude API keys have been revoked and urges OpenCode users to create new ones.</figcaption></figure><h3>A Massive Security Hole at OpenCode</h3><h3>$1,700 in Fraudulent Charges Overnight</h3><p>The story of OpenCode getting robbed blew up fast.</p><h3>At first, only a handful of programmers were posting their absurd API bills.</h3><p>While they slept, token usage from API calls shot up like a rocket on their dashboards and drained their balances to zero. It even triggered auto-reload, the feature that tops up prepaid credits automatically.</p><figure><img alt="Tweet and OpenAI usage chart: a huge Oct 17 spend spike from a deleted opencode API key" src="https://cdn-images-1.medium.com/max/1024/1*UqHnxrmps1RW-FXORdJ_9g.png" /><figcaption>One leaked opencode credential produced a single-day OpenAI spend spike of over $700.</figcaption></figure><figure><img alt="Tweet confirming drained API balances, with an OpenAI usage dashboard and a Google Cloud support email." src="https://cdn-images-1.medium.com/max/1024/1*FIHRDwudxp0XmCCy7uTgBA.png" /><figcaption>API keys leaked through the tool drained OpenAI and Anthropic balances and got a Google account restricted.</figcaption></figure><p>Then someone posted a screenshot: $400 had been racked up on their Anthropic account without them having any idea. Worse still, one person’s Google Cloud API bill went off the charts, with a sudden charge of more than $1,000.</p><h3>As the fire spread, OpenCode’s official account rushed out a tweet confirming what everyone had feared most:</h3><p>“We had a security breach last night related to a large scale distillation attack that impacted 1.4% of OpenCode Console users.”</p><h3>In an urgent email to users, OpenCode finally came clean.</h3><p><strong>The cause of the whole mess was “a legacy system that we were in the process of decommissioning.”</strong></p><figure><img alt="Reddit post showing OpenCode’s email warning that provider keys may have been exposed through a legacy system breach" src="https://cdn-images-1.medium.com/max/1024/1*kiWCUvJtQgrthR_P_DHMlg.png" /><figcaption>OpenCode’s email tells users that a legacy system it was decommissioning was breached and that stored provider keys may be exposed.</figcaption></figure><h3>Not Even the Founder Escaped</h3><p>OpenCode’s investigation turned up a surprise: hackers had broken straight into this legacy system, which the team hadn’t yet gotten around to shutting down. The email warned that “third party provider keys (eg OpenAI, Anthropic) saved in your workspace may have been exposed.”</p><p>For now, all OpenCode can do is urge users to protect themselves right away:</p><ul><li>Revoke and reissue their keys immediately.</li><li>Update their app settings.</li><li>Go carefully over their recent bills for API calls.</li></ul><p>The most ironic part is that in a hack that spread across the entire internet, <strong>not even OpenCode’s founder escaped</strong>. Shortly after the official statement was released, OpenCode co-founder Dax Raad posted on X, sounding both helpless and shocked:</p><blockquote><em>My personal accounts are being attacked now too; whole new world with ai powered hacking.</em></blockquote><figure><img alt="Dax Raad’s post1" src="https://cdn-images-1.medium.com/max/590/1*grDCnemWrw-AnzTalx2KHQ.png" /></figure><figure><img alt="Dax Raad’s post2" src="https://cdn-images-1.medium.com/max/601/1*EGjlkuU_e_fs5EmF84gqYQ.png" /></figure><h3>The Surprising Motive: Freeloading on the APIs to Train Their Own Models</h3><p>More dramatic still, the hackers didn’t steal the keys to resell them, or to grab people’s personal chat histories. <strong>They took the tokens and went off to “train their own models.”</strong> In a window of just a few hours, the hackers who targeted OpenCode turned every last cent of credit on these keys into high-quality tokens from top-tier LLMs.</p><figure><img alt="Hooded hacker at a desk with several monitors in a dark server room" src="https://cdn-images-1.medium.com/max/1000/1*kyvEBcLUbcFmNTEJMrQflw.png" /><figcaption>Attackers turned stolen API key quota into training data for their own models.</figcaption></figure><p>To make a small model smarter, the fuel that matters most is huge amounts of high-quality chain-of-thought (CoT) and code-generation data. Buying OpenAI and Claude API access outright to generate this training data could easily cost millions or even tens of millions of dollars. On top of that, the big AI companies have long banned training on their models’ outputs in their terms of service, and they have deployed extremely tight abuse-detection systems on the back end.</p><p>So the hackers came up with a particularly nasty way to make someone else foot the bill: <strong>distributed freeloading on developers.</strong> By breaking into a third-party AI tool aggregator like OpenCode, the hackers got their hands on the API keys of hundreds, even thousands, of real developers in one go. Every token they burned along the way was charged to the victims’ accounts.</p><p>TestingCatalog, one of the big names in AI news, nailed it: this is “a serious precedent,” and “using smaller providers is now less safe.”</p><figure><img alt="TestingCatalog post on X about OpenCode’s statement on a distillation attack affecting 1.4% of users" src="https://cdn-images-1.medium.com/max/1024/1*7BVWaLwL1MxTtu-TTivHEA.png" /><figcaption>OpenCode confirmed that a large-scale distillation attack hit 1.4% of its Console users.</figcaption></figure><p><strong>Don’t just change your password!</strong></p><h3>If the Old Keys Aren’t Revoked, the Money May Keep Burning</h3><h3>If you received OpenCode’s security email, the notice already spells out the most important step right now:</h3><p>Go to each model provider’s official console, then revoke and replace any API keys that may have leaked.</p><p>Changing your login password is no substitute for revoking your API keys. Deleting the configuration in OpenCode or shutting down the app on your machine won’t, on its own, invalidate credentials that have already leaked. You have to go to the platform that issued the key and revoke the old one there.</p><figure><img alt="Tweet warning that OpenCode was hacked and urging users to rotate keys and change passwords immediately." src="https://cdn-images-1.medium.com/max/1024/1*5DcM2EANMOPmQSZzQCzYeg.png" /><figcaption>A tweet warns that OpenCode has been compromised and urges users to rotate their keys now.</figcaption></figure><p>Then check your recent usage and bills. Keep records of the anomalous calls, when the charges occurred and any related notifications, and ask the provider to investigate and sort out the charges. At the same time, you need to update any legitimate apps that rely on the old keys.</p><h3>The One Thing to Remember</h3><p>This time OpenCode leaked the keys of 1.4% of OpenCode Console users. The share is small, but the lesson is big:</p><p>In the age of AI, an API key is a developer’s second ID card.</p><p>Don’t hand it over casually to any one console just because it’s “convenient.” After all, hackers will never lose sleep over your bill.</p><p>References:</p><ul><li><a href="https://x.com/opencode/status/2111519637049319818?s=20">https://x.com/opencode/status/2111519637049319818?s=20</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=07900f54e45e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How OpenAI’s Model Wrote 722 Math Manuscripts With Lean Proofs]]></title>
            <link>https://medium.com/@gene.rastignac8/how-openais-model-wrote-722-math-manuscripts-with-lean-proofs-05a90121faba?source=rss-5e601784a999------2</link>
            <guid isPermaLink="false">https://medium.com/p/05a90121faba</guid>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[mathematics]]></category>
            <category><![CDATA[ai4science]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Eugène Rastignac]]></dc:creator>
            <pubDate>Sun, 18 Oct 2026 02:08:33 GMT</pubDate>
            <atom:updated>2026-10-18T02:08:33.104Z</atom:updated>
            <content:encoded><![CDATA[<h4>An unreleased frontier model produced the papers and much of the proof code. Mathematicians now have to read them</h4><figure><img alt="A stack of manuscripts splits into two paths, one through a green box of Lean-checked proofs and one through a dashed amber box of unformalized results, both converging on a magnifying glass labelled peer review." src="https://cdn-images-1.medium.com/max/1024/1*-GbSAnUIJVI-1SpiYBKRDQ.png" /></figure><p>OpenAI has put 722 manuscripts from its unreleased internal frontier model on GitHub, organized into 372 groups of related results. Two highlights: <strong>pi’s irrationality exponent is exactly 2</strong>, and <strong>a fixed zero-free boundary at 7/8 for the zeros behind the Riemann hypothesis</strong>. Lean formalizations back both, but verification across the repository is uneven, and the evaluation still has to come from the mathematical community.</p><h3>TL;DR</h3><ul><li><strong>OpenAI released 722 manuscripts in 372 groups of results.</strong> One result can lead to corollaries or several proofs, and the model behind them hasn’t been released yet.</li><li><strong>Pi’s irrationality exponent is exactly 2.</strong> That is the theoretical minimum any irrational number can have on this measure.</li><li><strong>The quasi-Riemann manuscript sets a fixed zero-free boundary at 7/8.</strong> Earlier zero-free regions narrowed further out, and the Riemann hypothesis itself requires the 1/2 line.</li><li><strong>The model catches its own wrong turns.</strong> In the pi summary it noticed that its argument would contradict known counterexamples for other numbers and went back to check for a missed condition.</li><li><strong>Verification is uneven, and the evaluation falls to mathematicians.</strong> Only some results come with Lean formalizations, and AGMAI says advising OpenAI does not mean endorsing the results.</li></ul><p>Early this morning, OpenAI put <strong>722 math manuscripts</strong> produced by its internal frontier model on GitHub, all in one go.</p><p>The problems range from pi, which we all know, to questions tied to the famous Riemann hypothesis. The table of contents runs through number theory, geometry, theoretical computer science and mathematical physics, and just skimming it takes a good while.</p><p>Many of the manuscripts also come with proof code for Lean, a proof assistant that lets a computer check a derivation step by step. The papers and the code have come out first; the model that produced them hasn’t been released yet.</p><figure><img alt="OpenAI Research Catalog front page: 372 result families in 722 manuscripts, with a table of contents by subject." src="https://cdn-images-1.medium.com/max/1024/1*QPBhEobgMqrldz1Uvw9TZg.png" /><figcaption>The catalog’s front page lists 372 result families across 722 manuscripts.</figcaption></figure><p>The front page of the results catalog lists 372 result groups and 722 manuscripts</p><p>One result can lead to new corollaries or have several different proofs, so the 722 manuscripts are organized into 372 groups of related results. In August, OpenAI shared ten math advances; this time it has laid out a whole library of research material. With the AI writing this much, mathematicians have plenty of reading ahead of them.</p><h3>Even Familiar Pi Hides an Old, Hard Problem</h3><p>Pi is irrational, as we all learned in school: its decimal expansion never ends and never repeats. Approximating it with a fraction is easy, though. The fraction 22/7 is approximately 3.142857, and 355/113 is approximately 3.1415929, already very close to pi’s 3.14159265… Allow larger denominators and you have a chance at more accurate approximations. What mathematicians want to know is this: <strong>how fast can the accuracy actually improve?</strong></p><p>The measure of this kind of approximation is called the irrationality exponent, and the larger the exponent, the more often fractions of exceptionally high accuracy keep turning up. This manuscript concludes that pi’s irrationality exponent is exactly <strong>2</strong>, the theoretical minimum any irrational number can have on this measure. Most people only memorized a few digits of pi as kids, and it turns out the number still held a question this deep.</p><figure><img alt="Manuscript header: The irrationality exponent of π is 2, with abstract" src="https://cdn-images-1.medium.com/max/965/1*wbbsVsi7O5JCYWX-vaX7mQ.png" /><figcaption>The manuscript’s abstract states that the irrationality exponent of π is exactly 2.</figcaption></figure><p>Title, authors and abstract of the manuscript on the irrationality exponent of pi</p><p>Turn to the Riemann hypothesis, and the stakes get higher still. 2, 3, 5, 7, 11… The primes look scattered, but their distribution hides deep patterns. The Riemann hypothesis, closely tied to those patterns, asks where certain zeros of a mathematical function actually lie.</p><p>Think of these zeros as points scattered across a region. To prove the Riemann hypothesis, you would have to show that every one of them in fact sits on a single straight line down the middle.</p><p>This manuscript takes on the “quasi-Riemann hypothesis”: proving, as a first step, that these points can never get into a fixed region off to the side. The regions previously proven free of zeros grow narrower the further out they extend, while this paper gives a fixed boundary that always holds.</p><p>In concrete numbers, the paper puts the boundary at <strong>7/8</strong>, while the Riemann hypothesis requires the <strong>1/2</strong> line. That is still some distance from the final goal, but proving that this region stays empty all the way out is an important advance in its own right.</p><figure><img alt="Title page and abstract of the Quasi-Riemann Hypothesis manuscript, claiming a zero-free half-plane Re s &gt; 7/8." src="https://cdn-images-1.medium.com/max/910/1*9du8qrX9X7ynGtJcKv8K-w.png" /><figcaption>The manuscript proves a fixed zero-free half-plane, Re s &gt; 7/8, against the 1/2 line the Riemann hypothesis requires.</figcaption></figure><p>Title, authors and abstract of the quasi-Riemann hypothesis manuscript</p><p>The catalog also includes a result on the Hodge conjecture for a special class of geometric spaces, along with work on problems such as the four-dimensional Kakeya problem and the Unique Games Conjecture. Each of these problems anchors a string of follow-up research, and a substantial proof often opens up more leads worth chasing.</p><h3>An AI Doing Research Also Catches Its Own Wrong Turns</h3><p>Why did OpenAI pick such hard problems? The team says existing math benchmarks are close to saturated, so the model has to be tested further on open research problems.</p><p>Across the evaluation, the model was given about 4,000 problems, and some of its work built on results it had found earlier. The team then sorted the results by importance, grouped related ones together and kept this set of manuscripts. Most results came from the same internal model running a fixed pipeline. The quasi-Riemann work and the Hodge conjecture result mentioned above are the exceptions.</p><p>OpenAI also published 10 condensed reasoning summaries, which show more of the trial and error than the final papers do.</p><p>In the summary for pi, the model tried one line of argument and, following it through, hit a problem: applied to other numbers, the same argument would contradict known counterexamples. So the model went back to check whether it had missed a condition, and its later exploration picked up some of the ideas it had left behind earlier. These records show how it tries things, spots problems and changes direction. Before a conclusion makes it into a paper, it has often been through this kind of back-and-forth.</p><p>So how do you check the proofs it writes? This is where Lean comes in. Math papers sometimes wave a step through with a single “obviously.” In Lean code, the assumptions, every derivation step and every theorem used have to be spelled out, and then a computer checks them. The main result on pi and the 7/8 bound of the quasi-Riemann hypothesis both come with this kind of formalization.</p><p>That gives later readers another way to check the proofs. Still, verification is at different stages across the repository, and results that haven’t been formalized may need corrections. With several hundred manuscripts now public, the item-by-item review has only just begun.</p><h3>Reading Them Is Only the Start</h3><p>On the day of the release, AGMAI, an independent advisory group on mathematics and AI, responded. Its nine scholars include three winners of the Fields Medal, the top prize in mathematics.</p><p>The group sees this as a major event whose impact goes beyond any single result. Its members were also clear on one point: having advised OpenAI does not mean they endorse these results or the process that produced them, and the evaluation still has to come from the mathematical community.</p><figure><img alt="AGMAI’s statement on OpenAI’s release of mathematical results, dated October 6, 2026." src="https://cdn-images-1.medium.com/max/762/1*ahW8SGW4kTi_RtJPp1yBCg.png" /><figcaption>AGMAI says its advisory role is not an endorsement of OpenAI’s results, and that only the mathematical community can assess them.</figcaption></figure><p>On September 11, Terence Tao and other Fields medalists also took up this issue in a joint statement. Its point was that mathematical progress needs peer discussion, exposition and simplification, so that new ideas reach more people and can eventually be taught to the next generation. Once a result is written down, a long stretch of work still lies ahead. With several hundred manuscripts arriving at once, that follow-up work has to keep pace too.</p><p>The advisory group also cares about something else: can mathematicians use the same model to explore their own questions? If the model only ever stays inside the company, outside researchers can only follow along, reading the results it hands over.</p><p>OpenAI laid out its next steps.</p><ul><li><strong>It will fund events around the major results.</strong> These include workshops, conferences and special programs.</li><li><strong>It is working toward opening up access to the model.</strong></li><li><strong>The manuscripts will keep their version history.</strong> Revisions will be published when problems turn up, so that later readers can keep track.</li></ul><h3>The One Thing to Remember</h3><p>This time it really is mathematicians’ turn to work through AI’s papers!</p><h3>References</h3><ul><li>OpenAI’s announcement of this release and of the earlier ten math advances: <a href="https://openai.com/index/sharing-ai-progress-in-mathematics/">https://openai.com/index/sharing-ai-progress-in-mathematics/</a> | <a href="https://openai.com/index/ten-advances-in-mathematics/">https://openai.com/index/ten-advances-in-mathematics/</a></li><li>Results repository and manuscript catalog: <a href="https://github.com/openai/math">https://github.com/openai/math</a></li><li>Quasi-Riemann hypothesis manuscript: <a href="https://github.com/openai/math/blob/main/preprints/The-Quasi-Riemann-Hypothesis-September-30-2026/paper.pdf">https://github.com/openai/math/blob/main/preprints/The-Quasi-Riemann-Hypothesis-September-30-2026/paper.pdf</a></li><li>Manuscript on the irrationality exponent of pi: <a href="https://github.com/openai/math/blob/main/preprints/The-irrationality-exponent-of-pi-is-2-September-24-2026/paper.pdf">https://github.com/openai/math/blob/main/preprints/The-irrationality-exponent-of-pi-is-2-September-24-2026/paper.pdf</a></li><li>Response from the advisory group on mathematics and AI: <a href="https://agmai.org/">https://agmai.org/</a></li><li>Joint statement by Terence Tao and others: <a href="https://terrytao.wordpress.com/2026/09/11/a-severe-misalignment-of-ai-in-mathematics/">https://terrytao.wordpress.com/2026/09/11/a-severe-misalignment-of-ai-in-mathematics/</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=05a90121faba" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Can Anyone Check 722 Math Manuscripts Written by an AI Model?]]></title>
            <link>https://medium.com/@gene.rastignac8/can-anyone-check-722-math-manuscripts-written-by-an-ai-model-1c88c195ff2b?source=rss-5e601784a999------2</link>
            <guid isPermaLink="false">https://medium.com/p/1c88c195ff2b</guid>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[mathematics]]></category>
            <category><![CDATA[ai4science]]></category>
            <category><![CDATA[lean]]></category>
            <dc:creator><![CDATA[Eugène Rastignac]]></dc:creator>
            <pubDate>Sat, 17 Oct 2026 03:10:35 GMT</pubDate>
            <atom:updated>2026-10-17T03:10:35.492Z</atom:updated>
            <content:encoded><![CDATA[<h4>What OpenAI released, what counts as a proof artifact, and why mathematicians are pushing back on how it was published</h4><figure><img alt="A wide orange funnel crammed with dozens of AI-written manuscripts, many marked green as Lean-formalized, narrows to a single checked output at a neck labelled expert attention." src="https://cdn-images-1.medium.com/max/1024/1*dcQ2AiGFVPgyfXtVLHAl8Q.png" /></figure><p>On the morning of October 7, the AI and math worlds blew up. Researchers and people online sat slumped in their chairs, stunned by the intelligence explosion.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/592/1*kc8EHRAnLbhsaNzv99seKw.png" /><figcaption>OpenAI’s announcement tweet and the openai/math GitHub repository.</figcaption></figure><p>It’s hard to picture: a single GitHub repository holds not one or two heavyweight math results, but 722.</p><p>The repository is a list of math papers: <strong>722 manuscripts, grouped into 372 result families, all produced by an unreleased internal frontier model.</strong> So what did it claim, and can anyone check it?</p><p>The names on the list are enough to make anyone who has sat through graduate math courses stop and look twice. Some of the problems have stood open for more than half a century, and others are directions that extend Fields Medal-level work. Among them:</p><ul><li>The quasi-Riemann hypothesis.</li><li>The 4D Kakeya conjecture.</li><li>The Unique Games Conjecture.</li><li>The free group factor isomorphism problem.</li><li>Hilbert’s tenth problem over the rationals.</li><li>The Hodge conjecture for CM abelian varieties.</li><li>The Mahler conjecture.</li></ul><p>Sam Altman said, “We are entering a new era of discovery.”</p><figure><img alt="Sam Altman tweet announcing a new era of discovery, with an OpenAI card titled Sharing AI progress in mathematics" src="https://cdn-images-1.medium.com/max/799/1*kSfOYBNH75x-mp1tXFTCAg.png" /><figcaption>Figure: Sam Altman’s tweet announcing OpenAI’s mathematics results. — Source: 机器之心</figcaption></figure><h3>An Explosive Release</h3><p>OpenAI says the results came out of an evaluation of the internal model’s math abilities.</p><p>The model’s performance on existing math benchmarks had saturated, so the team extended the evaluation to open research problems. Some of the results build on ones the model produced earlier. Across the whole evaluation, the model was fed about 4,000 research problems.</p><p>The team grouped the outputs into result families by relatedness and then filtered them against a “significant enough” threshold. That left the current 372 families and 722 manuscripts.</p><p>A “result family” usually holds one main result along with supporting arguments, corollaries or alternative proofs. By OpenAI’s own subject labels, the 372 families cover 17 areas.</p><p>The largest are theoretical computer science (40 families), combinatorics (37 families), algebraic and complex geometry (36 families) and number theory (31 families). Probability and statistical mechanics, differential geometry, mathematical physics, operator algebras and others follow. Mathematical logic has the fewest, though it still has 6 families.</p><p>On compute, OpenAI says the vast majority of results ran through the same fixed pipeline, <strong>with each result using, on average, compute roughly equal to 3 hours of ChatGPT Pro thinking.</strong></p><p>Another detail stands out: the manuscript folder names end in a date, and almost all of them fall between September 23 and October 5. September 23 and 24 alone account for nearly 200 manuscripts each. That suggests <strong>the bulk of this list was turned out in batches in under two weeks</strong>.</p><p>Alongside the papers, OpenAI released three kinds of “proof artifacts”: Lean formal proofs, 10 condensed summaries of the model’s reasoning, and statistics on compute and the number of problems attempted. By our own count of the repository’s directories, 235 of the 372 families come with a Lean formalization notes page, about 63%.</p><blockquote><em>OpenAI is just as candid that not every manuscript has been formalized. The unformalized results “may have problems,” and the team says it will fix them as soon as possible and keep filling in the rest.</em></blockquote><p>Right now the repository’s GitHub stars are climbing like crazy: it passed 2,000 stars in just three hours.</p><figure><img alt="GitHub page of the openai/math repository showing 2.2k stars, 188 forks and a single initial commit" src="https://cdn-images-1.medium.com/max/1024/1*VTPRquKQ1ym7L8Lmavq8wQ.png" /><figcaption>The openai/math repository on GitHub, at 2.2k stars three hours after its initial commit.</figcaption></figure><p>Repository: <a href="https://github.com/openai/math">https://github.com/openai/math</a></p><h3>Riemann and Kakeya Take Over the Feeds</h3><p>Of the 722 manuscripts, the first to flood social media was <strong>№003: the quasi-Riemann hypothesis</strong>.</p><p>The Riemann hypothesis asserts that all nontrivial zeros of the Riemann zeta function lie on the line with real part 1/2. That goal is so far out of reach that number theorists settled for something weaker, the “quasi-Riemann hypothesis.”</p><p>Even proving there is some constant θ less than 1 for which the entire half-plane with real part greater than θ contains no zeros would be huge progress.</p><p>For more than 100 years, every zero-free region anyone could prove hugged the line with real part 1 ever more closely as the imaginary part grew. No one ever managed to carve out a vertical strip of fixed width.</p><p>OpenAI’s result this time is θ = 7/8, holding uniformly for the zeta function and all Dirichlet L-functions, and it already comes with a Lean formalization. The companion paper also gives a uniform exclusion result for Landau-Siegel zeros.</p><figure><img alt="Title and abstract of OpenAI’s paper The Quasi-Riemann Hypothesis: a zero-free half-plane Re s &gt; 7/8" src="https://cdn-images-1.medium.com/max/988/1*btE8ymwA7_2qayhZAwZrOA.png" /><figcaption>Title page and abstract of OpenAI’s paper on the quasi-Riemann hypothesis.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/595/1*T2dxgzSGJRxbBBfpOBZrBg.png" /><figcaption>A tweet reporting OpenAI’s claimed proof of the quasi-Riemann hypothesis.</figcaption></figure><p>Another result that set off heated discussion is <strong>№074: the Kakeya problem</strong>.</p><p>On July 23 this year, Hong Wang won the Fields Medal for the proof of the Kakeya set conjecture in three dimensions that she completed with Joshua Zahl in 2025. She is the third woman ever to win it. At the time, the 4D Kakeya conjecture still sat prominently on Quanta Magazine’s list of “next steps.”</p><p>This family in OpenAI’s list claims to settle both the 3D Kakeya maximal function conjecture and the 4D Hausdorff dimension conjecture.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/595/1*0bwAFM0vERnbRD7merBA9w.png" /><figcaption>A mathematician’s tweet on the Kakeya problem in 3D and 4D.</figcaption></figure><p>The heavyweight results go far beyond these two.</p><ul><li>In theoretical computer science, №102 claims to prove the Unique Games Conjecture, which Subhash Khot proposed in 2002. It uses the result to establish, in one stroke, optimal hardness-of-approximation thresholds for classic problems such as Max-Cut and Vertex Cover.</li><li>In algebraic geometry, №032 claims to prove the rational Hodge conjecture for all complex CM abelian varieties.</li><li>In number theory, Nos. 002 and 006 together give the full BSD formula on a density-one family of quadratic twists of every rational elliptic curve. №004 gives a negative answer to Hilbert’s tenth problem over the rationals.</li></ul><p><strong>These three results all touch on Millennium Prize Problems or landmark problems of the 20th century</strong>.</p><p>In operator algebras, №287 claims to settle the free group factor isomorphism problem, concluding that free group factors of different rank are isomorphic to each other. It comes with a Lean formalization.</p><p>The list also has counterexamples constructed for Kaplansky’s zero-divisor conjecture and the Hadwiger conjecture, along with results proving that:</p><ul><li>The irrationality measure of π is exactly 2.</li><li>Catalan’s constant is irrational.</li><li>The Mahler conjecture holds in every dimension.</li><li>The plane has no unit-distance coloring with five colors.</li></ul><p>As a supplement, below is a ranking of every discovery from the past three years by former OpenAI researcher Will Depue. He used GPT 6 Pro and Fable 5.1 to rank them. Blue marks human discoveries, red marks AI discoveries (before October 6) and green marks AI discoveries from the OpenAI/math repo. The stunning part: <strong>81% of these discoveries were released only today</strong>.</p><figure><img alt="Numbered list of recent math results, color-coded as human, AI or OpenAI-repo AI discoveries, with sources." src="https://cdn-images-1.medium.com/max/814/1*p3Nr_bdfq5sAJpi7opxfPg.png" /><figcaption>Will Depue’s ranking of math discoveries over the past three years: blue for human, red for AI, green for OpenAI/math repo AI.</figcaption></figure><h3>A Mysterious Internal Model</h3><p>The model itself never appeared anywhere in the release. OpenAI did not disclose its name, calling it only an “internal frontier model” in the blog post, and said it is working toward its “responsible release.”</p><p>On September 8, OpenAI published its Navier-Stokes announcement. According to the announcement, the work was done by an internal model “significantly stronger than GPT-6 Astra” that began training on August 28 and is still in training.</p><p>That model ran about 10,000 parallel agents over 88 hours. It proved that a three-dimensional incompressible fluid under a smooth external force can start from rest and develop a singularity in finite time. GPT-6 Astra then spent an extra 17 hours completing the Lean formalization.</p><figure><img alt="3D vortex of colored streamlines spiraling inward and stretching along a vertical axis" src="https://cdn-images-1.medium.com/max/1024/1*2QPQuncx4LMleJxzUrdYgg.png" /><figcaption>Visualization of a fluid vortex with inward spiraling flow and axial stretching.</figcaption></figure><p><em>Related link: </em><a href="https://openai.com/zh-Hans-CN/index/navier-stokes-solution/"><em>https://openai.com/zh-Hans-CN/index/navier-stokes-solution/</em></a></p><blockquote><em>Ahead of the release, an OpenAI spokesperson had already confirmed that the new internal model has solved more than 100 long-standing open problems “across most areas of mathematics.”</em></blockquote><h3>Why Mathematicians Are Up in Arms</h3><p>Over the past two-plus months, friction and pressure between AI labs and mathematicians have kept building.</p><p>On August 1, OpenAI announced ten advances in mathematics and theoretical computer science by its next-generation model Astra. Only a few days later, mathematicians pointed out that one of the groups in the counterexample meant to disprove the Connes rigidity conjecture did not satisfy the required conditions.</p><p>On August 11, Anthropic also disclosed that its unreleased model had made progress in directions related to the Riemann hypothesis.</p><p>According to Wired, OpenAI privately convened about 40 mathematicians in August to discuss what to do if AI surpasses humans at mathematics.</p><p>Attendees said OpenAI hinted at the time that its models had already solved hundreds of long-standing hard problems, and asked for input on how to release them. One attendee said the company also promised not to release everything at once. An OpenAI spokesperson said they were “not aware” of any such promise.</p><p>A Northwestern University mathematician recalled that the attendees felt a mix of excitement and fear. On the spot, they proposed that the results no longer be released through blog posts and tweets alone, as the ten problems had been in August. The results should instead be written up as papers so that mathematicians could digest and use them.</p><p>According to reports, OpenAI mobilized thousands of agents to attack the Navier-Stokes problem only after hearing “rumors” that someone was about to solve it. Tristan Buckmaster of New York University (NYU) accuses OpenAI of scooping work he had been doing privately with Levent Alpöge. Alpöge is himself an Anthropic employee. The two had not yet published their results, yet they had been using OpenAI’s tools to assist their research.</p><figure><img alt="WIRED headline: OpenAI Is Pissing Off a Bunch of Mathematicians—Again, by Isabella Ward" src="https://cdn-images-1.medium.com/max/721/1*FNnYdI0izZnfWheVAqqs2w.png" /><figcaption>WIRED report on OpenAI and mathematicians.</figcaption></figure><p><em>Related link: </em><a href="https://www.wired.com/story/openai-is-pissing-off-a-bunch-of-mathematicians-again/"><em>https://www.wired.com/story/openai-is-pissing-off-a-bunch-of-mathematicians-again/</em></a></p><p>Mathematicians say people feel there’s something mafia-like about the way these leading AI companies operate.</p><p>On September 11, 25 Fields medalists, including Terence Tao, signed a joint statement. It argued that cracking famous problems was always meant as a tool for driving conceptual understanding and as a proxy measure of that understanding. Rushed announcements leave no time to write good papers and cite properly, the statement said, and that creates serious attribution and plagiarism problems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/520/1*niWWwxeqPr1OPvW6s7_78w.png" /><figcaption>Terence Tao’s blog post announcing the Fields medalists’ statement, September 11, 2026.</figcaption></figure><p><em>Related link: </em><a href="https://terrytao.wordpress.com/2026/09/11/a-severe-misalignment-of-ai-in-mathematics/"><em>https://terrytao.wordpress.com/2026/09/11/a-severe-misalignment-of-ai-in-mathematics/</em></a></p><p>Before that, more than 4,000 mathematicians had signed the Leiden Declaration, which Bryna Kra helped draft. It called on AI companies to meet the mathematical community’s standards on disclosing AI use, peer review, citing prior work and open access.</p><p>On September 29, the independent advisory group on mathematics and AI at the Institute for Advanced Study (IAS) released recommendations for publishing AI-generated mathematical results, drawing on more than 600 community survey responses. OpenAI made a point of citing that very document in its tweets and blog post this time.</p><figure><img alt="Screenshot of the IAS advisory group’s ‘Responsible Release of AI-Generated Mathematics’ recommendations, dated 29 September" src="https://cdn-images-1.medium.com/max/802/1*1D2vNEzF8bbCAMwHJWAdEg.png" /><figcaption>The IAS advisory group’s recommendations on releasing AI-generated mathematics</figcaption></figure><p><a href="https://agmai.org/general-sep29/"><em>https://agmai.org/general-sep29/</em></a></p><p>The IAS recommendations come in two steps.</p><p>The first step is “initial release.” Under it, labs should:</p><ul><li>Search for and cite relevant existing work.</li><li>Write proofs up to the conventions of a traditional math paper.</li><li>Disclose the model’s name, the prompts, chain-of-thought summaries, time spent and estimated compute costs.</li><li>Formalize as much as possible.</li><li>Deposit the results in a neutral repository “not controlled by any AI lab,” with persistent identifiers and a revision history.</li><li>Report failure rates on comparable problems.</li></ul><p>The second step is “supporting understanding”: funding conferences, workshops and expository writing, and giving the global math community broad, equitable access to the models.</p><p>Graded against this checklist, OpenAI’s release earns what you might call a “selective pass.” It published reasoning summaries, compute estimates and the number of problems attempted. It committed to funding a series of workshops and conferences around major AI results, and promised versioned history, with each revision published as a new version. It also formalized more than 60% of the families.</p><p>The release also falls short of the recommendations in plenty of places, above all by not disclosing the model’s name. The repository lives under OpenAI’s own GitHub organization, and OpenAI says it is still exploring community-hosted options. OpenAI also admits in its blog post that future releases will keep improving on citations, mathematical exposition and presentation.</p><h3>The Bottleneck Moves From “Proof” to “Understanding”</h3><p>If September’s Navier-Stokes announcement proved that AI can crack one of the hardest problems, this release proves something else. Across a fairly broad stretch of the mathematical frontier, new, checkable results can now be produced in batches, at a pace of “a few hours per problem.”</p><blockquote><em>This sharpens Terence Tao’s point from September: </em><strong><em>the real conflict is the mismatch between an abundance of proofs and scarce expert attention, not human versus machine reasoning</em></strong><em>.</em></blockquote><p>Lean proofs for 235 of the 372 result families do make it far cheaper to check whether a proof is correct. But the Connes rigidity episode showed the limit: formalization guarantees error-free reasoning, not that the statement proved is the one mathematicians have in mind.</p><p>The Lean formalization notes page OpenAI prepared for each family spells out “what the formalization covers and what it doesn’t.” The quasi-Riemann family, for instance, notes that the later applications in the paper fall outside the formalization. That alone shows that someone still has to check the statements themselves.</p><p>The weightiest results that have not been formalized yet include 4D Kakeya, the Hodge conjecture for CM abelian varieties and Hilbert’s tenth problem over the rationals. Whether they hold up under line-by-line expert review over the next few months will be the real report card for this release.</p><p>A question worth pressing even harder: when new theorems are no longer scarce, how should mathematicians reallocate their most precious resource, the time it takes to understand them?</p><p>Sam Altman says we are entering a new era of discovery. As of today, the “new era” part has 722 pieces of evidence behind it. For the “discovery” part to hold, people still have to understand these results, absorb them and pass them on.</p><p><em>References</em></p><p><a href="https://openai.com/index/sharing-ai-progress-in-mathematics/"><em>https://openai.com/index/sharing-ai-progress-in-mathematics/</em></a></p><p><a href="https://github.com/openai/math/blob/main/lean/docs/003.md"><em>https://github.com/openai/math/blob/main/lean/docs/003.md</em></a></p><p><a href="https://huggingnews.com/ai/openai-math-release-includes-claimed-progress-on-three-millennium-prize-29a7cc5e"><em>https://huggingnews.com/ai/openai-math-release-includes-claimed-progress-on-three-millennium-prize-29a7cc5e</em></a></p><p><a href="https://quantamagazine.org/hong-wang-wins-2026-fields-medal-the-third-woman-ever-20260723"><em>https://quantamagazine.org/hong-wang-wins-2026-fields-medal-the-third-woman-ever-20260723</em></a></p><p><a href="https://www.remio.ai/post/terence-tao-ai-math-warning-challenges-the-benchmark-race"><em>https://www.remio.ai/post/terence-tao-ai-math-warning-challenges-the-benchmark-race</em></a></p><h3>The complete reference</h3><ul><li>OpenAI released 722 manuscripts in 372 result families, all produced by an unreleased internal frontier model.</li><li>The results came from an evaluation that fed the model about 4,000 open research problems after existing math benchmarks saturated.</li><li>Each result used, on average, compute roughly equal to 3 hours of ChatGPT Pro thinking, and most were turned out in less than two weeks.</li><li>235 of the 372 families come with a Lean formalization notes page; OpenAI says the unformalized results may have problems.</li><li>Headline claims include θ = 7/8 for the quasi-Riemann hypothesis, the 3D Kakeya maximal function and 4D Hausdorff dimension conjectures, and the Unique Games Conjecture.</li><li>25 Fields medalists, including Terence Tao, warned that rushed announcements create attribution and plagiarism problems.</li><li>The IAS recommendations have two steps, “initial release” and “supporting understanding”; OpenAI’s release earns a selective pass and still withholds the model’s name.</li><li>Lean can guarantee error-free reasoning; checking that a statement is the one mathematicians have in mind still needs people.</li><li>The scarce resource is now expert attention: the time it takes to understand the new theorems.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1c88c195ff2b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LLM Interview Note: How to Fix an Agent’s Overflowing Context?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@gene.rastignac8/llm-interview-question-how-to-fix-an-agents-overflowing-context-1ef656a351aa?source=rss-5e601784a999------2"><img src="https://cdn-images-1.medium.com/max/800/1*K7YTn8M-87YXUQuiRgfJ-g.png" width="800"></a></p><p class="medium-feed-snippet">6 compression techniques worth knowing when a bigger context window falls short</p><p class="medium-feed-link"><a href="https://medium.com/@gene.rastignac8/llm-interview-question-how-to-fix-an-agents-overflowing-context-1ef656a351aa?source=rss-5e601784a999------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@gene.rastignac8/llm-interview-question-how-to-fix-an-agents-overflowing-context-1ef656a351aa?source=rss-5e601784a999------2</link>
            <guid isPermaLink="false">https://medium.com/p/1ef656a351aa</guid>
            <category><![CDATA[llm-interview-questions]]></category>
            <category><![CDATA[harness-engineering]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[llm-applications]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Eugène Rastignac]]></dc:creator>
            <pubDate>Wed, 14 Oct 2026 11:48:25 GMT</pubDate>
            <atom:updated>2026-10-14T17:53:04.218Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Jev Beaten by an Open-Source Model? Here Is How to Read That Benchmark]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@gene.rastignac8/jev-beaten-by-an-open-source-model-two-years-of-efforts-vanish-overnight-fb098470497d?source=rss-5e601784a999------2"><img src="https://cdn-images-1.medium.com/max/917/1*DaLv_oKWP4c4Jk4etPMY9w.png" width="917"></a></p><p class="medium-feed-snippet">What the Jev and Laya comparison actually measures, and which numbers survive when you run the same decisions yourself.</p><p class="medium-feed-link"><a href="https://medium.com/@gene.rastignac8/jev-beaten-by-an-open-source-model-two-years-of-efforts-vanish-overnight-fb098470497d?source=rss-5e601784a999------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@gene.rastignac8/jev-beaten-by-an-open-source-model-two-years-of-efforts-vanish-overnight-fb098470497d?source=rss-5e601784a999------2</link>
            <guid isPermaLink="false">https://medium.com/p/fb098470497d</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[llm-reasoning]]></category>
            <category><![CDATA[jevs]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[laya]]></category>
            <dc:creator><![CDATA[Eugène Rastignac]]></dc:creator>
            <pubDate>Tue, 22 Sep 2026 23:45:22 GMT</pubDate>
            <atom:updated>2026-09-24T13:03:30.886Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Two Agents Can Argue Forever. Here Is Where the Stop Button Actually Belongs.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@gene.rastignac8/two-agents-can-argue-forever-here-is-where-the-stop-button-actually-belongs-1495e2038608?source=rss-5e601784a999------2"><img src="https://cdn-images-1.medium.com/max/902/1*f7loW9xV7yNPdYYwQqd8rQ.png" width="902"></a></p><p class="medium-feed-snippet">Why a simple prompt can never end a loop, and what an orchestrator has to own instead.</p><p class="medium-feed-link"><a href="https://medium.com/@gene.rastignac8/two-agents-can-argue-forever-here-is-where-the-stop-button-actually-belongs-1495e2038608?source=rss-5e601784a999------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@gene.rastignac8/two-agents-can-argue-forever-here-is-where-the-stop-button-actually-belongs-1495e2038608?source=rss-5e601784a999------2</link>
            <guid isPermaLink="false">https://medium.com/p/1495e2038608</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[multi-agent-systems]]></category>
            <category><![CDATA[interview-questions]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Eugène Rastignac]]></dc:creator>
            <pubDate>Sun, 20 Sep 2026 16:12:44 GMT</pubDate>
            <atom:updated>2026-09-22T23:50:23.714Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Why are LLM Sizes Commonly Designed as 6/7B, 13B, and 130B?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@gene.rastignac8/why-are-llm-sizes-commonly-designed-as-6-7b-13b-and-130b-0f2bbdbd6dc5?source=rss-5e601784a999------2"><img src="https://cdn-images-1.medium.com/max/600/0*_BdFga1DEsyIWXaA.png" width="600"></a></p><p class="medium-feed-snippet">Analysis of Scaling Laws in Large Language Models</p><p class="medium-feed-link"><a href="https://medium.com/@gene.rastignac8/why-are-llm-sizes-commonly-designed-as-6-7b-13b-and-130b-0f2bbdbd6dc5?source=rss-5e601784a999------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@gene.rastignac8/why-are-llm-sizes-commonly-designed-as-6-7b-13b-and-130b-0f2bbdbd6dc5?source=rss-5e601784a999------2</link>
            <guid isPermaLink="false">https://medium.com/p/0f2bbdbd6dc5</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[scaling-law]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Eugène Rastignac]]></dc:creator>
            <pubDate>Tue, 14 Jan 2025 16:14:41 GMT</pubDate>
            <atom:updated>2025-01-14T16:14:41.542Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[It has gone viral! Check 3Blue1Brown’s latest video on how LLMs store facts.]]></title>
            <link>https://medium.com/@gene.rastignac8/it-has-gone-viral-check-3blue1browns-latest-video-on-how-llms-store-facts-4eb3b7fbe75b?source=rss-5e601784a999------2</link>
            <guid isPermaLink="false">https://medium.com/p/4eb3b7fbe75b</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[tutorial]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[3blue1brown]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Eugène Rastignac]]></dc:creator>
            <pubDate>Wed, 04 Sep 2024 07:56:58 GMT</pubDate>
            <atom:updated>2024-09-04T07:56:58.472Z</atom:updated>
            <content:encoded><![CDATA[<p>When typing “Michael Jordan plays the sport of ___” into a Large Language Model (LLM), it will likely predict the next word to be “basketball.”</p><p>This indicates that the model’s billions of parameters contain knowledge related to this specific individual. Users might even feel that these models have memorized a vast amount of facts.</p><p>But what is actually happening?</p><p>Recently, 3Blue1Brown updated the 7th chapter of its course “Deep Learning”. In this chapter, they illustrated vividly the way how LLMs store facts through detailed animations. The video has garnered an impressive 180,000 views.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vk33nwgR7221YfhX" /></figure><p>In december of last year, researchers from Google DeepMind published a related paper, where one of the specific examples involved matching athletes with their respective sports.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4qCqAqCQcp10LHSn" /></figure><p>Although this paper did not fully answer the question of how LLMs store facts, it did yield some rather intriguing results. One key finding is that facts are stored in a specific part of the network, known as the Multi-Layer Perceptron (MLP).</p><p>In this newly updated 23-minute video by 3Blue1Brown, they demonstrate how LLMs store and process information, covering the following topics:</p><ul><li>what are the facts “hidden” in LLMs</li><li>quick review of Transformers</li><li>examples</li><li>multi-layer perceptrons</li><li>how to calculate the number of parameters</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F9-Jl0dxWQs8%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D9-Jl0dxWQs8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F9-Jl0dxWQs8%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/7096ff52eb08c4408e32f7811fb7a418/href">https://medium.com/media/7096ff52eb08c4408e32f7811fb7a418/href</a></iframe><p>In this demonstration video, the speaker’s articulate and precise language, along with high-definition graphics, gives the reader a good understanding of how LLM stores knowledge.</p><p>Many users were amazed by the teaching quality of 3Blue1Brown after watching the video:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vLl0XdYTCBCHRORV" /></figure><p>Some even mentioned that they had been eagerly awaiting this video update for a long time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YLNwxziPUUtvmTTq" /></figure><p>Let’s delve into the details of MLPs. In this article, I will provide a brief overview of the core content of the video. Interested readers can refer to the original video for the full explanation.</p><p>MLPs occupy a significant portion of LLMs, but their structure is much simpler compared to the mechanisms such as attention. However, they can still be somewhat challenging to understand.</p><p>To simplify, let’s continue with the specific example of “Michael Jordan plays basketball.”</p><p>First, let’s briefly explain how Transformers work. The training objective of a Transformer is to predict the next token (usually representing a word or phrase) based on existing tokens, with each token associated with a high-dimensional vector.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xOiMXBNk979rQnqk" /></figure><p>These vectors repeatedly undergo two main operations: attention (which allows vectors to exchange information with each other) and the MLP. Additionally, there’s a specific normalization step between them.</p><p>After several iterations of such operations, each vector is expected to have absorbed sufficient information. This information may come from general knowledge embedded in the model’s weights during training and./or come from the surrounding context. This knowledge is what the model relies on to predict the next token.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*XbxhFllNAKH0piQu" /></figure><p>It’s important to note that these vectors encode more than just a single word; they absorb richer meanings as information flows through the network, influenced by the context and the model’s knowledge.</p><p>In essence, each vector encodes far more than the meaning of a single word, enabling the model to predict what comes next. The MLP stores this information (while attention combines the context), which is why most of the model’s parameters (about 2 / 3) are located within the MLP.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JdZddhcSjqH1PIrx" /></figure><p>Let’s now see how the MLP stores this specific “Michael Jordan plays basketball” fact.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_qrY5270LG-giC8X" /></figure><p>Assume there are three different vectors in high-dimensional space, defining Michael Jordan’s last name (“Jordan”), his first name (“Michael”), and “Basketball.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QoFuoGg_QRcAiW4u" /></figure><p>If there’s a vector in this space whose product with the “Michael” vector equals 1, we consider that vector to encode the concept of “Michael;” If the product is 0 or even negative, it’s considered unrelated to “Michael.”</p><p>Similarly, we can compute the product of that vector with “Jordan” or “Basketball” to determine its relation to those concepts.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eNcqvQr3UNzsemb-" /></figure><p>Through training, if the vector’s product with both “Michael” and “Jordan” equals 1, we can consider that vector to encode the overall concept of “Michael Jordan.”</p><h4>Internal Details of the MLP</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/926/0*FBTLG5Bnvhh3utbP" /></figure><p>When this sequence of vectors (encoding the aforementioned text) passes through an MLP module, each vector in the sequence undergoes a series of operations:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HLOYWwizNLJxUJSZ" /></figure><p>After that, a vector with the same dimensions as the input vector is obtained. This resulting vector is then added to the input vector to form the output vector.</p><p>Each vector in the sequence undergoes these operations independently and in parallel, meaning they do not influence each other at this stage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6KyieshDMKYXGONe.png" /></figure><p>For “Michael Jordan plays basketball,” we hope that for the input “Michael Jordan,” after a series of operations, we can get the vector of “basketball.”</p><p>Let’s first look at the first step of the process. It is a linear projection process: the vector is multiplied by a large matrix, where the matrix’s data represents the model’s parameters. This can be thought of as a control panel filled with knobs — by adjusting these parameters, we control the model’s behavior.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*023oJFdQU4uIX50C" /></figure><p>For matrix multiplication, the video shares a perspective. We can think of matrix multiplication as treating each row of a matrix as a vector, and then performing a series of dot productions of those rows with the vector being processed (denoted here by E, meaning embeding).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*sckumOpdmD1qJz_E.png" /></figure><p>If we assume that the first row of the matrix encodes exactly the First Name “Michael” and that the processed vectors also encode it, then the resulting dot product is about 1.</p><p>If they encode both the first and the last names, then the result should be approximately 2.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-HIr6n-4EtOicO-9.png" /></figure><p>You can assume that the other rows of that matrix are working on other problems in parallel.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/805/0*tOT5uMMA05ko_9bh.png" /></figure><p>Additionally, this step usually adds another vector to the output, known as the bias vector, where the parameters are learned from the data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*dyMQcF_4E0FPZLWM.png" /></figure><p>In our example, this bias vector’s first element is -1, meaning the related dot product in the final output vector is reduced by 1. This operation ensures that only when the vector encodes the full name “Michael Jordan” will the first component of the output be positive; otherwise, it will be 0 or negative.</p><p>In practice, the size of the model’s matrix is massive. For instance, GPT-3’s matrix has 49,152 rows and 12,288 columns (The number of columns is the number of dimensions of the embedding space).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UuYmXh07oCS45oZX.png" /></figure><p>You may notice that the row count is exactly four times the dimensionality of the embedding space — this is actually a design choice that can vary.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qgAEsTkuGMH8zEe-.png" /></figure><p>We then represent this matrix and vector in a more concise way, as shown in the following animated diagram:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*r5oKTP74jHm0lUJ3.gif" /></figure><p>After the linear process described above, the output vector is processed by a simple function, typically the Rectified Linear Unit (ReLU), which truncates all negative and zero values to zero, leaving a clean result. This behavior is similar to an AND gate, where the output is 1 only when the input is the full name “Michael Jordan.” ReLU also has a smoother version called GeLU.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zKjiPByVMzjAnWlf" /></figure><p>Next comes another linear projection step, which is very similar to the first one: multiplying by a large matrix and adding a bias to produce the output vector.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-fq90mzukAFL_Xnv.png" /></figure><p>However, this time, we think of the matrix columns as representing various concepts. These columns have the same dimensions as the embedding space.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*dSTWT-45wMV32CkX.png" /></figure><p>If the first column represents “Basketball” and n_0 is 1 (indicating that the neuron is activated), that result will be added to the final result; otherwise it does not affect the final result. Of course, these columns can also represent any concept.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*BifbXUF5NVunEIHO.png" /></figure><p>Similarly, we denote this large matrix as W↓, denote the bias as B↓, and put it back in the diagram.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QUKhoquBLFRUCqbb.png" /></figure><p>If the input vector encodes both the first name “Michael” and the last name “Jordan”, then triggering the sequence of operations will result in an output vector pointing in the direction of “Baskerball.”</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xDhIydci6T3vdzOZ.png" /></figure><p>This procedure is executed in parallel for all vectors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6DrW_HNQO0dL0b3K.png" /></figure><p>This is how MLP operates: two matrix multiplications, each adding a bias. In earlier machine learning era, this kind of network has been used to recognize handwritten digits with reasonable success.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*R8JjfPQkhtgUcOov" /></figure><h4>How do we know that GPT-3 has 175B parameters?</h4><p>The next part of the video covers how GPT-3’s parameters are calculated and where do they come from.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/996/0*SWLvKbNVnkoPSiJo" /></figure><p>For GPT-3, the embedding space size is 12,288, and multiplying them gives over 600 million parameters for just this matrix. The downward projection (the second matrix) has the same number of parameters (with shape transposed), so together they account for about 1.2 billion parameters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fHU9aWLYdTAsnAdN" /></figure><p>The speaker also indicates that several other parameters need to be considered, but this is a very small and negligible portion of the total. More importantly, the sequence of embedding vectors flows through not one MLP, but 96 different MLPs, leading to over 100 billion parameters — about <strong>two-thirds</strong> of the total parameters in the network.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*S2Md4u584xnd0Iv1" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*7fR1vywmYAanU3Ss" /></figure><p>Finally, when we add up all the parameters from the attention blocks, embeddings, and unembeddings, the total comes to 175 billion parameters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*BdEgVt3ZoQwMEGLB" /></figure><p>It’s also worth mentioning that there is another set of parameters related to the normalization steps. However, these parameters were skipped in the video demonstration because they make up only a small fraction of the total parameter count.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1SrnQYdyywudcTLQ" /></figure><p>The video concludes by introducing the concept of Superposition. Evidence suggests that single neurons rarely represent a single feature such as “Michael Jordan.” This relates to a hypothesis known as Superposition, which may help explain why these models are particularly challenging to interpret and why they scale so well.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DEdZnjv4YwEUClD9" /></figure><p>For more detailed information, interested readers can refer to the original video.</p><p>3Blue1Brown (3b1b) is a channel dedicated to creating visualized educational videos, covering topics such as mathematics and artificial intelligence. Each lesson is accompanied by intuitive and lively animations that help viewers deepen their understanding of concepts and theorems.</p><p>In addition to its 6.4 million subscribers on YouTube, 3b1b also has an official account on Bilibili (video straming giant in China) with over 2.15 million followers. Each video garners over 100,000 views, and some teachers even play these videos in their classrooms. For a hardcore educational content creator, these achievements are nearly unparalleled.</p><p>As a site long recognized for its popular approach to understanding mathematical concepts, 3Blue1Brown has consistently excelled in visualizing ideas. In this latest video, we can clearly see this strength in action.</p><p>3b1b’s creator, Grant Sanderson, graduated from Stanford University with a degree in mathematics. He independently produces most of his videos and animation engines, which began as a side project during his time at Stanford.</p><p>At Stanford, Grant “took a slight detour into computer science,” and after graduating, he joined Khan Academy, where he worked as a math instructor for two years. In 2016, he decided to fully commit to 3b1b.</p><p>If you’re ever feeling uncertain about your learning abilities, watching 3b1b’s content might be a good idea. The fully animated demonstrations ensure you grasp the concepts thoroughly.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4eb3b7fbe75b" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>