In partnership with
In Today’s Issue:
🧬 GPT-6 Astra patches another AI's weights, no training needed
📄 Anthropic's IPO filing says future models may soon hit its acceleration threshold
🏠 OpenAI's LifeVal tests AI on real-life chores
📉 The chip rout reaches $3.7 trillion
🤖 Why US humanoid projects keep turning to Chinese hands
✨ And more AI goodness…
⚡ The Signal
AI working on AI moved from slogan to paperwork on Monday, in two very different documents.
At MIT, researchers had GPT-6 Astra diagnose a looping fault inside a small Qwen model and patch the numbers directly, with no training run. In an SEC filing, Anthropic told future shareholders that its expectations for future models indicate it "may imminently meet" its own definition of dramatically accelerated, AI-driven progress. A lab demo and a risk factor are different kinds of evidence, and what they share is a direction: models are starting to do the work of improving models. Investors, meanwhile, are asking what all of it costs, and chip stocks have lost about $3.7 trillion since October 6. AMD reports today after the close.
All the best,
Kim Isenberg
Excerpt from the risk factors in Anthropic's amended S-1 (SEC)
📄 Anthropic Warns Investors: Its AI-Acceleration Threshold May Be Close
Anthropic has told investors it may soon reach its own threshold for AI-accelerated progress. In an amended IPO filing dated November 2, it says expectations for future models, "rather than our current models," indicate it "may imminently meet" the Responsible Scaling Policy definition of "dramatically accelerated progress attributable to AI." It also concedes it has noticed past speedups only after they happened, and is weighing rules that would require human review of certain research decisions as Claude takes on more of that work.
👉 tl;dr: Anthropic says future models could speed up its AI research enough to trigger its own safety policy, and it is telling shareholders before the IPO.
LifeVal results: average score against estimated API cost per attempt (OpenAI)
🏠 OpenAI's LifeVal Tests Whether ChatGPT Can Handle Real Life
OpenAI has published LifeVal, a benchmark of whether AI can actually get everyday jobs done. Built from two large time-use studies and a new survey of 3,995 people, it covers areas from caregiving and personal finances to dealing with government services. The best average score rose from about 49% for GPT-5.2 to about 83% for GPT-5.6, with GPT-6 Astra at 82%. By OpenAI's own numbers, GPT-6 Astra is about 23% faster than GPT-5.6, but GPT-5.6 gets similar results at roughly half the estimated API cost.
👉 tl;dr: On OpenAI's own test of real-life errands, such as comparing kindergartens or planning meals after a diabetes diagnosis, its models have improved sharply, though GPT-6 Astra costs about twice as much as GPT-5.6 for a similar score.
Market value lost since October 6 (Yahoo Finance)
📉 Intel Sinks Again as the Chip Rout Reaches $3.7 Trillion
Chip stocks lost more than $400 billion on Monday even as about 7 in 10 S&P 500 stocks rose, by Yahoo Finance's count. Intel fell about 9%, after a nearly 29% plunge on October 30 on a weak forecast, its worst day since 1974. Memory and storage names dropped roughly 6% to 8%, and the group has shed about $3.7 trillion since October 6. Last week's tech slump came as investors scrutinized how quickly huge AI spending turns into profit.
👉 tl;dr: Investors are selling many of the companies that supply the AI boom's hardware, and AMD's results on Tuesday after the close are the next test.
Let world-class chefs answer your cravings.
Tired of the same meal prep on repeat? CookUnity can help. Because they connect you with 500+ meals handcrafted by over 130 world-class chefs, including Michelin-starred chefs and Food Network stars.
Explore 60+ different cuisines and diets, and choose dishes that fit your cravings and preferences. Your meals arrive fresh and ready to heat in minutes, so no prep or cleanup.
Whether you're in the mood for a comforting classic or a new flavor, there's always something new to discover because the menu rotates weekly, and their chefs are always creating new recipes.
Explore the menu available in your area and build a weekly plan you can skip, pause, or adjust anytime.
🎬 Watch This
Katelyn Lesse and Angela Jiang, Anthropic's heads of engineering and product, walk Latent Space through how Claude becomes a product, including rebuilding the Messages API from Python to Rust with Claude and checking the rewrite against shadow traffic. They also get into the work still left for human engineers. Published October 27, the episode opens with swyx on Anthropic's S-1 and its improvement loops, so it pairs well with today's filing story.
"AI was mostly not as smart as our smartest scientists, and now it's mostly smarter than them."
– Sam Altman, CEO of OpenAI, at ADIPEC in Abu Dhabi on Monday
That is his own account of this year's change. In the same conversation he insisted there "should not be a handoff to AI," and added: "I can't believe that this differentiates us from other AI companies, but it does."
Two frontier models may already be peeking out. gpt-6.2-astra appeared in some Codex users' usage analytics on October 31 while the model picker still said GPT-6 Astra, and a hidden Claude Mythos 5.5 mention was spotted on Anthropic's website. OpenAI has announced nothing; TestingCatalog reads it as quiet routing, either an A/B test or a staged rollout.
Become an email marketing GURU.
Is your email strategy due for a tune-up? GURU Conference (powered by Constant Contact) is back Nov 12–13, 100% free and virtual. You'll get email tactics you can steal the same day from top B2B and B2C marketers. Last year, 29,000+ marketers showed up. Save your free spot.
GPT-6 Astra Found a Flaw Inside Another AI, Then Rewrote Its Weights
The Takeaway
👉 MIT researchers asked GPT-6 Astra to read and edit the raw weights of other neural networks, without gradient descent or training.
👉 Astra changed 0.016% of Qwen3.5-0.8B's weights, lifting its MATH-500 score from 7.9% to 48.1%.
👉 It also wrote small game-playing networks from scratch that beat 100 million steps of conventional training in all three games.
👉 The Qwen fix removed a loop rather than adding math skill, and Gemma 3 270M, which rarely loops, saw no real gain. Still, the authors see a path toward recursive self-improvement.
This could become a huge breakthrough! Neural networks are normally trained, and their millions or billions of numbers are too opaque for people to edit directly. MIT researchers now report that GPT-6 Astra can do it. Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab call it Weight-Space Programming: a general-purpose model reads and changes a network's weights, the numbers that determine its behavior, without gradient descent, the usual training method. Reading only the weights and a plain description of the inputs, Astra named the game each of four networks had been trained to play, with no list of options. Databricks' Yuchen Jin summed it up on X: "It's like gene editing for neural networks."
Reading, fixing and improving networks: three results from the post (Takehi et al., MIT CSAIL)
Qwen3.5-0.8B is the headline case. The small Alibaba model tends to repeat itself until it runs out of room: 93% of its MATH-500 answers hit the token limit. Over 158 turns and about 20 hours, Astra traced the loop to one direction in two layers and removed it by editing 0.016% of the weights. Math scores rose from 7.9% to 48.1%, and instruction following tripled. Hugging Face's Ahsen Khaliq rebuilt the edit and measured 51.0%. Then Astra wrote networks from scratch for three small games, without ever playing them. Its Breakout network scored 517.4, against 25.3 for one trained with PPO, a standard method, for 500 million steps.
Astra's one-shot networks against PPO in three games (Takehi et al., MIT CSAIL)
For now, an edit can remove an obstacle hiding an ability but cannot add one the model lacks. Still, the authors believe this "marks the beginning of a dramatic shift." Models could one day add or remove capabilities on request, or audit a network for backdoors from its weights alone. The biggest prospect is recursive self-improvement: a model that can inspect, debug and improve its own weights, or a newer model's, offers what they call "a richer, more native path" there than writing code that launches training jobs. Voyage AI's Chandan Singh says the authors are "underselling" the approach, while PauseAI's Maxime Fournes calls it "a terribly reckless research direction."
Why it matters: If frontier models can diagnose and patch networks directly, fixing an AI starts to look more like debugging software than running a new training job. Done right, it is also a step toward recursive self-improvement, a more direct version of what Anthropic's IPO filing calls "closing the loop."
Sources:
The chart: Hours after the MIT post, Xingyu Dang ran a Writing Weights Index: models write small game-playing networks from scratch, scored over three games, where 50 equals PPO, a standard training method, after 500 million steps. GPT-6 Astra leads at 60.7, ahead of Gemini 4 Carbon at 57.7 and Claude Opus 5.5 at 51.2. Kimi K3.1 Pro reaches 29.8, Qwen3.8 12.5 and DeepSeek V4.1 Flash 10.0.
The lesson: Astra's result in the Featured Story is no one-model fluke: three frontier models out-score half a billion steps of conventional training here. The drop below them is steep, and no Chinese model on the chart reaches half of Astra's score. If writing weights becomes a practical way to improve models, the labs already in front gain another advantage.
The caveat: This is one person's run on X, separate from the MIT authors, who say their own benchmark is still being built.
🤖 America's Humanoid Push Keeps Reaching for Chinese Hands
⚡ Bottom line: Nvidia's new reference humanoid uses Chinese-made Wuji hands, and Meta reportedly built its prototype with Shanghai's Fourier.
💡 Why it matters: Hands are the hard part of making humanoids useful, and US projects still lean on Chinese suppliers for them.
🔎 What it means: US firms can lead on chips and AI software, yet the physical robot race still runs through Chinese makers.
A humanoid robot is only as useful as its hands, and the hands in some of America's biggest robot projects keep coming from China. When Jensen Huang presented Nvidia's new humanoid training system in Berlin on October 21, the slide behind him featured the Wuji Hand 2, from a startup with much of its team in Shenzhen, CNBC reports. Nvidia's first reference design, announced in May, paired a body from China's Unitree with hands from Sharpa, whose manufacturing R&D is in Shanghai.
The five-finger Wuji Hand 2 (Wuji Tech)
The reason is plain: walking gets the demos, but useful work means gripping, turning and placing things. Zhu Mingxuan of the humanoid startup JoyIn, who left Figure last year, told CNBC that hardware is limiting humanoids and that hands are the main challenge. Even the best-known makers have limits. Google DeepMind's Keerthana Gopalakrishnan said a Wuji hand has the strength of a ten-year-old, with Sharpa's a bit stronger.
The dependence goes beyond hands. Meta built a prototype called Max with Shanghai's Fourier in about three months, The Information reports, and a Meta vice president spent almost two months at Fourier's factory. Meta says Max is for internal research and not for sale, but work on a second version began in September.
Agility's Digit 5 humanoid (Agility Robotics via The Robot Report)
The American counterexample reaches Nasdaq this week. Agility Robotics is expected to start trading as AGLT on Thursday, raising about $465 million, down from the more than $620 million expected in June, with a private financing led by Foxconn. Its Digit robots already work in nine customer environments, including moving totes onto conveyors at a GXO warehouse in Georgia. The finer work Nvidia's platform targets relies on hands like Wuji's, and those come from China.
