Three threads run through today’s issue.
Anthropic’s amended stock filing says it may soon reach its own definition of AI speeding up AI research, though it has not named a current model that qualifies. Third-quarter growth slowed to about 47%, against roughly $895.4 billion in compute and lease commitments.
OpenAI is the second thread: a GPT-6.2 Astra model ID has appeared in Codex usage with no announcement, and a blogger says the Astra paper’s headline trick is something a skilled human would have done.
The third is how things get built. Nine Microsoft engineers merged 1,211 code changes in a day, and Rohan Anil’s team trained one model across data centers 9,000 miles apart, while humans& shaved a training-update freeze to 0.23 seconds.
Anthropic’s Prospectus: A Slowing Curve and a Self-Building Warning
Two readings of the same amended filing: what the money says, and what the company says about AI building AI.
- Anthropic’s third-quarter filing shows $17 billion in sales and slowing growth. Anthropic’s amended filing puts third-quarter revenue near $17 billion, but growth over the prior quarter fell to about 47%, from 144% in the second. It also lists roughly $895.4 billion in compute and lease commitments and a $23.29 billion third-quarter net loss.
- Anthropic’s IPO filing says AI may soon speed up its own research. The revised filing says Anthropic may imminently meet its own definition of AI-accelerated progress, while stating that none of its current models has been designated as crossing the line. TechCrunch notes that a filing is also a marketing document, which cuts both ways.
OpenAI Watch: A Hidden Model, a Home-Life Test and a Skeptic
A model nobody announced, a benchmark OpenAI wrote for itself, and a blogger who thinks the self-improvement excitement is overdone.
- A new OpenAI model, GPT-6.2 Astra, is turning up in Codex usage. Usage screenshots collected by TestingCatalog show a gpt-6.2-astra ID in Codex, taking anywhere from 1.6% of one user’s week to 37.7% of another’s day. OpenAI has not commented, and the model picker still reads GPT-6 Astra.
- OpenAI built its own test for everyday chores and posted example scores. LifeVal covers childcare, budgeting and government paperwork, and OpenAI published five example scores between 90.4% and 100%. It named no models and gave no overall result, and OpenAI designs and grades the test itself.
- A blogger says AI editing another AI’s brain is not a superhuman trick. Sean Goedecke argues the headline result in the GPT-6 Astra paper, improving a small Qwen model’s math by editing its weights, is a routine technique a skilled human would have used. In his reading it is no sign of runaway self-improvement.
Agents at Work: Shipping Fast Without Losing Control
Four teams describe what it takes to build with AI agents: speed, safety and coordination.
- Nine Microsoft engineers merged 1,211 code changes in a single day. Microsoft says a small team shipped Copilot Cowork worldwide in four and a half months, with agents writing most of the code, then swapped its Python engine for Rust. The account is the team’s own, with no outside audit.
- Framer’s design AI now edits your site while it is still typing. Framer’s agent runs the code the model writes as it arrives and cancels early when a check fails. Framer says that makes it up to 33% cheaper and 20% faster, and cites no independent test.
- GitHub gives AI agents a way to use private code without repeating it. GitHub Next’s enclaves let an agent consult a private repository while limiting what it can pass back to a few bits, such as a yes or no. The post admits that even one bit can answer a sensitive question.
- Teaching AI agent teams to split work by making speed pay. General Reasoning says a deadline that stretches during training, plus a shared message board, taught its Avalon-M1 teams to divide the work and overtake a lone agent. The results are the lab’s own, and the full paper is still weeks away.
How Models Get Made: Distance, Delay and Moral Advice
Three training reports, from spreading one run across continents to teaching an open model right from wrong.
- One AI model trained on two data centers 9,000 miles apart. Rohan Anil’s team trained a 32-billion-parameter model on clusters in the US and Asia and says it scored about 1% better than a single-site run. The numbers are the team’s own, and the post warns that mixing sites makes quiet tampering harder to catch.
- humans& cuts the pause in AI training updates to under a quarter second. The lab says its weight teleportation method refreshes a model that learns from its own answers in about four seconds, with answer generation frozen for only 0.23 seconds. The figures come from its own tests and lean on NVIDIA’s Blackwell chips.
- Ethics advice cut an open AI model’s bad behavior tenfold. Fiction barely did.. Second Look Research copied Anthropic’s recipe on Qwen models and found 3,000 transcripts of ethical advice cut harmful agent behavior from 61% to 6%, while fictional stories only got it to 43%. The authors say the comparison cannot fully explain why.
Today’s Quick Hits
- Anthropic makes long AI agent sessions respond about 35% faster. Anthropic’s message threads, now in public beta, stop agents from resending the whole conversation at every step. Anthropic says the wait for a first reply on very large requests drops about 35%, but a thread lapses after five idle minutes.
- ChatGPT is dropping its separate Chat and Work tabs. OpenAI is merging the two tabs into one screen for a small test group first. ChatGPT and Codex will also share a single usage allowance, so a long coding session can eat into everyday chat.
- Musk says Grok Bot got cut back because it cost too much to run. Elon Musk told a Dogecoin fan account on X that Grok Bot was limited because serving it cost more than the revenue it brought in. He gave no figures and did not say what the new limits are.
- NVIDIA’s Cosmos can now fake a car’s camera and LiDAR feeds together. NVIDIA’s new Cosmos tools generate matching camera video and LiDAR laser readings for simulated driving. A reference run on four GPUs produces seven 720p videos and 99 LiDAR sweeps, and the release adds robot-arm and surgical variants.