131,072 weights. According to MIT CSAIL researchers, that's how many numbers GPT-6 Astra changed inside a small Qwen model to lift its MATH-500 score from 7.9% to 48.1%. The same Monday, Anthropic told prospective shareholders it may soon cross its own threshold for AI speeding up AI research.
Neither story means the machines are building themselves. Together they make recursive self-improvement feel a lot less like a seminar topic. The rest of the day's news was about the limits the industry is putting around that capability: how much of it you can use, what you're allowed to build with it, and where its data comes from.
Anthropic put its RSI warning in the one document lawyers read#
The headlines on Anthropic's revised S-1/A were about money: revenue soaring, expenses soaring too. TechCrunch found the more interesting part in the risk factors. Anthropic already uses AI in its own R&D, and the filing now says this:
“...we have in the past not identified, and may in the future not identify, a significant acceleration in our rate of development until after it has occurred.” — Anthropic's S-1/A
It goes further. Anthropic says its internal estimates for future models, not current ones, indicate it “may imminently meet” its Responsible Scaling Policy's definition of “dramatically accelerated progress attributable to AI.” No current model has been designated as over that line. The company says it is considering changes to the policy to keep humans in control, and that it doesn't expect to safely let Claude fully design and build its successors without meaningful human supervision.
Yes, an S-1 is also a sales document, and “our models are almost too good” is a convenient risk to disclose. But risk factors are written by people who get sued over them. The sentence I take most seriously is the one about the past. If the lab with the most detailed scaling policy in the industry has already missed an acceleration while it was happening, nobody downstream of it will spot one in real time either.
The MIT CSAIL work reported by VentureBeat shows a much smaller and easier-to-read version of the same loop. Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab had GPT-6 Astra inspect the weights of Qwen3.5-0.8B and propose direct edits, with no gradient training. The results:
MATH-500: 7.9% → 48.1%
GSM8K: 13.8% → 67.1%
IFEval: 21.0% → 62.3%
Edit size: about 0.016% of weights, in layers 15 and 19
Read the fine print before you get excited. The gain came from removing a direction in the weights tied to repetition: the share of MATH-500 answers hitting the token limit fell from 93% to 33.1%. The model learned no new math. The same approach did nothing for Gemma 3 270M, which didn't have the looping problem, and the best run took 20.3 hours. Even so, one model diagnosing another's internals and shipping a working patch is a new kind of tool, and the logs are on GitHub. I'd file it under interpretability that pays off, not under doom.
Grok Bot's Halloween throttle shows what flat-rate agents really cost#
Over Halloween weekend, some Grok Bot users woke up to find their weekly usage had jumped from as low as 50% to 100%, without them using the agent overnight. Musk confirmed in an X reply that it wasn't a glitch: “The cost of serving Grok Bot was unsustainably high relative to revenue.” TechCrunch notes he hasn't said what the limits will be from now on. This is the product he called the fastest-growing he's ever seen.
OpenAI made a quieter version of the same move. It is merging ChatGPT's Chat and Work tabs on desktop, and ChatGPT and Codex will now draw from one shared usage allowance. OpenAI says most people won't notice, though some will hit their limits more often. If you live in Codex, your coding budget now competes with your chats.
Then there's Garry Tan, who told Walter Isaacson's podcast that he spends roughly $100,000 to $200,000 a year on tokens. He also says costs have to come down before personal agents go mainstream. Flat-rate subscriptions were priced for chat, and agents that run for hours break that math. Expect more limits announced in Sunday replies on X rather than on pricing pages.
If a workflow you depend on runs on a subscription quota, set up a metered API key as a fallback this week.
Publishers and Anthropic both draw lines around other people's work#
The New York Times sued Brave in federal court in Manhattan on Monday. The Daily News, Chicago Tribune and Orlando Sentinel filed a similar suit the same day. According to The Verge and TechCrunch, the publishers allege that Brave built an index of more than 30 billion pages with a crawler they can't block, then sold access to AI companies, naming Anthropic, Perplexity, Mistral, xAI and Cohere. The suits also go after the browser's paywall blocker. The publishers' attorney, Steven Lieberman, summed up the theory: “Buying stolen goods secondhand doesn't make them legitimate.” Brave didn't immediately comment, and Perplexity dismissed the suit outright.
For builders, this is the suit to watch. Many teams ground their apps on a search API because it feels like a clean, licensed layer between them and the open web. If a court accepts that the intermediary is laundering content, where your retrieval data comes from becomes your problem, not just your vendor's.
Anthropic is drawing its own line from the other side. Claude Code now carries a rule, found in version 2.1.314 per Ars Technica, telling Opus 5.5 not to build “a recreation of a specific existing game, even if it is renamed or called original.” Anthropic's Addy Osmani confirmed the change, said original and genre games should keep working, and asked people to send him false refusals. Ars points out that the AI decompilation wave is already big enough that decomp.dev shut down last month. A system-prompt rule is a speed bump, not a wall, but it signals that model makers now see IP misuse as their liability too. If you prototype games from your own designs, keep your assets and specs in the repo so the intent is obvious.
Also worth a click#
OpenAI's Grid Protectors pilot promises to free up to 1,000 megawatts during emergencies at a planned data center in Effingham County, Georgia.
Trajectory raised $45M at a $300M valuation to let non-engineers train models.
MIT Technology Review tracks who has adopted the White House's “Super Intelligence” rebrand. So far, mostly Zuckerberg and Musk.
All three of the day's threads are about control: how fast models improve, how much of them you can use, and what they're allowed to copy. My bet is that Anthropic's promised policy update will define “meaningful human supervision” before the IPO prices, and that this definition will matter more to people building on Claude than any benchmark released this month.