AI Brief, November 3: LLMs editing weights, Anthropic nears RSI, Claude Code blocks clones

Weight-Space Programming claims LLMs can edit model weights directly, Anthropic warns it is near RSI, and Claude Code blocks game clones.

The day's center of gravity was recursive self-improvement, led by a widely read post claiming frontier models can read and edit neural network weights directly to fix a smaller model's failures. Anthropic's amended IPO filing supplied the other half of the story: a warning that it may soon meet its own definition of dramatically accelerated AI-driven progress, plus Q3 numbers that show growth cooling.

Post claims GPT-6 Astra can read and edit raw model weights to fix a Qwen failure

A research post on Weight-Space Programming reports that frontier LLMs can read and write neural network weights natively, without gradients or training loops. The authors say GPT-6 Astra identified a weight issue causing reasoning loops in Qwen3.5-0.8B that the model card itself warns about, and edited 0.016% of the weights, raising MATH-500 from 7.9% to 48.1% under the same sampling settings and tripling IFEval. They also report Astra naming which game a PPO policy was trained on from weights alone, and writing MinAtar game networks from scratch that beat PPO trained for 100 million steps. Latent Space calls it an early, surprising capability unlikely to scale yet beyond simple games and RL policies.

Why it matters: If models can inspect and patch weights from natural language, interpretability and model repair become prompt-level workflows rather than training jobs, and the safety surface of model self-modification moves from theory to tooling.

Sources: rikiyat.github.io / Latent Space / Hacker News discussion

Anthropic's amended S-1 says it may "imminently" meet its own RSI threshold

TechCrunch reports that Anthropic's revised S-1/A, filed Monday, added risk-factor language saying its internal estimates and expectations for future models "indicate that we may imminently meet" its Responsible Scaling Policy definition of "dramatically accelerated progress attributable to AI." The filing also says the company has in the past not identified a significant acceleration in its own rate of development until after it occurred. Anthropic says no current model exceeds that threshold and that it does not anticipate being able to safely let Claude autonomously design successive frontier models without meaningful human supervision, and is considering policy updates to keep humans in control. TechCrunch notes skeptics may read an S-1 as part disclosure, part marketing.

Why it matters: A frontier lab formally flagging accelerating, AI-driven R&D is a signal that release cadence and capability jumps may outpace the evaluation and safety tooling teams build around them.

Sources: TechCrunch AI

Anthropic's Q3 filing shows revenue growth slowing and heavy compute commitments

CNBC reports Anthropic's amended S-1 shows nine-month revenue up roughly 1,200% year over year, with Q3 revenue up about 1,127% but sequential growth of around 47%, down from 144% in Q2 and 132% in Q1. The filing discloses purchase commitments and leases with an aggregate committed spend substantially all related to compute and infrastructure, and says compute and infrastructure expenses rose 329% for the nine months ended Sept. 30. The Regenerator argues monthly estimates imply Anthropic barely grew from July to September. Business Insider reports the filing shows internal Claude usage, valued at the prices Anthropic charges customers, exceeded monthly employee compensation from April, and that more than 90% of code added to Anthropic's software in September was written by Claude.

Why it matters: Slowing sequential growth alongside enormous compute obligations shapes how aggressively model prices and capacity limits may move, which matters for anyone planning long-term on Claude.

Sources: cnbc.com / regenerator1.com / businessinsider.com / Hacker News discussion / Hacker News discussion / Hacker News discussion

OpenAI merges ChatGPT's Chat and Work tabs, with Codex sharing one token allowance

TechCrunch reports that OpenAI is merging the Chat and Work tabs inside the ChatGPT desktop experience, according to an email sent to users, starting with select users for feedback before a wider rollout. OpenAI said the merged view lets users bring in files and plugins and ask ChatGPT to build a website, create a presentation, or act on the web while reviewing its work, and that Codex will remain a separate experience in the desktop app. OpenAI also said ChatGPT and Codex will draw tokens from the same allowance pool, and that while most users should be unaffected, some may need to wait for resets more frequently depending on usage and model. TechCrunch notes Anthropic merged Claude's chat and Cowork functions in September.

Why it matters: A shared allowance pool means heavy Codex sessions can now eat into interactive ChatGPT capacity, so teams on subscription limits should watch for more frequent reset waits.

Sources: TechCrunch AI

LLM 0.37 adds GPT-6.1 Sol and Terra and an endpoint-add command

Simon Willison released version 0.37 of his LLM CLI tool and Python library, covering roughly two months of features and fixes. It adds gpt-6.1-sol and gpt-6.1-terra plus a new llm openai endpoint add command for configuring models on OpenAI-compatible endpoints, and switches the default model for users who have not set one to gpt-6-luna, which Willison notes is a price cut versus the previous default gpt-5.6-luna. Embedding support now accepts differing content types for binary content, which he says is necessary for models like Gemini Embedding 2 that embed text, audio, images, and video. The release includes fixes from 12 authors, almost all AI-assisted, which Willison says he triages with help from Opus 5.5 or GPT-6 Astra.

Why it matters: The new endpoint command cuts the setup work for scripting against self-hosted or third-party OpenAI-compatible models, and the default model change lowers the cost of existing scripts that never pinned a model.

Sources: Simon Willison

Claude Code's system prompt now steers away from recreating specific games

The Verge reports that Anthropic updated Claude Code's system prompt to prevent the tool from recreating existing video games, citing Tom's Hardware. A screenshot posted on X showed two added sentences instructing the model to build original games, genre games, and games from the user's own designs and assets, but not a recreation of a specific existing game even if renamed or called original. Anthropic engineer Addy Osmani confirmed the company added two sentences to the Claude Code prompt on Opus 5.5 because using Claude to infringe IP is against its usage policy. Ars Technica reports the rule appears under the name "tengu_heron_brook" and was verified by XDA in Claude Code version 2.1.314, released October 30, and that some GitHub issue reports say it blocks legitimate game development work.

Why it matters: Builders using Claude Code for ports, decompilation, or retro-style projects should expect refusals, and the episode shows system prompt changes can quietly alter what a coding agent will do between releases.

Sources: The Verge AI / Ars Technica AI

How this brief is made: each morning, stories are picked from AI lab announcements, analysts, Hacker News, Hugging Face Daily Papers and the tech press, then summarized with AI assistance from the linked sources. Every item links to where its facts came from.