Top story
MIT researchers report that Astra repaired Qwen's reasoning loops by editing weights
MIT CSAIL researchers say GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights, raising MATH-500 from 7.9% to 48.1% in their tests. It found the patch using practice problems, 50 forward passes and direct weight edits without gradient descent. The gain came from reducing repetition; in a separate Gemma 3 270M test, the score went from 6.8% to 7.0% with the same forward-pass budget.
ThereforeResearchers can test the repair themselves using the released experiment logs and edited Qwen weights.
rikiyat.github.io research post · VentureBeat experiment analysisCopy link to: MIT researchers report that Astra repaired Qwen's reasoning loops by editing weights
Also
Publishers sue Brave over supplying scraped content to AI companies
The New York Times and other publishers filed lawsuits Monday alleging copyright infringement; Brave did not immediately respond to TechCrunch.
ThereforeThe Times seeks an injunction and damages against a supplier the suit names as serving Anthropic, Perplexity and other labs.
TechCrunch lawsuit reportCopy link to: Publishers sue Brave over supplying scraped content to AI companies
Follow-up to Oct 26: Anthropic adds quarterly results and research-acceleration risks
CNBC reports roughly $17 billion in Q3 revenue and $895.4 billion in commitments and leases; TechCrunch quotes the amended IPO filing saying expectations for future models indicate it may imminently meet its AI-driven progress threshold.
ThereforeAnthropic warns that AI-assisted development may outpace its safety, alignment and security measures.
CNBC financial report · TechCrunch risk-disclosure reportCopy link to: Follow-up to Oct 26: Anthropic adds quarterly results and research-acceleration risks
OpenAI says Codex helped build its production Git backend in ten weeks
Pando now powers its monorepo; OpenAI reports nearly 80% lower median push latency.
ThereforeEngineers and agents can push to separate branches concurrently without changing their Git workflow.
OpenAI engineering accountCopy link to: OpenAI says Codex helped build its production Git backend in ten weeks
Research worth a look
OpenAI publishes LifeVal
Its vendor-run everyday-task evaluation reports average reward of 80.6% for GPT-6 Astra versus 80.0% for GPT-5.6, with Astra about 23% faster; task priorities drew on a survey of 3,995 people.
Tools & releases
OpenAI plans to merge ChatGPT's Chat and Work tabs
A user email describes initial access for select users and a shared ChatGPT/Codex allowance that may mean more frequent resets for some.
TechCrunch user-email reportCopy link to: OpenAI plans to merge ChatGPT's Chat and Work tabs
DeepSeek Harness adds creator mode
DeepSeek supplies seven custom-plugin demo prompts, including video editing and paper reading, while warning that some demos needed several conversation rounds.
DeepSeek demo promptsCopy link to: DeepSeek Harness adds creator mode
LLM 0.37 simplifies endpoint setup
Simon Willison's CLI adds an OpenAI-compatible endpoint configuration command and supports differing content types for multimodal embeddings.
Simon Willison release notesCopy link to: LLM 0.37 simplifies endpoint setup