AI News — Afternoon — 2026-11-02

Editing a model to stop reasoning loops

• Researchers use one AI model to edit another’s weights — The experiment shows a possible way to fix a specific failure in a small model without retraining it.
  • MIT researchers asked GPT-6 Astra to improve Qwen3.5-0.8B by changing its weights, the numbers that determine its behavior. Astra used a Python session and practice problems to test edits.
  • With a budget of 50 forward passes, Astra changed 131,072 weights and reduced repetitive reasoning. The authors report a rise from 7.9% to 48.1% on MATH-500, a set of math problems.
  • Answers hitting the 32,768-token limit fell from 93.0% to 33.1%. A token is a unit of information a model processes or generates.
  • Astra’s edits did not meaningfully improve a separate Gemma model that rarely looped. The team released experiment logs and code.
  • Caveat: The Qwen result mostly reflects allowing interrupted answers to finish, not a demonstrated general ability to add new mathematical skills. (HN 1,133 points · 353 comments · researchers’ experiments)

AI and software work

• Linux company plans job cuts and cites AI and automation — SUSE’s memo offers a concrete example of management invoking AI while reducing staff.
  • The Register reports that CEO Dirk-Peter van Leeuwen announced cuts of approximately 10% of the global workforce.
  • The memo says changing workflows and flatter reporting lines have diminished demand for some roles. It does not identify those roles.
  • Affected staff are to be informed by November 4, with individual meetings starting the following week.
  • Caveat: The memo states management’s explanation; it does not establish how much work AI actually replaced. (Techmeme · reporting based on the staff memo)

Measuring AI adoption

• Census changes the wording of its business AI survey — A terminology change could complicate the next reading of a widely used adoption series.
  • Simon Willison’s analysis found that the survey began calling AI “Super Intelligence” in the October 19 collection. The Census Bureau attributes the wording update to an executive order.
  • The first results using the new wording are due Thursday, according to Willison. The latest published data still uses the prior wording.
  • His six-survey trailing average puts reported use at 40.7% among businesses with at least 250 employees. That reflects use in any business function.
  • Caveat: The question broadened in November 2025, so the earlier goods-and-services series is not directly comparable with the later series. (HN 74 points · 32 comments · analysis and Census documentation)

Models and products

• LongCat introduces a model that handles images and code — The launch offers a hosted model to try, with an unusually detailed account of its training setup.

  • LongCat-2.5 has 1.6 trillion total parameters, with approximately 48 billion active per generated token, according to its developers.
  • The team describes training on long coding tasks and converting internal agent failures into reproducible tasks. An agent is a model running a multi-step task with tools.
  • Its site links a chat service and API documentation, while its GitHub and Hugging Face links say “coming soon.”
  • Caveat: The published comparisons are developer-reported, with mixed testing setups and some scores taken from other sources. (HN 64 points · 20 comments · developer announcement)

• ChatGPT plans to combine its desktop Chat and Work tabs — The interface change would put conversation and delegated work in the same place.

  • TechCrunch reports that OpenAI emailed users about an initial rollout to a select group for feedback.
  • Codex will remain a separate experience in the desktop app, according to the email.
  • The related shared usage allowance was covered in November 1 Morning.
  • Caveat: The report gives no date for a broad rollout. (Techmeme · report on OpenAI’s user email)

AI infrastructure costs

• Anthropic updates its financial filing with third-quarter results — The new filing adds a quarter of financial data and a large future infrastructure commitment.
  • CNBC reports that the amended filing puts third-quarter revenue at roughly $17 billion and net loss at $23.29 billion.
  • The company reports $895.4 billion in purchase commitments and leases, substantially all for computing and infrastructure.
  • For the first nine months, it reports a $104 billion net loss. CNBC says most of that period’s loss came from changes in the fair value of convertible notes and derivative liabilities.
  • Caveat: The commitment total covers future agreements, while the loss is an accounting measure; neither is the same as cash spent in the quarter. (Techmeme · CNBC reporting on the amended filing)