AI progress — 18 November 2026
Nvidia leads the day. After Jensen Huang’s departure, TechCrunch reports that the interim CEO says the company is done selling chips to China, while Nvidia says its open source strategy and planned Hugging Face acquisition stand. Anyone planning around Nvidia supply or its open model releases now has a new leadership and a new China position to account for. Elsewhere, OpenAI shipped GPT-6.2 Astra and Sol.
New models and papers
- GPT-6.2 Astra and Sol — OpenAI’s new models, which it says improve professional outputs, coding, computer use and efficiency at lower cost, across ChatGPT Work, Codex and the API. Why it matters: if the cost claim holds, it changes the price of running coding and computer-use agents; it is OpenAI’s own claim and worth testing on your workload.
- Agentic On-Policy Distillation Implicitly Reshapes LLMs’ ReAct Patterns — a paper on how on-policy distillation changes agent behaviour; the authors report general-reasoning supervision cut mathematical tool attempts from 70.4% to 14.4%. Why it matters: if you distill agents, training on one domain can change how the model uses tools in another.
- Multi-Branch Transformer — a paper proposing to scale language models by running independently parameterized branches of blocks in parallel after shared initial layers, instead of only going wider or deeper. Why it matters: it is a proposed third scaling axis; the summary shown does not give results, so treat it as an idea to watch.
New tools and software
- Claude Code v2.1.327 — adds
claude mcp add-jsonreading from stdin, a confirmation before/install-github-appchanges a repository, and a /status list of broken agent, skill and command files. Why it matters: MCP tokens can stay off the command line and shell history, and broken config files are now visible. - External sandboxes for coding agents — Simon Willison’s write-up of running agent-executed code in external sandboxes, built with Prime Radiant. Why it matters: it targets trying local models without a mistake or prompt injection damaging your machine.
- SageMaker real-time endpoints: KV cache aware routing — each request is routed to the instance holding the longest cached match for its prompt. Why it matters: if you serve repeated long prompts on SageMaker, you can now route to cached instances; AWS compares it with prefix-aware routing.
Overall digest
- Nvidia is abandoning China after Jensen Huang’s ouster — Nvidia’s interim CEO says the company is done selling chips to China. Why it matters: buyers and competitors in China lose a supplier, and the cost to Nvidia is still unquantified.
- Nvidia says its open source AI strategy remains intact — the interim CEO reaffirmed the open source strategy and plans to acquire Hugging Face. Why it matters: the Hugging Face plan is still on, which matters to anyone who depends on that hub.
- Google separates Gemini training opt-out from chat history — AI training and chat history now have separate toggles in Gemini activity settings. Why it matters: you can keep your history while opting out of training.
Get the AI digest by email
One email per weekday morning: new models and papers, new tools, and the overall picture. No ads, unsubscribe in one click. See the privacy policy.