AI progress — 9 November 2026

Nvidia’s board has opened an independent investigation into CEO Jensen Huang’s relationship with a Megaspeed executive who, according to a US intelligence memo reported by TechCrunch, was recruited by China’s Ministry of State Security to spy on the company. The story is about governance at the company that supplies most AI training hardware. Elsewhere it was a quiet weekend for releases: a paper on exact arithmetic, a CPU-only run of a very large open model, and a retreat by Meta on AI-training terms.

New models and papers

• Sketchpad: certified exact arithmetic with a small vision-language model — the authors argue exactness at all input lengths becomes a checkable property when each step reads only a bounded view of an external working memory. Why it matters: if the result holds, arithmetic correctness could be verified for a network rather than estimated on samples; this is a paper claim, not yet independently tested.
• Confucius4-R2T2 — a streaming speech-recognition model that, per the authors, never revises earlier output while keeping accuracy high. Why it matters: voice agents and translators act on each token, so a revised token means wasted work or a wrong action.
• Exploring Weight-Space Programming — a study of agents that inspect and edit network parameters directly, across 16 editing trajectories on seven benchmarks and two model scales. Why it matters: it tests the capabilities and limits of an alternative to gradient-based adaptation that had been poorly characterized.

New tools and software

• llama-modchip — a repository whose title claims a 284B DeepSeek model running on 64GB of RAM with no GPU. Why it matters: if it reproduces, very large open models become testable on commodity machines; the claim is the author’s and I have not seen verification.
• kanzo — a self-hosted kanban board for humans and agents, shipped as a 4.7 MB binary; the author lists 5 MB of RAM and 10 to 60 tokens per agent reply. Why it matters: it gives agents a shared task board with no setup or service to run.
• @GitHub in Slack and Teams, generally available — GitHub’s changelog says you can turn conversations in Slack and Teams into code. Why it matters: teams already discussing work in chat can hand it to Copilot without switching tools.

Overall digest

• Nvidia investigates its CEO — the board announced an independent investigation into Jensen Huang’s relationship with Megaspeed executive Alice Huang, reportedly recruited by China’s Ministry of State Security. Why it matters: the outcome bears on leadership at the main supplier of AI accelerators; the allegations come from a memo and are not established findings.
• Meta drops its commercial terms update — Meta will not force Facebook page owners to allow AI training on linked content, Ars Technica reports. Why it matters: page owners keep control over whether linked content feeds Meta’s AI training.
• Rights holders of Elvis Presley and Eminem sue Suno — two new suits claim Suno generates music and lyrics that mimic those artists; the Presley suit cites a Tennessee law. Why it matters: it adds to the cases testing whether AI music services can imitate named artists.