AI progress — 13 November 2026

Nvidia CEO Jensen Huang has resigned, and TechCrunch reports the circumstances involve an alleged Chinese spy, with Democrats in Congress opening an investigation into the company’s China sales. Separately, MIT Technology Review reports Nvidia privately told a congressman that chip tracking against smuggling is technically possible. Together these put export-control enforcement back on the table for the biggest GPU supplier. On the model side, Google released Gemini 4 Hydrogen and put it into Search.

New models and papers

• Google rolls out Gemini 4 Hydrogen — Ars Technica reports it replaces 3.8 Flash and skips the slower, security-conscious rollout of Carbon and Argon. Why it matters: Simon Willison notes it is the cheapest Gemini 4 model yet priced not far below its larger siblings, so check the price list before assuming it is the budget option.
• On the Design of Hyper-Connections: Simplification, Stability, and Scalability — a paper examining the components of Hyper-Connections, an architecture the authors say is increasingly used in frontier language models. Why it matters: the authors say its components are poorly understood and carry non-trivial overhead; the paper targets simplifying and stabilizing them, which matters to anyone training with this design.
• Duo: A Reciprocal World Model for Full-Duplex Interaction — a report introducing Duo v1, which the authors describe as continuously perceiving audio and video while responding in real time. Why it matters: if the claim holds, it points toward voice and video assistants that can be interrupted mid-response rather than taking turns.

New tools and software

• Claude Code v2.1.323 — adds Code Mode, an experimental opt-in where Claude writes JavaScript to call its tools; organizations can turn it off. Why it matters: it is a different way of invoking tools that you can now test, with an admin switch for teams that do not want it.
• SGLang-Neuron on AWS Trainium — now publicly available through an AWS Deep Learning Container for serving models on Trainium. Why it matters: teams can serve on Trainium with SGLang without building their own stack.
• Track and cap Amazon Bedrock spend per IAM principal — an AWS walkthrough using CUR 2.0 and AWS Budgets for per-identity Bedrock cost visibility. Why it matters: it lets you attribute and cap model spend per user or service instead of seeing one account-wide bill.

Overall digest

• Democrats open an Nvidia investigation — a Congressional investigation into Nvidia’s China sales. Why it matters: buyers dependent on Nvidia supply face possible changes to how China-bound sales are handled; the outcome is not yet known.
• Nvidia privately confirmed it could track chips — per MIT Technology Review, Nvidia told Rep. Bill Foster tracking is technically possible, though the industry publicly questioned feasibility. Why it matters: it weakens the argument that location verification cannot be done, a point for any future export rule.
• FAA expands AI advice for air traffic congestion — the tool now covers all US airspace; airlines praised it, but the controllers’ union was not involved. Why it matters: it is a live government deployment of AI advice in safety-relevant operations, rolled out without one affected party’s input.