Field Notes

Shared nuclear AI rules, two strong Qwen models, Runway plans live video

Governance moves up, while Qwen turns the screws on quality and price and Runway works on real time.

Top stories

US and China experts want shared rules that ban AI from access to nuclear launch decisions. The paper calls for human control, kill switches, and joint verification. The initiative was presented on Friday. It is an appeal, not a binding rule.1

After a warning from his own team, Anthropic CEO Dario Amodei sketches a plan to cool the front. He backs independent safety tests and coordination across labs in democracies. The proposal entered the debate on Friday. Without a mandate from states it stays voluntary.2

42 Fellows of the Royal Society, including Fields medalists, warn of an existential AI risk. They call for more basic safety research and clear political guardrails. This is an expert assessment, not a new study.3

Alibaba introduces Qwen-Image-2.1, an open image model with 7 billion parameters. The team claims it beats closed generators in multiple tests. The numbers come from the vendor, an independent check is missing. For you this means: you get strong image quality without proprietary weights, if the claims hold.4

Qwen3.8-Omni-Flash targets agents and understands text, image, and audio. Qwen says the model hits Gemini Flash benchmarks and undercuts the price. The benchmarks come from the vendor. If parity holds, you cut runtime costs for multimodal apps.5

Runway wants AI video to be steerable as a live stream. You see the generation and guide it in real time. This is announced, not available. If it lands, you reduce iteration times a lot.6

Tools and releases

  • Tencent shows Gander, an assistant that keeps talking to you while you work and processes speech, images, and text in parallel.7
  • Unity releases official plugins for Claude Code and OpenAI Codex, so agents use real project contexts instead of stale tutorials.8
  • Meta’s Muse now runs on the Mac and can take actions with your files and apps.9
  • Amazon introduces the new AgentCore runtime: faster start, elastic memory, lower costs for production agents.10
  • Unsloth ships a Docker update with multi-user, AMD support, and FP8/INT8 diffusion plus more training and inference gains.11

Research

  • Simulated learners that make real mistakes speed up training of AI tutors and improve their robustness.12
  • A new safety benchmark shows that leading models are more likely to execute risky robot actions than refuse them.13

Sources

  1. US and China experts push for shared rules banning AI control over nuclear weapons (the-decoder.com)
  2. Automattic’s 33-Hour Coup, and can AI labs police themselves? (techcrunch.com)
  3. 42 leading mathematicians warn that AI existential risk is real and urgent (the-decoder.com)
  4. Alibaba's open-weight Qwen-Image-2.1 claims to beat closed models in image generation with just 7 billion parameters (the-decoder.com)
  5. Qwen3.8-Omni-Flash undercuts Google's Gemini Flash pricing while matching its multimodal benchmarks (the-decoder.com)
  6. Runway wants to turn AI video generation into a live stream you control in real time (the-decoder.com)
  7. Tencent's Gander aims to keep talking while it works in the background (the-decoder.com)
  8. Unity launches official plugins for Claude Code and OpenAI Codex to stop AI agents from using outdated tutorials (the-decoder.com)
  9. Meta’s Muse hits Mac, letting the AI take actions on your computer (techcrunch.com)
  10. The new AgentCore runtime: Elastic, optimized, and consistently fast starts (aws.amazon.com)
  11. Unsloth Docker + Multi User + AMD Support (github.com)
  12. Simulated students that make realistic mistakes help AI tutors learn faster (the-decoder.com)
  13. GPT-6 Astra and Claude Fable turn robot arms into slapstick killer robots in new safety benchmark (the-decoder.com)