This brief covers AI news from 2026-11-12 UTC.
An open-weight robotics model tops a leaderboard, a chipmaker writes two nine-figure checks, and Google turns Gemini into a decision engine.
Runway opens Praxis-1 robotics weights under Apache 2.0
What happened: Runway released Praxis-1, a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together, with weights and inference and fine-tuning code under Apache 2.0, including commercial use. It tops NVIDIA’s RoboLab-120 at 46.2% success across 1, 200 episodes, ahead of FLUX 3 Action at 42.9%.
Why it matters: Robotics teams can now fine-tune a leaderboard-topping action model on their own hardware, from H100 and B200 down to Jetson Thor, without a license negotiation.
TSMC board approves $1B split between Anthropic and OpenAI
What happened: TSMC’s board approved up to $1 billion in investments split evenly between Anthropic and OpenAI at its November 10 meeting, alongside a plan for a new semiconductor investment company capitalized at up to $500 million. The OpenAI portion buys preferred stock in OpenAI Group PBC.
Why it matters: The world’s largest contract chipmaker taking equity in the two labs ties compute supply and model development closer together, which shapes who gets capacity first.
Google launches Gemini Decision 1 and a Decisions API
What happened: Google released Gemini Decision 1, described as its first decision model, with a Gemini Decisions API for fast, calibrated multimodal decisions. It launched on OpenRouter the same day.
Why it matters: A model tuned for picking between options rather than generating text gives builders a cheaper primitive for routing, triage and agent branching logic.
Gemini app arrives on Apple CarPlay
What happened: The Gemini app is now available on Apple CarPlay, letting drivers tap the icon to continue phone conversations or talk with Gemini Live hands-free from the car display.
Why it matters: Voice agents get a new surface with real constraints: short turns, noisy audio and no screen attention, which is a different test than a phone app.
Judge clears 56 amicus briefs in OpenAI copyright case
What happened: Judge Sidney Stein approved filing of 56 amicus briefs in In re: OpenAI Copyright Infringement Litigation, which consolidates suits including the Authors Guild and New York Times complaints. The court will take up whether training on copyrighted works is fair use without a trial.
Why it matters: A ruling on fair use for training data would set the terms for what builders can ingest, license or have to pay for, so watch the briefing list closely.
Huawei details Peerium interconnect, 256, 000-card system in test
What happened: Huawei unveiled Peerium, an architecture combining nested parallelism, shared memory addressing and direct chip-to-chip links, claiming up to 800GB per second between racks versus 200GB per second for its existing setup. A 256, 000-card SuperCluster is being deployed and tested.
Why it matters: Interconnect bandwidth, not accelerator count, is the binding constraint on training scale, so any credible alternative to the dominant fabric matters for capacity planning.
Cognichip says its chip design models cut months to days
What happened: Cognichip’s Chief Product Officer told Tom’s Hardware that its Artificial Chip Intelligence models, trained on more than ten million samples including synthetic ones, feed ACI Enterprise, a platform pairing them with agentic workflows and project memory for chip design.
Why it matters: If design iteration compresses from months to days, hardware teams can try more architectures per quarter, which changes how silicon roadmaps get planned.