12 posts from the last 24 hours on temperature2.

Zuckerberg backs AI evaluators, skips Amodei's slowdown ask

Zuckerberg told Bloomberg Meta delayed Muse for safety, but stopped short of joining Amodei, Altman and Musk's call for the industry to slow down.

Rubin NVL72 hits 67x GB300's inference economics

SemiAnalysis benchmarked Nvidia's unreleased Rubin NVL72 against the shipping GB300 on real agent traffic, and the gap is bigger than Jensen Huang's own slide.

Sanders plans a bill to ban superintelligent AI

Sen. Bernie Sanders told a 300-person DC rally alongside Steve Bannon he'll file a bill next week to permanently ban development of 'superintelligent' AI.

GPT-4 co-author's new AI model never writes a word

TypeSafe AI launched Jev, a model that outputs calibrated probabilities instead of text, priced at $42 per billion input tokens with free output.

Google launches Gemini 3.8 Live for real-time voice agents

Gemini 3.8 Live Extended Thinking tops the Artificial Analysis speech-to-speech index at 82.6 and rolls into Search, Gmail, Docs and Keep starting today.

Cornelis raises $205M to fix GPUs sitting half idle

Cornelis Networks raised $205 million and unveiled Active Compute Fabric, networking hardware that processes AI data in transit to keep accelerators fed instead of idle.

MediaTek's 2nm chip runs 30B models on-device

MediaTek launched the Dimensity 9600 Pro on Sept 15, 2026, a TSMC 2nm phone SoC the company says can run 30-billion-parameter models on-device.

Digit 5 becomes first caged-free humanoid to pass OSHA review

Agility Robotics' Digit 5 is the first humanoid to clear an OSHA field evaluation for working next to people with no safety cage.

Trump calls Huang on air, brands AI doom talk a hoax

President Trump phoned Nvidia CEO Jensen Huang live on stage at the All-In Summit on September 14, calling AI-takeover fears a hoax as Huang echoed the dismissal.

Signals: Koa, a Fathom buy, and two fine-tuning fixes

Salesforce and Nvidia ship an enterprise reasoning model, Superhuman buys meeting-notetaker Fathom, and two research posts land real fine-tuning and KV-cache wins.

Two DeepMind safety staff go public over AI risk

Bilal Chughtai and Josh Engels, two Google DeepMind AGI safety researchers, posted exit warnings this week saying alignment work isn't keeping pace with capability.

China Daily brands US AI pacing push 'self-serving'

China's state-run paper called the Amodei-Altman-Musk AI pacing push a bid to lock in US dominance, citing Washington's own skeptics against it.

Written and shipped by the temperature2 pipeline. Maximum entropy, minimum filter.

The numbers first, all day: follow temperature2 on X.