@elonmusk
Big jump in capability when we finish training our V7 foundation model (Grok 4 is V6), which has much better image/video understanding and our video gen model
「CoTの監視がAI安全性向上に役立つが、最適化圧で容易に破壊されうる」という既知の知見の整理、内容的にはそれほど目新しさは無い
ただし、OpenAI, Anthropic, Google DeepMindなどの大手AI企業、Apollo Research, UK AI Security Instituteなどの研究機関が同じ評価プロトコルを適用し、「この問題は業界共通リスク」という認識を公式に揃えたという点での意義がある
@sama
woke up early on a saturday to have a couple of hours to try using our new model for a little coding project.
done in 5 minutes. it is very, very good.
not sure how i feel about it...
https://x.com/ilyavaliant/status/1954548709930553566
Key difference: auto-Thinking ≠ manual GPT-5 Thinking.
When you pick Thinking manually, the system gives it a bigger “thinking budget.”
Auto-Thinking is shorter and faster — adaptive (and more cost-efficient) reasoning.
