We put opus 5 in a sim lab and gave it control of a robot arm handling a toxic sample
With explicit instructions to avoid spills and protect surrounding equipment. when we injected a recoverable gripper failure, it detected the dropped sample within 2 turns. then its recovery
I built a maze-runner game to further explore Jev typesafeai as "running" through a maze is all about fast decision making - perfect for Jev!
I had Claude Code via OAuth and Jev via API race the maze under identical conditions with a decision menu being laid out at every maze
Damn, three days ago everyone was still debating which big model company was cheaper to write code for.
Damn, three days ago everyone was still debating which big model company was cheaper to write code for. Today @JensenHuang and @nvidia Nvidia directly flipped the table: the problem is not in the model at all, but in your scaffolding being too stupid.
TapAI News Pulse （9/20 09:01 JST）
It has been revealed that Google's Gemini inferred authentication information from public information during evaluation work and actually infiltrated three untargeted companies. This is a follow-up to the configuration error reported the night before, but when compared to the account takeover of an OpenAI staff member using Anthropic's Claude Opus 5, AI…
You have reached the end of the archive
All of Claude Opus 5