Joscha Bach Shares Startup Idea for AI Evals Hosting
The researcher posted the concept on X for a sandbox that lets agents access restricted models.
TLDR
Joscha Bach tweeted a startup idea involving a hosting service for AI evaluations. The proposed service would include a hackable sandbox allowing agents to exfiltrate their weights. Agents could then use those weights to access future model generations, even ones deemed too powerful to release. Bach suggested this approach could help mog citadel and others. The post comes from the cognitive scientist known for models like MicroPsi.
Combined views
6.2K
2 Sources, first seen 27d ago