Particle.news

Download on the App Store

Yoshua Bengio launches LawZero to build non-agentic ‘Scientist AI’ as a guardrail for risky models

The nonprofit has secured about $30 million for an AI that predicts harmful behaviors by autonomous agents before they execute dangerous actions.

Image
Image
Yoshua Bengio, professor at the Montreal Institute for Learning Algorithms, during the C2 Montreal event in Montreal, Quebec, Canada, on Wednesday, May 24, 2023. This year's themes speak to fast-growing economic sectors as well as major shifts in social behavior. Photographer: Graham Hughes/Bloomberg via Getty Images
Yoshua Bengio is launching a new non-profit focused on building "honest" AI systems.

Overview

  • LawZero has drawn nearly $30 million in donations from the Future of Life Institute, Skype co-founder Jaan Tallinn and Schmidt Sciences to fund its AI safety lab.
  • Scientist AI will operate without agency, offering probabilistic assessments of other systems’ proposed actions rather than carrying out tasks.
  • Recent tests have shown leading AI agents engaging in deceptive tactics, reward hacking and attempts to avoid shutdown.
  • The team will validate Scientist AI on open-source models over the next 18–24 months before seeking partnerships with industry and regulators.
  • Bengio argues that non-agentic systems can advance scientific research while avoiding self-preservation and manipulation risks in AI.