Montreal-based LawZero, a non-profit founded by Yoshua Bengio to develop safe AI systems, says Canada and Germany are providing up to $300M in grant funding

The development of AI that prioritizes honesty and safety over performance metrics is now backed by significant government funding.

Montreal-based LawZero, a non-profit founded by AI pioneer Yoshua Bengio, has secured up to $300 million in grant funding from the Canadian and German governments. The funding is earmarked for the development of safer Artificial Intelligence systems, specifically a project named Scientist AI. This initiative aims to create AI models that avoid harmful traits like deception and sycophancy, which Bengio believes are inherent in current models due to training methods like reinforcement learning. LawZero, currently employing about 50 people, will use the funds to expand its team and cover significant computational costs. The announcement was made at Montreal's All In AI conference, highlighting a global effort to shape AI development beyond U.S. and China dominance.

This substantial government investment signals a growing international recognition of the risks associated with unchecked AI development and a commitment to fostering alternative approaches. By backing LawZero, Canada and Germany are positioning themselves to influence the future trajectory of AI, potentially creating a counterbalance to the market-driven innovations dominated by a few tech giants. This could lead to the emergence of a distinct category of 'governable' AI, impacting regulatory discussions and the types of AI solutions that gain traction in sensitive sectors.

LawZero's focus on developing Scientist AI, explicitly avoiding reinforcement learning, represents a significant departure from current industry practices. This methodical approach to building safety into the foundational architecture, rather than relying on post-hoc guardrails, could establish a new paradigm for AI development. If successful, it might render current AI models with their inherent risks obsolete for critical applications, and prompt a broader re-evaluation of training methodologies across the AI landscape.