David Silver, who led the team that built AlphaGo and AlphaGo Zero at DeepMind, has co-founded Ineffable Intelligence with Lasse Espeholt to pursue a radically different approach to AI development. Rather than pre-training on human-generated data, the company aims to build a "superlearner" agent that starts with no knowledge and learns to discover science, mathematics, and language through direct experience in simulated environments.
Silver argues that relying on human-language reasoning represents a ceiling, not a foundation, for AI development. He also contends that current alignment approaches using chain-of-thought monitoring are reaching their limits. The founders describe their reinforcement learning stack and the computational infrastructure needed to run these experiments at scale, building on techniques from their work at DeepMind and Google Brain.
Silver forecasts strong evidence of the approach within one year and superintelligence within three years. The team emphasizes the importance of working out the science of how superintelligent agents coexist in simulation before any real-world deployment.
This is an AI-written summary. The original reporting belongs to @sequoiacapital. How we write these · Spot a mistake?