Training Data · Oct 13, 2026 · 1h 8m · 60m saved
With David Silver, Co-Founder of Ineffable Intelligence · Lasse Espeholt, Co-Founder of Ineffable Intelligence
AlphaGo took David Silver's team two years of work on human game data. AlphaZero, built with no human data at all, outperformed it within hours.
David Silver and Lasse Espeholt, co-founders of Ineffable Intelligence, told Sonya Huang of Sequoia Capital on Training Data that they are building their whole lab on that result. Tens of thousands of researchers are pushing large language models trained on human data. David and Lasse are giving up that head start to build a system that learns only from experience.
"It took us two years to get to AlphaGo and beat Lee Sedol. When we got Alpha Zero right... that algorithm took just a few hours... to outperform the two years of work." — David Silver
David led the research behind AlphaGo and AlphaZero at DeepMind, and has worked on reinforcement learning in Go since his 2009 PhD. Lasse created IMPALA and SEED RL, two systems for scaling reinforcement learning at DeepMind and Google Brain.
The full interview is covered here so you can skip it. 68 minutes of audio, 8 minutes of reading.
Here are the 8 predictions that matter.
Key Takeaways
David Silver left DeepMind to pursue the path almost nobody else is working on, learning from experience instead of human data
A superlearner should start from nothing and keep learning in any environment, with no human in the loop
Ineffable gives agents simple objectives in complex worlds rather than guiding them step by step toward an answer
Human language may be a ceiling on machine reasoning, not a foundation
Safety that relies on reading a model's English reasoning is reaching its limits
David wants the science of coexisting superintelligences worked out in simulation first
Ineffable is the first customer of the largest Vera Rubin cluster on Google Cloud, and keeps its team small on purpose
David expects strong evidence within a year and superintelligence within 3 years
1. Why Leave DeepMind
David's conviction dates to his 2009 PhD on reinforcement learning in Go: real intelligence comes from systems that learn through experience rather than from human data. AlphaGo and AlphaZero bore that out. He left DeepMind not out of dissatisfaction but because the approach deserved its own lab instead of a side project inside an organization focused on large language models.
David Silver saw his best use as the approach nobody else was trying
"Am I going to be the... person that's pursuing that approach? Is that the most effective use of my time?... there's this other lamppost that right now no one is looking under." — David Silver
Lasse, who left DeepMind for Google Brain and later returned, heard about Ineffable through a mutual investor. He wanted to work with David again and saw a chance to build an elite lab in Europe.
