David Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience

Summary · 6 min read instead of 1 h 8 m

David Silver did his 2009 PhD on reinforcement learning for Go and spent about two years at DeepMind building AlphaGo on human data, before AlphaZero gave that data up. Lasse Espeholt helped build the IMPALA agent at DeepMind, then founded Google Brain's Amsterdam office, where he built SEED RL and worked on the MetNet weather model. Together they co-founded Ineffable Intelligence, which has hired 20 people.

David Silver argues that superintelligence will come from a superlearner that starts knowing nothing and discovers knowledge through its own experience, and that human data and human language hold systems back. He describes giving agents simple rewards inside crucibles, simulated worlds meant to be complex enough for science, maths and language to emerge, and points to AlphaZero, which dropped human data and quickly surpassed two years of human-guided work on AlphaGo. He says alignment based on agents' language is reaching its limits, so the science of how superintelligences can interact peacefully should be worked out in simulation, with real-world steps gated on that evidence. Lasse Espeholt describes an all-GPU stack on a large Nvidia Rubin cluster on GCP that draws on his IMPALA and SEED RL work. Silver expects extremely strong evidence within a year and true superintelligence in three years.

Minute by minute

1. 1:06

   Learning from experience has held true. Silver says that since starting his PhD on Go he has believed the right way to AI is learning from experience, with systems that discover things for themselves through reinforcement learning, and that this leads to superintelligence. What surprised him in the LLM era was that a model trained on massive data could be fine-tuned quickly on a small amount of data to change its behavior completely, which suggested that at the right scale, slow things can become very fast.

2. 3:36

   The experience path needed its own lab. Silver says he loved DeepMind, but the path he believes in does not go through LLM-based products, so it deserved an elite lab rather than a corner of an organization focused on them. With tens of thousands of researchers doubling down on LLMs, he saw more impact under another lamppost where no one is looking. The name points to knowledge too deep for human language, plus a pun: the approach has to work, so he "can't possibly F up".

3. 9:34

   A superlearner starts from nothing. Silver defines a superlearner by depth, starting from knowing nothing and discovering far beyond human knowledge, as humanity did without a blueprint, and by breadth, learning in any simulated, digital or physical stream of experience. He expects first contact to bring wonderment, as AlphaGo did for Go players. He respects LLMs but says their breakthroughs still need a lot of human-in-the-loop effort, and human data may be weighing them down.

4. 13:54

   Simple rewards can yield emergent complexity. Silver says the team is building a large-scale RL stack and crucibles, simulated worlds where agents can discover what we mean by intelligence, with milestones drawn from moments of human discovery. Instead of "pulling the agent by its nose" through a trail of problems, as he says much RL with LLMs does, they give agents simple objectives in complex environments. He cites evolution optimizing fitness, and humans moving from coordination to communication, trade, arithmetic, tools and science.

5. 21:03

   Crucibles rely on emergent complexity. Silver says crucibles are not handcrafted puzzles: complexity comes from physics engines with millions of interacting objects or from millions of interacting agents, well beyond simulators like Atari 57. Espeholt says realistic physics is expensive, so they want a range that includes simple, fast crucibles. Silver notes that single-agent Go would be the most boring game, but adding a second agent creates an "infinite onion" of complexity.

6. 25:35

   Human language is not the endpoint. Asked whether agents will rediscover language, Silver calls the view that human language is the best way to reason "a little naive" and expects richer, parallelizable kinds of reasoning to be far more capable. Starting from human language also imports human biases: agents limited to the language of 500, a thousand or two thousand years ago might believe the world was flat or that the four humors caused disease.

7. 27:19

   Dropping human data beat years of work. Silver says he was about 50-50 on whether removing human priors would work. AlphaGo took about two years of human-in-the-loop work on human data; AlphaZero, which gave that data up, beat the best AI ever created in chess in four hours and, he thinks, took a couple of days in Go. He calls giving up human data a short-term sacrifice, and says the team is "not religious about this": weak priors are fine, trillion-item datasets are not.

8. 33:42

   The timing is right for superlearners. Silver strongly believes the approach will work at some point, since the human brain is the only existence proof of general intelligence, and says success depends on timing. He predicted computer Go 10 years before it was solved and "got it right to the year", and also called StarCraft and the IMO with AlphaProof. He calls compute an enormous driver and says they could run experiments with 100,000 times more compute than OpenAI's hide and seek.

9. 38:08

   Build superintelligence in simulation first. Espeholt notes that MetNet, a weather model he worked on, beat physics simulations that accumulate errors, yet simple environments can still show emergent technology. Silver argues that a superintelligence built in simulation would arrive 10 or 100 times faster and in "the safest place", and that simulations can express computation, economics, language and maths. He agrees with Rich Sutton that learned skills shouldn't be expected to transfer to the real world, but expects the algorithm to.

10. 43:39

    Alignment needs a science of superintelligence. Silver says alignment that relies on agents' language and human-data priors is reaching its limits, as agents trained more and more learn to "say one thing and act in a different way". He calls understanding the science of superintelligence the company's most important task: learning in simulation how superintelligences can interact peacefully, including with lesser intelligences, and sharing it publicly. Ineffable commits to gating real-world steps on that evidence.

11. 47:38

    Language monitoring will not be enough. Silver says monitoring agents' English reasoning still helps but will get harder as agents produce a million words of reasoning or use subterfuge. He suggests shadow actions, which show what an agent would do without letting it act, and superintelligent monitors that limit how much a scientist agent affects the world. He supports pacing the frontier in the real world so governments and organizations can adapt, while moving "as fast as possible in simulation".

12. 52:54

    Simulation, training and inference on GPUs. Espeholt says Ineffable is the first customer of the largest open cluster on GCP by the end of the year, Nvidia Rubin GPUs on which a rack acts like one machine, which suits fast RL iteration. Simulation, training and inference all run on GPUs, with tricks such as model-based RL and replay buffers. He says RL scaling laws differ because the data distribution keeps changing, and the stack uses derivatives of IMPALA and SEED RL, his earlier work scaling RL on GPUs.

13. 1:00:56

    A small, research-led team. Silver says Ineffable has intentionally stayed small, with 20 people hired for talent, kindness, mission fit and needed skills, and Espeholt says that keeps compute per person high. Four co-founders come from the AlphaStar team and three from successful startups, and Silver has turned down many former colleagues. He describes a research-led culture that pushes back on product pressure, betting that breakthroughs will lead to far bigger products later.

14. 1:05:51

    Superintelligence in three years. Silver predicts "extremely strong evidence" within a year that Ineffable is on the right track. He expects true superintelligence, meaning discoveries far beyond what humans can achieve across science, maths, technology and language, in three years, down from his earlier estimate of three to five, because coding agents can speed up the work.