David Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience
David Silver and Lasse argue that superintelligence will come from agents that learn from their own experience in rich simulated worlds, not from human-data LLMs, and that simulation also allows safer, faster progress.
Watch on YouTube ↗Charlie's summary
- ▶ 9:43
A superlearner should discover everything for itself, with depth beyond human knowledge and breadth across any stream of experience.
David defines a superlearner by depth and breadth. Depth: it starts knowing nothing and ends up beyond what humans know, as humanity did, starting from gray mush and discovering language, science and mathematics with no blueprint or dataset. Breadth: it can be dropped into any simulated, digital or physical world and keep learning. He says no one has cracked this in AI. Current LLM breakthroughs still need heavy human involvement.
- ▶ 16:20
Give agents simple objectives in complex worlds and let complexity emerge, rather than breadcrumbing them toward a target.
David says RL practitioners often agonize over reward functions that capture science or interesting maths. He prefers an easy-to-describe objective in a rich environment. Evolution optimized a simple fitness signal and produced all of life's complexity. Human goals led from coordination to communication, trade, arithmetic, tools, science and technology. Ineffable does not craft rewards for science or language. It expects agents to discover them and to surprise the team.
- ▶ 21:30
Crucibles: environments with emergent complexity, where a second agent turns a boring game into an infinite onion of depth.
Ineffable builds crucibles rather than hand-crafted task sets. Complexity comes from physics engines, millions of interacting objects, and many interacting agents. David's example is Go. With one agent it would be the most boring game imaginable. Add a second agent and each must reason about the other's behavior, which yields a learning curve that never ends. Lasse adds that they want a range, from fast simple crucibles to more realistic and expensive ones.
- ▶ 28:08
Letting go of human data is a practical bet: AlphaZero beat two years of human-in-the-loop AlphaGo work in hours.
David says Ineffable accepts a short-term performance sacrifice by not relying on trillion-item human datasets, because the more you lean on them the less you discover new algorithms or new fields of knowledge. AlphaGo took about two years with human data and human-in-the-loop improvement. AlphaZero, with no human data, beat the best chess AI in four hours. He also argues human language carries our biases, such as flat-earth or four-humors beliefs. The stance is pragmatic, not religious: weak, environment-agnostic priors are fine.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.