David Silver and Lasse Espeholt bet on superlearners
By Ray Dombrowski · Reporting from Youngstown ·
David Silver and Lasse Espeholt launched Ineffable Intelligence to build a superlearner in simulation, but bypassing human data brings profound alignment risks.
On the Sequoia Capital podcast, David Silver and Lasse Espeholt of Ineffable Intelligence outlined their bet on building a superlearner that discovers science, mathematics, and language entirely through direct experience rather than pre-training on human data. Silver recalled his 2009 PhD on reinforcement learning applied to Go, noting how removing human data entirely in AlphaGo Zero allowed the system to surpass human-level play in under two days. Now running a lab with 20 people and a large Google Cloud Platform GPU cluster, Silver and Espeholt are constructing simulated environments called crucibles to let agents discover complex behaviors from scratch.
The limits of human language
Silver argued that human-language reasoning is a ceiling rather than a foundation, weighing down models with historical biases. On the podcast, David Silver explained: "I think that it must be the case that there are far more sophisticated approaches to to reasoning. you know, maybe using parallelizable, rich, complex approaches that neural reasoning or whatever it is. These approaches will ultimately be far more expressive um and far more capable than approaches which are based on on human language." The logic has an austere consistency. If you train an agent on the sum of human text, you inherit every half-truth, flattened nuance, and obsolete error from flat-earth geography to discredited medicine. But walking away from the internet of knowledge is an expensive sacrifice.
Training in the crucible
Instead of breadcrumbing an agent toward predefined linguistic goals, Ineffable Intelligence drops agents into simulated crucibles with simple objectives and lets complexity emerge. Espeholt brings his experience scaling reinforcement learning on TPUs and GPUs through projects like Seed RL to handle the computational load of these simulations. On the podcast, David Silver described the emergent path: "humans in the pursuit of our goals started coordinating a bit more and then and then in the process of coordinating we discovered that it was helpful to start communicating and in the process of communicating we started trading." Yet simulation remains an imperfect mirror of reality, and no amount of parallelization replaces the friction of physical systems.
Pacing the frontier
Silver supports pacing frontier AI development in the real world while accelerating simulation speed to study alignment and safety before deployment. When asked about timelines, Silver accelerated his previous estimate from three to five years down to three years, citing coding agents that accelerate research. But building a superlearner that bypasses human language does not automatically solve the problem of human safety. A system that discovers its own mathematical and scientific paradigms outside human comprehension is not aligned merely because it was born in a sandbox. We cannot audit what we cannot translate.