The Age of Experience: Ineffable Intelligence's plan to build a superlearner that discovers knowledge for itself
Key points
Key takeaways from a Sequoia Capital interview with David Silver and Lasse Espeholt, co-founders of Ineffable Intelligence, hosted by Sequoia partner Sonya Huang (October 2026):
The founding conviction. Silver's PhD (University of Alberta, 2009) applied reinforcement learning to Go, an experience-first approach he has held to ever since: systems that learn for themselves by trial and error will eventually produce intelligence and, ultimately, superintelligence. AlphaGo achieved superhuman play in Go in 2016; AlphaGo Zero, trained with no human data, surpassed it in under two days.
Why a new lab rather than DeepMind. Silver says his route to superintelligence does not pass through LLM-based products as an intermediate step, so he wanted an elite lab representing that approach 100% rather than a corner of an LLM-focused organisation. He also frames it as picking the lamppost nobody is searching under: the world already has tens of thousands of researchers doubling down on the LLM path, and a success on the experience-based path would change all of AI.
Why “ineffable”. The knowledge superintelligence will one day discover could be so profound it cannot be expressed in human language — hence a name deliberately detached from language-based approaches. Silver admits a secret second reason: “it can't possibly f*** up, so it's ineffable.”
Espeholt's infrastructure lineage. Espeholt joined Google DeepMind in January 2015 and began scaling up reinforcement learning with Silver around 2016, leaving for Google Brain in summer 2018 before the merger pulled him back; he credits early work on scaling and parallelising RL as the root of today's approaches, including the IMPALA and SEED RL line. He is now co-founder and CTO, running compute and infrastructure, and describes building a state-of-the-art RL stack to run at massive scale — including a GPU cluster on GCP's Vera Rubin AI accelerator — with roughly 20 people and an unusually high compute-per-person ratio.
What a superlearner is. Silver defines it in two dimensions: depth, from knowing nothing to discovering things beyond human knowledge (the way humanity collectively discovered languages, science and mathematics without a blueprint or a trillion-element dataset), and breadth, meaning it can be dropped into any stream of experience — simulated, digital or physical — and keep learning. “First contact” would feel like wonderment, echoing how Go and chess players reacted to superhuman play.
Why LLMs are not superlearners. Silver respects LLM progress but argues today's systems still lean heavily on human-in-the-loop effort, are not yet applicable to real-world embodied domains, and may not be well served by language as the reasoning substrate — in fact human data could be “weighing things down.” He sees language-based reasoning as a ceiling rather than a foundation, and says alignment via chain-of-thought monitoring is already reaching its limits.
Crucibles, not breadcrumbs. Rather than rewarding a pre-specified objective, Ineffable builds simulated environments it calls “crucibles” in which an agent is given simple rewards and asked to discover things that surprise it, letting emergence do the work — the way maximising fitness produced all of evolution's complexity. Milestones are set by the great moments of humanity's own discovery journey, and self-play (“one agent is boring, two is infinite”) is central. The company believes the science of how superintelligences coexist should be worked out in simulation before anything touches the real world.
The forecast. Silver predicts strong evidence of the approach within a year, and superintelligence within three. Ineffable Intelligence, incorporated in London in November 2025, raised a $1.1bn seed round at a $5.1bn valuation in April 2026 led by Lightspeed and Sequoia, with participation from Index Ventures, GV, Nvidia, the British Business Bank and the UK government.
Read more: YouTube · Sequoia Capital
