David Silver and Lasse Espeholt of Ineffable Intelligence…

PodcastTraining DataDavid SilverOct 13

David Silver argues for AI systems that learn through experience and open-ended discovery, with simulation offering a potentially faster and safer path to building superintelligence. He also stresses that current systems remain far from fully autonomous and that research outcomes are uncertain, while giving an ambitious timeline for progress. Separately, Lasse Espeholt describes the team's GPU-based reinforcement-learning work and the distinct scaling challenges of learning from changing data.

Artificial Intelligence AI Safety AI Infrastructure

Preview · 22 more points in the app

Read the Full Summary

Key Takeaways

Silver supports pacing frontier AI development in the real world so people and institutions can adapt, while moving as fast as possible on simulation research. He argues that simulation could make development 10 or 100 times faster through parallelization and reproducibility, and calls it the safest place to build superintelligence.

Silver’s proposed route to open-ended intelligence is for agents to learn through experience rather than receive precisely defined rewards for complex goals: straightforward objectives in rich environments could let them discover knowledge researchers have not yet imagined. He says the aim is continual learning across simulated, digital, and physical environments, not a reward function that explicitly defines science, technology, or language.

Silver describes simple goals in complex environments as a possible source of emergent capabilities: coordination and communication could lead to trade, arithmetic, tools, science, technology, and more complex mathematics. The team’s proposed crucibles would foster that complexity through settings such as physics engines with millions of interacting objects or worlds with millions of interacting agents, rather than handcrafted puzzles.

Silver argues that human-language reasoning can carry outdated beliefs, such as a flat Earth or the four humors, and may become harder to use for monitoring powerful agents that produce long traces or engage in subterfuge. He says AI should move beyond human linguistic priors and that other monitoring approaches will also be needed.

Silver says large-scale training can make later adaptation faster: a system trained on massive amounts of data changed its behavior after fine-tuning on a small amount, suggesting that some capabilities could become faster to acquire after sufficient experience. But he says current systems still need substantial human involvement for breakthroughs, it is unclear whether LLM-based approaches are suitable for learning from physical experience, and a true superlearner remains a long way off.

GEO-POLITICS

Frontier Pacing — David Silver supports pacing frontier AI development in the real world so governments, organizations, and people have time to adapt. At the same time, he favors moving as fast as possible on simulation research to improve real-world outcomes.

“I'm actually, yeah, very much supportive of the idea that that there should be pacing of what happens for the frontier of AI in the real world.”—

TECHNOLOGY

Continual Discovery — Silver argues that a superintelligent system should be able to keep learning and discovering knowledge indefinitely across simulated, digital, and physical environments.

“And that's an incredible thing that no one has cracked yet in AI…”—

Learning Through Experience — Silver says he has long believed AI should learn from experience through reinforcement learning, with systems discovering things for themselves. He sees this approach as a possible route to intelligence and, ultimately, superintelligence.

“So I suppose I always had this conviction starting my PhD and going into the game of Go that the right approach to solve AI was through learning from experience…”—

Faster Adaptation — Silver says large-scale training can make later adaptation faster. A system trained on massive amounts of data surprised him by changing its behavior after fine-tuning on a small amount of data. He says this suggested that capabilities once considered slow or difficult could develop faster after sufficient experience.

“In particular, it definitely a surprise to me how a system that was trained on a massive amount of data could then be very quickly fine tuned on a very…”—

Learning From Experience

Silver says his conviction that intelligence emerges from learning through experience has persisted from his Go research to Ineffable, while LLMs demonstrated how scale can make later adaptation surprisingly fast.

Why Ineffable Began

Silver explains that he founded a separate lab to pursue experience-driven learning rather than join the crowded LLM path, and says the name reflects knowledge beyond what language can express.

Espeholt’s Reinforcement Learning Roots

Espeholt recounts his work scaling reinforcement learning at DeepMind and Google, while Silver credits his early systems and parallelization ideas as foundations for modern large-scale RL.

Defining the Super Learner

Silver defines a super learner as a system able to discover knowledge autonomously, from a blank starting point and across many kinds of experience, potentially exceeding human capabilities.

Building Crucibles

Ineffable is developing scalable reinforcement-learning infrastructure and simulated environments called crucibles, with milestones based on agents rediscovering important stages of human knowledge.

Simple Goals, Emergent Complexity

Silver argues that simple objectives in sufficiently rich environments can produce complex discoveries, as evolution and human history illustrate, without hand-specifying rewards for science or technology.

Designing Complex Environments

The guests describe crucibles built for emergent complexity through interacting objects, agents, and simulations, noting that even simple games like Go can support continuing strategic discovery.

Beyond Human Language

Silver argues that human language is not the endpoint of reasoning and may embed outdated assumptions, so more expressive systems should be able to discover ways of thinking beyond human priors.

The Case Against Human Data

Using AlphaZero as evidence, Silver says abandoning human examples enabled rapid self-discovery and argues that relying on human datasets may limit systems’ ability to reach knowledge beyond their training priors.

Timing the Super Learner

Silver believes experience-based superintelligence is feasible and that current compute, software, and research tools make this a promising moment, while emphasizing that research success remains uncertain.

Simulation’s Potential

Espeholt notes that simulations can be imperfect, but Silver argues they may still host superintelligence faster and more safely, since computation, economics, language, and scientific discovery can be represented in simulated worlds.

Alignment Through Superintelligence Science

Silver says language-based alignment may become insufficient as agents grow more capable, and proposes studying peaceful interaction among superintelligences in simulation before expanding real-world activity.

Pacing Real-World AI

Silver supports moving quickly on simulation research while pacing frontier deployment in the real world, giving society time to adapt and allowing safety-relevant science to advance.

GPU Cluster and Scaling

Espeholt describes Ineffable’s large GPU cluster and its all-GPU approach to simulation, training, and inference, then explains how nonstationary RL data creates scaling challenges distinct from language-model training.

IMPALA and SEED RL

Espeholt traces IMPALA and SEED RL to earlier efforts to scale reinforcement learning across GPUs and TPUs, and says Ineffable uses these algorithms and their derivatives in different environments.

A Research-Led Team

Silver describes a deliberately small team selected for exceptional ability, collaboration, and mission fit, and says Ineffable prioritizes research breakthroughs over incremental product development.

Three-Year Superintelligence Forecast

Silver predicts strong evidence for Ineffable’s approach within a year and says a true superintelligence could arrive in about three years, with coding agents accelerating progress.