
David Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience
- 1h 8m
- Transcript
Summary
Written by a model from the transcript; not yet reviewed.
This episode explores the founding and mission of Ineffable Intelligence, a new AI lab aiming to build a 'super learner' that discovers knowledge from scratch rather than relying on human language data and large pre-trained datasets. The guest founders explain why they left DeepMind and Google to pursue a reinforcement learning path outside the dominant LLM paradigm. They discuss creating simulation 'crucibles' for emergent complexity, the parallels to their work on AlphaGo, and their unique high-compute setup for scaling reinforcement learning.
- Abandoning massive human datasets in favor of reinforcement learning from scratch can ultimately unlock far deeper capabilities and faster progress. 31:32
- Simple environment dynamics and agent interaction can give rise to endless layers of complex behavior without hand-crafted rewards. 16:51
- Simulation offers a safe, infinitely parallelizable way to develop superintelligent systems before ever introducing them to the real world. 40:22
- Keeping a small, highly talented team ensures a high compute-to-person ratio, allowing individual researchers to run massive experiments. 1:02:23
Chapters
- 0:00Podcast Intro and Host Greeting
Hosts introduce the episode and welcome guests.
- 0:49Guest Introduction and Background
Hosts introduce Dave Silver and Lasse Espeholt, co‑founders of Ineffable Intelligence, and discuss their early careers.
- 2:06Why Ineffable Pursues Learning from Experience
Dave explains his conviction that learning from experience, not language data, is the path to superintelligence.
- 3:51Founding Ineffable and Naming the Lab
Dave and Lasse discuss why they left DeepMind, founded Ineffable, and the meaning behind the lab’s name.
- 9:48Defining a Super Learner and First Contact
The founders describe what a super learner is and how they envision first contact with superintelligence.
- 15:13Building Crucibles for Emergent Discovery
They explain how they design complex simulation environments—crucibles—to enable agents to discover knowledge autonomously.
- 28:35Removing Human Priors to Enable Autonomy
Dave discusses the importance of eliminating human data biases to allow agents to discover new knowledge beyond human limits.
- 34:13Timing, Compute, and the Future of Superintelligence
The founders reflect on the timing of their work, the role of compute, and their confidence in achieving superintelligence.
- 40:16Simulation vs. Real‑World Deployment
They debate whether a superintelligence can be built entirely in simulation and when to transition to real‑world environments.
- 44:10Alignment and the Science of Superintelligence
The conversation turns to alignment, the limits of language‑based approaches, and the need for a scientific understanding of superintelligence.
- 47:00Agents’ Language and Monitoring Challenges
Discussion of how advanced agents may bypass human‑language monitoring and the need for alternative safety measures.
- 52:55Pacing AI Development and Simulation Acceleration
Exploration of balancing rapid AI progress with careful real‑world deployment, emphasizing simulation as a safety tool.
Highlights
- 18:18
Simple Goals Create Complexityinsight
Optimizing very simple quantities in a complex environment can naturally generate extreme complexity, much like evolution.
- 31:08
After abandoning human data, the AlphaZero algorithm learned chess and beat the world champion in just four hours.
- 41:07
Turing Complete Simulationinsight
Any Turing complete simulation can provably express everything needed for computation, economics, and language.
- 53:42
The lab secured the first large Rubin cluster on GCP, featuring high bandwidth between racks tailored specifically for fast reinforcement learning iterations.
- 1:01:06
Small Team Size Strategystrategy
The founders deliberately capped the company at twenty people to maintain an exceptionally high talent bar and a massive compute-to-person ratio.
- 1:06:33
Superintelligence Timeline Predictioninsight
The guest predicts seeing strong evidence of their approach within a year and achieving true superintelligence within three years.
Quotes
“I think that L L M's are doing a lot of great things and I have a lot of respect for what's being achieved.”
Unnamed speaker · 12:57 “Does that too count as a super learner?”
Unnamed speaker · 12:49 “What I mean by a super learner is something which first of all discovers completely autonomously.”
Unnamed speaker · 13:03 “So if it is possible for it to be built in simulation, I believe it will happen ten times or a hundred times faster.”
Unnamed speaker · 40:22 “Science in some sense is the study of our world.”
Unnamed speaker · 42:02 “We're hurtling towards an outcome with lots of interacting superintelligences.”
Unnamed speaker · 47:25 “Like it or not, I believe that the approach of monitoring the human language that these systems use is going to be increasingly difficult to, as agents get more powerful.”
Unnamed speaker · 48:28 “I do think we're going to end up in a world with lots of complex, interacting super intelligent beings.”
Unnamed speaker · 50:24 “So I think we will see some extremely strong evidence that we're on the right track within a year.”
Unnamed speaker · 1:06:33 “My timelines have accelerated a little bit, so I'm going to say three years because of what can be done with coding agents to accelerate things.”
Unnamed speaker · 1:07:02
About this episode
David Silver spent two years building AlphaGo on human data. When he removed human data altogether building AlphaGo Zero, the AI surpassed AlphaGo in under two days. He left DeepMind to run that same bet in pursuit of superintelligence. David and Lasse Espeholt, co-founders of Ineffable Intelligence, explain their plan to build a “superlearner”—an agent that starts knowing nothing and discovers science, mathematics, and language for itself by direct experience instead of through a pre-training run. Silver argues that human-language reasoning is a ceiling, not a foundation, that alignment via chain-of-thought monitoring is already reaching its limits, and that the science of how superintelligences coexist should be worked out in simulation before anything touches the real world. Espeholt describes the GPU cluster and the lineage from IMPALA and SEED RL that makes the experiment possible. Silver’s forecast: strong evidence within a year, superintelligence in three. Hosted by Sonya Huang, Sequoia Capital 0:00 – Introduction 1:09 – What held true from Go to today, and what the LLM era changed 3:35 – Why leave DeepMind to start a new lab 5:58 – Why "Ineffable" (and the secret second reason) 7:03 – Lasse's path: scaling RL at DeepMind and Google Brain 9:31 – Defining a super learner and first contact 12:35 – Super learners vs. LLMs that make discoveries 13:59 – Day to day: the RL stack, crucibles and human-journey milestones 16:13 – Don't breadcrumb the objective: simple rewards, emergent complexity 21:16 – What a crucible looks like 24:14 – Go: one agent is boring, two is infinite 25:41 – Will agents rediscover language? 27:22 – Removing human priors: AlphaGo to AlphaZero 32:05 – Keeping human bias out of the crucibles 33:35 – Confidence, timing and super forecasting 36:24 – Is compute the driver? 38:13 – The ceiling on simulation and Rich Sutton's objection 43:44 – Alignment and the science of superintelligence 47:45 – Monitoring agents that don't think in English 51:01 – Pacing the frontier 52:54 – The compute setup: Vera Rubin on GCP 55:18 – Training, inference, simulation and RL scaling laws 57:40 – IMPALA and SEED RL 1:00:58 – 20 people and the compute-per-person ratio 1:04:49 – A research-led culture 1:05:47 – Predictions
Transcript
12,496 words · 3 speakers
Topics
- simulation
- compute
- reinforcement learning
- DeepMind
- GPUs
- AlphaGo
- LLM
- deep learning
- Google Brain
Page updated · transcript and summary are machine-generated