David Silver

1. Key Themes

The Super Learner Defined: Depth and Breadth of Autonomous Discovery

Ineffable's goal is a system that learns like humanity collectively did: starting from nothing and discovering everything, with no human blueprint or giant dataset. David Silver frames it along two axes. Depth is learning from zero to beyond human knowledge. Breadth is being droppable into any stream of experience (simulated, digital or physical).

David Silver: "we started with this kind of gray mush in our head... everything we learned, we we discovered it for ourselves and that's an incredible thing that no one has cracked yet in AI." 00:00:00 On breadth: "it should be possible to drop it into any stream of experience. So whether, you know, it's it's like a simulated world, or a digital world, or a or a physical world... it should be something that can just learn and learn and learn." 00:10:52

Letting Go of Human Data Is a Deliberate Bet, Not a Religion

The core strategic wager is that giving up the "whole internet of knowledge" costs short-term performance but unlocks a far higher ceiling. The AlphaGo-to-AlphaZero arc is the evidence: two years of human-in-the-loop work were beaten in hours once human data was dropped. Silver stresses the position is pragmatic, not ideological.

David Silver: "we are making a a bet at Ineffable, and the bet we're making is that we can go further by giving up like on on the priors that everyone's relying on today, like the human data." 00:28:23 On the AlphaZero result: "It took us 2 years to get to AlphaGo and beat Lee Sedol. When we got AlphaZero right... it took just a few hours." 00:30:45 "in the game of chess, it took 4 hours to beat the world champion." 00:31:08 On pragmatism: "It's not like a religion that you can't have any human prior anywhere... It's more like a pragmatic position." 00:32:51

Simple Objectives in Complex Worlds Beat Hand-Crafted Reward Functions

Rather than "breadcrumb-trailing" agents toward specific outcomes, Ineffable gives agents simple objectives in rich environments and lets complexity emerge, analogous to how evolution's fitness signal produced all of life's complexity.

David Silver: "A lot of the way things people do RL with LLMs at the moment is to kind of say... the agent should be learning about this problem, and then this problem... we're kind of pulling the agent by its nose towards a particular desired outcome. We're not taking that approach." 00:16:22 And: "the pursuit and optimization of of very simply easy-to-describe quantities can often give rise to incredibly hard-to-describe complexity and emergence." 00:18:10 His chain of emergence: coordinating, then communicating, then trading, then arithmetic, then tools, then science, then more complex maths. 00:19:08

Crucibles: Environments Where Complexity Comes From Interaction, Not Handcrafted Puzzles

The second pillar, alongside the learner itself, is a set of simulated worlds ("crucibles") engineered for emergent complexity through physics engines and multi-agent interaction, a contrast to today's benchmark-style task sets. The Go example shows how adding a second agent turns a trivial game into an "infinite onion."

David Silver: "It's not like we've kind of handcrafted a whole load of puzzles for the agents... There's just a huge amount of emergent complexity there." 00:21:34 On Go: "If this was a single agent game, it would be like the most boring game you could imagine... The moment you put a second agent into the environment, you get this emergent complexity." 00:24:21 Lasse Espeholt adds the practical design principle: "we want to like create this range of crucibles where we have like simple ones that can execute quickly, we can iterate on, but where we can still see like very complex uh behavior emerging." 00:23:03

Simulation as the Fastest and Safest Path to Superintelligence

Silver argues superintelligence can be built entirely in simulation, which is faster (10-100x via infinite parallelism and reproducibility) and safer. He supports the claim by showing that computation, economics, language, math and science are all expressible inside a sufficiently rich simulator.

David Silver: "if it is possible for it to be built in simulation, I believe it will happen 10 times or 100 times faster, just because simulation can be infinitely parallelized, you get perfect reproducibility... It's also the safest place for a superintelligence to be built." 00:40:15 "If you have a Turing complete simulation, suddenly, everything that we mean by computation... can be expressed completely inside simulation." 00:40:15 He concedes Rich Sutton's point: "you shouldn't expect that the things that you learn in simulation to like transfer, like sim to real," but argues the algorithm, not the learned policy, is what transfers. 00:39:08

Timing Is the Real Variable, and Compute Is What Makes Now the Moment

Silver frames the bet as a timing call, backed by a forecasting record, with compute as the largest enabler. He estimates Ineffable can run experiments with roughly 100,000x the compute of prior self-discovery efforts like OpenAI's Hide and Seek.

David Silver: "whether this succeeds or fails, I think is about timing... For example, with Computer Go... I made a prediction, I think, 10 years before... and I got it right to the year." 00:34:13 On compute: "I think, you know, we could probably run experiments that like are use 100,000 times more compute." 00:36:42 And: "it's kind of amazing that people haven't come back and really pushed on these types of approaches, given the, the, the availability of compute." 00:36:42

Alignment: Language-Based Monitoring Won't Scale, So Build the "Science of Superintelligence" in Simulation

Silver argues that chain-of-thought and human-prior-based alignment is already hitting limits as agents grow more powerful. His proposed answer is to study superintelligent agents in simulation first, learn what conditions yield peaceful interaction among many superintelligences, and gate real-world deployment on that evidence.

David Silver: "As agents are trained more and more, they, they are kind of, um, going to start to overwhelm their priors. They're going to start to learn to say one thing and act in a different way." 00:44:24 On the commitment: "we will gate anything we do in the real world on having sufficient, um, evidence from this kind of science of superintelligence that we are really confident that things are going to go well." 00:44:24 His summary stance: "accelerate in simulation... careful in the real world," confirmed with "Yeah. I think that would be my summary." 00:51:07

A Small, Research-Led Team With a High Compute-per-Researcher Ratio

Ineffable has deliberately hired only about 20 people, optimizing for talent, kindness, mission alignment and skill fit, and keeps compute per head high so researchers don't fight for resources. Silver says the company pushes back on product pressure on purpose.

David Silver: "We have hired 20 people in total... it's not just about talent for us. It's also about bringing together people who are... kind, collaborative human beings." 01:01:06 Lasse Espeholt: "this relatively small amount of people so far means that we have a lot of compute per person... So, we're not like fighting for resources uh from day one." 01:02:13 Silver: "we push back on on on product pressure, because we believe that if we make these breakthroughs, it will lead to, you know, far bigger products downstream." 01:04:58

Compute Stack: GPU-Native RL, Vera Rubin Racks, and the Simulation Bottleneck

Espeholt explains that Ineffable is the first customer of the largest Vera Rubin cluster on GCP and builds everything GPU-native (training, inference and simulation) to avoid the CPU-sandbox scaling problem that LLM labs now face. The remaining bottleneck is that GPUs have been tailored to training and inference, not simulation, which pushes them toward model-based RL and replay approaches.

Lasse Espeholt: "we are the first, first customers of the largest Vera Rubin cluster on GCP." 00:53:37 "one rack can basically be considered like one machine. There's like a very high bandwidth between the nodes in one rack, um, which is very important for reinforcement learning where you have like really fast iteration cycles." 00:53:37 On the CPU problem: "I think like the LLM companies now have kind of like the problems that we had uh many years ago where they need like sandboxes running on CPUs... and we're trying to not put ourselves in that situation." 00:53:37 On simulation: GPUs "have been more and more tailored towards training... and inference, but less about simulation." 00:55:51

2. Contrarian Perspectives

Superintelligence Does Not Route Through LLM Products

Most of the industry treats LLM products as the on-ramp to superintelligence. Silver argues the path is orthogonal, and that thinking in human language may actively hold agents back.

David Silver: "the path that I believe in, um, doesn't actually go through LLM-based products as an intermediate point." 00:03:57 On language: "it might just be that all of this human data that's being used in LLMs is kind of weighing things down." 00:13:30 He also says "anyone who thinks that human language is the end point of... the best way to to reason... that perspective is a little naive." 00:25:50

The "Lamppost" Argument: Hundreds of Thousands of Researchers Are on One Approach

Silver's decision to leave DeepMind rested on a Thiel-style scarcity argument: when the whole field crowds one paradigm, the marginal researcher is more valuable on the neglected one.

David Silver: "the world already has tens of thousands if not hundreds of thousands of really talented researchers all pursuing exactly the same approach... Am I going to be the 100,001 person that's like pursuing that approach?" 00:04:58 "there's this other lamppost that right now, no one is looking under." 00:05:29

LLM-Style RL Is "Semi-Autonomous, Human-in-the-Loop" With Thousands of People in the Loop

Silver reframes the leading labs' RL pipelines as human-labor-intensive systems that cap out well short of true learners, drawing the parallel to AlphaGo's own two-year human-in-the-loop phase.

David Silver: "what I see in a lot of the kind of LLM companies today is they're basically, you know, building uh semi-autonomous, human-in-the-loop kind of learning algorithms with thousands of people in the loop of the learning progress, which is, you know, it works and and a lot of respect for how far it gets, but... until we step back from that... I don't believe we'll get all the way to something which is truly a super learner." 00:31:33

Hard Research Should Be a 50-50 Bet, and Safety Research Is Under-Diversified

Silver says he deliberately works only on questions where he's "about 50-50," because 95%-certain research is boring. On safety, he argues that if the field relies only on language-based monitoring, "we're in trouble."

David Silver: "I think it's not very interesting to do research where you already know the answer... good research is about asking questions where you're really not sure what the answer is in advance." 00:27:55 On alignment: "if people aren't looking at this because they're only looking at the language-based approach, I think we're in trouble, because I think that language-based approach is not enough in itself." 00:48:19

Pace the Real World, Not the Simulation

Where much of the AI debate frames "pace the frontier" as a single dial, Silver splits it: slow real-world deployment, but sprint in simulation, because simulation work produces the science that makes real-world deployment safe.

David Silver: "I think we should move as fast as possible in simulation... I think that kind of work should happen as fast as possible to make sure that when things really happen in the real world... things have the best possible opportunity to go well." 00:51:07

3. Companies Identified

Ineffable Intelligence

A new European AI lab founded by David Silver, Lasse Espeholt and others, building a "super learner" via reinforcement learning from experience in simulated "crucibles." Mentioned as the subject of the episode: a roughly 20-person, research-led lab with a very large GPU cluster and a mission to make "first contact with superintelligence." David Silver: "we're building out environments which are are capable of super learners like discovering everything... We call those crucibles." 00:14:52 On timeline: "I think we will see some extremely strong evidence that we're on the right track, um, within a year." 01:06:32

DeepMind

Google's AI lab, where both founders spent years (AlphaGo, AlphaZero, AlphaStar, AlphaProof). Mentioned as the origin of the founders' RL work and of the culture they respect, and as the place where Ineffable's approach would have been "the corner of an organization" focused on LLM products. David Silver: "I had a fantastic time at DeepMind... But at the end of the day... this deserved to be more than just the corner of a of a organization that's focused on on LLM-based products." 00:03:57

Google Brain

Google's research group, whose Amsterdam office Espeholt founded. Mentioned as the setting for SEED RL and MetNet work. Lasse Espeholt: "some of my work at Google Brain, after I left DeepMind to found the Google Brain Amsterdam office, was to scale to like racks of TPUs. And that produced called like SEED RL." 00:58:16

Google Cloud (GCP) and Google TPUs

Cloud provider for Ineffable's cluster, and the TPU platform that SEED RL was designed to scale across. Mentioned as a key compute partner. Lasse Espeholt: "we are the first, first customers of the largest Vera Rubin cluster on GCP." 00:53:37

NVIDIA

Supplier of the GPUs (Vera Rubin) powering Ineffable's training, inference and simulation. Mentioned for its rack-scale architecture suiting fast RL iteration cycles. Lasse Espeholt: "we have like this like Vera Rubin cluster which is like based on GPUs, right, like NVIDIA GPUs, which is uh really fortunate for us, because it means that we can easily do simulation, um, and like training and inference." 00:53:37 Sonya Huang noted both Google and NVIDIA "have been great partners for you." 00:53:03

OpenAI (Hide and Seek)

Mentioned for its early multi-agent "Hide and Seek" experiment as the nearest prior example of self-discovered tool use, but one Silver believes Ineffable can exceed by orders of magnitude of compute. David Silver: "something like Hide and Seek from OpenAI which... was like an early example where... agents discovered, um, kind of some basic tool use. Um, I think... we could probably run experiments that like are use 100,000 times more compute." 00:36:42

Google Research (MetNet)

Partner on MetNet, the short-term weather forecasting model built in Google Brain Amsterdam, cited as an example of ML beating physics-based simulation. Lasse Espeholt: "this is like a weather forecasting model we did in the Google Brain Amsterdam office together with Google Research, where like we're doing like a machine learning model to predict short term like weather forecasts. And this was like way better than the physics-based simulations." 00:39:08

Anthropic

Not named directly, but Silver refers to its CEO (see Dario below) when expressing support for pacing frontier AI.

Sequoia Capital (implied by hosts' firm)

Not directly named in the transcript; omitted from further analysis.

AlphaGo / AlphaZero / AlphaStar / AlphaProof (DeepMind projects)

Reference projects behind Silver's forecasting record and the human-data-removal evidence. AlphaStar was the team that connected several co-founders. David Silver: "I think, actually, four of our co-founders come from from that team" (AlphaStar). 01:02:59

4. People Identified

Co-founder of Ineffable Intelligence; former DeepMind researcher who led the AlphaGo, AlphaZero, AlphaStar and AlphaProof lines of work, with a 2009 PhD on RL applied to Go. Mentioned as the guest and as a self-described forecaster whose timelines have repeatedly proved right. Sonya Huang: "you demonstrated for the world for the first time that superhuman in one domain machine intelligence was possible, and I think you were really the figure that inspired the entire current generation of of reinforcement learning." 00:01:12 On timelines: "I originally thought within 3 to 5 years, um, my timelines have accelerated a little bit, so I'm going to say 3 years." 01:06:32

Lasse Espeholt

Co-founder of Ineffable Intelligence; creator of IMPALA, SEED RL and MetNet, and founder of the Google Brain Amsterdam office. Mentioned for foundational large-scale RL work and as the lead on Ineffable's RL stack. David Silver: "he actually worked on some of the original approaches that that, um, were the first attempts to actually scale up reinforcement learning... all of the ideas that people use today kind of, um, sort of hail back to to those origi- that original work that that Lasse did." 00:08:46 "he was the the first to figure it out." 00:09:12

Wojtek (Wojciech Marian Czarnecki) and Chris

Co-founders of Ineffable whom Espeholt knew "from back in the day." Named only by first name or partial name in the transcript. Lasse Espeholt: "I knew Wojtek and and Chris, uh, from back in the day, and I thought like the team was a great." 00:08:18

Rich Sutton

Reinforcement learning pioneer, referenced as a friend of Silver's and as a skeptic of simulation. Mentioned for his argument that sim-learned skills won't transfer to reality, which Silver partially concedes. Sonya Huang: "we re- recently had your, your friend, Rich Sutton, on the show, and I asked him about simulation and he said, yeah, he was very dismissive of simulation." 00:38:13

Yannis (Yannis Assael)

Former AlphaGo team member backed by the hosts' firm. Mentioned for his view of Silver's ability to ground ambitious visions in practical milestones. Sonya Huang: "he he told us that, you know, one of the things that's distinct about you is you can have these incredibly ambitious visions, but you also ground the team in very practical uh research milestones." 00:13:59

Demis (Demis Hassabis)

CEO of Google DeepMind, referenced by first name as one of the leaders advocating pacing of the frontier. David Silver: "I'm supportive of, of the, yeah, suggestions, like the kind of direction that, that people like, um, you know, Demis and, and Dario and so forth are, are suggesting." 00:51:07

Dario (Dario Amodei)

CEO of Anthropic, referenced by first name as an advocate of pacing the frontier. David Silver: "people like, um, you know, Demis and, and Dario and so forth are, are suggesting. I think it's, um, generally a good idea." 00:51:07

Lee Sedol

World-champion Go player defeated by AlphaGo, mentioned as the milestone that the two-year human-in-the-loop effort achieved. David Silver: "It took us 2 years to get to AlphaGo and beat Lee Sedol." 00:30:45

John von Neumann (implicit via "von Neumann architecture")

Referenced as the origin of the computer architecture Ineffable does not want to bias agents toward reinventing. Lasse Espeholt: "we don't want to like bias it towards uh building a machine based on like the von Neumann architecture." 00:32:20

Isaac Newton (implicit via "Newtonian physics")

Referenced only in an analogy about open-minded discovery in science before quantum physics and relativity. David Silver: "imagine you were back in, you know, the days of Newtonian physics, and and no one even knew that that like quantum physics was a thing." 00:20:43

Sonya Huang

Host of Training Data. Mentioned as the person who cold-emailed Silver as the show's first outreach. Sonya Huang: "Dave, you were the first person I cold-emailed to come on Training Data when we started the show 100-ish episodes ago." 00:00:49

5. Operating Insights

Keep the Team Tiny and the Compute-per-Head Ratio High

Ineffable hired only 20 people and explicitly manages for abundant per-researcher compute so experimentation is never resource-starved. The tactical implication is that headcount is a lever you can deliberately hold down to keep iteration velocity high. Lasse Espeholt: "these days, you can you can do a lot of experiments just like one person. So, it's really also to keep like the compute to person ratio quite high. So, we're not like fighting for resources uh from day one." 01:02:13

Hire on Four Axes and Say No to Friends

Silver's hiring filter has four parts: talent, kindness/collaboration, mission alignment, and skills that fit current needs. He says he's turned down people he's worked with before because the bar is so high. David Silver: "all of those four things are really important and, um, and yeah, we kind of have very high standards... I've had to say no to a lot of people I've worked with in the past just because our, our bar is so super high." 01:01:06 01:02:59

Structurally Insulate Research From Product Pressure

Rather than letting product timelines dictate research, Ineffable's culture is built so researchers pursue breakthroughs and product pressure is actively pushed back, on the thesis that downstream products will be far larger. David Silver: "We're creating... a haven for research, you know, where where we push back on on on product pressure, because we believe that if we make these breakthroughs, it will lead to, you know, far bigger products downstream." 01:04:58

Build a Portfolio of Fast and Slow Experiments (Cheap Crucibles for Iteration)

Rather than going straight to expensive realistic physics, build a range of environments, including cheap ones that execute quickly, so research can iterate fast while still exhibiting emergent complexity. Lasse Espeholt: "going all the way to like 100% realistic physics will be like very computational expensive, right? So, we want to like create this range of crucibles where we have like simple ones that can execute quickly, we can iterate on, but where we can still see like very complex uh behavior emerging." 00:23:03

Anchor a Vague Mission in Milestones Borrowed From a Known Historical Trajectory

To ground a "discover everything" mission, the team uses the human history of discovery as a milestone ladder, asking whether agents rediscover those landmarks. David Silver: "we ground it in in in milestones, and the milestones are kind of analogous to, you know, what happened with the human um journey... those become natural milestones for us to for our agents to rediscover the great moments that that humans went through." 00:15:50

6. Overlooked Insights

Simulation Speed Is Gated by GPU Design, Which Is a Hidden Bottleneck for the Whole Paradigm

Espeholt drops, almost in passing, that GPUs have been progressively tailored to training and inference and "less about simulation." This quietly identifies the true scaling bottleneck for experience-driven learning: the data-generation side, not the learner. It implies a hardware opening for simulation-optimized silicon, and explains why Ineffable invests in model-based RL, replay buffers and GPU-resident simulators. It also means the "CPU sandbox" problem that LLM labs are now hitting with agentic RL is a preview of a structural constraint the experience-first approach has already anticipated. Lasse Espeholt: "they've been more and more tailored towards training, right? And, and maybe in the last like few years, also inference, but less about simulation, right?" 00:55:51 And: "LLM companies now have kind of like the problems that we had uh many years ago where they need like sandboxes running on CPUs... and we're trying to not put ourselves in that situation." 00:53:37

Reinforcement Learning's "Non-Stationary" Scaling Laws Are an Unsolved Research Frontier, and Ineffable Is Scaling Into It Blind

When asked directly whether they see scaling laws, Espeholt answers that RL scaling is fundamentally different because the data distribution shifts as the agent progresses, calls it "a very interesting research direction," and says they're "actively pushing towards" it as they move to a larger cluster. That is a notable admission: the central empirical bet (that more compute buys more discovery) has no established, predictive law yet. If Ineffable finds one, it becomes both a moat and a forecasting tool for the whole field. Lasse Espeholt: "with reinforcement learning, it's a little bit different, because uh the problem is like non-stationary. So, as you progress in the problem, the data changes. The data distribution changes." 00:57:05 "especially as we like transition from like a smaller cluster to the bigger cluster, we want to be ready for that." 00:57:42