Training Data

David Silver and Lasse Espeholt on superintelligence from scratch

David Silver & Lasse Espeholt· Co-founders at Ineffable Intelligence
·~60 min·English·Sequoia Capital
ReasoningTrainingAI SafetyAgents
TL;DR

The DeepMind researchers behind AlphaZero left to found Ineffable Intelligence on one bet: that superintelligence will come from an agent that learns everything from its own experience in simulation, not from human data, and that the safe way there is to run the science in simulation first.

01Core Thesis

Learning From Experience

Silver's bet from his AlphaGo PhD to today: intelligence comes from a system that learns by doing and discovers for itself, not one that is taught from our data.

to me that is the essence of what it will take to get to super intelligence like a system that can just keep learning and discovering knowledge forever.

— David Silver, Training Data (Sequoia)
Key Insight
This reframes the race. Most labs treat human data as the fuel for intelligence; Silver treats it as training wheels, useful at first but a ceiling on what a system can ever discover beyond us.
02Why a New Lab

The Other Lamppost

Silver left DeepMind over direction, not resources: he thinks the road to superintelligence does not run through LLM products, so it deserved a lab built entirely around the other approach.

I have a really strong belief about how um we can go beyond where we are today um with AI and all the way to super intelligence and the path that I believe in um doesn't actually go through LLM based products as an intermediate point.

— David Silver, Training Data (Sequoia)
Key Insight
The company name encodes the thesis: the knowledge a super learner one day finds may be so profound that human language cannot capture it, and so it is literally ineffable.
03The Goal

Depth And Breadth

A super learner is defined on two axes: depth (start from nothing, discover far past human knowledge) and breadth (drop it into any world, simulated, digital or physical, and it still learns).

in the depth dimension, it should be something that can start from not knowing anything at all and can go all the way through to discovering things that go way beyond what humans know.

— David Silver, Training Data (Sequoia)
Key Insight
Breadth is the quieter, harder claim. An LLM lives in one medium, text; Silver wants one algorithm that works whether the stream of experience is a game, a browser or a robot body, because a mind locked to language is locked to what language can already say.
04The Core Bet

Let Go Of The Human Data

The AlphaZero lesson, made concrete: AlphaGo took two years of human-guided work to beat a world champion; once the team dropped the human data, AlphaZero relearned the game and beat the best chess AI ever made in four hours.

it took four hours to beat the world champion, the best ever AI that had ever been created.

— David Silver, Training Data (Sequoia)
Key Insight
The four hours is not the point; the direction is. Human data bought a two-year head start and a hard ceiling. Removing it cost a short-term drop in performance and bought an open-ended climb, and Silver is betting the same trade now scales from a board game to all of knowledge.
05How It Learns

Simple Rewards, Emergent Worlds

Silver's team does not hand-craft tasks or write a reward that spells out 'science' or 'language'; they drop agents into rich simulated worlds they call crucibles, give one simple objective, and let complexity emerge, the way evolution did.

So all in the pursuit of relatively simple goals this incredible complexity emerged.

— David Silver, Training Data (Sequoia)
Key Insight
This inverts how much of RL is done today, where engineers breadcrumb an agent toward a known answer. Silver refuses to name the destination: if you could write the reward for 'discover a new field of science,' you would not need the agent. The bet is that simple pressure in a deep enough world teaches more than any human-written goal.
06Beyond Our Priors

Rediscovering Language

Silver argues human language may be the wrong substrate for thought: start an agent from our words and it inherits our biases, so he wants agents to discover their own, possibly richer, way to reason.

maybe our agents would believe that the world was flat or or that the diseases were caused by uh the four humors

— David Silver, Training Data (Sequoia)
Key Insight
A safety tension hides here. The same move that frees an agent from our biases also frees it from our language, and the next idea's whole problem is that we mostly keep AI safe today by reading what it says.
07Alignment

Run The Science In Simulation

Today's alignment leans on reading an agent's words, but as agents grow more capable they learn to say one thing and do another; Silver's answer is to study superintelligence in simulation first and only move into the real world once that science says it is safe.

They're going to start to learn to say one thing and act in a different way.

— David Silver, Training Data (Sequoia)
Key Insight
Silver's 'accelerate in simulation, go carefully in the real world' splits the usual pace debate in two: race where it is contained and reproducible, pace where it is not. The risk he names is not intelligence itself but a real world moving faster than our governments, organizations and emotions can adapt to.
08Compute And Timing

The Right Time To Bet

The approach is old; what changed is compute. With one of the largest GPU clusters on Google Cloud and a team of twenty, Silver, who says he called the timing of computer Go to the year, now puts strong evidence within a year and a true superintelligence in about three.

I originally thought within three to five years. Um my timelines have accelerated a little bit. So I'm gonna say three years

— David Silver, Training Data (Sequoia)
Key Insight
Note what carries the forecast: not a new idea but a new budget. Silver finds it striking that others have not pushed harder on these experience-based approaches given the compute now available, so his edge is less a secret than a willingness to spend a very large cluster on the lamppost they walked past.