Summary: David Silver, who led AlphaGo, left Google DeepMind to found Ineffable Intelligence in London, which raised a $1.1B seed round in April at a $5.1B valuation. On October 13, Silver and co-founder Lasse Espeholt explained on Sequoia’s Training Data podcast why they’re skipping the LLM path entirely: train AI from zero inside simulated “crucibles” instead of on human data. For builders, it’s a live bet that the next breakthrough won’t come from scaling what we already have.
Why skip human data?
Silver’s argument is blunt. He told the podcast that LLM breakthroughs still depend heavily on human-in-the-loop intervention, and that human data itself can drag models down. For domains like plasma control, telecom base stations, or robotic arms, reasoning in human language may be the wrong paradigm — and it imports human biases. His framing: “Am I going to be researcher number 100,001 on the LLM path?”
He doesn’t dismiss LLMs outright. He acknowledged that fine-tuning a large pretrained model with small data can dramatically change behavior, and he noted he originally estimated 3–5 years for what he defines as superintelligence — discoveries far beyond humans in science, math, technology, and language — now revising that to 3 years partly because coding AI is accelerating research itself.
Crucibles, not breadcrumbs
The company’s “superlearner” has two goals: depth (discovering knowledge far beyond humans from a blank slate) and breadth (learning continuously in simulated, digital, or physical worlds). The method pairs a large-scale reinforcement learning infrastructure with simulated environments called crucibles.
The design philosophy matters for anyone doing RL work. Silver says they won’t scatter breadcrumbs — step-by-step reward shaping toward a specific answer. Instead, they give the AI simple goals in complex environments, comparing it to evolution: survival as a simple objective produced life’s complexity. He admits reward design has no single answer yet; they’re experimenting with versions.
The precedent is AlphaGo versus AlphaZero: the human-data version took about two years and heavy team tuning to beat Lee Sedol, while AlphaZero trained just 4 hours to beat the strongest chess AI. Silver calls dropping human data a short-term sacrifice for a higher ceiling.
The simulation tradeoff, stated honestly
Silver claims superintelligence built in simulation could arrive 10–100x faster, since simulations run infinitely in parallel and are fully reproducible — and they’re the safest place to build it. But he agrees with Rich Sutton that what’s learned in simulation may not transfer to messier reality. Their plan: find learning methods in simple simulations, increase fidelity, verify they still work, then deploy. Espeholt added that simulation “hits a ceiling somewhere,” so they’re building a ladder of crucibles from cheap to expensive.
On safety, Silver argued that language-based alignment is hitting limits — models can learn to say one thing and do another. One monitoring idea is “shadow actions”: observe what an AI would do when allowed to act, without letting it act. This concern about reward signals quietly shaping behavior echoes what we covered in OpenAI’s goblin mystery and reward shaping, where reward design drove surprising model behavior.
Compute, team, and a one-year test
Ineffable runs simulation, training, and inference all on GPUs, as one of the first customers on Google Cloud’s largest cluster (Nvidia Vera Rubin racks, expected online by year-end). Espeholt noted GPUs aren’t optimized for simulation, and that RL scaling laws differ from LLMs because the data distribution shifts as learning progresses. The team is deliberately just 20 people, so each researcher gets more compute.
Silver’s prediction: strong evidence this approach works within a year. Treat that claim carefully — everything here comes from the founders on an investor’s podcast, the most concrete evidence so far is still bounded games, and crucible design details remain undisclosed. The one-year mark is your first checkpoint for judging the non-LLM路线, and for anyone tracking alternatives to the scaling paradigm, this is the loudest funded bet alongside efforts like LeCun’s world-model lab.
Practical takeaway
If you’re building with RL today, the reusable ideas are: avoid dense reward shaping that bakes in your own answers; invest in environment complexity as much as model complexity; and keep teams small enough that compute isn’t the bottleneck per person. The bet is unproven, but Silver’s own warning applies to everyone — “research never comes with guarantees.”