AlphaGo Architect Raises $1B to Build Superintelligence Without LLMs

David Silver, the researcher behind AlphaGo, AlphaZero, and some of the most consequential AI breakthroughs of the past decade, is betting $1 billion that the path to superintelligence does not run through large language models.

His new London-based startup, Ineffable Intelligence, is raising what would be Europe's largest seed round ever - $1 billion led by Sequoia Capital at a roughly $4 billion pre-money valuation. The pitch: reinforcement learning agents that learn from experience, not text.

TL;DR

~$1B seed round - the largest in European history, led by Sequoia Capital
~$4B pre-money valuation with Nvidia, Google, and Microsoft in talks to co-invest
Approach: reinforcement learning + world models, explicitly rejecting LLMs
Goal: "endlessly learning superintelligence" that discovers knowledge humans never produced
Founded November 2025 in London by former DeepMind lead researcher David Silver

"We want to go beyond what humans know... AIs need to discover new things humans don't," Silver said in a vodcast before his departure from DeepMind.

Why Silver Thinks LLMs Are a Dead End

The intellectual foundation for Ineffable Intelligence is a paper Silver co-authored with Richard Sutton, the godfather of reinforcement learning, titled "Welcome to the Era of Experience." The thesis is blunt: AI systems trained on human-generated data are hitting a wall.

The Data Problem

Most high-quality text sources have already been consumed by frontier models. Progress driven by supervised learning alone is, in Silver and Sutton's words, "demonstrably slowing." If the current crop of LLMs - the GPT-5s, Claudes, and Geminis that dominate today's rankings - are fundamentally limited by the data they were trained on, then making them bigger will only yield diminishing returns.

The Four Pillars

Silver and Sutton sketch an alternative built on four pillars:

Streams of lifelong experience - AI agents that never stop learning, accumulating knowledge over their entire operational lifetime
Sensor-motor actions - Direct interaction with environments, not passive consumption of text
Grounded rewards - Feedback signals tied to real outcomes, not human ratings
Non-human modes of reasoning - Discovery of strategies and knowledge that no human has ever produced

This is the architecture that produced AlphaGo's famous Move 37 - a play so unconventional that professional Go players initially thought it was a mistake, before realizing it was brilliant. Silver wants to generalize that to everything.

What They Measured (and What They Didn't)

Silver's track record with reinforcement learning is extraordinary in domains with clear rules: Go, chess, StarCraft, mathematical proofs. AlphaGo beat Lee Sedol. AlphaZero taught itself chess in four hours and crushed Stockfish. AlphaProof solved International Mathematical Olympiad problems.

But every one of these successes has something in common: well-defined win conditions, bounded state spaces, and fast simulation. The open question - the one Silver has not yet answered publicly - is how you define "winning" when the goal is general intelligence. What does the reward function look like when you're trying to cure cancer or write legislation?

The Competitive Landscape

Silver is not the only prominent researcher betting against LLMs-as-the-only-path. The superintelligence startup space is getting crowded:

Company	Founder	Approach	Funding	Valuation
Ineffable Intelligence	David Silver (DeepMind)	Reinforcement learning + world models	~$1B (seed)	~$4B
Safe Superintelligence (SSI)	Ilya Sutskever (OpenAI)	Undisclosed, safety-first	$3B	$32B
Reflection AI	Various (DeepMind alumni)	Undisclosed	Undisclosed	Undisclosed
AMI Labs	Yann LeCun (Meta)	World models	Raising ~$500M	Undisclosed

Ilya Sutskever's SSI has been the most visible comparison. But where Sutskever has been tight-lipped about methodology, Silver is loudly publishing his theoretical framework. And where SSI has explicitly avoided shipping products in favor of pure research, Silver's reinforcement learning background suggests he may have a more tangible path to demonstrable results - or at least, more demonstrable intermediate milestones.

The involvement of Nvidia, Google, and Microsoft as potential co-investors alongside Sequoia is also notable. Google investing in a company founded by one of its own former star researchers signals either deep confidence in Silver's vision or a hedging strategy against the possibility that their own LLM-centric Gemini bet isn't the whole answer.

The London Factor

Ineffable Intelligence is deliberately staying in London, where Silver holds a professorship at University College London. The company has been recruiting former DeepMind colleagues and researchers from other leading labs.

An investment banking executive quoted by Sifted called the round "further evidence that the UK and wider European ecosystem can produce globally significant companies." At $4 billion pre-money, Ineffable Intelligence would immediately become one of Europe's most valuable AI startups, competing in the same valuation tier as Mistral AI.

The choice matters for practical reasons too. London's AI talent pool is disproportionately deep in reinforcement learning - a direct consequence of DeepMind being headquartered there for over a decade.

Should You Care?

Here is what I know after years covering infrastructure and model releases: every generation of AI progress has been driven by a paradigm shift, not by scaling the previous paradigm harder. Neural networks replaced hand-crafted features. Transformers replaced RNNs. Scaling laws replaced architecture search.

Silver is making a specific bet that the next shift is from learning-from-text to learning-from-experience. The intellectual pedigree is impeccable - this is the person who actually built systems that discovered superhuman strategies from scratch. The financial backing is serious. The theoretical framework is published and internally consistent.

But there are real reasons for skepticism. Reinforcement learning, for all its game-playing triumphs, has a well-documented history of struggling in open-ended environments. Defining reward functions for real-world tasks remains an unsolved problem. And the current generation of LLMs - augmented with reasoning, tool use, and agentic capabilities - keeps surprising people with how far text-based learning can go.

The $1 billion question is not whether Silver is smart enough to pull this off. It is whether the problem he is solving - general-purpose reinforcement learning without clear win conditions - is tractable with current compute and techniques. If it is, Ineffable Intelligence might produce the most important AI system since AlphaGo. If it isn't, it will be the most expensive seed-stage lesson in the history of European tech.

Either way, the LLM monoculture just got its most credible challenger.

Sources: