David Silver’s $1B Bet That Reinforcement Learning Beats LLMs

In March 2016, an AI system played a Go stone that no human had ever considered. Move 37 — given a 1-in-10,000 probability by expert commentators — won the game and cracked centuries of Go theory. The man who built that system, David Silver, is now raising $1 billion on the bet that the same principle extends to everything.

Ineffable Intelligence, Silver’s reinforcement learning startup founded in London in November 2025, is raising approximately $1 billion at a $4 billion pre-money valuation. That would make it Europe’s largest-ever seed round — for a company with no product, no public roadmap, and a thesis that directly challenges the LLM paradigm powering most of the AI industry.

The raise is extraordinary on its face. But the real story sits in two layers: what Silver actually believes about reinforcement learning and superintelligence, and why the exact companies most committed to LLMs are reportedly writing him seed checks anyway.

Europe’s largest seed round — for a company with nothing to show

Ineffable Intelligence was formed in November 2025. Silver was appointed director on January 16, 2026. Fortune broke the story of his departure from Google DeepMind on January 30. Three weeks later, on February 19, The Decoder reported that Sequoia Capital — with partners Alfred Lin and Sonya Huang personally leading — is anchoring a $1 billion seed round at a $4 billion pre-money valuation, with Nvidia, Google, and Microsoft in active negotiations to participate.

That number makes Ineffable Intelligence’s seed larger than Humans&’s $480 million round in January 2026, and comparable in scale only to Safe Superintelligence’s $1 billion initial raise in September 2024 — another no-product, pure-thesis bet on a legendary researcher.

For context on how lopsided the AI capital map remains: Anthropic’s $30B Series G closed on February 12, 2026, at a $380 billion valuation. OpenAI raised $40 billion in March 2025. The entire post-LLM rebel cohort — SSI, Thinking Machines Lab, Ineffable Intelligence, AMI Labs, Core Automation — totals roughly $6-7 billion combined. Microsoft’s quarterly AI infrastructure spend exceeds that figure.

A billion dollars for a thesis and a track record. The thesis is what makes the check rational.

What Silver’s track record actually proves

Silver’s resume is short but unmatched. Every entry follows the same template: start from zero human knowledge, then surpass human performance through experience alone. AlphaGo defeated world champion Lee Sedol 4-1 in March 2016. AlphaZero achieved superhuman chess, shogi, and Go within 24 hours of training, starting from random play. MuZero mastered games in 2020 without even knowing the rules.

Then the pattern broke domain boundaries. In July 2024, AlphaProof scored 28 out of 42 at the International Mathematical Olympiad — one point short of gold medal standard — and solved the competition’s hardest problem, which only 5 of 609 human contestants answered. RL was no longer confined to board games. Silver received the 2019 ACM Prize in Computing ($250,000) for these contributions.

Silver is not the only DeepMind veteran to leave and raise large rounds on RL credibility. The founders of AlphaChip departed to start Ricursive Intelligence, raising $300M for AI-driven chip design. The common thread: researchers who built DeepMind’s identity around reinforcement learning are now leaving to pursue that original mission with venture capital, after the lab pivoted toward shipping Gemini and commercial LLM products.

What “Ineffable Intelligence” actually means — and why the name is the thesis

Ineffable: incapable of being expressed in words. The company name is not branding — it is Silver’s philosophical argument condensed to two words. If the most important knowledge is inexpressible in human language, then LLMs face a hard ceiling that is not one of engineering but of medium. They can only operate within the domain of human text. Move 37 is the concrete proof of concept: no human had language for why it was correct. It existed outside the vocabulary of Go theory entirely.

Silver laid out the technical case in April 2025. On the Google DeepMind podcast, he put it plainly: “We want to go beyond what humans know, and to do that we’re going to need a different type of method… our AIs [must] actually figure things out for themselves and discover new things that humans don’t know.”

The same month, he and RL pioneer Richard Sutton co-authored “Welcome to the Era of Experience,” arguing that current LLM approaches including RLHF have “side-stepped the need for value functions by invoking human experts in place of machine-estimated values.”

The paper outlines four pillars for what comes next: streams of lifelong experience, sensor-motor actions, grounded rewards, and non-human modes of reasoning. Each pillar is designed to break a specific dependency on human data. The company’s stated mission distills the ambition to a single sentence: “an endlessly learning superintelligence that self-discovers the foundations of all knowledge.”

As Pedro Domingos, the University of Washington professor and author of The Master Algorithm, quipped on X: “David Silver’s new startup is called Ineffable Intelligence. In that spirit, I’m calling mine Imaginary Intelligence. By the logic of AI startup names, it’ll be the first to achieve real intelligence.” The name landed — which, for a philosophical argument disguised as a company name, is the point.

Ineffable Intelligence, reinforcement learning, and the superintelligence hedge

Here is the number that reframes every headline about a “paradigm rebellion.” The full roster of post-LLM startups — SSI ($3 billion total), Thinking Machines Lab ($2 billion), Ineffable Intelligence ($1 billion), and a handful of smaller ventures — has raised roughly $6-7 billion combined. That is less than a single hyperscaler spends on AI infrastructure in a quarter. It is a rounding error on the LLM economy.

The investor list is the tell. Google is DeepMind’s parent company and the maker of Gemini. Microsoft has invested tens of billions in OpenAI. Nvidia’s revenue depends on LLM training demand. All three are reportedly in talks to co-invest in a startup whose founder’s explicit thesis is that their core AI approach cannot reach superintelligence. SiliconAngle noted the investor composition “suggests Ineffable Intelligence’s technology won’t directly compete with their offerings.”

The logic is straightforward: for less than 0.1% of their combined AI spending, the hyperscalers get a front-row seat if the paradigm shifts. This is cheap insurance, not a vote of no-confidence in LLMs.

Silver himself spent 15 years at DeepMind, co-authored the Gemini research paper in 2023, and argued on a public podcast that LLMs can never reach superintelligence. The contradiction was unsustainable — but for his former employer and its peers, the rational response is not to pick a side. It is to own both sides of the bet.

The case against Silver — and why it matters

Every major RL breakthrough on Silver’s resume was achieved in a closed, rule-based domain with perfect information and an objective reward function. Go has a clear winner. Chess has a clear winner. Competition math has verifiable proofs. The real world does not. “Write a good legal brief” and “design a useful product” have no formal reward signal, and how an endlessly learning superintelligence would define its own training signal in open-ended domains is the question Silver has not yet answered.

The simulation-to-reality gap compounds the problem. World models trained in simulation — including Minecraft, where DeepMind’s Dreamer V3 collected diamonds from scratch — do not automatically transfer to physical environments.

And while Silver frames RL and LLMs as divergent paradigms, the field is converging: RLHF, reinforcement learning with verifiable rewards (RLVR), and inference-time scaling show the two working in tandem, not opposition. As Yuxi Li’s analysis of the Silver-Sutton paper observes, “Achievements in games, maths and coding may not be straightforwardly generalizable/transferable to other problems.”

Meanwhile, the paradigm Silver is challenging refuses to plateau. GPT-5 runs autonomous biology labs. Gemini 3 Deep Think scores 84.6% on ARC-AGI-2. The target is moving — and it is moving fast.

The question a billion dollars cannot answer

Can reinforcement learning systems that discovered superhuman knowledge in board games, video games, and competition math do the same in domains where nobody can define what “winning” means? That is the honest, unanswered question at the center of Silver’s $1 billion bet — and the benchmarks that would prove it are ones nobody has agreed on how to measure yet.

The most telling signal in this story is not that Silver raised $1 billion. It is that Google, Microsoft, and Nvidia — the companies spending billions per quarter to prove LLMs are the path forward — thought it was worth paying for a front-row seat to watch him try to prove them wrong.

At DeepMind, Silver averaged one major system every two to three years. If that pace holds, Ineffable Intelligence’s first results could land by late 2027. The question those results need to answer is not whether RL can beat humans at another game — but whether it can beat them at something nobody knows how to score.