- David Silver, who led AlphaGo’s development at Google DeepMind, has founded Ineffable Intelligence and closed $1.1 billion in seed funding at a $5.1 billion valuation.
- The company is betting on reinforcement learning—not large language models—as its path to superintelligence, positioning itself against the prevailing approach at most frontier labs.
- Silver plans to train AI agents inside simulations where they pursue objectives and collaborate with other agents, without dependence on human-generated training data.
- Silver has committed to donating his entire personal equity proceeds from the company to high-impact charities.
What Happened
David Silver, the researcher who led development of AlphaGo at Google DeepMind—the first AI to defeat a professional Go player at the highest level, in March 2016—has founded a new London-based AI lab called Ineffable Intelligence, which has raised $1.1 billion in seed funding at a $5.1 billion valuation, according to a Wired report published in April 2026. Silver left DeepMind to build a lab dedicated entirely to reinforcement learning (RL), a training paradigm in which agents learn by interacting with an environment and receiving reward signals, rather than by processing human-generated text. Lightspeed Ventures and Sequoia Capital are among the investors.
Why It Matters
Silver’s position directly challenges the architectural consensus driving the largest AI investments in 2026. OpenAI, Anthropic, Google DeepMind, and Meta are all building toward general AI on large language model foundations—a strategy grounded in the empirical observation that scaling compute and training data produces consistent capability improvements across reasoning, coding, and scientific tasks. Silver argues that this path is bounded, because any LLM’s capabilities are capped by the intelligence already encoded in its human-generated training corpus. The RL approach received institutional recognition in 2025, when Rich Sutton and Andrew Barto—whose foundational RL algorithms informed AlphaGo’s design—won the Turing Award for their contributions to the field.
Technical Details
Silver’s critique rests on a distinction between learning representations of existing human knowledge versus acquiring new knowledge through direct world interaction. “Human data is like a kind of fossil fuel that has provided an amazing shortcut,” Silver told Wired. “You can think of systems that learn for themselves as a renewable fuel—something that can just learn and learn and learn forever, without limit.”
He illustrated the proposed ceiling of LLM-based systems with a thought experiment: an LLM trained in a society that believed the Earth was flat would reproduce flat-Earth claims indefinitely, even while improving in other areas, because it cannot test beliefs against physical reality. Silver’s proposed technical solution is to embed RL agents inside simulations where they can pursue objectives and collaborate with other agents—acquiring capabilities that no human training corpus contains. Ineffable Intelligence has not disclosed the specific architecture or scope of these simulation environments, and Silver acknowledged that scaling from constrained, rule-defined settings like board games to real-world complexity is an open research problem that has not yet been solved.
Who’s Affected
The most immediate effects fall on AI research talent markets. Silver has already recruited researchers from Google DeepMind and other frontier labs, and his investors cited his professional reputation as a competitive advantage in hiring. “There’s only a very, very small number—less than a handful of people—who have done truly foundational work,” Sonya Huang of Sequoia Capital told Wired. “Dave is one of them.” For the broader industry, a well-capitalized lab explicitly rejecting LLM architectures introduces a credible institutional counterpoint to the dominant scaling hypothesis. Ravi Mhatre of Lightspeed Ventures described Silver’s career as “basically a single, coherent argument for being able to scale intelligence without human priors.”
What’s Next
Ineffable Intelligence’s central technical challenge is generalizing reinforcement learning from bounded, rule-defined environments to the open-ended complexity of real-world domains—a problem the RL research community has not solved at scale. Silver has said the simulation-based approach will also function as a safety evaluation mechanism: by observing how agents interact with one another, including with less capable agents, the company aims to identify behavioral risks before any deployment. “We can actually see what kind of behavior emerges from this,” Silver said. Silver has pledged to donate his entire personal equity stake—potentially worth billions of dollars if Ineffable Intelligence succeeds—to high-impact charitable organizations.
