David Silver, one of Google Deepmind’s earliest and most influential researchers, has left the lab to build a new AI company in London. The move matters because Silver is not simply changing employers; he is backing a different path toward superintelligence than the one currently driving much of the large language model race.
According to Fortune, Silver’s new company is called Ineffable Intelligence. It was registered in November 2025 and is currently seeking investors and AI researchers.
A major Deepmind departure
Silver was one of Google Deepmind’s first employees and played a central role in projects including AlphaGo, AlphaZero, and MuZero. Those systems helped define Deepmind’s reputation for AI that can learn powerful strategies through interaction, search, and trial-and-error methods.
A Google Deepmind spokesperson confirmed Silver’s departure to Fortune and called his contributions "invaluable." The company loses a researcher closely associated with some of its landmark work at a time when the AI field is arguing over what comes after today’s dominant model designs.
Silver’s stated focus is superintelligence, described in the source as AI that surpasses humans in every meaningful way. His view is that large language models are fundamentally limited because they are built on human knowledge. Instead, he is betting on reinforcement learning, where AI learns by trying actions, observing results, and improving through experience.
The split is about the route, not the goal
Deepmind itself is not opposed to superintelligence. The source notes that it has been Deepmind founder Demis Hassabis’s lifelong goal. The disagreement is more likely about how to get there.
Some researchers at Deepmind still see large language models as capable of going a long way. A recent discussion between Deepmind’s Adam Brown and former Meta AI leader Yann LeCun showed that this view remains active inside the field. Many at Anthropic and OpenAI also continue to see significant potential in LLMs.
Silver’s departure therefore highlights a strategic divide. One side continues to push language models and related systems further. The other argues that models trained mainly on human-created data may not be enough for systems that can exceed human knowledge in a deeper sense.
What the "Era of Experience" means
In April 2025, Silver co-authored a paper with renowned AI researcher Richard Sutton calling for a fundamental shift in AI development. Their proposal moves away from training on human knowledge and toward AI systems that learn from their own experience.
The authors call this the "Era of Experience." The idea centers on world models, which are simulations that let AI agents predict the consequences of their actions. Instead of only absorbing patterns from existing human data, an agent would build knowledge by acting, testing, and adapting.
A key part of that vision is continuous learning. Rather than being trained once and then deployed, AI agents would keep adapting to their environment over months or years, much like humans or animals. That approach treats learning as an ongoing process rather than a separate phase that ends before use.
Silver says Ineffable Intelligence aims to build an "endlessly learning superintelligence that self-discovers the foundations of all knowledge." That sentence captures the difference between his approach and the current LLM-centered model: the goal is not just to reproduce or reorganize what humans have already written, but to create systems that can discover more through their own interaction with the world.
Why skepticism around LLMs is growing
Silver is not alone in questioning whether the current Transformer architecture can reach superintelligence. The source describes a growing number of leading AI researchers who doubt that this architecture can achieve that goal, or even move far beyond what it can currently do, though some headroom remains.
Ilya Sutskever, former chief scientist at OpenAI, founded Safe Superintelligence to explore new approaches beyond large language models. Jerry Tworek, who helped develop OpenAI’s reasoning models, recently left the company and founded Core Automation.
Like Silver, Tworek sees continuous learning as one of the final missing pieces before true AI emerges. In a recent podcast interview, he explained that humans do not have a separate learning mode; everything happens at once. The implication is clear: if models do not learn directly from data as they operate, their capabilities may remain limited.
Tworek also argued that this kind of fundamental research is no longer possible at a heavily commercialized company like OpenAI. That point adds a practical dimension to the technical debate. Researchers who want to pursue new foundations may decide that startups are the better place to test those ideas.
What to watch next
Ineffable Intelligence is still at an early stage, seeking investors and AI researchers. Its importance lies less in what it has already built and more in what it represents: a bet that superintelligence will require systems that learn continuously from experience.
The current AI market remains heavily shaped by large language models, but Silver’s move shows that some of the field’s most experienced researchers are looking beyond them. If the next phase of AI depends on reinforcement learning, world models, and ongoing adaptation, the race may shift from scaling language systems to building agents that can keep learning over time.
For now, the split is not settled. LLMs still have supporters at Deepmind, Anthropic, and OpenAI. But Silver’s departure gives the alternative view a prominent new company and a clear thesis: the path to superintelligence may depend on experience, not just human data.