David Silver, a well-known Google DeepMind researcher who played a critical role in many of the company’s most famous breakthroughs, has left the company to form his own startup.
Silver is launching a new startup called Ineffable Intelligence, based in London, according to a person with direct knowledge of Silver’s plans. The company is actively recruiting AI researchers and is seeking venture capital funding, the person said.
Google DeepMind informed staff of Silver’s departure earlier this month, the person said. Silver had been on sabbatical in the months leading up to his departure and never formally returned to his DeepMind role.
A Google DeepMind spokesperson confirmed Silver’s departure in an emailed statement to Fortune. “Dave’s contributions have been invaluable and we’re grateful for the impact he’s had on our work at Google DeepMind,” the spokesperson said.
Silver could not immediately be reached for comment.
Ineffable Intelligence was formed in November 2025 and Silver was appointed a director of the company on January 16, according to documents filed with U.K. business registry Companies House.
In addition, Silver’s personal webpage now lists his contact as Ineffable Intelligence and provides an ineffable intelligence email address, although it continues to state that he “leads the reinforcement learning team” at Google DeepMind.
In addition to his work at Google DeepMind, Silver is a professor at University College London. He continues to maintain that affiliation.
A key figure behind many of DeepMind’s breakthroughs
Silver was one of DeepMind’s first employees when the company was established in 2010. He knew DeepMind cofounder Demis Hassabis from university. Silver played an instrumental role in many of the company’s early breakthroughs, including its landmark 2016 achievement with AlphaGo, demonstrating that an AI program could beat the world’s best human players at the ancient strategy game Go.
He also was a key member of the team that developed AlphaStar, an AI program that could beat the world’s best human players at the complex video game Starcraft 2, AlphaZero, which could play chess and shogi as well as Go at superhuman levels, and MuZero, which could master many different kinds of games better than people even though it started without any knowledge of the game, including not knowing the games’ rules.
More recently, he worked with the DeepMind team that created AlphaProof, an AI system that could successfully answer questions from the International Mathematics Olympiad. He is also one of the authors on the 2023 research paper that debuted the Google’s original Gemini family of AI models. Gemini has now Google’s leading commercial AI product and brand.
Looking for a path to AI ‘superintelligence’
Siliver has told friends he wants to get back to the “awe and wonder of solving the hardest problems in AI” and sees superintelligence—or AI that would be smarter than any human and potentially smarter than all of humanity—the biggest unsolved challenge in the field, according to the person familiar with his thinking.
Several other well-known AI researchers have also left established AI labs in recent years to found startups dedicated to pursuing superintelligence. Ilya Sutskever, the former chief scientist at OpenAI, founded a company called Safe Superintelligence (SSI) in 2024. That company has raised $3 billion in venture capital funding to date and is reportedly valued at as much as $30 billion. Some of Silver’s colleagues who worked on AlphaGo, AlphaZero, and MuZero have also recently left to found Reflection AI, an AI startup that also says it is pursuing superintelligence. Meanwhile, Meta last year reorganized its AI efforts around a new “Superintelligence Labs” that is headed by former Scale AI CEO and founder Alexandr Wang.
Going beyond language models
Silver is well-known for his work on reinforcement learning, a way of training AI models from experience rather than historical data. In reinforcement learning, a model takes an action, usually in a game or simulator, and then receives feedback on whether those actions are productive in helping it achieve a goal. Through trial and error over the course of many actions, the AI learns the best ways to accomplish the goal.
The researcher was often considered one of reinforcement learning’s most dogmatic proponents, arguing it was the only way to create artificial intelligence that could one day surpass human knowledge.
On a Google DeepMind-produced podcast that was released in April, he said that large language models (LLMs), the type of AI responsible for most of the recent excitement about AI, were powerful, but they were also constrained by human knowledge. “We want to go beyond what humans know and to do that we’re going to need a different type of method and that type of method will require our AIs to actually figure things out for themselves and to discover new things that humans don’t know,” he said. He has called for a new “era of experience” in AI that will be based around reinforcement learning.
Currently, LLMs have a “pretraining” development phase that uses what is called unsupervised learning. They ingest vast amounts of text and learn to predict which words are statistically most likely to follow which other words in a given context. They then have a “post-training” development phase that does use some reinforcement learning, often with human evaluators looking at the model’s outputs and giving the AI feedback, sometimes just in the form of a thumbs up or thumbs down. Through this feedback, the model’s tendency to produce helpful outputs is boosted.
But this kind of training is ultimately dependent on what humans know—both because it depends on what humans have learned and written down in the past in the pre-training phase and because the way LLM post-training does reinforcement learning is ultimately based on human preferences. In some cases, though, human intuition can be wrong or short-sighted.
For instance, famously, in move 37 of the second game of AlphaGo’s 2016 match against Go world champion Lee Sedol, AlphaGo made a move that was so unconventional that all the human experts commenting on the game were sure it was a mistake. But it wound up later proving to be a key to AlphaGo winning that match. Similarly, human chess players have often described the way AlphaZero plays chess as “alien”—and yet its counterintuitive moves often prove to be brilliant.
If human evaluators were passing judgments on such moves though in the kind of reinforcement learning process used in LLM post-training, they might give such moves a “thumbs down” because they look to human experts like mistakes. This is why reinforcement learning purists such as Silver say that to get to superintelligence, AI will not just have to get beyond human knowledge, it will need to discard it and learn to achieve goals from scratch, working from first principles.
Silver has said Ineffable Intelligence will aim to build “an endlessly learning superintelligence that self-discovers the foundations of all knowledge,” the person familiar with his thinking said.
