David Silver, a prominent Google DeepMind researcher who played a key role in many of the company’s most famous breakthroughs, has left the company to found his own startup.
Silver plans to launch a new startup called Ineffable Intelligence, which will be based in London, according to a person with direct knowledge of Silver’s plans. The company is actively hiring AI researchers and is also seeking funding from venture capital.
Google DeepMind informed staff of Silver’s departure earlier this month, the person said. Mr. Silver took several months off before retiring and did not officially return to his role at DeepMind.
A Google DeepMind spokesperson confirmed Silver’s departure in an emailed statement. luck. “Dave’s contributions have been invaluable and we appreciate the impact he has had on our work at Google DeepMind,” the spokesperson said.
Mr. Silver could not be reached for comment.
Ineffable Intelligence was formed in November 2025 and Mr Silver was appointed as a director of the company on January 16, according to documents filed with Companies House, the UK company registrar.
Additionally, Silver’s personal webpage now lists his contact information as Ineffable Intelligence and provides an email address for Ineffable Intelligence, but continues to state that he “leads the reinforcement learning team” at Google DeepMind.
In addition to his work at Google DeepMind, Mr. Silver is also a professor at University College London. He continues to maintain that relationship.
Important people who supported DeepMind’s many breakthroughs
Mr. Silver was one of DeepMind’s first employees when it was founded in 2010. He has known DeepMind co-founder Demis Hassabis since college. Silver played a key role in many of the company’s early breakthroughs, including a 2016 breakthrough at AlphaGo that demonstrated that an AI program could beat the world’s best human players at the ancient strategy game Go.
He was also a key member of the team that developed AlphaStar, an AI program that can beat the world’s best human players at complex video games. StarCraft II; AlphaZero can play not only Go but also chess and shogi at a superhuman level. And MuZero was able to master many types of games better than most people, even though he started with no knowledge of the game, including not knowing the rules of the game.
Most recently, we worked with the DeepMind team that developed AlphaProof, an AI system that successfully answered International Mathematics Olympiad questions. He is also one of the authors of the 2023 research paper that premiered Google’s original Gemini family of AI models. Gemini currently leads Google’s commercial AI products and brands.
Exploring the path to AI “superintelligence”
Mr. Silver told a friend that he wanted to return to “the awe and wonder of solving the most difficult problems in AI,” and sees superintelligence, or AI smarter than any human and potentially smarter than all of humanity, as the field’s biggest unsolved challenge, according to people familiar with his thinking.
Other prominent AI researchers have also left established AI research labs in recent years to found start-up companies dedicated to the pursuit of superintelligence. Ilya Satskeva, former chief scientist at OpenAI, founded a company called Safe Superintelligence (SSI) in 2024. The company has raised $3 billion in venture capital so far and is reportedly valued at $30 billion. Several of Silver’s colleagues who worked on AlphaGo, AlphaZero, and MuZero also recently left to found Reflection AI. The company is a startup pursuing superintelligence. Meanwhile, Meta last year reorganized its AI efforts around new Superintelligence Labs, headed by former Scale AI CEO and founder Alexandr Wang.
Beyond language models
Silver is best known for his work on reinforcement learning, a method of training AI models based on experience rather than historical data. In reinforcement learning, a model performs an action, usually in a game or simulator, and receives feedback about whether the action helps achieve a goal. Through trial and error over the course of many actions, the AI learns the best way to accomplish a goal.
The researcher is often considered one of the most opinionated proponents of reinforcement learning, arguing that it is the only way to one day create artificial intelligence that surpasses human knowledge.
In a Google DeepMind-produced podcast released in April, he also said that large-scale language models (LLMs), the type of AI responsible for much of the recent excitement about AI, are powerful but limited by human knowledge. “We want to go beyond what humans know, and that requires a different kind of method. That kind of method requires AI to actually figure things out for itself and discover new things that humans don’t know,” he said. He called for a new “era of experience” for AI based on reinforcement learning.
Currently, the LLM has a “pre-training” development phase that uses so-called unsupervised learning. They ingest vast amounts of text and learn how to predict which words are statistically most likely to follow other words in a given context. Then there’s a “post-training” development phase that uses reinforcement learning, where human evaluators often look at the model’s output and give feedback to the AI, sometimes in the form of a simple thumbs up or thumbs down. Through this feedback, the propensity of the model to produce useful output is strengthened.
However, this type of training ultimately relies on what humans know. This is because it depends on what humans have learned and written down during the pre-training phase, and because the LLM post-training reinforcement learning method is ultimately based on human preferences. But in some cases, human intuition can be wrong or short-sighted.
For example, in a famous story, on move 37 of the second game of AlphaGo’s 2016 Go World Champion Lee Sedol, AlphaGo made a move so unconventional that every human expert commenting on the match was convinced it was a mistake. However, it later turned out to be the key for AlphaGo to win that match. Similarly, human chess players often describe AlphaZero’s way of playing chess as “alien,” yet its counterintuitive moves often prove to be masterful.
However, if a human rater is passing judgment on such a move, such as in the reinforcement learning process used in LLM post-training, the human expert may consider it to be a mistake and give such a move a lower rating. This is why reinforcement learning purists like Silver argue that for AI to reach superintelligence, it not only needs to surpass human knowledge, it also needs to abandon human knowledge and learn to achieve goals from scratch based on first principles.
Mr. Silver said Ineffable Intelligence aims to build “an infinitely learning superintelligence that discovers for itself the basis of all knowledge,” a person familiar with his thinking said.
