David Silver, the scientist from DeepMind behind AlphaGo’s landmark victory over world Go champion Lee Sedol in 2016, has successfully raised $1.1 billion to start a venture with the belief that future AI will not rely on current dominant technology.
Ineffable Intelligence, his startup launched in January valued at $5.1 billion, is focusing on reinforcement learning, where AI systems enhance themselves through trial and error. Silver posits this method as more viable than large language models for achieving superintelligence.
“Our mission is akin to first contact with superintelligence,” Silver told Wired. “I envision something truly remarkable that could independently uncover new scientific, technological, governmental, or economic paradigms.”
The concept of superintelligence was popularized by philosopher Nick Bostrom in his 2014 book ‘Superintelligence’ and denotes AI surpassing human intelligence across almost all areas, whereas artificial general intelligence (AGI) pertains to systems that can match human reasoning over a broad spectrum.
Silver contends that large language models are inherently constrained as they rely on data generated by humans rather than acquiring understanding through direct experience.
“Human data is similar to fossil fuel, offering an incredible shortcut,” he explained. “Self-learning systems can be likened to renewable energy—capable of continuous learning without boundaries.”
Silver’s career has been dedicated to this philosophy. AlphaGo, which merged human training data with reinforcement learning and self-play, devised strategies that even top players found surprising, demonstrating AI’s potential to surpass humans in specific fields.
“It’s crucial for there to be a premier AI lab wholly devoted to this methodology,” he said to Wired. “This shouldn’t just be a niche within another organization focused on LLMs.”
Ineffable Intelligence aims to create ‘superlearners’—AI agents situated in simulations where they can pursue objectives, encounter failure, adapt, and enhance without the confines of static human datasets. While Silver didn’t detail these simulations, he noted that this method would allow for autonomous collaboration and capability development.
Silver argued that large language models are hindered by their training data, suggesting that a model educated in an era where people believed the Earth was flat might retain such misconceptions unless it could verify reality independently. He believes a self-learning system could uncover the truth otherwise.
Ineffable Intelligence did not provide immediate comments to Decrypt.