Turing Winner Richard Sutton Launches Oak Lab

One of the most respected minds in AI just made a pointed statement about where the field is headed, by walking away from the company he was helping build AGI at. Richard Sutton, the University of Alberta professor whose work laid the mathematical foundations for reinforcement learning, has founded Oak Lab, a Canadian-incorporated research company built with former student Khurram Javed, according to The Logic's reporting on the launch.

The move is notable specifically because of where Sutton is coming from. He joined John Carmack's Keen Technologies in September 2023, after Google's DeepMind shuttered the Edmonton lab Sutton had helped lead. Keen, backed by Shopify CEO Tobi Lütke, was itself pursuing artificial general intelligence. Sutton's departure to start something new signals a genuine disagreement about the right technical path forward, not just a change of scenery.

Sutton's core argument, laid out in The Next Web's coverage of Oak Lab's launch, is that intelligence comes from run-time experience, not from being distilled out of a clean, human-curated dataset. Today's large language models learn from data that people have already collected, cleaned, and filtered. Real-world experience is messier, and Sutton's research argues that standard optimization techniques like SGD and Adam can't distinguish genuine signal from noise, so they end up quietly absorbing errors rather than learning from them.

This isn't a fringe position from an outsider. Sutton co-wrote "Reinforcement Learning: An Introduction," the standard textbook that has trained a generation of AI researchers and been cited more than 175,000 times, and won the 2024 ACM Turing Award, often called the Nobel Prize of computing, alongside longtime collaborator Andrew Barto. His 2019 essay "The Bitter Lesson," arguing that general methods leveraging computation ultimately outperform hand-crafted domain knowledge, is quoted constantly across the AI research community, a foundational idea worth understanding alongside our broader explainer on what a large language model actually is and how it learns.

Sutton's skepticism about the LLM scaling race puts him in genuinely notable company. Meta's former chief AI scientist Yann LeCun made a similar case before leaving Meta's orbit to bet $1 billion on world models rather than bigger chatbots. AlphaGo creator David Silver has placed his own related bet on a different technical route entirely. All three researchers share a core belief that machines should learn the way a child does, through direct experience, rather than from a frozen snapshot of internet text.

The timing fits a broader shift happening across the AI industry right now. The race has quietly stopped being purely about who can build the biggest model. Cost and efficiency now matter as much as raw scale, and researchers across major labs are digging into how models actually reason rather than simply making them larger, a dynamic we've tracked in our coverage of AI inference and the economics behind running models efficiently at scale.

I've advised companies on AI adoption for four years, and Sutton's move is worth watching closely, not because Oak Lab will necessarily produce a breakthrough tomorrow, but because it signals genuine technical uncertainty at the highest levels of AI research about whether the current scaling approach is actually the right long-term bet. Businesses making multi-year AI infrastructure and vendor decisions based on the assumption that today's LLM architecture will keep improving in a straight line should understand that some of the field's most credentialed researchers disagree.

For companies building AI strategy, this is a useful reminder to stay technology-agnostic where possible, rather than locking into infrastructure or partnerships that assume one specific research direction is guaranteed to win.

Watch what Oak Lab actually publishes and ships. Sutton has previously estimated a one-in-four chance of achieving human-like intelligence by 2030, a genuinely bold near-term forecast for a field where many researchers hedge much further out. If Oak Lab's always-learning agent approach produces even modest, verifiable results within the next year, expect renewed debate across the industry about whether the scaling-first approach dominating current AI investment is actually the fastest path forward.