Turing Award winner Richard Sutton has founded a new AI lab called Oak AI to develop agents that learn independently, marking a decisive shift away from scaling large language models.
Sutton, who received the 2024 Turing Award alongside Andrew Barto for foundational work in reinforcement learning, announced the lab along with three founding papers. The company aims to build AI systems that learn through trial, error, and feedback, rather than relying on massive datasets.
The new lab is already staffed with 10 researchers at a dedicated facility in Edmonton, Canada. Oak AI is also accepting applications for remote workers, internships, and PhD students.
A Direct Rejection of Scaling Laws
Sutton has long argued against the prevailing industry focus on scaling up transformer models. In a famous 2019 blog post, “The Bitter Lesson,” he wrote that “the biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin.”
Oak AI’s approach directly applies this lesson. The company says large language models are not the path to general intelligence because they rely on static, pre-collected data.
The Core Problem with Current AI
The core issue is that modern LLMs learn from fixed datasets and cannot produce new behavior or continuously improve through interaction. Oak AI describes this as “learning from someone else’s experience, which is limited and non-interactive.”
Sutton’s alternative focuses on agents that learn in a continual, online fashion. These systems would generate their own data through action and observation, allowing them to adapt to new situations.
The company’s first stated goal is to “achieve continually optimal behavior from both a learning system that never stops learning and never stops improving its performance over its entire lifetime,” according to a founding paper.
Three Research Papers Set the Foundation
Oak AI published three papers outlining its scientific vision.
- The “Annotated” algorithm and continuing optimal selection reveals a new formal approach to decision-making under uncertainty, specifically tackling the exploration-exploitation dilemma.
- The “Lapsed” algorithm and provably efficient learning with partial models introduces a method for efficient learning even when the agent’s internal model of the world is incomplete.
- The “Ribbit” algorithm and continuing optimal learning in nonstationary environments shows how an agent can thrive in a world where the rules keep changing.
These papers provide theoretical backing for agents that continuously adapt without resetting their learning process.
A Different AI Economy
Sutton believes AI systems should be judged by the computational cost of their learning, not just raw performance. He argues that expensive training processes make models obsolete quickly, while a system that improves continuously can compound its returns over time.
“In contrast,” Sutton’s research states, “a conventional system’s performance is not improving over time and, therefore, cannot compound.” This introduces a concept known as “continuing optimality,” where a system always makes the best decision it can at a given moment.
What This Means for the AI Landscape
Oak AI directly challenges the transformer-based paradigm that dominates companies like OpenAI and Google. While other labs pump billions into larger models and more data, Sutton is betting on agents that learn cheaply and continuously.
If successful, this approach could create more robust, adaptable systems that don’t break when faced with novel situations. The lab’s bet is that true intelligence emerges from interaction, not from digesting static text.
The company is currently hiring and seeking venture capital to scale its operation. For Sutton, this represents a chance to build the kind of AI he has championed throughout his career — one that learns like a living thing, not a statistical model.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.