Richard Sutton launches Oak Lab to transform continuous learning in AI

Richard Sutton, a Turing Award-winning luminary in artificial intelligence, has unveiled a new research venture, Oak Lab, in Toronto, Canada. Co-founded with his former student Khurram Javed, the lab aims to develop AI systems capable of learning continuously from their environment, marking a significant departure from current AI models reliant on static datasets.

A Bold Bet on Reinforcement Learning

Sutton’s initiative is rooted in his belief that reinforcement learning—where AI systems adapt based on feedback from their actions—will drive the next wave of innovation in artificial intelligence. He has been critical of the inefficiencies in deep learning methods and large language models (LLMs), which dominate the field today. Sutton contends that these systems, while powerful, are limited by their dependence on pre-existing data and lack the ability to evolve dynamically.

Oak Lab’s research will focus on creating algorithms that learn in real time without storing or replaying data, a move aimed at reducing computational demands while enhancing adaptability. The lab also plans to develop internal ‘world’ models that enable AI agents to independently handle variation, evaluation, and selection.

Ambitious Goals for Energy-Efficient AI

The lab’s long-term vision includes building an AI agent, named Oak, with a trillion parameters that learns and plans in real time using just 20 watts of energy—comparable to the power consumption of a standard light bulb. This agent will continuously evolve through experience, embodying Sutton’s vision of AI systems that are both efficient and capable of genuine discovery.

Departure from Keen Technologies

Sutton’s move to establish Oak Lab follows his recent departure from Keen Technologies, a Dallas-based startup focused on artificial general intelligence (AGI). Founded by gaming legend John Carmack, Keen Technologies counts Shopify CEO Tobi Lütke among its high-profile backers. Sutton and Javed left the company in pursuit of what Sutton describes as “a slightly different path,” one that challenges the foundational assumptions of deep learning.

Growing Criticism of Current AI Models

Sutton’s critique aligns with a broader wave of skepticism among AI researchers, including Yann LeCun, who argue that scaling laws—the principle that larger models yield better results—are no longer effective. Sutton has previously stated that generative AI models, such as ChatGPT, merely imitate their training data and lack the capacity for real discovery or self-evaluation.

Why It Matters

Sutton’s work has already shaped the AI landscape, with reinforcement learning playing a pivotal role in training LLMs like ChatGPT and Claude. His new venture signals a potential paradigm shift, challenging the dominance of static, data-driven models and advocating for systems that learn dynamically and efficiently. If successful, Oak Lab’s research could redefine the future of AI, making it more adaptable, energy-efficient, and capable of genuine innovation.

Source: Indian Express