Richard Sutton, 68, Launches Oak Lab to Build Real‑Time Learning Trillion‑Parameter Agents

Reinforcement‑learning pioneer Richard Sutton, together with his former student Kuram Javed, has founded Oak Lab (Oak) to develop autonomous agents that can learn continuously from real‑world interaction, directly challenging the prevailing focus on ever‑larger static large‑language‑models (LLMs).

The AI industry today is obsessed with scaling model parameters and text‑generation capabilities, treating LLMs as the sole path to artificial general intelligence (AGI). Sutton argues that this approach suffers from two fundamental bottlenecks: models stop learning once training ends, and they lack a trial‑and‑error feedback loop that mirrors biological learning.

Sutton’s credentials underscore his authority: a lifelong leader in reinforcement learning, co‑author of the seminal textbook Reinforcement Learning: An Introduction , Turing‑award recipient in 2024, and mentor to figures such as AlphaGo creator David Silver. After leaving Keen Technologies—a DeepMind spin‑off backed by investors including Shopify CEO Toby Lutek—he announced Oak Lab, emphasizing a shift away from text‑centric pretraining toward interactive, self‑evolving intelligence.

In his 2019 essay “The Bitter Lesson,” often cited by LLM practitioners, Sutton warned that merely increasing compute and data cannot replace systems that learn autonomously from the environment. He contends that LLMs merely ingest massive human text, reproducing existing knowledge without genuine experience‑driven learning.

The core of Oak’s approach is the Options‑and‑Knowledge (OaK) architecture, derived from the University of Alberta’s long‑term “Alberta Plan.” It rests on three principles that deliberately separate it from LLMs:

General‑purpose without preset knowledge: agents start with no domain‑specific rules or textual priors.

Experience as the sole source of knowledge: all cognition arises from interaction and trial‑error, not from human‑annotated data.

Reward‑driven continual evolution: long‑term cumulative reward guides the autonomous generation of new behaviors.

OaK consists of two fundamental modules:

1. Options (Temporal Action Policies)

Rather than single actions, Options represent extended behavior strategies with explicit start and termination conditions, mirroring long‑term biological decision making.

2. Knowledge (World Model)

After each interaction, the agent abstracts environmental regularities into an internal world model, enabling prediction of long‑term outcomes for future actions.

The learning loop proceeds as follows: perceive environment → extract features → abstract high‑level cognition → generate complex long‑term behavior → execute and receive feedback → update world model and create new Options. This loop is theoretically limited only by compute, not by static text corpora.

Sutton acknowledges two major technical hurdles that currently prevent large‑scale OaK deployment: catastrophic forgetting (new learning overwriting old knowledge) and plasticity decay (loss of learning ability after prolonged training). Although no mature OaK system exists yet, he remains confident that the route will eventually succeed.

Parallel to Sutton’s effort, his former student David Silver founded Ineffable Intelligence in 2025, securing $1.1 billion in funding and pursuing a similar RL‑centric AI strategy, further illustrating a coordinated “reinforcement‑learning rebellion” against the dominant LLM paradigm.

The article also highlights ongoing debates: while RL excels in games with clear reward signals, open‑world environments present ambiguous feedback, making pure RL deployment challenging. Moreover, many recent breakthroughs in reasoning and mathematics arise from hybrid LLM + RL systems, suggesting the two tracks are not strictly mutually exclusive.

In conclusion, Sutton’s venture raises a fundamental question for the field: are we building sophisticated text‑copying machines, or are we engineering agents that can truly explore, learn, and adapt like living organisms? The future may split into two divergent tracks—static generative models versus interactive, self‑evolving reinforcement‑learning agents—leaving the ultimate answer to time.