Aug 18, 2026, 8:00 AM ETFireside Chat, Interview

Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again

Rich Sutton asserts that "continual learning" is the default state of intelligence, arguing that the field is "weird" for isolating it as a special phase rather than treating all learning as continuous.
Sutton posits that human-centric "bitter lesson" logic (relying on computation over human priors) is often misinterpreted, with current LLM trends representing a reliance on finite human knowledge (the internet) that will hit a ceiling.
He defines "The Bitter Lesson" concisely: do not be distracted by human knowledge; focus on learning methods that scale with computation (search and learning) rather than scaling with data or human input.
Sutton contends that current Large Language Models (LLMs) are "inconsistent" with true intelligence because they freeze their weights after training, lacking the ability to update their internal world models through ongoing experience.

The "Big World Hypothesis" states that the world is infinitely complex and contains infinitely many things to learn, making any finite synthetic dataset inherently insufficient for general intelligence.
Sutton rejects the current industry push for synthetic data generation, arguing it remains bottlenecked by human expertise (experts must define what data is "good") rather than removing the human in the loop.
He argues that simulating the physical world (e.g., for drones or robots) is microscopic compared to reality, where factors like friction, wear, and unmodeled physical interactions cannot be perfectly replicated by programs.
Sutton emphasizes that true intelligence requires agents to learn from their own experience in the real world to correct their own models, a capability absent in systems trained solely on curated simulations.

Oak Lab is founded on a 12-step research agenda, with "Continual Deep Learning" identified as the most critical step to unlock the ability to update world models continuously.
Sutton identifies "catastrophic forgetting" as the primary technical barrier to continual learning, noting that naive weight updates based on single examples destroy prior knowledge.
The proposed algorithmic solution involves:
- Step-size optimization: Assigning individual learning rates to every weight, where most weights have tiny step sizes to preserve prior knowledge.
- Meta-learning: The system must learn how to learn, updating its own learning parameters alongside its knowledge base.
- Generate and Test: Injecting new units with random weights into the network (Continual Backprop) to test new features without relying solely on slow gradient descent.
Sutton predicts that with the right algorithms, a trillion-parameter model could operate on roughly 20 watts of energy within five to ten years, a feat currently impossible due to energy inefficiency and the lack of new paradigms.

Sutton argues that current LLM-based assistants fail to be true "experiential learners" because they cannot retain personalization or updates without retraining on massive batches of data from all users, rather than learning specifically for the individual.
He distinguishes between "prior knowledge" (culture/schooling) and "learning" (acquiring new knowledge), asserting that while humans utilize both, current AI focuses too heavily on the former (training data) and neglects the latter (runtime learning).
The interview highlights that "paradigm shifts" (e.g., Einstein discovering new physics) require agents to self-discover new abstractions and planning models, a capability currently missing in systems that rely on pre-defined human abstractions.
Sutton notes that major labs are "stuck in a local minimum," unwilling to pursue new algorithms where performance may initially drop before achieving a superior, energy-efficient paradigm.

Oak Lab is hiring a small, highly aligned team of researchers who have long thought about these problems, avoiding rapid scaling to ensure everyone shares the same philosophical vision.
The company's long-term goal is a "single design" of mind that can form abstractions across a spectrum of knowledge, from low-level sensory-motor tasks (walking, grasping) to high-level planning (space flight, math).
Sutton envisions a future where multiple distinct instances of this mind exist, each learning unique experiences from their specific environments, rather than one monolithic system attempting to learn everything.
While acknowledging LLMs as a "scientific breakthrough" in language use, Sutton warns they may become "irrelevant" if the field does not transition to systems capable of genuine, continuous, and self-consistent learning.
The ultimate objective is to build agents that are "self-maintaining," capable of keeping their internal knowledge coherent and correct through continuous interaction, avoiding the drift into incoherence common in static models.