rich sutton

Current LLMs stop learning after pre-training (weights frozen), creating a fundamental gap: they cannot adapt to new experiences or personalize to individual users. Sutton and Javed argue the solution is algorithmic — continual backprop with per-weight step sizes and generate-and-test in feature space — enabling models to learn continuously without catastrophic forgetting.

Synthetic data requires human experts to design and validate, creating a fundamental bottleneck. True scaling requires agents that learn autonomously from their own experience in the infinitely complex real world, not from human-curated simulations.

Sutton and Javed argue current LLMs fundamentally stop learning after training (weights frozen), and propose continual backprop with per-weight step-size optimization and generate-and-test to enable models that learn continuously from single-stream experience without catastrophic forgetting.

Current LLMs freeze weights after training, preventing true continual learning. Oak Lab's continual backprop algorithm with per-weight step sizes and generate-and-test enables models that continuously update from experience without catastrophic forgetting, potentially obsoleting the massive pre-training + frozen inference paradigm.

Sutton acknowledges LLMs as a major scientific breakthrough in neural language use, but argues they represent only a fraction of intelligence (sensory-motor, planning, abstraction). The field mistakenly treats frozen-weight language models as complete AI, ignoring the need for continual learning, model-based planning, and self-discovered abstractions.

Current foundation model labs are locked into a local optimum where models stop learning after deployment; Sutton argues this paradigm cannot achieve general intelligence and will be superseded by continually learning systems, implying today's massive capex on static models may face obsolescence.