The OaK Architecture: A Vision of SuperIntelligence from Experience

Summary

OaK is Sutton's proposed model-based reinforcement-learning architecture for agents that learn continually from their own experience. Its components learn online, its parameter step sizes are meta-learned, and it creates state and time abstractions through the FC-STOMP progression: Feature Construction, SubTask, Option, Model, and Planning. This is an invited keynote and research program rather than a conventional peer-reviewed paper. Oak Lab's companion technical post demonstrates one piece of the motivation: adaptive credit assignment can isolate predictable signal in noisy online streams where the shown SGD baseline absorbs noise. The full OaK architecture remains a proposed system, not a validated end-to-end result.

Why George Should Read It

It offers a deeper model of agents than “LLM plus tools”: intelligence as a persistent process that discovers concepts, turns them into reusable skills, learns models of those skills, and plans with them. It also creates a technical reading path through continual learning, options, world models, meta-learning, and online credit assignment.

Best For

• Understanding experience-grounded alternatives to static pretrained models
• Building a research map for continual and model-based reinforcement learning
• Evaluating whether an agent is actually accumulating reusable skills

Notes

• Source type: RLC 2025 keynote / architecture proposal, not a peer-reviewed paper.
• Key mechanism: FC-STOMP builds a ladder from learned features to planning.
• Oak Lab's 20-watt, trillion-parameter agent is an explicit “holy grail,” not a current result.
• Technical companion: https://oaklab.ai/posts/learning-from-experience-instead-of-curated-datasets

Next Reading Action

Read Oak Lab's noisy-stream credit-assignment post, then watch the keynote and draw the FC-STOMP loop before judging the larger superintelligence claim.

Copyable Markdown

# The OaK Architecture: A Vision of SuperIntelligence from Experience

Year: 2025
Category: Continual Learning / Reinforcement Learning

Why it matters:
Sutton proposes an agent architecture that learns continually from experience,
constructs reusable temporal abstractions, and plans with learned option models.

Best for:
- continual-learning research maps
- model-based reinforcement learning
- thinking beyond static LLM agents

Next step:
Read Oak Lab's credit-assignment post, then map the five FC-STOMP stages while
watching the keynote.