Continual learning under non-stationarity
Distribution shift and changing dynamics are treated as the default setting, not an edge case.
True Intrinsics
An independent AI research lab.
Overview
Learning beyond a fixed model
True Intrinsics studies small, growable agents that acquire and verify knowledge through interaction. We treat the Big World Hypothesis—the premise that an agent's environment is too large and non-stationary to be captured by any fixed model—as a systems constraint. Our research investigates continual, reward-driven adaptation, explicit memory module design, and conditional computation, with the goal of improving capability through accumulated experience rather than relying solely on parameter count, training-data volume, and static pretraining.
Research
What we work on
Long-horizon credit assignment
Learning useful behavior when rewards are delayed, sparse, or only available through verification.
Memory module design
Selective storage, retrieval, consolidation, and controlled reuse of experience under distribution shift.
Adaptive computation
Allocating search, reasoning, and verification effort according to task difficulty.
Foundation models serve as structured interfaces for interaction; the agent's core learning is driven by reward and verification-derived signals.
Blog
Recent notes
The Big World as a Design Constraint
Why non-stationarity should be treated as the default setting for adaptive agents.
Memory Modules Under Distribution Shift
Selective retrieval, consolidation, and controlled experience reuse as design problems.
Computation Should Follow Difficulty
Why an agent should spend more inference effort only when a task demands it.