Ineffable Intelligence
Research Areas
Three interconnected questions about how a system can acquire intelligence without the crutch of human-labelled training data.
Self-Supervised Representation Learning
How can a system learn to represent the world — its structure, its regularities — from raw signal alone? We study objectives, architectures, and training dynamics that lead to representations useful for downstream tasks without ever seeing a labelled example.
Explore this areaReward-Free Exploration
Curiosity, surprise, and competence-seeking are intrinsic to intelligent behaviour. We study agents that learn rich world-models by exploring their environments without external reward, then apply that understanding to new tasks with minimal guidance.
Explore this areaContinual Learning Without Forgetting
Intelligence accumulates. A system that discards prior knowledge each time it learns something new cannot be intelligent in any meaningful sense. We investigate architectures and training regimes that allow models to grow their understanding over time.
Explore this areaWritten output
Publications
Technical reports, preprints, and research memos from Ineffable Intelligence, documenting methods, partial findings, failure modes, and open questions. We publish negative results.
View Publications