Independent AI Research · London · Est. 2021
Learning without a teacher.
Ineffable Intelligence is a London research laboratory studying how AI systems can acquire genuine understanding from raw, unlabelled experience — drawing on self-supervised representation learning, intrinsic motivation, and reward-free exploration to investigate what comes after supervised training.
Research Programme
Three interconnected questions
Self-Supervised Representation
How does a system learn to represent the world's regularities from raw signal alone — without a single labelled example to guide it?
Reward-Free Exploration
Agents driven by intrinsic motivation — curiosity, prediction error, competence-seeking — that build world-models without any externally specified reward.
Continual Learning
Overcoming catastrophic forgetting — building systems that accumulate knowledge across a sequence of experiences rather than resetting to the last task.
“
The hardest problems in intelligence are the ones where no one has written down the answer. That is where we work.
Ineffable Intelligence research philosophy
Our Approach
The self-play intuition, generalised
In closed game environments, self-play generated an endless curriculum of difficult problems without a human ever labelling a single position as good or bad. The agent was its own teacher. We investigate whether this principle — an agent constructing its own learning signal from structure already present in its experience — can extend to open-ended domains where no rulebook exists and no oracle records correct answers.
Our Approach
Writings
Recent Thinking
Beyond the Label
Five years into building AI that learns without human-labelled data, here is where we think the field stands — and where the real difficulties lie.
Scaling Without Supervision
Scale has powered much of recent AI progress, but almost all of it has been scale in supervised settings. What happens when you scale the other kind?
Continual Learning in Practice
Catastrophic forgetting is well-understood theoretically. In practice, making a model accumulate knowledge over time without regression is still an open engineering challenge.
Work with us or join us
We collaborate with university research groups and organisations with genuine interest in unsupervised and self-supervised AI. We also hire researchers and engineers who want to do foundational work without a product deadline.