About Us
An independent laboratory for foundational AI.
Founded in London in 2021, Ineffable Intelligence is a small, independently funded research group pursuing foundational questions in unsupervised AI — questions that require a long time horizon and no pressure to demonstrate results on labelled benchmarks.
Who we are
Ineffable Intelligence was founded in London in 2021 around a single observation: that the self-supervised and self-play methods which had produced the most striking AI results to date — systems that learned to master complex tasks without human-labelled training data — had not yet been taken seriously as a general research programme. Most of the AI community was still optimising supervised learning at scale. We thought the more important question was what learning looks like when the supervision is removed entirely.
We are a small, independently funded research group. We work on self-supervised representation learning, reward-free exploration, and continual learning. We publish what we find, including the parts that do not work.
The name reflects the nature of the problems we care about. Ineffable: too great or extreme to be expressed in words. In AI, this describes the kind of knowledge that cannot be conveyed by example — the understanding of a domain so rich and open-ended that no finite set of annotations can capture what matters. This is where we work.
Why independent
We are independently funded by design, not by default. Foundational research takes years. The questions we are pursuing do not resolve on product timelines. We are not measuring progress in benchmark points on labelled evaluation sets — the evaluation itself is part of what we are trying to understand. Organisations under commercial pressure cannot sustain this kind of work; we chose independence specifically to protect the time the problems require.
Independence also means we select questions by their importance, not by the availability of labelled datasets, the interests of commercial sponsors, or the requirements of a product roadmap. We are not building toward a deployment. We are building understanding — the kind that does not yet exist, about learning that does not yet work reliably.
How we work
Ineffable Intelligence is a small group and we intend to remain one. Each researcher works on a defined strand of our programme while staying in close contact with colleagues on adjacent problems. We write down what we try and what fails. We publish findings as they become useful rather than holding them for competitive advantage — because there is no competitive advantage in hoarding negative results.
We collaborate selectively with university research groups and with organisations that have genuine interest in the underlying science. We have no commercial product and no self-serve offering. Partnerships are built around shared research questions, not around licensing or API access.
What we stand for
Principles that guide our work
Long-horizon thinking
Foundational research resolves on a timescale of years, not quarters. We structure our work — the questions we accept, the collaborations we take on, the milestones we set — to protect that timeline against the pull of short-term results.
Transparency in method
We document what we tried, what failed, and why. Negative results are results. The field advances faster when researchers publish what did not work than when they publish only what did.
Independence of judgement
We choose questions by their importance, not by the availability of a benchmark or the convenience of an existing dataset. This sometimes means working where progress is slower and harder to demonstrate.