We spent longer than we should have picking the name. Not because we lacked options, but because most names available to an AI research lab in 2021 seemed to describe the answer before the question had been properly asked. Words like "deep", "clarity", "insight" — they all carry a premature confidence, a suggestion that the hard work of understanding has already been done. The word ineffable does the opposite.
Ineffable: too great or extreme to be expressed in words. The OED lists a secondary sense — not to be uttered. Originally theological, describing qualities of the divine that language cannot contain. It is an admission that something real exists and resists description at the same time.
We chose it because it names, quite precisely, the class of problem we care about most.
The problem with labels
When we train an AI system on labelled data, we are encoding a prior claim: that a human, at some prior moment, understood the situation well enough to annotate it. That annotation becomes the teacher. The system learns to reproduce what the annotator saw.
This is enormously useful for a large class of problems. But it means the system can only learn what was already understood — and expressed — by whoever did the labelling. The teacher's knowledge is the ceiling.
Genuinely hard problems in intelligence are not like this. They involve situations where no one has written down the right answer because no one knows it. They involve domains where understanding is still forming, where the structure of the problem is itself part of what needs to be discovered. These problems are, in the original sense, ineffable: the knowledge that would constitute good performance cannot be fully articulated in advance.
When we founded this laboratory in 2021, this was the distinction we kept returning to. Not supervised versus unsupervised as a technical taxonomy. Something more like: problems where the answer is known versus problems where it is not.
What self-play demonstrated
The clearest existence proof that AI can exceed human understanding in its own domain came from game-playing agents that learned through self-play. The key feature of self-play — and the reason it interests us deeply — is that it generates a curriculum of problems without any teacher who knows the answers. Two agents play against each other. Each victory by one agent becomes a training signal for the other. The signal does not come from a human who annotated correct moves. It comes from the structure of the game itself.
Over many iterations, the quality of play rises beyond anything a human annotator could have described at the outset. The knowledge the agents acquire is, in a real sense, ineffable — it outstrips what any person had articulated. The game had structure that the agents discovered, not structure that was handed to them.
This observation has been central to our thinking since we started. The self-play intuition is not a narrow technique for board games. It is a pointer toward a class of learning methods in which the signal for improvement comes from within the system's own experience, not from external annotation. Extending that intuition beyond the clean formal structure of games — into the messy, open-ended world — is one of the central problems we are investigating.
Why "ineffable" is not a mystical claim
We are not saying, by using this word, that the knowledge AI systems might acquire is permanently beyond human understanding. We are not making a mystical or anti-scientific argument. What we are saying is narrower and more specific: at the moment a training signal is required, that knowledge may not yet exist in articulable form. The gap is temporal and epistemic, not fundamental.
A system that can learn without the prior existence of labelled data can, in principle, discover understanding that is later articulated, understood, and used by humans. The ineffability is the condition at the start, not the end.
This is important because it separates our claim from the kind of "black box" mysterianism that sometimes surrounds AI. We are interested in systems whose learned representations can eventually be interpreted, studied, and built upon. But we do not want the requirement of prior interpretability to be a constraint on what can be learned in the first place.
Where this leaves the research agenda
The name commits us to a specific type of problem. Not: how do we improve performance on benchmarks where good answers are known? But: how do we build systems that can form understanding in domains where no one yet knows what good performance looks like?
This requires different machinery. Representations that are not trained against a labelled target, but against the structure of the data itself — its statistical regularities, its symmetries, its predictable and unpredictable parts. Exploration strategies that generate informative experience without an external signal saying which experiences are valuable. Learning dynamics that can accumulate knowledge over time, across shifting distributions, without collapsing prior understanding.
We are at an early stage on all of these problems. The word "ineffable" in our name is not a solved claim. It is a statement of what we are chasing: the capacity to learn in the space that annotation cannot reach.
Most of the interesting things humans know were, at some point, ineffable — not yet writable, not yet teachable in the standard sense. Mathematics before formal proof systems. Physical intuition before equations. The feel of a language before grammar. Labelled data is a wonderful tool for the parts of knowledge that have already crossed into articulable form. We are interested in what happens on the other side of that boundary.
The laboratory as a bet
Running an independent research group in London in 2021, focused on foundational questions about learning without labels, was a choice that required justifying. The easy path, for anyone working in AI at that time, was to direct effort toward systems that already worked impressively and make them work more impressively still. Scale the architecture, clean the dataset, push the benchmark.
We think that path is well-resourced and does not urgently need another group working on it. The path toward systems that can genuinely acquire knowledge beyond the reach of human annotation is less explored, harder to measure, and — we believe — more likely to produce the kind of understanding that earns the word "intelligence" in its strongest sense.
The name is a reminder of that bet. Every time we encounter a problem where the difficulty is precisely that no labelled data exists, no benchmark captures the right answer, no human expert can write down what correct performance looks like — we are in the territory the name points toward. That is the territory we wanted to be in when we started, and it is where we remain.
