The problem is the absence of labels.

The annotation bottleneck

The dominant paradigm in machine learning asks a human to look at a datum — an image, a piece of text, a sensor reading — and to assign it a label. That label becomes the target for a learning system. The implicit assumption is that intelligence can be transferred from humans to machines by making human judgements legible enough for gradient descent to act on.

This assumption has been extraordinarily productive. Supervised learning has produced systems that, in specific narrow domains with ample labelled data, match or exceed human performance. But it has a structural cost: the scope of what AI can learn is bounded, in practice, by the set of problems where human annotators can be deployed at scale and where their judgements are consistent enough to form a reliable training signal. Where annotations are sparse, unreliable, or impossible to obtain, supervised methods fail by design — not by accident.

The hardest problems in intelligence are the ones where no one has written down the answer.

What self-play taught us

Some of the most striking demonstrations of AI learning without human guidance have come from game-playing systems trained through self-play. By competing against itself — generating an endless curriculum of increasingly difficult positions — such a system can improve substantially without any human ever labelling a move as good or bad.

We think this result is significant not because of the game, but because of the principle it demonstrates: a system can construct its own learning signal. The environment, and the agent's history within it, contain enough structure to drive genuine improvement. The game's rules happen to provide a natural reward; this is what makes game-playing a special case. Our question is what plays that role when the rules are absent — when no score is kept and no win condition is defined.

We are not a games AI group. We use the self-play insight as a starting point for thinking about learning in open-ended environments where no oracle exists to record correct answers.

Intrinsic motivation as a curriculum

One promising direction is intrinsic motivation — learning signals derived from the agent's own state and predictions rather than from any external evaluator. Curiosity, surprise, and competence-seeking are all candidates. Each can be formalised as an objective that an agent pursues in the absence of task instruction.

The appeal of intrinsic motivation is that it does not require a human to specify what the agent should be curious about. The agent is drawn toward novelty and toward the edges of its current understanding. In sufficiently rich environments, this can produce genuinely exploratory behaviour and, over time, a broad and useful world-model.

We are investigating what makes intrinsic motivation signals robust — which formalisations hold up across environments with different degrees of stochasticity, and which degenerate into reward hacking or trivial solutions.

Representation as the foundation

Underlying all of these approaches is a question about representation. An agent that encounters the world as raw high-dimensional input — pixels, sensor readings, token sequences — cannot efficiently reason about it without forming some internal structure: a compressed, organised representation of what it has observed.

The quality of this representation determines almost everything else. A representation that captures the genuine causal structure of an environment will support transfer to new tasks far better than one that encodes surface statistics. Our self-supervised representation learning work is investigating how to build the former without access to labels that would tell the system what structure matters.

Where this leads

We do not have a complete theory of intelligence without supervision. We have a set of connected questions, a set of methods worth taking seriously, and a conviction that the annotation bottleneck is the correct problem to be working on. The work is investigative rather than applied. We are not optimising for benchmark performance on labelled tasks. We are trying to understand what it would take to build AI that genuinely learns from experience in the absence of a teacher — AI whose competence is not a function of how many labelled examples it was given.

We publish our findings as technical reports and preprints, including work that does not reach positive conclusions. We believe that shared knowledge of what does not work is at least as valuable to the field as descriptions of what does. The accumulation of partial, honest, openly shared knowledge is how this kind of foundational research advances.