Ineffable Intelligence

Writings

Essays, research notes, and perspectives from our team — on the ideas behind our work, the problems that motivate it, and the questions we have not yet answered.

Abstract representation of labelled versus unlabelled data — scattered unmarked points gradually revealing structure
ResearchRetrospective

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.

David Silver 13 min
Exponential growth curve without labelled axis points — suggesting scale in an unsupervised domain
ScalingInfrastructure

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?

Priya Narayanan 10 min
Layered rings representing accumulated knowledge over time without forgetting
Continual LearningEngineering

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.

Tobias Wenzel 9 min
An expanding spiral of forms representing open-ended growth and emergent complexity
Open-EndednessPhilosophy

On Open-Endedness in AI Systems

Open-ended evolution has generated biological complexity without an objective function. AI systems, trained against fixed objectives, produce convergence rather than open-endedness. We examine what the difference tells us.

David Silver 14 min
An abstract agent exploring a space, moving toward novel regions with no external reward marker
CuriosityExploration

Intrinsic Motivation, Revisited

Count-based novelty, prediction error, information gain — curiosity has been formalised in many ways, with different failure modes. We compare them and argue for a notion of intrinsic motivation grounded in prediction complexity rather than state visitation.

Priya Narayanan 12 min
Abstract representation of a reward signal — a sparse jagged waveform against a quiet background
RewardResearch

Designing Reward Signals from First Principles

Before you can have reward-free learning, you have to understand what makes a reward signal work — and why most of them are poorer guides than we assume.

Tobias Wenzel 8 min
Abstract neural representation — a high-dimensional cloud of points condensing into a structured form
RepresentationTheory

Representation Without Supervision

What does it mean for a model to understand something, as opposed to pattern-matching against a labelled training distribution? We examine what unsupervised representations capture that supervised ones do not, and why that difference matters.

Amara Osei 10 min
Two abstract agents in a feedback loop, representing the self-play learning principle
Self-PlayResearch

Self-Play, Generalised

Game-playing agents learned to exceed human performance without any human labels, by competing against themselves. Can this principle be taken beyond games?

David Silver 11 min
Stack of documents with labelling marks, symbolising the manual effort of data annotation
Self-SupervisedData

The Hidden Cost of Annotation

Human labelling is treated as a data pipeline problem. We think it is an architectural constraint — and one that fundamentally limits what AI can learn.

Amara Osei 9 min
An abstract visual of something beyond description — faint form emerging from noise
PhilosophyFounding

What 'Ineffable' Means to Us

A word that means 'too great or extreme to be expressed in words' turns out to be a precise description of the problems we care about most in AI.

David Silver 7 min