Far Akin Research Labs studies how minds, natural and artificial, learn, represent, remember, perceive and act, and how far they are akin. A model's hidden state and a person's inner life are both known only from outside. The lab's programme is to read both under one standard of evidence, so that a claim that a model and a mind are alike can be tested rather than asserted.
A model's hidden state and a person's inner life share one condition: neither can be read directly for what it holds. Every activation in a model can be recorded, yet what the activations represent is not written on them. Each is known from outside, through what it does, what can be decoded from it and what changes when it is disturbed.
That shared condition is where the lab starts. The same questions recur on both sides: what is represented, what is held steady, what is remembered and what is only reported. A method built for one kind of system can sometimes be carried to the other: the cybernetician's disturbance, the psychometrician's test-retest check, the interpretability researcher's probe.
Resemblance is tested, never assumed. Behaviour that looks human, or a feature that can be decoded from a model, is not evidence of a shared mechanism or of experience. Where a resemblance holds, the work is to show it. Where it breaks, the break is a result.
A probe reads a model the way a sounding reads a seabed: exactly where it touches, by inference everywhere else. The figure shows the seabed half of that comparison.
Hidden. Reveal it to compare with the chart.
6 soundings · average error against the hidden seabed: 5.6 m
Select a point on the chart, or focus the chart and move with the arrow keys, then press Enter to sound there. Contour lines mark every 10 metres.
How the figure works: the hidden seabed is a fixed surface of six smooth rises and hollows on a 48 by 28 grid. Each sounding reads the exact depth at one point. The chart fills every other point by inverse-distance weighting, each sounding weighted by one over its squared distance. The figure opens with six soundings; Start again clears them all. The error is the mean absolute difference between chart and seabed over all grid points. It illustrates inference from sparse measurement, and models neither a network nor a mind.
Each programme follows one long-horizon question about how minds and models come to hold what they hold. Its inquiries are the current steps on that question. The research page shows the wider map of subjects and lenses behind them.
What does a model hold, which of it does it use and report, and how does what it holds stay correct as the facts it rests on change?
Minds and models carry internal state and stored knowledge that decide their behaviour but cannot be read directly. This programme asks what can be read from outside, which readings reflect what a system uses, and how a memory can be revised and checked as the world changes.
Where should what a model knows live, in its weights, in a store it consults or in checks on its output, and when does a small specialist judge better than a general model given the same knowledge?
Much real work needs consistent judgement inside one domain, under that domain's own rules. This programme asks where specialised knowledge should live, in weights, in retrieval or in validators, and what a small specialist must show before it is worth building.
What does an agent hold steady, and what must a learned model of its world keep consistent, under action and over time, before its predictions can stand in for that world?
World models learn to predict what follows from an action, and they are now used to train agents and to judge policies. This programme asks what such a model must hold constant to be trusted, and what an agent defends when it is disturbed.
When a language model stands in for a person or a population, which parts of its answers are evidence about people and which are artefacts of the model, the prompt or the simulation itself?
Language models are now used to simulate respondents, users and interviewees, and agreement with published human data is the usual evidence that this works. This programme designs tests that separate an estimate from an echo, and a stable simulated person from one the procedure has quietly changed.
Interpretability has begun to open artificial networks. Psychology, neuroscience and cybernetics spent more than a century learning to read minds from outside. Both now ask the same questions: what a system represents, what it uses, what it can report, what it keeps and what it holds steady when disturbed. The agenda sets out how the lab asks them of both kinds of system, and how each question becomes an experiment.
Read the agendaThe agenda and research pages now organise the lab's work as four programmes on the interpretability of minds, natural and artificial, with the subjects and lenses each draws on.
The agenda sets out what the lab aims to build and how it will know whether it worked. Research is now organised in four areas, two new inquiries are designed, and every design states its definitions and the rival explanations it must rule out.
Eight inquiry designs, three notes and the research pages are published. None of the designs has run.
Essays and explainers on the questions behind the programmes. Each is labelled for the kind of writing it is.
A system that controls a perception acts to cancel disturbances to it, so what it controls shows in what is restored, not in what it does. The test for the controlled variable, and what carrying it from people to learned policies would and would not license.
A depth sounding is one exact reading. A chart is inference between readings, and can be confidently wrong where nothing was measured. The same distinction separates what probing a model's interior shows from what it only suggests.
Why one lab studies learning machines, human minds and simulated worlds together, what it would take for them to be akin, and how that claim could fail.
The lab seeks to work with researchers in interpretability, cognitive science, neuroscience and measurement: to critique a design before it runs, to co-author an inquiry, or to share data and instruments.
Write to the lab