Eusophic
Try NY ↗

We study how much a small system can come to understand when everything it knows has to come from what it has lived through.

Grounded language

Words from situations

Learning what words mean from the situations they occur in, not from a text corpus.

Memory

And forgetting

Long-term memory in which forgetting is a behaviour of the mind, not a storage limit.

Goals

Formed from within

Goals that form inside a system and can be deferred, instead of being assigned to it.

Provenance

Where a belief came from

Beliefs that keep track of their origin and of how many independent sources agree.

Priors

Knowledge that transfers

Passing confirmed knowledge to a new system while leaving it open to contradiction.

Worlds

Worth learning from

Consistent simulated worlds that set the rules and never leak the answers.

···Approach

Why we start from zero

A system that has read everything can say almost anything. It is much harder to tell what it actually knows.

We take the opposite route. The systems we build begin without a vocabulary, without a map and without memories. Each thing they later know has a history: when it was first met, how often, from whom, and what contradicted it. That history is stored next to the knowledge and can be read.

This is slower, and the result is far narrower than a language model. We accept that trade. A little knowledge with a full account of where it came from tells us more about learning than a great deal with none.

NY, the system this research runs on

06Questions

Figures measured on NY

Method

Thresholds fixed in advance

The numbers that decide when something counts as learned are frozen before an experiment runs, and stay frozen.

Evidence with every claim

Each result is tied to machine-readable artefacts and to the commands that reproduce them.

Limits stated plainly

We list what is unverified beside what is verified. We do not report IQ, age equivalents or claims of awareness.

Research questions and collaboration

hello@eusophic.com