A new kind of language network, built to capture meaning the way a photograph captures light.

A photograph, not a phrasebook

Most language models learn something like a huge phrasebook of what usually comes next. I'm building one that works more like a photograph of meaning.

Geometric representation means meaning captured as a continuous structure — the way brightness varies across a photograph, or loudness across a sound — rather than as statistics about word order. It's a different foundation, not a tweak to existing models.

I don't share architecture details publicly. Technical detail is available to qualified partners under NDA.

Goals, not results

Everything here is a goal. I'll only claim results once they've been tested and published.

• Aims to learn structure efficiently — including from rare examples, not only from huge amounts of data.
• Aims to be inspectable — so you can see why it gave an answer, not just what it gave.
• Aims to say when it doesn't have enough information — instead of confidently guessing.

Where it's at

Research prototype. Results will be published when properly tested.

I'll pre-register hypotheses and test on held-out data against baselines. Held-out data is data kept out of training and only used for testing. A baseline is a standard model to compare against.

There are no performance numbers here and no comparisons with other models. When there are tested results, they'll go here.

Patent pending (Canada).