SkinDeepRESEARCHSteve Seguin

Research

Test results

Choose a question. Each comparison explains the method, the data and what happened.

Learning what you like

Can a model learn your taste?Tests of learning from ratings and changing a design.

Decisions and computation

How much processing does a decision need?Stopping at different layers, checking confidence, trying a smaller model, and training changes.

Finding a place to click

Can a model return a point directly?Screenshot targets, neighbouring image patches and missing buttons.

Choosing an action

Can a small model finish a maze?More training examples, legal moves and the failures that remain.

Earlier combined report · Evidence and limitations · Research FAQ

Working past the context limit

A 121K-token ledger with under 9K active context, faster repeat questions, and the limits of disk-backed memory.

Plain-language overview · Measured results · Common questions

Results for practical tasks

Search ranking · Request routing · Finding names · Receipt totals

Saved browser runs

BERT method comparisons · Qwen execution check