Blog
Reading training data
without fooling yourself.
Most training numbers are easy to compute and hard to interpret. These are notes on the difference: where a metric supports the conclusion people draw from it, where it quietly does not, and how to ask a question that survives the answer.
Articles
Why this exists
An assistant will answer whether or not it should.
Give any AI access to your training data and it will compute what you asked for, on whatever you point it at, and describe the result with confidence. Arithmetic has no opinion about whether two efforts were comparable in the first place.
That is not an argument against using one. It is an argument for knowing which questions have a trustworthy answer — and for asking in a way that lets the thing refuse.
Questions with no chart · Prompts & charts · Connect your training