
Michael Podgortsev
I measure what it costs to build systems for an average user. Fifteen years building data and AI functions, and now the research that puts a number on the assumption underneath them.
Measuring the cost of the average
Open papers with the data, the code and the failed experiments included. Every number in them regenerates from files in the repository.
Single Human-AI Fit
Does a language model treat a person worse when one detail about them changes and everything else stays byte-identical?
The Average-Fit Loss
What does it cost to serve a whole population with one design tuned to its centre?
Written elsewhere
The same argument, made for the people who have to act on it: executives, operators, and the teams who own the systems.


