Newsletter · Engineering Frontier
How to tell when a model is lying to you
One essay at a time, about eight minutes to read, on supply chain and distribution and the places where AI and data analysis quietly stop being trustworthy. Sent when a piece is finished. Free.
Why there is no subscriber count here
Every page like this one quotes a number. The honest reason this one does not is that the number would be small, and a site whose argument is that claims should be measured is a poor place to start dressing one up.
So here is the evidence instead, which is better than a number anyway: the archive is already public and free. Nothing is held back for subscribers. Read one before you decide whether you want the next one.
What has already gone out
- The training set moves the error 390x. Nothing predicts it. Across 231 training-set choices the worst case moved by a factor of 390. I tested nine cheap ways to pick a good one. None of them held up.
- The metric was measuring my model, not the problem A transferability score held at 0.54 across two independent budgets, then collapsed to 0.14 when I changed the model class. It was tracking the algorithm.
- You cannot count the sale you did not make Sales data records what you sold, not what people wanted. The two differ exactly on the days they matter most, and the gap teaches a forecast to shrink.
Who is writing it
I work on supply chain and distribution, and I spent a year in flower supply selling business to business, which is where most of the operating detail in this writing comes from. Alongside it I publish computational research with DOIs, indexed on Google Scholar, and I am a mechanical engineering undergraduate at Tel Aviv University. The pieces that cite my own work link the data and the code.