The CFO points to a line item in the quarterly forecast. “This new personalization model,” she says, tapping the screen. “We spent two million on the training compute and another million on the data licensing. What’s its useful life? Five years? Ten?” Across the table, the VP of Engineering feels a cold sweat. He knows the truth: the useful life could be six months, or it could be six hours. It depends on whether a new product goes viral on TikTok tomorrow.
We are building a new class of technical asset, and we are trying to manage it with the intellectual tools of the factory floor. It does not work. A software license has a clear expiration date. A server rack can be depreciated on a predictable five-year schedule. A trained neural network is different. It is not a durable good. It is a snapshot of a world that is already gone, a statistical photograph whose subject is walking out of the frame.
The value of a model does not decay; it curdles. Its performance is not a gentle slope downwards, but a cliff edge you discover only after you’ve fallen off. A fraud detection system trained on transaction data from 2022 is a liability when faced with a novel scam cooked up last week. A customer support bot fine
Generated by Reportify AI — Automate your team's status reports, standups, and weekly updates. Try free →