Skip to content
Reza Motaghi
Lab Notes3 min read

What does not depreciate when the model changes

By Reza Motaghi

Everything you rent from a model vendor depreciates within a year. Three things you make yourself do not.

I run every AI call in my pipeline through one alias. A config file maps the alias to whichever model answers today. When a better model ships, I change one line and every agent runs on a different brain with no code touched. I did it this summer, when the model I had picked a few months earlier was already superseded.

The adapter is the opposite. An adapter is the small set of weights you add to a base model when you fine-tune it on your own examples. It is shaped to one base and does not port. When the base changes, the adapter is retrained, never patched. So on both sides the weights are the part that gets thrown away.

What does not depreciate is what the weights were trained and graded against.

My signed reads, taken to consensus with a second reader, with a version number. The written rules a read has to satisfy, in plain language a colleague could check. Every correction I made to a model's output, kept as a case. A new base model is a config change. Those three carry over unchanged, and every model jump makes them worth more, because generation gets cheaper with each release and verification does not.

The published tests say the same. In 2026 the three strongest general models each passed fewer than half of the critical items on clinician-written rubrics across four specialties, and more than half of those items were met by none of them (arXiv, 2026). In rheumatology, a small open model given the current guidelines to check against outperformed larger general models that had none (Frontiers in Medicine, 2026). What the model was given to check against mattered more than which model it was.

Own what does not depreciate: graded truth, explicit rules, your own corrections. Everything else, route through an alias and expect to replace.

The check

The depreciation test, for anything in an AI stack:

  1. If the best model in the world shipped tomorrow, would this asset be worth more, the same, or less? Graded truth and written rules are worth more. Weights are worth less.
  2. Can it be swapped by a config line? If not, it is coupled to a model that will be gone in a year. Decouple it or accept that it is disposable.
  3. Is there a version number on your answer key? Without one, your own corrections cannot compound, because nothing can be cited.

Sources

  • arXiv 2607.02175 (2026). Frontiers in Medicine, DOI 10.3389/fmed.2026.1817215 (2026).

About the author

Reza Motaghi is an oral and maxillofacial radiologist and Chief Innovation Officer who reads every day and builds, evaluates, and trains the imaging AI for it. He built CBCTScope, the first CBCT viewer with native AI-agent control, and writes about what a model did on a real case and what it did not.

More about RezaSubstackLinkedInGitHub

The newsletter

Get the essays as they’re written.

Imaging AI, evaluation, and building alone, from a radiologist who reads every day and tests before he trusts. A few essays a month, nothing else.

Join for FreeFree · Unsubscribe anytime

Related

  1. Read a paper for the experiments it lets you skip

    One paper read twice struck a planned experiment and handed me a step. How to read a paper for the experiments it lets you skip.

  2. Three training runs, one lesson

    Three fine-tunings on my own radiology reads learned my report and not the image, and no training dial showed it. The three checks that catch it.

  3. Write the missing page

    The first Unsloth fine-tune of a vision mixture-of-experts model: the thirty-second gradient check and the two silent defaults that would have ruined the run.