What happens when a new model ships

Model launches read as consumer events. A new frontier model ships, the demo videos circulate, the benchmark charts get argued over, and everyone with an opinion has one before lunch. What happens inside the labs during that same window is different, and it is the part of the story that never surfaces publicly.

Every new model is a new evaluation problem. The capabilities it unlocks need new benchmarks. The failure modes it introduces need new adversarial tests. The training data domains it opens up need new expert reviewers. The pattern has been consistent for two years and it is not slowing down.

The eval work does not exist yet

A model can only be improved against evaluations that already exist. When a new model shows unexpectedly strong performance in medical reasoning, or legal argumentation, or graduate-level mathematics, the internal question is immediately: how do we evaluate it accurately at that level. The old benchmarks are saturated. The next set has to be built.

Building the next set is expert work. It requires physicians authoring clinical vignettes at a specificity previous benchmarks did not reach. It requires lawyers writing case questions dense enough for the current models to fail on. It requires mathematicians constructing problems that separate genuine reasoning from pattern-matching against training data.

That work is contracted out. It is where the largest expert-hour budgets go in the weeks after a major model release.

Adversarial testing scales with capability

The more capable a model becomes at a domain, the more subtle its failures become. A model that gets basic medicine right will still fail on a specific interaction between two medications, or on a rare presentation of a common disease, or on a jurisdictional nuance in a treatment guideline. Finding those failures requires practitioners who have spent years watching real cases go sideways.

Adversarial testing is one of the fastest-growing categories of expert work. Not because labs want their models to fail, but because they need to know exactly where they fail before deployment. Every new model gets a wave of expert red-teaming across every high-stakes domain it might touch. Physicians, lawyers, financial analysts, teachers, and safety-critical engineers are all part of that wave.

The training data domains keep expanding

Each release opens up domains the previous generation could not credibly enter. Medical AI, legal AI, education, scientific research, and applied engineering are all fields where the previous generation was not good enough to be useful and the current generation increasingly is. The moment a field crosses that threshold, the labs need experts in it.

This is why the expert-work catalog looks different every three months. It is a leading indicator of where the frontier is moving.

What this means for you

If you are a domain expert reading this, the practical implication is that demand for what you know how to do is increasing with each major model release, not decreasing. The framing that AI will make expertise obsolete has been running in one direction. The reality inside frontier labs is running the other way. They need more specialists, not fewer, because the surface area of what they are trying to teach models is expanding at every scale.

The work is here now. Access is free. The catalog updates hourly against the current openings across the labs.

If you have not started, the moment a new model is announced is a good time to. That is when the next round of expert work is being sourced.