[1attice]

The main problem with AI-gen solutions is validating the solutions. Across all industries, the most automatable jobs are the ones where the correctness of the AI's solution are most easily validated, and the situation is especially acute where the validation itself can be either done automatically (e.g. CI/CD), or outsourced to Mechanical Turk or Upwork.

So: anything that is impractical to validate in an automatic way, or, better, validate at all, is going to fare better.

Careful, however: many things that do not appear easily validatable actually are! For example, you might not be able to validate (with a GitHub workflow) that users prefer (say) a new UI design, but of course, you can check the design against well-understood usability principles, A/B test, and, soon, probably, user simulation, etc.

So: in the short and medium term, work that requires careful human attention or domain knowledge to validate (e.g. whether or not an interface design or system design or architectural plan is fit-for-purpose) will last longer.

In the long term, I don't think any programming work is safe.

[sirspacey]

I consider this a first principle of “GenAI-Native” solutions - workflows become eval systems on a task engine.

I hadn’t considered building a classification model for “how easy is this to validate.”

Thanks for the inspiration!