@paulgp.com I do agree that moving from an equilibrium of "zero tests (and not even thinking about them...)" would be an improvement!

FWIW a fair number of academics write useful "assert" statements in their code. When introducing more rigorous testing to academics, that's a useful reference point.

I taught my intro health econ students yesterday that:

Not having disease D > Having disease D, diagnosed and treated > Having disease D, undiagnosed (which is self-reported as no disease)

But there's a 4th worse option: Having disease D, but confidently believing you don't because of faulty tests

That lack of experience is one reason why academics could be especially vulnerable to the appealing idea: “I’ll get the AI to write my tests for me”.

Designing good tests requires knowledge/taste! But if the LLM vibecodes the tests, you’re back to a “who will watch the watchmen” problem.

I'll be over here sorting my chats into the "kiki" and "bouba" folders.

The sources section on the @tpwky.bsky.social HIV episode could have some great pointers:

thispodcastwillkillyou.com/2018/02/10/e...

@khoavuumn.bsky.social "have you considered grad school?"