It sucks. I see itnas the snake eating its tail or whatever. Hopefully when the internet is full of useless ai slop people will realize we cant acrually replace humans for creative tasks.

After a lengthy argument with my boss, I was allowed to “only” force the AI-first model onto half of my small dev team of 8 engineers and 4 QAs. After about a year, the results are obvious: the engineers who went AI-first are now AI-fluent: they know how to prompt the model, develop agents, AI-automate workflows, and push a TON of code. Like almost an order of magnitude more code into prod. However, to say their skills have stagnated would be generous. In some cases an engineer I could count on to debug a tricky issue is completely not able to anymore. Meanwhile, the non-AI team is now essentially my code quality and bugfix team. Their PRs are small and concise, their changes are understandable, and most importantly, their skills have remained the same (and in the case of the 2 juniors, have improved).

If we hadn’t kept a few people in reserve, we’d be fucked, because there have been multiple times where an AI-first engineer has tried debugging an issue (and with big models - Astra, Fable, Opus) that either totally failed and gave up, or made things worse by introducing “fixes” that made problems elsewhere.

So this is a small sample, and YMMV, but for junior engineers in particular (who usually have the least amount of say in pushing back against bad ideas) this is – not “is going to be” – a major problem.

This is a real problem, I already see it with CS students.

AI is useful in some limited scenarios, the problem is that when you’re learning CS all of the scenarios that you’re given as exercises are almost trivial to solve and fall within AI’s capability to do for you. That completely handicaps a student in ways they don’t even understand until much later.

It isn’t until they get into more complex projects, where AI flounders, that they recognize that they have rarely had to troubleshoot an issue so they don’t have the mental framework to do so. They’re not comfortable with the tools, and they haven’t internalized the methodology so they hit a brick wall.

If you’re a student now, avoid it at all costs. It’s easier to use AI, of course, but facing the difficulty and overcoming it is probably the most important part of learning anything.

I’ve added this as one of my interview questions for people who clearly use AI.

“How has mentoring junior engineers changed since the advent of AI? How do you guide juniors to ensure they are still learning, even if they want to use AI heavily?”

It’s a good way to ask if they like AI slop or not without asking the question directly

If you’re learning, you need to be writing. You can read a book, say, The Go Programming Language, you can read that front to back and you may understand the code in there. But after, when you sit at a keyboard, you probably won’t be able to write a significant program in Go. Not because the book is bad, but because you didn’t PRACTICE. You have to write code. Lots of it.

All that to say, you have to put in the work to become good.

I worry the same thing happening in all fields, especially medical. we train the models on experts, then give the newbies the model… so when and how do we improve and retrain these models, and ok what data?