Post

At the risk of starting the flame war to end all flame wars... Modern LLMs (GPT-5.1, Claude 4.5, Gemini 3) produce excellent code and can be a significant productivity boost to software engineers who take the time to learn how to effectively apply them - especially if used with coding agent tools

it's also true I wouldn't accept code from someone who couldn't write that code without an AI assistant and who doesn't know how it works, especially if it's a hit-and-run PR that they won't have to maintain

That looks like a very intermediary state of the world. At some point I expect we will just commit the prompt + hash of model, whose output has been proven to work by downstream validation, and only look at the "intermediary assembly" in unusual cases.

Ok I mean it's pretty easy to say this, but rarely can someone actually explain every layer of dependencies underneath the code they CAN explain. Comprehension is always shared across people for complex work.

True, but they're mostly not responsible for those layers. It's basically the same as before LLMs, right? My PR must be human-understood at the usual levels... but it may be mostly LLM written. ...it is quite the quirk, how modern sw dev is a very high % of code review.

I mean same as before, when? It's not the same understanding as it was in the 70s-80s, from which era I was just reading papers claiming that the only people who should be allowed to create code should pass timed math challenges

That's true, I may be able to reason my way through Python libraries I'm using but once you get down to the cPython interpreter I'm mostly out of my comfort zone, and once you get to the operating system calls or hardware I'm completely lost

What level of comprehension/and fixing we want to expect from people, in order for them to do certain kinds of work, is such an interesting and important question! And has been a bit ignored in software so far in a way that isn't helping us now, imho.

As someone in a leadership role in our software engineering organization, this is actually the sticking point for me: there are often perceived or actual (in our case mostly just perceived) incentives to SHIP FAST in ways that very directly undermine the commitment to understanding generated code.

In my experience, how well the tools work is also (a) very context-dependent, not only in the sense of “how much of this is in the distribution?” but also “what kind of work am I doing?” and (b) depends on how comfortable you are with the totally unsurprising tendency to regress toward the mean.

There are ways to deal with both of those, but they aren’t always worth the cost, and *one* part of the push-back is because there are folks saying, “This is a useful thing that you should have in your toolbelt”… and then there are many, *many* folks saying “all software MUST be written by AI!”

I want to briefly name that you, Simon, and @unstable-idiot.bsky.social are saying two different things that I think have two very different implications. “have to be able to write the code without an LLM” and “can explain how the code works” are, in fact, pretty different things.

Good point - yeah, I do disagree slightly there I'm OK with code you couldn't have written without AI assistance provided you completely understand it and can explain how it DOES work

Ope, let’s keep going! What does “completely understand” mean vs “understand”? And I imagine you mean “explain how it will actually work in this given context” when you say “explain how it DOES work”?

“Code you couldn’t have written without assistance” is - in my opinion - essentially equivalent to “code you don’t understand”