[zozbot234]

Because it's not "PhD-expert level" at all, lol. Even the biggest models (Mythos, GPT-Pro, Gemini DeepThink) are nowhere near the level of effort that would be expected in a PhD dissertation, even in their absolute best domains. Telling it to work out a plan first is exactly how you would supervise an eager but not-too-smart junior coder. That's what AI is like, even at its very best.

[TeMPOraL]

That's not the best framing, IMO. More important is, even a PhD expert human wouldn't one-shot complex programs out of short, vague requests. There's a process to this. Even a thesis isn't written in one, long, amphetamine-fueled evening. It's a process whose every steps involves thinking, referencing sources, talking with oneself and other people, exploring possibilities, going two steps forward and one step back, and making decisions at every point.

Those decisions are, by large, what humans still need to do. If the problem is complex, and you desperately avoid needing to decide, then what AI produces will surprise you, but in a bad way.

[ModernMech]

I understand that but 1) expert-level performance is how they are being sold; but moreover 2) the level of hand-holding is kind of ridiculous. I'll give another example, Codex decided to write two identical functions linearize_token_output and token_output_linearize. Prompting it not to do things like that feels like plugging holes in a dyke. And through prompting, can you even guarantee it won't write duplicate code?

I'll give a third example: I gave Codex some tests and told it to implement the code that would make the tests pass. Codex wrote the tests into the testing file, but then marked them as "shouldn't test", and confirmed all tests pass. Going back I told it something to the effect "you didn't implement the code that would make the tests work, implement it". But after several rounds of this, seemingly no amount of prompting would cause it to actually write code -- instead each time it came back that it had fixed everything and all tests pass, despite only modifying the tests file.

In each example, I keep coming back to the perspective that the code is not abstracted, it's an important artifact and it needs/deserves inspection.

[ai-tamer]

Note to self: don't go away for the weekend without HN ;-)

Personally I run several agents. At minimum Codex and Claude, so they cross-check each other. Exactly to avoid what you describe. Duplicate functions or "tests all pass" when nothing was implemented is the kind of thing a second model, reading the diff with fresh eyes, tends to flag immediately.

But it takes coordination and solid skills + rules to make it work. Tomorrow's battle -> AI skillers.

[ai-tamer]

https://arxiv.org/abs/2604.04721 I like it when researchers confirm my intuition

[zozbot234]

> the code is not abstracted, it's an important artifact and it needs inspection.

That's a rather trivial consideration though. The real cost of code is not really writing it out to begin with, it's overwhelmingly the long-term maintenance. You should strive to use AI as a tool to make your code as easy as possible to understand and maintain, not to just write mountains of terrible slop-quality code.