Скачать или смотреть Why LLMs Cant Really Build Software

All right, we’re going to talk about a few things today, starting with why LLMs are not great at building software, even though they often write code that runs. Conrad Erwin explains that hiring software engineers is tough, and there’s no magic answer, skills matter, but fit and communication probably matter more in an interview. You want to know if you’ll work well together; technical chops can be proven later.

Figuring out what software engineers really do comes down to what Conrad calls the “software engineering loop”: first, you make sense of the requirements; second, you write code you think matches those needs; third, you figure out what the code actually does; and fourth, you spot the differences and fix things. Great engineers are good at keeping this whole flow in their heads. LLMs can write and even update code pretty well if you give them direct problems to solve, like bug reports or errors, but they struggle to keep the whole mental map of the project and requirements at once.

They get confused by unclear tests, broken code, or when the goals change, while humans can rely on memory and intuition to figure out where stuff went sideways. LLMs also run into trouble with context, they forget details or get sidetracked if a session runs too long or tasks pile up, and their answers lose quality the deeper you go in one conversation. To get the best results today, you need to be specific about what you ask, give lots of detail, and keep to one task per chat.

Problems like context omission, recency bias, and hallucination can be worked around if you’re careful, but right now, big models still cannot keep track of enough moving parts to build nontrivial software solo. Humans have to check the code, keep requirements clear, and keep things on track, LLMs are a great tool, but you’re still in charge. Overcoming these problems means breaking down work into focused sessions and remembering that the fun part of coding can shift from typing code to figuring out how best to ask for what you want.

So, LLMs are helpful, but you have to put in effort and pay attention if you want good results. As for writing “cursed” Python code, it’s funny to look at how different models have handled this prompt over the years. GPT-1 did not know what to do, while GPT-2 wrote strange code that looked sort of right but did not run.

Text DaVinci 001 just made a vague comment about Python being cursed. GPT-4 went too far with safety and refused to give anything that could be called cursed, thinking it meant something harmful. But GPT-5 gets it and outputs actually “cursed” but working Python, exactly what was asked.

It shows that models today understand prompts better, though the earlier generations were not even close. All in all, with LLMs, the key is to be clear, specific, and to manage your own expectations, they are powerful, but they are just tools, and you still need to do the real thinking.