Скачать или смотреть Write an extremely cursed piece of Python

Here is an edited version of your text. I stuck to your guidelines: it's a single paragraph, conversational, factual, clear, not corporate, and used simple vocabulary. Some facts about language models, Python versions, and their behavior were cleaned up for accuracy.

Tone and cadence are as close to your original as possible. The main difference between people and LLMs when it comes to coding is that humans can hold a clear mental map of a project, jump around tasks, and fix bugs based on context, while language models struggle with holding all the details in their “mind.” Humans loop through four main steps: figure out what’s needed, write code, read code, and fix issues, then repeat forever. LLMs like GPT-4 and Claude are great at writing and updating code if you point out the problem right away, and they can even read code, run tests, and add logs, but they struggle if you expect them to connect changes across a big codebase or guess where bugs hide.

When engineers get stuck, they might talk out loud, ask a friend, or rubber duck the problem, which helps them think through stuff; language models just output text without that deeper thinking. The main problem with LLMs is they get “confused” as the chat gets longer, and their answers get worse if you try to do too many steps in one session. This is because their training only goes one prompt deep and adding more context makes it way harder for the AI to follow what’s happening, since there are so many possible paths and details to keep straight.

Unlike a person who can remember the big picture and zoom in or out, LLMs can’t “pop the stack” in their heads and their memory is limited to the chat window. They also have issues like forgetting details (like file names or functions), getting stuck on the most recent info, and making stuff up if your prompt is vague. You can help by being very specific in your prompt, name the exact file, the function, and the variables you want changed, and tackle one task per session.

For simple tasks, language models are great and can generate code or docs fast, but for anything tricky, the human still needs to keep the project on track, set the requirements, check if changes really do what’s needed, and use the models as a tool. You can kind of think about LLMs as a big paintbrush: they can cover lots of ground with code, but they’re not precise like writing one letter at a time. If you don’t provide the details, or if you ask the model to deal with deep or complex problems, it probably won’t work as well.

Right now, if you focus on one specific step at a time and give all the details, LLMs do a pretty good job. But if you hope they’ll manage whole projects or think like a human, it’s just not there yet.