Prevent cognitive debt by manually retyping LLM-generated code

Yes, of course, but a lot of it is boilerplate and just figuring out how some function is called or how exactly it is exposed in the library, dealing with some frustrating but trivially solved inconsistencies or incompatibilities or oh you first have to initialize this or that, or actually this feature is missing from version X but can be emulated by setting an environment variable to this other value and blablabla.

> all programmers are saying that the code is bad.

Ridiculous. When did you last use AI for actual work where you wanted it to actually work? I know it may sound like saying James Randi's negative attitude blocks my telepathy or dowsing skill, but I get this impression of many people. They tried agentic coding a year ago, decided that it doesn't work and if they tried it ever since, they didn't provide context, didn't set it up for a possibility of success then immediately closed the program and sighed a sigh or relief that they are safe. It's not like that. The current frontier is very capable on higher level design tasks as well. If your problem is spiritual or social or economic, then let's skip this "not productive enough" alibi talk.

Machine translation today may not be up to the levels of the best professional translators who work on something with full focus and full effort, given all the time they want to polish it, but it's better than an ordinary (not particularly the top of the top) translator working multiple project with tight deadlines. You have to compare like for like and take cost into account too. There is an immense amount of mistranslations by "professional" translators, not sure if you've ever seen movie subtitles or TV series dubs. Lots of paid professionals hired by production studios make huge mistakes, clearly miss cultural references even sometimes grammatical structures, which modern machine translation aces and even chat models are great at it and know basically all obscure pop culture references that e.g. a regular American audience member is expected to know but a translator working in Estonia may not get. Yes you should have a guy checking the output and doing some light touchups, given the draft and explanations and references provided by the AI, but the time of doing it fully manually and being competitive on the real world translator market has ended.

I've been trying to address this by telling LLMs primarily how the code should be structured, rather than only what it should do. Still, any design I hand to an LLM will be underspecified in one way or another (if it were fully specified, it would just be code), and the LLM fills those gaps somehow - which adds to the cognitive debt, slowly but surely.

Retyping LLM-generated code is an interesting solution. You'd certainly end up understanding the generated code better than if you merely reviewed it, but I doubt it produces a mental model as reliable as the one you'd build writing the code yourself. The longer you think, the better your mental model gets - and outsourcing the thinking to an LLM means thinking less.

That said, I've started to wonder whether I'm solving the wrong problem. Should I really insist on an accurate mental model of the code I own? We'd find it strange for a non-engineering manager to try to fully understand every piece of code their reports produce. If that's the right analogy, then as LLMs' agentic capabilities improve, maybe we should stop treating LLMs as tools that boost our own productivity as a software engineer and start treating them as independent agents we manage and steer.

NPM decencies is a bit of dumpster fire, supply chain attacks galore! A well known python (think numpy) certainly is less discomforting, and a "industry standard" C library like GSL or SDL3, I just use it and move on