LLM-assisted legacy code porting: promising but imperfect
A detailed post about using an LLM to port a 1993 Amiga game from 68000 assembly to Godot generated a lot of discussion. The technical result was genuinely impressive: the LLM read assembly, understood the domain, and made reasonable translation decisions. Commenters were surprised at how much obscure computing history the model had internalized.
But the thread also flagged a real problem: the prose in the writeup was obviously AI-generated, and at least one commenter said they couldn't force themselves to keep reading because of it. This is a useful split signal. The AI did the hard technical work well and then undermined the result by producing writing that signals 'this was generated' in a way that erodes trust.
The pattern here connects to the broader AI content quality debate on HN. The technical capability is advancing faster than the judgment about when to use it and when to write like a human.
So what?
If you're using AI for technical work and then using AI to write about that work, you're compounding the credibility problem. The technical output may be excellent; the AI-written explanation of it will make readers trust the technical output less. Write the prose yourself, or at least make it sound like you did.