alt Hacker NewsI don't see why this wouldn't just lead to model collapse:
https://www.nature.com/articles/s41586-024-07566-y
If you've spent any time using LLMs to write documentation you'll see this for yourself: the compounding will just be rewriting valid information with less terse information.
I find it concerning Karpathy doesn't see this. But I'm not surprised, because AI maximalists seem to find it really difficult to be... "normal"?
Rule of thumb: if you find yourself needing to broadcast the special LLM sauce you came up with instead of what it helped you produce, ask yourself why.
Replies
The article is not on training LLMs. it is about using LLMs to write a wiki for personal use. The article assumes a fully trained LLM such as ChatGPT or Claude already exists to be used.
Edit for context: the sibling comment from karpathy is gone after being flagged to oblivion. Not sure if he deleted it or if it was just removed based on the number of flags? He had copy-pasted a few snarky responses from Claude and essentially said “Claude has this to say to you:” followed by a super long run on paragraph of slop.
————
Wow, I respect karpathy so much and have learned a ton from him. But WTF is the sibling comment he wrote as a response to you? Just pasting a Claude-written slop retort… it’s sad.
Maybe we need to update that old maxim about “if you don’t have something nice to say, don’t say it” to “if you don’t have something human to say, don’t say it.”
So many really smart people I know have seen the ‘ghost in the machine’ and as a result have slowly lost their human faculties. Ezra Klein, of all people, had a great article about this recently titled “I Saw Something New in San Francisco” (gift link if you want to read it): https://www.nytimes.com/2026/03/29/opinion/ai-claude-chatgpt...
I did a proof of concept for self-updating html files (polyglot bash/html) some weeks ago. It actually works quite well, with simple prompting it seems to not just go in circles (https://github.com/jahala/o-o)
also my experience. it can't even keep up with a simple claude.md let alone a whole wiki...
Here in 2026, many forms of training LLMs on (well-chosen) outputs of themselves, or other LLMs, have delivered gigantic wins. So 2024 & earlier fears of 'model collapse' will lead your intuition astray about what's productive.
It is unlikely you are accurately perceiving some limitation that Karpathy does not.