Post #3810180
2026-07-14 09:57 UTC
@gunstick@mastodon.opencloud.lu @bagder@mastodon.social people who tell you they can write code don't realise it's because they regurgitate bits of code from training models – stolen data btw – and that it's just the same way as LLM chats, probabilities of what's going to be the next possible token.
Basically, they give you the same as if you were copy pasting code from other codebase except that when you do that there's a chance you understand it yourself and start learning.
Replies (1)
-
@troed@swecyb.com 2026-07-14 14:08
@kaneel@mas.to That's not how LLM-training works. It's known as "overfitting" and is quite useless since the models would then only be able to perform tasks that exist in their training data. https://www.ibm.com/think/topics/overfitting