Hacker News new | ask | show | jobs
by derefr 1199 days ago
It can't access the internet now. But it was fed a snapshot of the Internet during training. So if you tell it a URL, it can potentially remember what the page looked like, if it happened to see it during training.
3 comments

In either scenario, it's hallucinating. Either JSON queries were supported in that snapshot of the documentation, or they weren't.
That's very unlikely. ChatGPT doesn't contain a full copy of everything it saw when it was trained - it has model weights that were updated by that training process.

It's very unlikely the model weights would contain enough detail to let it match content to the URL that the content originally came from.

You’re saying ChatGPT has access to a an older snapshot of the internet with URLs and everything?