[revertmean]

It's not just about IP. Have you considered how much an LLM trained on scraping websites might know about you? Would it reveal that information if I prompted it with something like "Create a short story about HN user jMyles and his home life"?

It might be best to know what these models have actually been trained on.

[zmnd]

100s of millions of people used it for quite some time now, many on daily basis. Do we have any evidence that it “knows about you”? Why can’t we say that the “internet” knows too much about you?

[revertmean]

To your latter question, you very much could. But pretty much the whole point of these models is to create connections between various bits of information. It obviously knows about some people (otherwise how could you ask it about famous people?). If it's been trained on everything you can trawl from the net, then why wouldn't it know about you?

But I don't know if people have been asking these bots about themselves. I don't have access to ChatGPT4. Anyone checked this?

I suspect the ChatGPT models haven't been trained on everything available online, so it wouldn't know much, but perhaps the next generation will.

[zmnd]

Internet is basically a bunch of nodes with interconnected data. Any search index might be considered as an interconnected graph. Same way with LLMs, it’s nothing new, might be considered faster depending on the definition.

Comparing to celebrities, data about people are so sparse that it would look like noise. I would be surprised if it encoded anything useful.

Half of the internet and all media were obsessed with making it say the f-word and tricking it into saying that it would kill all the people. Attacking from the privacy is quite obvious but I didn’t see it mentioned. I asked about 10 random friends and myself and it didn’t recognize the names despite having plenty of search results.

In one of the interviews they mentioned that it was trained on 10% of the web. So should have enough data.