Hacker News Reader

I think this assumes there's no other AI spewing data on the Internet, and that people continue to share information.

The amusing thing is that controlling nonlinear multivariate feedback systems is an unsolved problem.

But what if it eats Bard data, which is maybe even worse?

Or more generally, you end up with a cycle:

ChatGPT trains Bard which trains Bing which trains ChatGPT, etc.

Nailed it!

> If ChatGPT is able to emit output that is watermarked

Then the watermark should be made public. We need a way to tell ai-generated content from human-generated content without actually having to read it.

I wonder how difficult it will be to apply the filter to human generated content.