At least with where AI is now, it’s basically an incredible data compressor. My local copy of gemm4:31b takes up less than 50GB iirc, and I can retrieve information from the entire Internet with it without an internet connection.
Using an AI to train an AI is like taking a jpg of a jpg. You’re going to lose information eventually. Hallucinations will become worse like in a game of telephone
I mean sure, you can ‘retrieve information’, with no way of knowing where the information came from, whether the source was accurate, or whether what you’ve retrieved is even remotely faithful to the source material. So basically you can’t actually retrieve anything, because it’s just mashing words together in a way that happens to sound correct most of the time.
At least with where AI is now, it’s basically an incredible data compressor. My local copy of gemm4:31b takes up less than 50GB iirc, and I can retrieve information from the entire Internet with it without an internet connection.
Using an AI to train an AI is like taking a jpg of a jpg. You’re going to lose information eventually. Hallucinations will become worse like in a game of telephone
I mean sure, you can ‘retrieve information’, with no way of knowing where the information came from, whether the source was accurate, or whether what you’ve retrieved is even remotely faithful to the source material. So basically you can’t actually retrieve anything, because it’s just mashing words together in a way that happens to sound correct most of the time.
You can ask it to provide a source.
Sometimes it will.
Sometimes it will make one up.
Lossy compression with no internal mechanism for detecting information corruption.
We did it! Progress! Let’s go back and show Claude Shannon!