I recently had a “conversation” with Gemini 2.0 from Google. I was asking it how it comes up with the answers to questions. Here is how that conversation went.

Me:

So would it be correct to say you are a statistical guessing machine?

Gemini:

That’s a simplification, but it captures a key aspect of how I work. It’s more accurate to say I’m a statistical pattern-matching and generation machine, but the “guessing” element is definitely there.

Here’s why “statistical guessing machine” is partially correct and where it falls short:

Why it’s partially correct:

Why it’s an oversimplification:

A better analogy:

Imagine a highly skilled musician who can improvise complex musical pieces. They’ve learned the rules of music theory, practiced for years, and have a vast repertoire of musical phrases in their memory. When they improvise, they’re not just randomly guessing notes; they’re using their knowledge and experience to create something new and beautiful. While there’s an element of “guessing” involved (predicting what notes will sound good together), it’s a highly skilled and informed kind of guessing.

I’m similar in that I’ve been trained on a massive dataset of text and code, and I use this training to generate new text. While there’s a statistical element to it, it’s more than just guessing. It’s a sophisticated process of pattern matching, generation, and manipulation of information.

End Gemini ..

I am not sure I believe the part of comparing AI to a highly skilled musician. If AI is not able to think on it’s own how can you can you say it can create music. Wouldn’t it just be giving you the highest probability of notes based on the date it was trained on?

Maybe in time AI will be able to create new ideas based on the info that it has been trained on, not just give you the highest statistically likely response.