It is black magic, I tell ya. It knows.. everything.
I know about tokens and I'm familiar with the term "stochastic parrot", but...
I opened Gimp and drew "A B C D" on the left.
On the right, I drew 4 fish of different colours and connected them to the letters with tangled lines of the same colour.
I asked ChatGPT 4 to solve the puzzle. It got it wrong on the first try, matching "A" with the first fish, "B" with the second, and so on.
I said "No".
It apologised and said "let me have some time to work on it". It actually stopped writing for a few seconds and then got it right!!!!
I don't get how it does it.
nobody really does, but the mechanics at the core of it are basically probability.
"hallucinations" are a PR term, the fact is that it has no idea what is true or what isn't. the program has calculated what is the most probable words to come next.
the generating new images is sort of just super hyper compressing images, and then uncompressing them. they claim its new stuff, its not.
the image editing, im not sure how that works. but LLMs having intelligence is a PR move, its not intelligent, it just predicts the most likely thing from what it has been trained on.
Yeah, noone does. Although in this case, the engineers probably have it run a chain of thought or program of thought type of thing. Chain of thought is when it has to explain its process in thorough details *before* producing an answer, and program of thought is when it generates a plan for solving the problem, then goes through it step by step.
It's nice, but very costly. I've been using them at work, and one thing I noticed is, no matter how good I think it does, it often looks very poor compared to human annotators, especially if you haven't put a lot of time into crafting the best possible prompt.