☭
Technical question about the evil LLMs
Apparently, the ChatGPT and such have some capability of processing images, either to generate slop images or to describe existing images? How does that work?
My understanding of an LLM was that it is trained on a lot of text and does fancy maths on tokens that describe text and then samples a statistical distribution to produce more text.
So how does interacting with images work? Does it call a different program for that? One of those neural networks that specialise in handling images?
re: Technical question about the evil LLMs
@vaporeon_ ChatGPT is multiple diffurent machine learning systems in a trenchcoat and part of what makes it tick is that it’s designed to forward requests to one of many diffurent subsystems to get around the things LLMs are bad at doing
re: Technical question about the evil LLMs
@aescling So they do forward it to a neural network that handles images? The language model itself doesn't do it? Makes sense.