Hacker News new | past | comments | ask | show | jobs | submit
Sure, explain to me how frontier multimodal models, which can interpret images and video frames directly and "listen" to music via various modes of analysis, are just text models.
Sure, we're all familiar with MLLM and shared space vector encoding.

I interpreted the discussion to be about LLM capabilities, reasoning, and whether they are doing something more, perhaps through emergent characteristics. Maybe that was a misunderstanding.

But anyway, sure, change "just text prediction engines" to "just token prediction engines".

Again, they are literally just that. I don't think there's anything else going on.

"just token prediction" is doing an insane amount of heavily lifting to the point where I could similarly just wave away all of neurology, psychology, sociology, etc. as useless and simply say, "it's simply quantum physics."

Token prediction is the initial way we train these models, after which we train them on much more specific tasks. And it works well enough, as you can see. It becomes a substrate on which higher-order logic and world representation emerge.

That's an area of pretty active debate, even amongst top researchers.

I guess it's pretty obvious where I stand. I understand them reasonably well from the math up and have implemented some toys. I don't see any magic, other than how interesting it is that much of human knowledge seems to be align along compressable patterns that can be reasonably predicted to a degree.

loading story #50585103