They acquire and somewhat understand it but it doesn’t get saved. It only lives in the context window.
The things the big ones can do now are amazing when you have tasks that can be iterated upon until completion. And then there’s Deepseek V4 flash, a much smaller model with a still huge context window that costs an order of magnitude (nearly 2) less than the American frontier models and managed to do some spectacular things for me with essentially no input from me beyond the original prompt. It took hours, but for me to learn the tools and everything would’ve taken days or weeks.
Still not intelligent the way humans are. Next session starts, it’s all wiped clean until it reads the last session’s notes. But in terms of understanding information and reasoning about it, it’s going to be better than a human that isn’t a domain expert.
But going devils advocate you could say the training is doing both, in a sense. And if it’s allowed to use the web it can also acquire and utilize (hesitant to say understand because it’s not the same) new information.
I personally would also expect to save and utilize all information permanently and to have own motivations but that is more sentience and dynamic memory.
/s
My point being that an LLM will never be more “intelligent” than the lump of text that it is.
It’s not intelligent, like an encyclopaedia is not intelligent.
Oh I do agree with you, but it sadly is, by definition, artificial intelligence.
But LLMs are incapable of acquiring or understanding information.
They acquire and somewhat understand it but it doesn’t get saved. It only lives in the context window.
The things the big ones can do now are amazing when you have tasks that can be iterated upon until completion. And then there’s Deepseek V4 flash, a much smaller model with a still huge context window that costs an order of magnitude (nearly 2) less than the American frontier models and managed to do some spectacular things for me with essentially no input from me beyond the original prompt. It took hours, but for me to learn the tools and everything would’ve taken days or weeks.
Still not intelligent the way humans are. Next session starts, it’s all wiped clean until it reads the last session’s notes. But in terms of understanding information and reasoning about it, it’s going to be better than a human that isn’t a domain expert.
I agree, it also does not meet my expectations.
But going devils advocate you could say the training is doing both, in a sense. And if it’s allowed to use the web it can also acquire and utilize (hesitant to say understand because it’s not the same) new information.
I personally would also expect to save and utilize all information permanently and to have own motivations but that is more sentience and dynamic memory.
I trained my book by putting words in it!