>ChatGPT has a model of all the text information on the internet, but it remains to be seen what the hard limits of this model are.
Before training is complete and loss is maxed, there will be limits on what the "learned so far" model can do that say absolutely nothing about the limits of a perfect(or very close to it) model.
It really looks like anything will converge with enough compute. I don't think architecture is particularly important except as "how much compute will this one take?" question.
https://nonint.com/2023/06/10/the-it-in-ai-models-is-the-dat...
>Does this model let it do logic or predict the future well, or will no amount of training give it those abilities?
There's nothing special about logic. Basically, any sequence is fair game. It literally does not matter to the machine.
Boolformer: Symbolic Regression of Logic Functions with Transformers(https://arxiv.org/abs/2309.12207)
That said, GPT-4 can already do logic. It's not perfect but if perfect logic were a requirement then humans cannot do logic either.
>Simply being good in one task doesn't imply a general ability to do everything, or even most of everything.
It's not one task. It's one modality (text) that a plethora of tasks could be learned in.
Coding and playing chess did not suddenly become a single task just because we found the common ground that allows a machine to learn both.
The text, image, video and audio data we could feed a transformer will cover anything we care about.