My question is whether what the AI startups are selling is fundamentally snake oil or not. Essentially they are all in some way pitching a story of applying wha

My question is whether what the AI startups are selling is fundamentally snake oil or not. Essentially they are all in some way pitching a story of applying what people see in ChatGPT in some way, but usually based on some in-house LLM, custom training or tuning etc, nearly always at vastly smaller scale. Which all begs the question, how much of what people are seeing is result of the sheer incredible model sizes in ChatGPT - 170b parameters for 3.5 and people are saying it could be 100 trillion in ChatGPT4. If network sizes of that scale are fundamental to achieving the kind of results people are expecting then all of these startups are going to fail. The most interesting discussion in the post to me is about nVidia. I am curious why, or how long, it will take for what happened with crypto to happen with Transformers - that is, why don't we have custom training hardware yet? Are we just too early? Or is it because the operation is fundamentally so memory intensive that the economics are totally different to crypto, where it's all about the computation and not about storing massive amounts of state? If it's true that the fundamentals of this are such that custom silicon can't help, nVidia looks to be a huge winner. Who knows which of these AI startups are going to win, but they are all going to buy GPUs from nVidia (or rent them from the cloud). nVidia is going to be the arms dealer in the coming war.