Hacker News new | ask | show | jobs
by smt88 1181 days ago
AIs are already incredibly expensive. There's no reason to think the current best AIs will be democratized across economic classes. M
2 comments

You can use chatgpt and many others for free. Doesn’t seem expensive at all actually.

I’m sure a super premium category will emerge, but the 80-20 rule likely applies here.

Paid users get priority access, with quotas. Free users get to are often shown a "Sorry, at capacity" type screen, that I suspect will increase through time, suggested by the fact that paid quotas are already shrinking, with ChatGPT4.

I suspect these models will usually have a "free", highly limited, option. It's the perfect bait! (worked for me)

Given how easy it is/will be to set up your own AI SaaS chatbot over time, I expect cost to consumer to trend very close to cost to provide the service. Many companies already have offerings with only a few to few dozen employees.

A few hours a year of compute seems a realistic median usage, so you can do the math. But it's not much. Certainly a premium AI that takes longer to process will cost commensurately more. But again, likely 80-20 rule applies. These more advanced AI aren't likely to be orders of magnitudes better than the cheap/common ones.

It will likely also be common to tier usage such that you only employ more expensive/computationally intensive models when needed. GPT-3.5 can already adequately cover a large number of use cases.

I don't follow. You will be able to set up a chatbot, but I can't imagine a home grown solution will ever be as useful as a private one, backed by hundreds of millions of dollars in training. They'll be increasingly, and exceedingly, incomparable, as they become more useful/profitable, as more resources are dumped in.

What path do you see that could make open models competitive?

was said about computers only 50 years ago, it's worth thinking about (tho probably too speculative at this point)
Sure, computers are, but the really important ones - the ones that can perform the biggest calculations, make the fastest trades, mine bitcoin fastest, and now run the heaviest AI models - aren't. Capital always gives advantages, many of which are sizeable. There's a reason that the new AI hotness has investments in the billions.
On other hand we are currently hitting limits or at least we see them. And we have some picture how much resources is required to train and run these models. So I think we could draw some conclusions with spending let's say 1000 and a million to run one or have one available.