More grist from the rumor mill:
If GPT-4 proves to be “too” good. Is it possible that it won’t be released at all but rather sold/licensed exclusively, or that it will be “rented” as an end-user based set of applications rather than via developer access?
I suspect that there’s more money to be made by selling it as a service to everyone than giving anyone exclusivity.
I’m not so sure. Microsoft gave a billion dollars for privileged access last time. I’d like to hope that this won’t be the case, but it’s not inconceivable. Would not have predicted an exclusive API last time. Fingers crossed it rolls into an API as that’s probably the best long-term revenue bet, but they have a lot of compute bills to pay, so a large structured arrangement may be hard to pass up.
Microsoft only exclusively licensed the source-code for GPT-3 AFAIK. Other entities, such as Eleuther, have demonstrated that the results are reproducible so the value of the source-code is probably already falling. I wouldn’t be surprised if Microsoft regretted that expenditure - it wouldn’t be the first time they regretted such a gamble. But I suspect they were betting that GPT-3 would remain unique and unreproducible. But it was still a gamble. When you look at the research all over the world of trillion-parameter models, we are clearly at an inflection point for large language models. I doubt Microsoft will sink another billion into something that will be outdated in literally less than 12 months