I mean, GPT3 requires some 800GB of memory to run, do we all have gazillion dollars supercomputers at home?
I think, unless there's some real breaktrough in the field or in the hw acceleration, this kind of model is going to stay locked behind a pricy API for quite some time.
GPT-3.5 requires less. And neither model is considered size-optimal. It's just that with Microsoft's money, it's easier for OpenAI to move fast by throwing said money at more hardware rather than trying to optimize for size.
And yeah, I wouldn't expect them to share any model that is competitive with their current offering. But it can leak, and the copyright situation around that is very unclear at the moment.
