Azure gets a hell of a lot more out of OpenAI than OpenAI gets out of azure. I’ll bet you GPT4 runs on nvidia hardware just as well regardless of who resells it.
I think the situation is tough because I can't imagine there aren't legal agreements in place around what OpenAI has to do to access the funding tranches and compute power, but who knows if they are in a position to force the issue, or if I'm write in my supposition to begin with. Even if I am, a protracted legal battle where they don't have access to compute resources, particularly if they can't get an injunction, might be extremely deleterious to OpenAI.
Perhaps Microsoft even knows that they will take a bath on things if they follow this, but don't want to gain a reputation of allowing this sort of thing to happen - they are big enough to take a total bath on the OpenAI side of things and it not be anything close to a fatal blow.
I was more skeptical of this being the case last night, but less so now.
> I’ll bet you GPT4 runs on nvidia hardware
Yes but they'll need to convince someone else like Amazon to give to them for free and regardless what happens next Microsoft will still have a signficant stake in OpenAI due to their previous investments.
Hypothetically he might also have very little trust in the decision making abilities of the new management and how much their future goals will align with those of Microsoft.
When I see it, it has always been “Amazon is a competitor and we don’t buy from competitors”.
I find the outputs of LLMs to be quite organic when they are given unique identities, and especially when you explore, prune or direct their responses.
ChatGPT comes across like a really boring person who memorized Wikipedia, which is just sad. Previously the Playground completions allowed using raw GPT which let me unlock some different facets, but they’ve closed that down now.
And again, I don’t really need to feed my unique thoughts, opinions, or absurd chat scenarios into a global company trying to create AGI, or have them censor and filter for me. As an AI researcher, I want the uncensored model to play with along with no data leaving my network.
The uses of LLMs for information retrieval are great (Bing has improved alot) but the much more interesting cases for me are how they are able to parse nuance, tone, and subtext - imagine a computer that can understand feelings and respond in kind. Empathetic commuting, and it’s already here on my PC unplugged from the Internet.
13B and 7B models run easily and much faster.
Which one side or the other would declare terminated for nonperformance by the other side, perhaps while suing for breach.
> and one way or another, everyone would get a stay on everyone else
If by a stay you mean an injunction preventing a change in the arrangements, it seems unlikely that "everyone would get a stay on everyone". Likelihood of success on the merits and harm that is not possible to remediate via damages that would occur if the injunction wasn't placed are key factors for injunctions, and that's far from certain to work in any direction, and even less likely to work in both directions.
> and nothing would happen for years except court cases
Business goes on during court cases, it is very rare that everything is frozen.
Ive found that benchmarks are great as a hygiene test, but pointless when you need to get work done.