The Zitronites will say that the data is fraudulent, and OpenAI must have classified some of their inference as marketing, or R&D, or some other wacky theory of the week. But the actual data does not show that. It is made up.
If you want to cherry-pick the worst parts from the leak and disbelieve the more positive ones, it feels like you're not in a great place epistemically...
They don't. They show that OpenAI need to people to draw that conclusion, because of course they do.
> The Zitronites will say that the data is fraudulent, and OpenAI must have classified some of their inference as marketing, or R&D, or some other wacky theory of the week. But the actual data does not show that.
It doesn't?
It shows a marketing budget so absolutely mahoosive that it's almost completely implausible, which does make you think — have a percentage of marketing-driven free plan tokens been hidden in there? If not, what the hell is in there? Because it's an insane figure for a company that has benefited from a level of word of mouth that makes "ChatGPT" broadly synonymous with "AI".
Fraudulent is a big claim, of course. I didn't say it.
Which, look, could be true! But it's currently speculation only.
Given that I don’t know how you(royal you, not you specifically) can constrain an analysis of inference profitability to just the literal running of inference and not include training costs. That’s where it all breaks down for profitability.
I think it’s reasonable to draw the conclusion that they are folding inference subsidies (for both paying and non-paying users) under this category. Frankly I think occam’s razor demands it because where else would all that money have gone? Fancy trips for enterprise clients?
