>
I take this as a contrarian signal that Meta has hit serious roadblocks improving their AI despite massive data advantages
Just a thought:
Assuming that Meta's AI is actually good. Could it rather be that having access to a massive amount of data does not bring that much of a business value (in this case particularly for training AIs)?
Evidence for my hypothesis: if you want to gain a deep knowledge about some complicated specific scientific topic, you typically don't want to read a lot of shallow texts tangentially related to this topic, but the few breakthrough papers and books of the smartest mind who moved the state of art in the respective area. Or some of the few survey monographs of also highly smart people who work in the respective area who have a vast overview about how these deep research breakthroughs fit into the grander scheme of things.
There’s been a lot of research on the necessity of singular geniuses. The general consensus (from studies on Nobel Prizes and simultaneous patent rates) is that advances tend to be moved by the research community as a whole.
You can get that technical or scientific context for a lot less than $250 million per head.
250 million probably gets you a couple of research labs and incremental field advances per year?
Assuming a lab has 20 phds/postdocs and a few professors, call it 25 people per lab, and you're compute / equipment heavy, getting you up to an average of 1M per person per year in total fully loaded costs (including facilities overhead and GPUs and conferences and whatnot), then you're looking at 200 PhD researchers. Assuming that each PhD makes one contribution per 4 years, then that's 50 advances in the field per year from your lab. if only 10% are notable, that's 5 things you've gotten that people are going to get excited about in the field. You need 2% of these contributions to be groundbreaking to get a single major contribution per year.
So 250M for a single person is a lot, but if that person is really really good, then that may be only expensive and not insane.
Just a thought:
Assuming that Meta's AI is actually good. Could it rather be that having access to a massive amount of data does not bring that much of a business value (in this case particularly for training AIs)?
Evidence for my hypothesis: if you want to gain a deep knowledge about some complicated specific scientific topic, you typically don't want to read a lot of shallow texts tangentially related to this topic, but the few breakthrough papers and books of the smartest mind who moved the state of art in the respective area. Or some of the few survey monographs of also highly smart people who work in the respective area who have a vast overview about how these deep research breakthroughs fit into the grander scheme of things.