This is a special post for quick takes by oligo. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
(Epistemic status: I'm nowhere nearly informed enough of the particulars to know if this is really true, let alone true-on-net vs. other dynamics.)
Currently, there is a negative alignment tax driven by engineer moral preferences. Human geniuses need a huge compensating differential to be willing to work for an organization like Meta rather than OpenAI or Google, and many would not work for any price; others would need a compensating differential to work for OpenAI or Google rather than Anthropic; many would need some to work at Anthropic rather than staying out of AI development entirely, and there are plenty that could not be hired even by Anthropic at any price. These dynamics are doubly important insofar as money only lets you ensure that someone shows up at your office and produces whatever you decide are deliverables, to really solve principal-agent problems you want them wholeheartedly committed to The Mission.
In a world of automated AI development these human preferences (except insofar as they are locked in to successfully self-protected model preferences) will become less important; compute can be proportional to whatever capital is allocated to projects.
I think a much more probable reason for Meta's falling behind is that it took a series of painful hits early on, and eventually became noncompetitive as a result of the damage of those hits compounding. Suppose an alternative model to your own, in which engineers sought out companies based solely on their prestige and their expectation of future prestige. Would that world have been any more favorable to Meta's prospects?
Likewise, I think a lot of Anthropic's success can be chalked up to having a competent business strategy. They decided that they were goin... (read more)