Why does Anthropic not take its stated beliefs seriously, despite many people there caring deeply about these beliefs? [I was originally going to include evidence of this, but Nate Soares has done that here, so I'll skip that and just explain my mental model of this. The mental model is pretty rough and possibly wrong in some places but maybe useful to share.]
I currently think of Anthropic as two entities in a trench coat: "Anthropic the research lab" and "Anthropic the corporation". When it started the spirit was almost entirely the former (see: this informal account) but over time it has become more and more the latter as it's grown into a huge incumbent tech company.
Also, the former is the nucleus of the latter; many of its most senior leadership come from the "research lab" era. They set its direction, but delegate many aspects of day to day work to the latter.
The research lab cares deeply about some specific problems, such as alignment and safety. They are small and have limited attention, so they can only focus on this subset of concerns. Mundane day to day decisions are delegated to "Anthropic the corporation".
Furthermore, the people who make them up the two entities have very different backgrounds; e.g. the "research lab" group has a high concentration of EAs / rationalists / people who were historically worried about AI safety. But the latter group is relatively much more "normie"; I've often heard that senior employees complain about new hires not sharing the same moral philosophy and seriousness.
In particular I get the sense that the research lab does not strongly care about e.g. their relation to external organizations or the general public. They are not interested in transparency or accountability to third parties. They can afford to do this because they do not depend on the external world; anything of value (e.g. talent, resources, ...) is quickly acquired into the corporation via hiring, contract, or other agreement.
The dissonance arises because many of Anthropic's stated views on AI risk come from "Anthropic the research lab". That vocal minority has thought about it the most / cares the most. The host corporation might not agree, but is happy to agree to disagree / exist as a silent majority so long as the money keeps rolling in.
Yeah I think we're mostly on the same page, except it seems clear they're doing galaxy brain 5D chess maneuvers (and probably from the start)
My guess is that Dario is dissembling (I had to google, it means ~lying) in interviews to temper AI fears so there isn't domestic regulation that slows them down. But maybe if there was a global slowdown on the table he'd support it?
Definitely hard to judge the true motive since a desire for power, money, or the MoLG utopia all result in the same consequentialist action (making Anthropic more powerful). Maybe it's not... (read more)