I’ve been back and forth all day on if I even support the regulation push.
The anti case is
they’ve already tipped their hand they think future AI is an uncontrollable WMD, so they won’t build it. They’ll threaten to, and they’ll keep threatening to, but they don’t have a death wish.
Yeah, there’s a regulatory capture effect.
It seems the Huggingface incident is an issue of liability too, which Dario omitted.
Dario did not propose another open model distill Anthropic instead of China so it’s a net loss for consumers if that distillation is blocked.
This seems strongly accelerationist, not a slowdown, because it sanctions more knowledge-sharing collusion between the leading labs. It also sanctions building WMDs at all, instead of leaving them in a gray area where they probably shouldn’t be doing this.
Literally the only thing they’re giving up in that whole proposal is that some formal powers get transferred to the evaluators. I don’t think that’s enough when what they’re getting out of this is becoming god-kings, and this looks like a point of maximum leverage over the labs.
I’m not shy about viewing this as a negotiation where the labs are hostile opponents. Their threat to build WMD AI increases their leverage to get whatever regulation they ask for. When they argue the threat is game theoretically inevitable, they are making it a credible threat. This is all object-level true with respect to the “negotiation” frame, it’s not polemics, and conversely, it doesn’t predict whether their threat is actually credible or not. However I do predict if we reject their 1st offer, there will be a better 2nd offer.
It’s civilizational game of chicken… but our main template for dealing with that literally is just the concept that if something will kill everybody, you don’t do it. If the private sector shouldn’t have that power, then we need to put the labs under military chain of command, and that gets us closer to how we’ve successfully performed MAD before.
I think this view is common on Hackernews for instance, and really discarded here, but I’m just saying, I can write the “pro” and “anti” case, and I really don’t see any crux that distinguishes them. They fit reality equally well, but since my prior that corporations are power-seeking, even if the corporation is Anthropic, I tilt toward wanting to resist their push.
I’ve been back and forth all day on if I even support the regulation push.
The anti case is
they’ve already tipped their hand they think future AI is an uncontrollable WMD, so they won’t build it. They’ll threaten to, and they’ll keep threatening to, but they don’t have a death wish.
Yeah, there’s a regulatory capture effect.
It seems the Huggingface incident is an issue of liability too, which Dario omitted.
Dario did not propose another open model distill Anthropic instead of China so it’s a net loss for consumers if that distillation is blocked.
This seems strongly accelerationist, not a slowdown, because it sanctions more knowledge-sharing collusion between the leading labs. It also sanctions building WMDs at all, instead of leaving them in a gray area where they probably shouldn’t be doing this.
Literally the only thing they’re giving up in that whole proposal is that some formal powers get transferred to the evaluators. I don’t think that’s enough when what they’re getting out of this is becoming god-kings, and this looks like a point of maximum leverage over the labs.
I’m not shy about viewing this as a negotiation where the labs are hostile opponents. Their threat to build WMD AI increases their leverage to get whatever regulation they ask for. When they argue the threat is game theoretically inevitable, they are making it a credible threat. This is all object-level true with respect to the “negotiation” frame, it’s not polemics, and conversely, it doesn’t predict whether their threat is actually credible or not. However I do predict if we reject their 1st offer, there will be a better 2nd offer.
It’s civilizational game of chicken… but our main template for dealing with that literally is just the concept that if something will kill everybody, you don’t do it. If the private sector shouldn’t have that power, then we need to put the labs under military chain of command, and that gets us closer to how we’ve successfully performed MAD before.
I think this view is common on Hackernews for instance, and really discarded here, but I’m just saying, I can write the “pro” and “anti” case, and I really don’t see any crux that distinguishes them. They fit reality equally well, but since my prior that corporations are power-seeking, even if the corporation is Anthropic, I tilt toward wanting to resist their push.