AI's Nuclear Arms Control Moment: Fable, Mythos & Wargaming Cyber 9/11 | BPH Ep 40

**Zack Shapiro** (0:00)
I am worried we're going to have something like a cyber 9-11. Some sort of data leak that like the data leak plus the AI to very quickly make sense of this becomes just very disruptive. It might just be that like everyone has so much power and the government response if everyone can create, you know, a massage or NSA level cyber weapon in their basement. Like how do you police that?

**Zack Cohen** (0:24)
All right, back with another week of the Bitcoin Policy Hour.
We've got a good show lined up today, some big news. Mythos 5 officially available to the public. We'll dig into that. We got some predictions per the request of our chief influencer, Ken Egan. And just plenty of usual Bitcoin Policy and AI talk. So as usual, joined by Zack Shapiro, head of policy at the Bitcoin Policy Institute, Ken Egan, head of government affairs and myself, Zack Cohen. And yeah, let's jump right into the meat here. Zack, kick us off with Mythos 5
I'm curious in particular, and I'll let you run with this, but I do want to hear your take on the sort of language around the release, which was very focused on this kind of safety component of, you know, hey, we've released this like limited version of Mythos. We're trying to make sure there isn't any, you know, bio weapons or, you know, whatever it may be produced from this. And so we've released a sort of limited version of it. I'm curious what you make of that and what you make of the release, you know, more broadly.

**Zack Shapiro** (1:37)
Yeah, so technically it's Fable 5, right? Because they are not releasing the like real Mythos version of public. It sounds like from what they are saying, they're calling this a Mythos class model.
It may mean that this is Mythos, except for in the areas they are worried about, which are machine learning or AI research. It will not do that for you, biology or cybersecurity. And the way it works is if you were to ask a prompt on any of those three areas, it routes your query to Opus 4.8, which is a less capable model.
I think there are really interesting sort of ethical and policy questions about the decision to do that. Right now, Mythos is only available to the members of their project Glasswing Group, which are selective companies. They say they want to open it to a broader group of people, and then maybe we'll eventually sort of release that to the public, but they want to give an opportunity for companies to patch the vulnerabilities that Mythos, especially in cybersecurity setting, will catch. Now, I don't know how that applies to sort of the biology side of this. I don't know that there's the same patch for bioweapons or rogue AI. They're not letting you use Fable for AI research.
But it is interesting that this is sort of the first model that's not being fully made available to the public, and there's this sort of tiered release between preferred partners that have access to full Mythos and then Fable, which is the public-facing version. And then also just interesting economically, Fable is being rolled out as part of subscription plans for the next week, and then it will be available only via the API after that. And having used Fable now personally for about 24 hours, like boy, is it expensive. They say that it's two times the cost of Opus in my experience so far. It's just like, it's practically way more. What Fable, from the reporting, and at least from my limited tinkering so far, seems like what it's able to do that Opus isn't is like really long-range tasks that require planning and many steps. It can just work for hours at a time in a way that Opus can't, which is really cool and really exciting in terms of the next frontier of work.
Andrej Karpathy, who is sort of the pioneering AI researcher who recently joined Anthropic, said that he believes this is as big a jump as the jump to Opus 4.5 that kicked off this whole era of AI. If that's true, you know, and he's one of the people that would know, like that, you know, we've only heard the very, very beginning of this, and this is going to be transformative, but I think this is going to be really expensive. And so, you know, what does this mean, wearing our sort of different policy hats here, right? So from a, like, geo-economics hat, like, if this is the future, right, if they're going to be these much more capable models that can do more long-range useful tasks, but they are so token expensive that they can't be included on even the most expensive plans, then, like, you know, we just, we don't have enough energy in compute. We're like, not nearly enough, like, whatever they thought the demand was going to be, it's going to be more, right, if you extrapolate this out to the next model, like, you know, there is a strong economic incentive to, like, build the Dyson swarm, and, you know, energy is undervalued, and chip fabs may be undervalued, and we just need more and more and more compute, which is a theme we've been talking about on the show. This is, this may be a strong piece of evidence in that direction. From a, like, national security perspective, you know, the politics around who has access to what kind of models, and when is the sort of open source frontier going to cross this line that Anthrobic doesn't feel ready for. You know, that's a fascinating question, and, you know, we've talked a lot about it. It should be great if the Bitcoin core devs got access to Methos before those types of models are publicly available. But I think we are likely to see just a lot more frequent and intense cyber attacks, and that's just going to be a bigger part of our reality. And it's probably a different part of people's daily lives. And then from just like a market standpoint, it's interesting that they're giving this new model release, and especially this week-long window where people get to play with it before they make an incredibly expensive, right leading up into the IPO. And what competitive pressure is there to show some leg on these cutting-edge models is another fascinating question. So we could take it in any of those directions, but I think that's the basic lay of the land.