You wrote about Anthropic on both Tuesday and Wednesday, and this Mythos model, reading about it was pretty unsettling earlier in the week. But, what do you think of what's happening here? Why is it unsettling? We just discussed it on Sharp Tech a week ago.
Well, exactly. What we discussed and the threats to security that we discussed now appear to be imminent, albeit private for the time being. So, I guess we can take solace in that.
Yo, maybe that's why we're already at stage five. No, I think a very timely discussion we had last week about the reality that, I mean, this actually ties into a long-running discussion that we've had, particularly this year, about the uniqueness of programming and code and its suitability for large language models. And the fact that, you know, how do you program? You put a bunch of words and symbols together in sort of arcane ways that it can be difficult for a lot of humans to sort of do, but computers are quite good at it. Particularly large language models can handle large amounts of language, which at the end of the day all of software is just massive amounts of language. And again, that language may not be very understandable to you or I, but it is very predictable and understandable. And to that end, given what we've talked about, this should not be a surprise to us or to our listeners. So yeah, here we are. Now, how here are we? This is, it's hard to say, right? Because this is Anthropic. These are the same people who going back, it's funny, people are like, 'Oh, OpenAI did this, too.' No, the Anthropic people did this at OpenAI.
Where they're like, you know, like, 'Why is OpenAI not open?' Because GPT-2 posed too many dangers to the world, so it's like, 'Yeah, we're not going to be open anymore.' And it just so happens that not being open is actually good for business. And the dangers at that point were what? Basically, potentially flooding the world with misinformation? Because I mean, that was purely text generation at that point, right?
Very poor text generation, but yes. You know, it's funny, we had zero problems with misinformation until GPT-2 came along, and then the world has... Still trying to recover from that 2022 release. No, I think this is, this is like 2019, actually. So, this has been a thing for a long time. And it's also very good for business. And what you think back to the, so let me just back up. No one get mad at me until we finish this whole segment, okay? Because we're going to cover lots of different areas. I already see our first email, someone that's very mad at me. So, Mr. Relax, we're going to get there, okay? You go back to 2019. I think it was 2019 when GPT-2 came out. And there's a, this is dangerous. And there's also a, maybe it's not the best thing in the world if we're on the leading edge to give everyone our weights. Because then they, like, they can just run the model themselves, right? The equivalent here, and by the way, I think another area where we were very early, what was one of the points that we brought up with DeepSeek a year and a half ago? Well, DeepSeek looks like it's kind of distilled from leading US models. And everyone just sort of takes it as a given or they hold it up as an excuse when these labs are complaining about distillation, which we've talked about, this idea that you basically query the API a gazillion times for all sorts of things, and you get your own data from the model to train your own model. Like, how do you get these, oh, OpenAI is or open source is only 6 months behind. Well, because it's about 6 months that it takes to query these models a gazillion times.
Successfully distill them. Well, can I ask one question on that? Because this came up on Sharp China, and it's come up a couple different times on Sharp China, and I don't have a good answer. And as a tech podcaster, I feel like I'm failing Bill Bishop in the course of these conversations. Is there a way...
A tough gig for you. You have to be the dumb normie on this podcast that you have to be the brilliant tech understander on... Wearing many hats.
But, is there a way to reliably prevent distillation in the future? Because distilling a model that's as powerful as Mythos seems like it could be a problem going forward.
Yeah, well, I mean, a distilled model is never going to be quite as good as the regular one. And it's much more jagged, there's much more holes. It's much more, much less comprehensive. So, just in general it's a bit where they're always going to be behind to a certain extent. But, that doesn't change the fact that if they're more than good enough, and these leading edge models are very expensive, it's a great alternative if you, you know, want something else. So, maybe just to go back to this story, there's a very good business reason for not making this available, just like there's a good reason for not making open weights available. This is like, it's the same story. And if you think about these companies wanting to have market power / pricing power in the long run, well, making sure there's not nearly as good models to the extent you can is a way to do that. And the challenge is if you have a self-serve walk-up API that anyone can use, yeah, it's pretty hard to stop. Like, I mean, we've all pirated music. It's not like the same story other than to say like trying to stop people doing stuff on the internet when there's open APIs and things that you can access is a tough game. Effectively impossible. You can make it harder, but not impossible. Right. And you know, it's one of those things you often find it after it's happened. Like, you go through your logs and say, 'Wow, we're getting hit on this endpoint from this set of IP addresses a gazillion times, which have been routed through a gazillion points, and they're, you know, it's not like they're coming from like the Forbidden City IP range and like accessing the model. Like, they're spinning up cloud servers on DigitalOcean or on AWS or whatever, and like doing this. Probably AWS, probably be too expensive. But, like, there's a, it's not, yeah, it's not easy. But, basically, just like, it is a rough analogy, like policing chips is a lot harder than like policing uranium, for example, right?
Which you can see from satellites, and it's much easier to track all over the world.
Like if someone like breaks into like OpenAI and exfiltrates the weights, you know, very clear thievery going on. If you're going on and just sort of asking a bunch of questions at a very high rate of speed, which computers are very good at, it's a lot tougher to stop. Yeah. So, you have this sort of business issue. You also have Anthropic can barely stay online right now. You know, the people, it is this massive upsurge in revenue, in users. They're doing this weird rationing thing like these 5-hour blocks, which aren't really 5 hours, because like the 5 hours is shorter than 5 hours during certain times of day, and then it's longer at other times. Like, and then people are complaining about, 'Oh, they're purposely reducing the model quality.' There's definitely like they're serving distilled models themselves, and you can distill much more effectively if it's your model, and you have like full access to it instead of just using the API. And they're quantizing, but also they're doing lots of things like trying to leverage cache and doing sort of like trying to batch a bunch of stuff together. And all these optimizations sort of layer on each other to really diminish the experience to the extent that it's very hard to... So, you have this new model comes out that is extremely computationally expensive and intense. And you just look at the API pricing, which is like 5x what Opus is. And by the way, Opus is significantly more expensive than say GPT-4, which is an even smaller model. And so, it's like if we could limit it not to the hoi polloi, but to people who will actually pay us real money, also a good sort of business justification, right? And so all of this is making me feel much better as we read about a potentially existentially dangerous model here. There are lots of rational reasons to approach it this way.
