ivan

An abliterated top-tier open weights model today can tell you how to build a nuclear weapon. The challenge to building a nuclear weapon isn't knowledge. It's not even really resources. Iran has knowledge and resources. They can't build one. Why? A lot of other entities with knowledge and resources have a vested interest in making sure Iran doesn't have a nuclear weapon, and they get in their way. If you wanted to build a nuclear weapon in your basement, to vibecode one with abliterated GLM 5.2, you would find that your efforts are stymied at every turn. Even with step by step instructions, even if it was good enough at day-trading (it's not) to make you a tidy profit to finance your operations, you would find you would quickly attract the attention of law enforcement and the intelligence community. It turns out they really don't want you to have a nuclear weapon! If GLM-5.2 were good enough to tell you how to build a superplague (it's not), you would quickly learn a similar lesson in biology. It turns out the challenge to building a superplague isnt really knowledge. Just as there are thousands of nuclear physicists with the theoretical knowledge to build nuclear weapons, there are thousands of biologists who have the theoretical knowledge required to make a superplague. But if they wanted to do it, or if you wanted to, with your abliterated GLM-5.2-Bio-Edition, you would find that you would be stymied at every turn. First you would design your virus particle or virus-like-particle. You could do this in the computer. But now you need to make it. Traditionally you would order it from a company like Thermo-Fischer scientific or IDT. They'll synthesize a DNA strand for you and send it back so you can inject it into a cell of your choice to get it to produce the protien you want. For a fee, they'll even get an organism to express it for you and send *that* to you. But, it turns out they're not stupid. And if you engineer a highly virulent strand of ebola and send it to them, before they help you make that, they're gonna say "hey wait a second, this kind of looks like a highly virulent strand of ebola!" They will get upset, law enforcement will be contacted, the cops will show up, and so on and so on. So maybe you'll do it yourself. You have GLM 5.2-Bio-Edition and GLM 5.2-Trading-Edition so you make a bunch of money to order the expensive, highly specialized equipment you need to do this complex synthesis at home in your basement. So you go online with your roughly $300k and start trying to order a Mermade DNA synthesizer and assorted sundries and suddenly they're asking all these strange questions like "what's your .edu email" or "what institution are you a part of?" and you don't really have answers to those questions because you're just some guy trying to make a superplague in your basement. And actually it turns out the Government, intelligence community, and Bio Research technologies company *also* have GLM-5.2-Supply-Risk Edition and so they notice you trying to order all the parts to make a superplague in your basement. They get very upset and the cops get called and so on and so on. And *actually* it turned out *everyone* had access to GLM-5.2 Trading edition, so it wasn't so easy to make a killing on the market to finance your operation in the first place because your agent wasn't placing trades against human rubes, but against a whole hoard of *other* agents every bit as intelligent as itself, so the net result is that the market became slightly more efficient, but no one was really able to lock in asymmetrical gains. Unless, however, *one* person had access to the GLM-5.2-ASI-Edition. Because if that person had access, they would be trading against rubes-by-comparison, and they would make a killing in the market to finance whatever operations they please. And their ASI would be smarter than anything anyone else had, so it could engineer a VLP that thermo-fischer wouldn't be able to detect and reject on site. Or it would be able to deftly bypass supply chain safeguards, hacking into MIT or something (who has no ASI defender) to get you a .edu email and the appearance of institutional backing. See, in a world where everyone has ASI, it's as though no one does. It's just normal society, where everyone's wishes and capabilities are mediated by group dynamics, everyone kind of has to stay within the overton window, and deviations are policed largely by the community, and then by the state when need be. But in a world where only anthropic has ASI, or only openAI, or only the US government, then the *first* bad idea they have is immediately implemented with no resistance, with no hope of stopping it. And on a long enough timeline, its only a matter of when, not if, they will have a bad idea. Pluralism has served America well for a long time, and when it comes to ASI, I'm a pluralist too.