AI出口管制或重塑互联网格局

Ilia, uh, last week, Entropic removed access to Fable 5 after the US government issued an export ban. Uh, they cited safety concerns about possible jailbreaks. Is this the last models that are going to be available to the general public? Are there no more models that are coming to me in my claude because the government is saying no?

So, I mean, there's a lot there's a lot here to unpack, right? First, Entropic released this model and kind of neutered it on on a number of dimensions and everybody was outraged about that because effectively if you asked it about cryptography, if you asked it about some bioscience, uh it would you know fall back to the kind of opus and initially even not very transparently right so

so I think like we getting to a stage where there is different parties who affect effectively deciding how you can use models, right? And so realistically this goes back against you know some of the fundamental like policing the the thoughts not policing the actions. M

and so I think like you know I I actually predicted that government will start stepping in because like they they they kind of feel they need to even though it's violating some of the fundamental like freedom of speech uh uh kind of basis. So, so yeah, I think the the we we'll start seeing a very strange things and and and already this is unrolling now where every like so there's a few things that are that are interesting here.

One is entropic effectively now dictates how you can use these models, right? They also changed their privacy policy for the fable, right? Which says, hey, even though even even though you're paying us to not keep your data, we'll keep your data for safety purposes, right? So that was kind of one one side of this. Then the you know if you read the reasoning you know from David Sax and kind of some of the other folks from the government side they're like hey they said it's super dangerous.

We have like a reasonable ex like a way to use it for effectively finding exploits like they did they didn't patch it. Now the fact that they also then enacted uh this expert control is a very kind of obviously strange reaction. It's also now puts every single other government in the world on alert that at any time not just AI effectively every single internet service can be shut down. Like if you think of it right now like if you take your phone out of their pocket it runs on a lot of services that are provided by us and if government enacts export control on any of them it's not it's not just it can be shut down like for for your country it's also like you can shut down everything like you can shut down everything because it's also an internet hard like if you as US citizen go abroad like you Did you give your phone to somebody else?

Like how are they going to detect that right? So so it's a very strange precedent honestly for effectively internet because it kind of again violates the whole premise that internet is borous. So, so I think that that was a very strange action to like use that methodology. Um, even even if they were kind of trying to like hey you guys should fix this this uh problem right that you said yourself is is you know hey security and safety is a big deal.

So now going back to the models I think the Pandora box is open. I think people who are I think like obviously you know misses came out but it was only like available to a few parties but we in crypto saw hack after hack after hack it wasn't because people were using misses it was because the models are there a lot of them are open source now or open weights you can put them in the right harness where you effectly inspect you know every line for invariance for all the properties and you can find vulnerabilities like this this is happening right know the obviously like I mean from kind of experiments and from what I've heard of people fable is like a good step forward on capabilities but I I like I really believe you know it's like a whatever uh 4-minute mile right as soon as one party can do it right everybody else kind of like pulls behind it and a lot of it is just like how much compute you can put behind it to really keep uh like reinforce learning and and updates.

So, so I think like to your point is this the last model you'll be able to access? I think we will have this level of model from open source one way or another within next 3 to 6 months. Right? We just saw JLM52 came out at least like from you know benchmarks and corsery usage it looks kind of like 4.7. So like one generation down from Opus. still still kind of couple months behind but like this is already something that you know two months two months ago was actually like hey this is a great product so I think like Pandora's box is open the this intelligence at scale is coming I think we we as crypto actually always thought in this way where we assume there is somebody in very intelligent on the other side who's trying to break in the rest of the world hasn't built like that.

Mhm.

And so I think this mindset of effectively like hey we have like a government adversary who's equipped with you know potentially quantum computers trying to break us we need to kind of apply that across a society right I've I've talked about this like I call it dodoing the government right like right now government is not designed to actually withstand somebody who has you know intelligence and an immense amount of patience right and it's a fixable by effectively having AI on the other side, right, to also respond to this and like some some of the services are doing like FDA for example getting so much filings and the filings are so detailed that they now using AI to analyze the findings before any like expert come in because like it's just like the amount of cont content and context is exploding on both sides.

So we just need to keep like ramping up and equipping government and services and kind of with better better tools and start using some of the more fundamental solutions like for code we need for modification right there's no way around it like you know otherwise you can and mouse right the better model will always find issues that the previous model didn't find if you spend more compute you'll find more things if you if you spend it more you know intelligently in a proper harness you'll find more things right so you need a more fundamental like hey this is actually mathematically proven so so I think like models are coming I think the the expert control is honestly was more of a like a really bad precedent that now I'm going to activate actually even more AI development outside of US now like I mean I've heard at least multiple countries where people like hey we should have our own sovereign lab right um that you know is able to do this and they will allocate capital to do this compute etc which is by the way good for crypto for you know we have projects doing decentralized training we have projects trying to kind of facilitate that I think those as well will be

uh like before they were kind of like hey we have open source models why do we need this now it's like well what if we don't have now we need actually something that is decentralized that is

able to extend it and then at the same time we need to assume everybody's equipped with like all the knowledge of the world and advanced intelligence. We should we should be able to build the society that kind of has that in in not police the thoughts like not police what you're thinking and what you're asking but what actions you you know you're doing.

So why did the government actually decide to put an export control on it? Like what does what how does that benefit them? Because do do they just want access to it before the rest of the world? kind of like how Mythos was given to a select group of insiders, call them elites. So the government needs to have the tools before everyone else because certainly the government knows that the all the other models in the world are also go

原文超出正文长度上限，此处截断——上游还有内容，完整版见上方「原文 ↗」。