The US Has About An 18 Month AI Lead On China Says Allen

It's based on the explosion in revenue that they're generating. Anthropic has disclosed some of their run rate numbers, and they are jaw-dropping. End of last year, they were at a $9 billion run rate, which means if you took their December monthly revenue and turned it into a 12-month annual, it would have been $9 billion. Now it's reportedly north of $47 billion. That is, the company's revenue in five months. I think that bodes well for the future of this company. But as you started out by saying, there's so many different things going on, there's still so much uncertainty in this. So I can't imagine why someone would want to open an AI research company right now. There's a lot of unanswered questions, and I'm happy to try and listen.

Well, let's start with sort of what's reasonable in what the president is saying. He's talking about competitiveness mostly with China, which actually has AI models. Some of them are pretty good. And depending on how you sort of measure how good the AI models are, you can say roughly that the United States has something like a 12-month to 18-month lead on China in terms of where their models are today and where our models are today, and sort of where their models are compared to where we used to be. Now, the draft executive order, which was leaked because all these executives had a copy — they had been told in advance they were in the air, as you pointed out — and that was talking about a 90-day pre-release review period. So think about if you're President Trump, okay, 90 days and we only have an 18-month lead. So we're going to consume, you know, a month and a half of that just for risk reduction. Is that a reasonable trade-off? Well, here's the thing. It wasn't even mandatory. As I said, we have the text of the executive order, and it was voluntary only. So if there really was some kind of concern, right, that this was going to take too long, you didn't even have to ask permission to not go through this sort of framework. And so you have to ask yourself, why is the government opposed to building the kind of capacity that we need to ensure that these models are safe, that they're going to work as intended, and they're not going to get in the hands of the wrong people? Let's not forget sort of how we got here. The Trump administration took office saying they wanted to dismantle all kinds of AI regulations, but the risks were not that significant, right? Then they met Claude Mythos. And according to some of the reporting on this, half of all the cybersecurity vulnerabilities that have been discovered ever were discovered by Claude Mythos just in the past handful of months. I mean, this thing is a chainsaw to most organizations' cybersecurity. Now, imagine if Anthropic had taken a different approach to releasing that model, just throwing it out into the wild and saying anybody who wants to use it can use it. I think a lot of banks would have been opposed to that kind of outcome. A lot of health care institutions would have been opposed to that kind of outcome. And so Anthropic was kind enough to go off and do this. But how do we know that future executives of future companies are going to have that approach? And how do they even go to the government for help? Right. One of the things we need the government to do is to invest, to get good at these types of skill sets, so that there are people who companies can call to check their homework and understand whether or not they've done a good job securing these models.