Transcription
There's been a big development in the AI race recently. Not, not the race between Google and Anthropic and OpenAI, etc. But what I think of is honestly the real AI race between the US and China. And you know, until recently, there's been this consensus that the US is quite comfortably ahead. Six to twelve months, people say. But lately, that's started to change a little bit.
This model, this model, this Chinese model GLM 5.2. Yes, I know the names. I know they're maddening. But GLM 5.2 came out. It caused quite a stir, but at the same time, it wasn't quite leading the the benchmarks, the tests. So, you know, people didn't worry too much. And then this one came out, Kimi K3. Suddenly at the top of some leaderboards, and that is being taken very seriously.
This is David Sax, former White House AI adviser, talking about a similar benchmark. "This is concerning. This is how you lose the AI race." And it's worth paying attention to, in particular, what he's saying because recently the White House and the AI companies have been restricting who can access the most powerful models. Sax is calling for that to end. He's calling for a return to permissionless innovation.
So, how does Kimi K3 really stack up? Because, you know, the funny thing with this stuff is that we don't actually really have a good measure of artificial intelligence. You, you think we would, right? Because it's just a machine. But actually, in terms of measurement, it's more like real intelligence. There are many different types and many different ways of measuring it. And so, in that case, then I think the best thing to do isn't just to take one measurement like that one I showed you before, but a whole basket of them.
In this case, from these guys, Artificial Analysis. And so, look, this is where Kimi was until recently. And this is where it's shifted to now. And I think this gives you a really clear picture of what is going on. The Chinese models aren't first, but they're not far behind. And what that means is the US isn't losing the AI race, but it is losing control of it. And that poses some enormous questions that honestly could shape so much in our world.
I'll give you three, three major questions I think this poses. First, about the big AI companies. Cuz, you know, there's a funny thing in AI, as in many other things, the Chinese ones are way cheaper. Look, this is a chart showing the cost for one benchmark task, one test. This is Kimi over here. This is the best ChatGPT model, and these are Anthropic's Claude models. You can see Kimi is a lot cheaper, right? It makes you wonder, well, if you can basically get something similar for much cheaper, why wouldn't you use that? And what would that do to the valuations of these companies? What would that do to the US stock market that rests so much on the valuations of these companies? Yeah, I mean, like I say, many, many big questions.
This is the set of questions, number two. What does this mean for cybersecurity? Because there's another funny thing about the Chinese models. They're what's called open weight, which means that anyone can download them, download and run them, potentially without safeguards. And look, in practice, this would be incredibly impossible to download and run Kimi unless you had like a data center in your garage. But even so, it is possible and could have big implications.
This is a chart of critical and high-severity vulnerabilities found since 2022. And you can see, look, for for ages, the number of vulnerabilities found ticked along about roughly the same rate until AI arrived. And until, in particular, Mythos arrived. You remember Mythos? That was the model that caused that big cybersecurity scare. Well, now, through Chinese AI, everyone can get access to that if they've got enough compute. Cybercriminals, but also banks and cybersecurity institutes that aren't part of the special American club where they're releasing the models to. Like I say, it feels like the US is losing control of this race.
And that brings me to my third set of questions about the politics. Because the US has been trying to restrict China's ability to build frontier AI, in particular with export controls on chips. That's what this chart shows. Who's got the AI chips? And by this measure, they've been very successful because you can tell, look, most of the AI chips in the whole world belong to big American companies. And the Chinese don't have that many. Yet, they are managing to build near-frontier AI. And honestly, no one knows how they're doing it. There are many rumors, many theories. Somehow, with this amount of chips, they are managing to build something that is so close to the frontier to be indistinguishable. So clearly, the restrictions aren't enough.
What does that mean for the US? Do they somehow change their restrictions, do something different? Well, look, people are putting forward options. So I'll give you the options. Okay, one set of options is being put forward by Anthropic and the big AI companies, who are understandably keen to restrict Chinese AI. This chart shows, uh, the, um, number of distillation attempts that Anthropic say Chinese companies have been making on their models. Basically, Anthropic are saying the Chinese are copying our models, stealing our stuff. You've got to stop that. We need to ban these people. And yes, it is quite ironic that Anthropic are complaining about someone else copying their models when their model, when their AI is based on millions of copyrighted books. And yes, it is quite self-interested for Anthropic to say that, given that this would give them complete freedom to price their models as high as they want. Nevertheless, a very influential and important argument.
Here's another one, may in fact recently by President Xi. He suggested some kind of deal. Uh, to paraphrase, he says roughly, why don't you give us more AI chips, and then we'll strike a deal on safety and create a joint licensing arrangement. That's one possibility here. Although I don't know about you, but I'm not sure that seems like the kind of thing the Trump White House would go for.
Which leaves us with what David Sax talked about, a return to permissionless innovation, a return to the AI race, no holds barred. Is that what we're going to see? I think that might be quite likely here because, like I say, the US is losing control of this race, and they are determined to win at all costs. That situation could have huge implications, really, for everyone.