00:00Okay, Selena, can you start by telling us what are some of the biggest misconceptions about the U.S.-China AI
00:04race today?
00:06I think people tend to think that, you know, both the U.S. and China, when they use the race
00:11metaphor, that they're very much, one, racing down the same lane towards the frontier, and then, two, that they are
00:17racing towards the same finishing line.
00:19And I think in the U.S., that finishing line is very much, whether you call it AGI, artificial general
00:24intelligence, or if you call it, you know, reaching the moment of recursive self-improvement that can unlock, you know,
00:31great gains for society and for the economy.
00:33In China, I think it's quite different, and part of it is because of export controls and compute constraints that
00:40has forced them to develop a very different kind of approach.
00:43So China, one, I think their dominant vision of AI is very much physical, so they're very focused on integrating
00:50AI into hardware, and then to accelerate that, they have very much adopted open-source models to accelerate the adoption
00:59of AI across industries and across sectors.
01:02So efficiency is a huge thing because they don't have enough chips.
01:06According to Stanford's 2026 AI Index report, the gap in performance between the U.S. and China's large language models
01:13has essentially closed.
01:14Do you believe that China has truly caught up?
01:17I think the gap has definitely narrowed from a few years back.
01:22So the gap was really wide right after ChatGPT came out, and China really struggled to catch up with its
01:29initial models like Baidu's Ernie.
01:31If you look at the more recent Chinese models like DeepSeq V4 or Kimi K2.6 or Z.AI's GLM
01:40models, it does seem like if you trust the benchmark reporting, it is generally somewhere around like under six months.
01:48But if the broader claim or if you want to make a bolder claim of Chinese AI is now neck
01:55and neck with U.S. frontier models, I think that is probably not true as of this point.
02:00It is not that their performance are totally on par, especially on some of the areas like coding capabilities or
02:07agentic capabilities.
02:08But it's more that, hey, maybe they have other metrics, right?
02:12Like DeepSeq is so much cheaper right now than Claude.
02:15Lastly, if you had to pick just one, whether it's chips, large language models, physical AI, infrastructure, etc., what's the
02:23most important frontier that you see determining who leads the U.S.-China AI race and why?
02:28I think the most important indicator might be which economy can first see like real productive impact from AI.
02:39So the U.S. obviously its economy is so deeply leveraged on on AI right now, whether you can see
02:48real returns on productivity without seeing very widespread, you know, displacement of, for instance, jobs that would harm, you know,
02:57people's ability to consume and buy AI services.
02:59Right. So I think that that is one thing to note in the U.S. context and the China context
03:05is very much more of the Chinese government now is so all in on AI because their old growth drivers
03:11like property and like infrastructure investment no longer work that well.
03:15And they want to pivot to these what they call new quality production forces, which include AI, robotics, clean energy,
03:22batteries and everything.
03:23In the U.S., there's so much anti-AI sentiment.
03:26Token costs are going through the roof.
03:29There is a memory squeeze.
03:31Like all these problems is like before you talk about reaching, you know, super intelligence, there are these really just
03:37like real issues about negotiating just how exactly you even bring that about in the first place.
03:44Is that I think is is very top of mind for policymakers and I actually think is the more real
03:49issue if you talk about a race like which country can actually bring about the fruits of AI without their
03:54societies breaking down.
Comments