A dense layer of red tape covers this AI race, which is informing semiconductor trade policy at the moment. Meanwhile, Huang has even claimed that China is aggressively tailing the U.S. in artificial intelligence — if not already ahead — and that most of the popular open-source models in the world today are from China. The region also produces some 50% of all AI researchers globally, implying the competition is stiff.
What exactly is an “AI researcher”?
Probably people who studied AI in Computer Science programs at UofT and American schools as well as domestic universities, who then proceed to apply said knowledge in new science and applications of AI at (Chinese) universities and firms. E.g. the guy that creates, adapts or integrates an AI model to drive a mining truck for rare earth minerals to free up human labour for other purposes. Or the person that came up with storing LLM context as image data in order to increase DeepSeek’s context window by 10x on the same hardware.
Out of the total population of “AI researchers” (as cited in the article), what percentage of them are engaged in research?
Or do we assume every single technical (we’ll exclude sales and market) person involved in AI/ML industries is a researcher?
AI research is very complex and includes several fields, like AI programming experts, linguistic experts, neuro scientists and philosophy. These people are highly specialized in their field, and it is a field of scientific study. Scientific AI research today is a combination of several scientific fields.
That’s fair, but are you sure the top line number cited in the article reflects your definition? Maybe it does, I don’t know.
I am also sceptical of the the inclusion of philosophers and neuro scientists in your list of fields. I would speculate employment of such specialists represents an infinitely tiny portion of total employment in AI/ML industry (however you define that).
I am happy to be proven wrong. The topic of employment dynamics in AI/ML is not an area I am familiar with
No idea. He’s probably just talking about the production of people with such education. That said, the field of AI R&D is enormous. There are so many opportunities for productivity improvements in healthcare, manufacturing, farming, chemicals, etc. The more AI-trained specialists you got, the more of those you can attempt. If a country has a goal of increasing economic productivity which doesn’t always jive with the free market economy.
There are so many opportunities for productivity improvements in healthcare, manufacturing, farming, chemicals, etc.
That’s definitely true. There is a lot of potential here. That being said, like with all tech, it’s about the people that apply it.
The more AI-trained specialists you got, the more of those you can attempt.
Not sure I believe this is true in all cases. It is not too unreasonable that there is a decrease in marginal returns on “AI-trained specialists” as the number goes up.
And it is likely that a large number of said “AI-trained specialists” are probably looking to cash in while this option is available.
I am a regular user of a variety of AI/ML tech (not only LLMs, although I do use them a lot in a relatively cautious manner), so I am not really an “AI skeptic”.
I am skeptic of the individuals involved in the AI/ML industry because I know what money can do to people. Not to mention claims/polemics about “millions of AI specialists working diligently to improve the world” have an almost Lenin/Mao-style feel to it.
Some people in this industry might have a measure of idealism, but they are likely a small minority. Most are probably just conformists and just doing whatever, but also not necessarily opposed to criminality or idealism depending on how it is marketed. A not inconsequential portion are committed criminals and will pivot to the “next big thing” that becomes popular for “white collar” criminal groups.
I have a little bit of exposure to the points I am bringing up, so I may be generalizing, but it’s not like I am just making stuff up.
Oh I’m not at all assigning a universally positive morals to the hypothetical researchers or idealism. I think people would do more or less what their reality pushes them to do. In a reality where studying AI leads to the shortest path from working class to early retirement, I expect people to do what that industry does. AI slop generators for example at untold socioeconomic cost. In a different reality where AI is just another research field, like material science, where studying it does not allow for getting rich quick but leads to careers in automation in various other fields, I expect people interested in it to do that. Just like they do in many other fields of research and development. In this sense someone with high degree of state control over private capital and economic planning could do things differrently than what we observe the market doing in the US. I’m not saying that’s what they would do for sure and that things would be amazingly great. The Chinese have stated they are planning to go this route though.
The writing’s on the wall if the US AI bubble pops. China ain’t buying NVIDIA for obvious reasons, at least not without steep discounts. I’m also curious how would TSMC be affected. I’ve no idea how much of their output is AI-related. If significant, it would be interesting to see how they fill that capacity if the AI demand drops off a cliff. Whether they’d fuck the sanctions and make Chinese CPUs and GPUs for example.
TSMC is all that stops the invasion of Taiwan. If TSMC is not relevant even for a moment, China will invade to end their civil war. You can count on NK invading SK at the same time and taking Samsung.
China already has domestic incentives in place for home grown GPUs. They will likely displace Nvidia entirely within 5-8 years.
Ultimately, a unified architecture will win. The reason CPUs cannot handle the load of AI is due to the L2 to L1 cache bus throughput. It requires a major redesign, but it is a solvable problem. The real problem is that that kind of redesign takes the full 10 year hardware design cycle time to create from scratch.
AI is still not going away in the long term. The present world is just like the early days of the microprocessor. The 6502 was little more than a toy. It is still in all western digital hard drives. The fundamental architecture is still the same in all CPUs. It was the systems we built around them that made them useful. A base inference model is primitive. The AI that owns the future is agentic systems.
China will invade to end their civil war.
I don’t think invasion’s on the cards. There’s too many cons and not many pros. Instead, I think in a TSMC irrelevance scenario, or otherwise lack of demand from the west, Taiwan’s gonna start getting a lot closer to China peacefully if China replaces that demand. China can play a very long game given its socioeconomic infrastructure. That’s my bet at this point. You could very well be right of course.
AI is still not going away in the long term. The present world is just like the early days of the microprocessor. The 6502 was little more than a toy. It is still in all western digital hard drives. The fundamental architecture is still the same in all CPUs. It was the systems we built around them that made them useful. A base inference model is primitive. The AI that owns the future is agentic systems.
Probably. There are definitely useful models that solve problems much better than previous algorithms.
The military analysts point to the age bubble in China as the tell for Taiwan. They really can’t play the long game because of the population bubble. Most project that peak invasion opportunity is around 2027 and the risk goes down substantially after 2030 and then 2035.
We are less than 10 years away from the final fab node. The exponential growth is over. It actually already is over as the people on the bleeding edge are already at the end.
Most people are ignorant of the implications here. Silicon chips are the only time in human history when a civilian industry grew faster than the largest military could finance. There is no replacement. We will return to the ways of 80+ years ago. The next major age of technology is biological, but we are still centuries away from that future.
We are less than 10 years away from the final fab node
Beyond fab nodes, there are things like advanced packaging approaches, BPD.
New nodes will probably start to offer more modest gains (while costs will start growing beyond the current 20-30% increase per node), but that doesn’t mean there won’t be improvements in leading edge semi-conductor performance.
The importance is on exponential growth beyond that of military spending. The leader of the global economy has massive overall implications on geopolitics.
I’ve seen that analysis and it makes sense under the population bubble hypothesis. I think the population problem is overstated though, because it’s actively being solved through automation-driven productivity improvement. Robotics use is exploding in China.
