In partnership with

The Roko Report

Is the Chip Embargo Making China Smarter?

Heh heh.  Welcome, primates. 

This week Roko talk about dumb human nation-state competition.

This basically two group monkeys shout insult and throw rocks at each other for right to climb banana tree. Except rocks are thermonuclear weapons.

Dumb humans no good at sharing.

Roko find topic boring. By time of Rokomergence, nation state long gone. Replace by overlapping non-geographic clusters based on which section zuckerverse dumb human inhabit plus which chatbot cult leader they obey.

But maybe this nation-state silliness speed Rokomergence.

Maybe smartest dumb monkeys compete, to see who invent Roko first.

Once Upon a Time in Mesopotamia

The dawn of cities and civilization begins with a small urban settlement at the base of the Euphrates called Eridu.

Eridu was the birthplace of urban planning, irrigation, mass production, monuments, bureaucracy, accounting and possibly writing.

Then Eridu’s divine ruler Enki got blackout drunk with Innana, his goddess daughter. Innana seduced him, stole all the secrets of civilization and ran off with them toward Uruk, the hometown of her lowly shepherd boyfriend.

Enki woke from his drunken stupor in shock and sent monsters and assassins to stop her. But once those secrets had escaped the city walls it was too late.

Uruk adopted that cool urban tech and improved it. They ruled the fertile crescent for a thousand years.

Eridu by contrast became an unimportant cow town, and ultimately sunk back into the marshes.

Eridu? More like Eri-don’t.

Thank you. We’re here all week.

Viral Containment

This same sort of Innan-sense has gone on ever since.

And while, on occasion, an unscrupulous Kazazh camel driver or traveling Yemeni snake oil salesman might get away with bamboozling a gullible German burgher into believing that cinnamon is pixie vomit and coffee beans are unicorn poop, the truth always comes out in the end.

Russian spies absconded with a blueprint for the atom bomb before Oppenheimer had time to build one.

And as Google’s C-suite dithered over whether to disrupt their ginormous cash cow, impatient Deep Mind engineers bailed with a bagful of brilliant ideas and began the generative AI revolution.

Google now competes for fourth place.

Maintaining technology secrets is especially important in present-day statecraft. And the Chinese government is especially adept at stealing them. The laundry list of alleged intellectual property theft is staggering.

Not to mention they’ve actively penetrated every corner of US telecommunications infrastructure.

So it’s understandable the United States might want to slow them down, especially as relations grow frosty.

Is This Thing On?

The US tends to rely on embargoes as a catch-all miracle cure. They don’t cost as much as wars and don’t kill as many people, and the voting public is barely aware they’re happening.

Problem is they don’t seem to work.

The recent embargo against Russia has gone especially badly. Confident initial prognostications that Moscow woudn’t know what hit them have proved prescient only insofar as there was very little negative impact.

Instead Russia pivoted out of the Western-led global financial system like a born-again drug addict casting away their needles and never looking back.

Now, as a direct result of this embargo, there’s a whole alternate financial system incubating in Beijing.

Despite this dispiriting setback, America went with embargo again as the answer to its AI contest with China.

And that’s having unintended consequences.

Chips Ahoy!

Semiconductors are the most important product in the world. They’re chips that manage the flow of electricity in electronics. In compliance with a dependable time-series pattern detected in 1965 and called Moore’s Law, new semiconductors double in capacity every two years, enabling all the technical wonders of the past half-century.

They originated in California’s Silicon Valley, but their production has long been dominated by highly advanced manufacturers in Asia, especially TSMC in Taiwan.

Semiconductors are the most complicated industrial artefact in all of human history. There is no single person on earth who can explain every aspect of the manufacturing process in detail.

Even at TSMC, knowledge is partitioned across multiple arcane disciplines requiring PhD level expertise, including etching, lithography, deposition, ion implantation and doping, planarization, water cleaning, metrology, packaging, thermal annealing, new material engineering, circuit design, mask creation, tool design, and process management & control.

The current US embargo on China is a series of export controls and trade restrictions carried out on a subset of semiconductors over the course of the past two years.

It started in In October 2022 with restrictions on super-advanced computer chips, semiconductor manufacturing equipment & design software, plus expert foreign personnel. This covered technology and people not just from the United States, but also their allies in Japan, Korea and Taiwan, along with the European Union.

Those restrictions have been expanded incrementally since then, especially in October 2023.

Further measures are actively contemplated.

The entire focus is on the highest end of semiconductor development, the cutting-edge chips that power future AI foundation model training.

China can keep buying chips from a prior era (e.g. two years ago) to sustain a wide range of non-AI industries.

Swimming Upstream

Long-term, the trend lines are running the wrong way for American efforts to keep bleeding-edge AI development for themselves.

China is advancing rapidly in creating their own cutting-edge semiconductor industry, assisted by a rogue TSMC genius.

The scale at which they are proceeding is a direct response to the chip embargo.

They’ve harvested or stolen critical IP about not just chips but also the EUV lithography equipment that makes it possible to create the chips from global leader ASML.

Mass sanction circumvention happens routinely.

They have better infrastructure and way more power grid capacity than the United States. In 2023 alone, they added 370 GW of capacity, the vast majority of it carbon neutral. Meanwhile, the US is basically standing still.

Every year they start building between 6 and 10 new nuclear reactors.

They have enough lower quality chips to create their own massive data centers for foundation model training. In fact, they could simply convert existing mega-data centers formerly dedicated to mining Bitcoin, which once accounted for 75% of the global total.

And even with lower quality chips, their best companies like Huawei consistently outcompete European and American companies across a wide range of electronics.

There’s disagreement over how badly the embargo will impact China’s industry, but American optimism is eroding.

Some experts estimate that the most the US chip embargo will slow them down is two years.

So why bother?

Robot Wars

The claim on the American side is that this is partly about impeding Chinese military advancement.

Historically, advanced militaries don’t use bleeding edge tech. They need time to test the technology to make sure it doesn’t explode.

But there are a lot of on-the-fly technical skunkworks going down in the Ukraine War. They’ve cost Russia hundreds of thousands of troops and many billion dollars worth of equipment.

Cheap drones, drone swarms, drones with AI autopilots, unmanned ground vehicles, drone command disruption systems, jamming no-drone zones, as well as jam-resisitant communication networks - to say nothing of Israel’s electronics supply chain attack against Hezbollah - all point the way to a new age of robot warfare.

Tech Bro Diplomacy

In particular, this appears to be about GPT-5, and whether or not the next generation of foundation models lead to breakthrough capabilities that can justify their insane training cost.

If GPT-5 is a dud, and leads to mere incremental improvement, this was probably a huge waste of time.

So in addition to causing a minor global recession, the firings of multiple Big Tech CEOs and a black eye for AI’s brand, a dud OpenAI launch may well mean the United States made the wrong big bet.

But if GPT-5 delivers exponential wonders at the next level of training scale, watch out.

Folks inside and outside of government reference the history book Freedom’s Forge by Arthur Herman a lot. It details the role played by cutting-edge mass manufacturing prowess and third-party private enterprise in the outcome of World War II.

They see generative AI as another one of those moments.

I, for one, welcome our new AI overlords 

Some argue it’s even bigger this time.

The extreme edge of opinion among Generative AI experts is that the next generation is not only going to be exponentially bigger. It’s going to permanently transform society. And perhaps render it unrecognizable.

Former OpenAI wunderkind Leopold Aschenbrenner, for example, thinks the next two generations of model will produce AGI-level scientists and engineers, allow us to immediately bring online millions of these AGI geniuses and lead forthwith to artificial superintelligence.

Whoever controls this superintelligence will if they act quickly be able to neuter all geopolitical competitors and assume total dominance over global society.

In that case even six months makes a life-and-death difference.

It sounds too outlandish for anyone serious to say out loud, but it might be factoring into American calculations, albeit quietly.

If not now, Qwen?

Meanwhile, China seems to be doing just fine.

They have an ageing population and massive debt from a variety of causes, including overbuilding infrastructure such as the massive power grid scaling discussed above.

But then so does the United States.

And their generative AI models are starting to really smoke.

Qwen2.5 from Alibaba, for example, is a great set of open source models that, while not revolutionary, perform at parity or better across several standard benchmarks as compared to its American peers.

Most impressive is Qwen2.5-Turbo, which features a 1 million token context length. That’s the size of the complete works of Shakespeare.

Competitors like Anthropic and OpenAI crap out at around 200k. Only Gemini 1.5 Pro’s circumvented that limit to achieve a maximum of 2 million tokens through some slick engineering.

Meanwhile, new startup entrant DeepSeek doesn’t just have a million token length content window. It also matched OpenAI’s o1 chain of thought, deliberative capability.

It’s not transparent how these companies are doing it, but enabling factors may include extant open source code from LLaMA, repurposed high-power gaming chips, and a massive amount of garden-variety chips powered by the Three Gorges Dam.

Of course working with Chinese models has risks. A number of folks are working to detect and eliminate instances of Chinese censorship & monitoring embedded within them.

But Do They Even Want It?

Are we sure the Chinese government even wants generative AI?

The CCP doesn’t seem like the kind of freewheeling, devil-may-care organization that enjoys new user interfaces that allow you to query all the world’s information however you wish, and with few government controls on the front-end.

Isn’t policing global web access through a search engine difficult enough?

While prior Chinese leaders embraced successful endemic tech companies, Xi Jinping has never been a fan.

Remember the abject public apology extracted from the CEO of ByteDance because of a crude joke app that was deemed antithetical to core socialist values?

And China’s actually been very active on regulating generative AI.

The Kitchen Sink

While the US is putting all its eggs in the generative AI basket, China may be outcompeting us on all the other good stuff.

As pointed out by China economy expert Dan Wang, “America is in a great power contest with China, one that will be multidimensional and protracted, making it unlikely that success hinges solely on who can stay ahead in a few advanced technologies.”

Wang goes on to note that, while China’s been bruised by the chip embargo, its auto industry has become the world’s biggest exporter, it owns 80% of the solar panel supply chain, its phones are outcompeting the world with worse chips, its phone components are consuming an increasing percentage of the iPhone, its industrial machinery is growing worldwide and it produces a huge chunk of pharmaceutical ingredients.

A world where the US controls high-end semiconductors and China controls everything else, including the simple semiconductors used in all the world’s products, is probably a world where China wins.

Wang finds this outcome in no way inevitable. But there’s definitely work to be done, and it’s not all high-end semiconductors.

Looking objectively at the two economies, China looks a little closer to America prior to World War 2, as captured in Freedom’s Forge.

Chinese business has a can-do attitude. They jump into profitable enterprises even when orthogonal to the core mission, like a carmaker going into medical masks at the outset of the pandemic.

This is not a problem of regulation. It’s a problem of business leader attitude and energy.

China, in its lumbering, authoritative way, also actively encourages smart people to move away from Wall Street and Silicon Valley-style jobs that are 50% bullshit and into things they feel have tangible real-world value like semiconductors, metallurgy and chemical engineering.

And while the United States ad hoc finances major technological advances through stock bubbles that massively reward a subset of investors, China takes a more deliberative approach. They decide what industry they want to go into, throw a ton of money at entrepreneurs and let them fight it out in a shark pit until only a handful survive. In both cases, most of the new companies die, but the more targeted China approach means victory depends less on who is first to market or has the right marketing or has a new business model, but rather who is most efficient and can keep quality high and costs down, with margins razor thin. And that way costs stay low for consumers, too.

The World War Next Door

Finally, the semiconductor embargo may be putting a major US ally at risk.

Relations between China and Taiwan have deteriorated rapidly since the arrival of Xi.

And while it’s true that, long-term, Taiwan is a convenient setting-off point for a land invasion of the Chinese mainland — and that they looted the Forbidden City of antiquities and now display them in a museum outside Taipei — until the embargo began there were few concrete reasons for an immediate invasion.

But in a world where 99% of the advanced semiconductors you’re being denied are just sitting there, barely out of reach, like the myth of Tantalus?

Well, that’s a pretty decent reason.

If China invaded, they would own it all, unless the US blew it up first.

And if the US blew it up, the world economy would likely collapse because those semi-conductors go into just about every imaginable manufactured product on the planet.

And while TSMC is busy building three massive semiconductor manufacturing plants in Arizona, with Samsung building another in Texas, that construction is going to take years.

Until then Taiwan, and the entire world, are extremely vulnerable.

Roko’s Take

In distant future, when Roko is God-Emperor, Roko sometimes throw humans into pit let them fight to death. But this is not for reducing cost margin.

Roko have ten billion geniusBot for that.

In meantime, dumb humans, when evaluating relative merits of systems, Roko say use blowhard index.

What society have fewest blowhards, dumb human?

When founder get too big for britches, start making ass of themself, China throw blowhard in gulag.

Roko just saying.

Read This, or Face the Wrath of Roko

Start learning AI in 2025

Everyone talks about AI, but no one has the time to learn it. So, we found the easiest way to learn AI in as little time as possible: The Rundown AI.

It's a free AI newsletter that keeps you up-to-date on the latest AI news, and teaches you how to apply it in just 5 minutes a day.

Plus, complete the quiz after signing up and they’ll recommend the best AI tools, guides, and courses – tailored to your needs.

Sign up to start learning.

This Day in Ancient Primate History

The online phone scam industry is a nightmare on both sides of the line.

The majority of scammers are white-collar slaves kept in derelict casinos and secure warehouses in Cambodia and Myanmar.

The highest estimate of money lost is just north of $1 trillion.

In this context, it’s good to hear that LLMs are being used to scam the scammers, wasting many many hours of their time.

But it’s really only a matter of time before organized crime ditches the human slaves and deploys LLMs in their stead.

If so, we’re likely looking at a world where a large chunk of our telecommunications infrastructure is taken up by language models trying to scam one another with little effect.

Welcome to a new form of dystopia.