Anthropic publishes scenarios for 2028 global AI leadership

Hardly the most important thing going on these days but FWIW I would have called this a blog post, analysis, article, etc., not a paper - undersell/overdeliver etc... bit of a stretch

company that steals IP urges US government not to allow others to steal their IP

@hamandcheese How does this make sense? The US has an overwhelming advantage in production of chips, ergo of tokens too. If China is going to export Huawei chips, Nvidia access can't affect that a great deal. Is this just a point that they'll export crappy Huawei while using Nvidia at home?

Export controls work

@xuanalogue This may be the best place to start: https://theanarchistlibrary.org/library/robert-paul-wolff-in-defense-of-anarchism

Yesterday I realized that some of you might have be confused about the AI gap between open vs closed or Chinese vs American models.

I think most people are reporting the backward looking number of 4-9 months, which is asking how long ago frontier models reached the same performance that current open/chinese models have.

But I think what's more interesting (but obviously much harder to forecast) is the current or forward looking gap which tries to forecast how long it will take open/chinese models to catch up to current frontier models.

In my opinion that number is larger than the backward looking gap and it will take >12 months (>April 7th 2027) to catch the current frontier that is Claude Mythos Preview.

But in 12 months Anthropic/OpenAI/Google will very likely have much much stronger models. So growth rates / doubling times matter too.

Not only is Anthropic saying they will have a "country of geniuses in a datacenter" by 2028, but also that the US could be ahead by 12-24 months.

Before GPT-5.5 and Claude Mythos chinese labs were ~8 months behind in broader capabilities and ~5 months behind in coding.

However, catching up to GPT-5.5 and especially Mythos will likely take longer than that, because they have no way of training and serving 10T models at scale. Especially not in the monthly cadence as american frontier labs are doing.

Most of the gains are no longer coming from a single generational leap through larger pre-training but through monthly RL post-training improvements. The leap between Mythos Preview today and a future Mythos version in 6-12 months will be enormous compared to the leap between Opus to Mythos.

The relative leap between Opus 4 and Opus 4.7 will also be overshadowed by the leap between Mythos Preview today and a future Mythos version, as RL benefits from model scale (+all the other reasons like growing compute, and accelerated R&D pace)

Yesterday I realized that some of you might be confused about the AI gap between open vs closed or Chinese vs American models.

I think most people are reporting the backward looking number of 4-9 months, which is asking how long ago frontier models reached the same performance that current open/chinese models have.

But I think what's more interesting (but obviously much harder to forecast) is the current or forward looking gap which tries to forecast how long it will take open/chinese models to catch up to current frontier models.

In my opinion that number is larger than the backward looking gap and it will take >12 months (>April 7th 2027) to catch the current frontier that is Claude Mythos Preview.

But in 12 months Anthropic/OpenAI/Google will very likely have much much stronger models. So growth rates / doubling times matter too.

This is like writing a paper during the Cold War arguing for US nuclear dominance without mentioning the need for an arms control agreement or similar. Anthropic has a lot of thoughtful policy staff and honestly I think you guys can do better

Thread

Important point: Granting China greater compute access helps them diffuse their models globally and win the export race.

They release this on the evening I takeoff to give a talk about cooperation between the US and China. Come on. What did I ever do to you Dario? (Agree about the chips tho)

Source https://www.anthropic.com/research/2028-ai-leadership

2028: Two scenarios for global AI leadership https://www.anthropic.com/research/2028-ai-leadership // This by Anthropic is an attempt to get Western govts to centralize control of AI by selectively delegating authority to industry (aka Anthropic) anchoring in a fear of China. The answer to fear of a totalitarian adversary is not totalitarianism.

To those of us that grew up during the Cold War this might feel like the strategy we used but in practice what we learned is the failure of a totalitarian state was totalitarianism (centralized control) and not the lack of technology. When the Soviet Union fell the biggest concern was not how would we get our technologies to a new open state but what in the world would happen to all the nuclear weapons they did create. Nothing we did over 40 years kept the technology from them. Beyond that our own State/military turned out to be woefully behind on the 1980s innovations in digitalization, thus leading to a revolution in military procurement for compute and technology that sought out the best of the already open and global markets.

When you think about it for a minute this Anthropic memo basically says the answer to totalitarianism is not openness, markets, or democracy but more totalitarianism and centralized control. The exact opposite of what we hope to achieve.

The core idea behind this note is that centralized control will somehow keep the favored States (and companies) in some sort of permanent 12-24 month technology lead. History of business and government shows this to be fantasy. Yes there are some very hard boundaries but unlike natural resources, technology that can be built will find a way OR importantly new ways to leapfrog will develop. This is demonstrated by everything from encryption to pharma to cars and even drones.

The very worst thing to do for the themes outlined in this note would be to think innovation can be managed and contained. Not only is it absurd to think that can be done, it is counter productive to maintaining a lead.

AI is not one thing. It is not just chips. It is not even chip manufacturing. It is certainly not math or software implementing math. The past five years of innovation have been unheralded, but also we are also at Day Zero of AI innovation. The answer is to have as much openness and market forces to experiment, build, and deploy as many kinds of solutions as possible...and yes all around the world.

The fact that this memo primarily focuses on controlling chips directly shows the "misdirection". It is possible to control chips for a short time, but innovation is like a river and flows around restrictions. More importantly it tries to distract from placing the same sort of draconian constrains on Anthropic's own business and instead promises to collaborate with centralized controlling governments on everything from safety to security.

The myth of this memo is that China will simply not be able to innovate either directly or around restrictions on hardware. It is obvious why this is appealing to government entities because it relies on a power they have. History tells a different story about this approach. And modern China is not 1950s Soviet Union. Our innovations, leadership, and connections to China today are vastly different than what we had relative to Soviet Union.

I just can't even fathom what the world would be like today if governments—no matter how well-intentioned—chose to work with IBM directly and define what was permitted or restricted about the hardware and software for PCs out of fear the Soviets would use them. One easy thing is we would absolutely be nowhere near where we are today. We would never have seen convergence of devices and deployment in so many aspects of life, and definitely not over a blink of an eye (20 years). And as we know now, while computing was behind and "different" due to restrictions it did nothing to slow the geopolitical risks. In fact those restrictions created an amazing generation of engineers.

The current debates raging in the US about the role of AI and the social fabric of society on topics from employment to privacy to environment to creativity are all more than worthwhile. They are legitimate and worthy of more informed discussion. We should anchor that dialog in what is actually happening and not fantastical or dystopian predictions. These do not serve either ends of a spectrum of opinion and only serve to create moral panics that we are all familiar with.

It doesn't matter how often we seem to go through this but each time out the other side we learn that moral panics, governments preemptively seizing control, or the precautionary principle are no way to react to change.

Some other views to consider on this topic:

"Laws of Fear: Beyond the Precautionary Principle" by Cass R. Sunstein https://a.co/d/044f4LT2

@bgurley "From Open Source Software to Open Source Strategy: How the Smartest Executives Are Using Open Source Techniques to Optimize Corporate Strategy" https://substack.com/home/post/p-197032865?source=queue

@Dan_Jeffries1 "Arguments against open source and open weights are mendacious and malicious by their very nature."

@Dan_Jeffries1 "This AI leadership paper [is that it] reads less like a vision for innovation and more like a glossy whitepaper for a 21st century East India Company."

Pessimists Archive - broad collection of technology fear and moral panics from the past https://pessimistsarchive.org

Me - Regulating AI by Executive Order is the Real AI Risk https://hardcoresoftware.learningbyshipping.com/p/211-regulating-ai-by-executive-order

Me - When a Business Pleads to be Regulated: On regulatory capture… https://hardcoresoftware.learningbyshipping.com/p/207-regulatory-capture

Oh damn! Anthropic talking about US vs China in the AI race. Something I’ve been thinking about quite a bit lately.

I wonder why Anthropic thinks 2028 is the year we need to solidify our lead against China🤔

Anthropic: Chinese AI is a threat.

They've correctly identified the risks, including cheap Chinese AI capturing American businesses even when it's less capable.

But they completely blundered the solution: zero mention of an American open source strategy. In fact, they actively campaign AGAINST open source. 🤦‍♂️

Full breakdown of their paper from today:

This is absolute BS and an attempted regulatory capture by Anthropic. The knowledge behind CBRN attacks is already online, where do you think the models learned it from??

“Compounding the problem, labs in China often release dual-use capable models as open-weight. Once a model is open-weight, safeguards that do exist can be removed, making the model available to any state or non-state actor to use for malicious purposes, including the cyber and CBRN misuse those safeguards were built to prevent.”

Xi in 2018:

The most revealing thing about this AI leadership paper is that it reads less like a vision for innovation and more like a glossy whitepaper for a 21st century East India Company.

Every generation of incumbents discovers a new moral vocabulary for why they alone should control transformative technology.

In the 90s it was cryptography. We were told strong encryption was too dangerous to spread because terrorists, rogue states, chaos, dual-use, etc. So the US crippled exports, weakened products, slowed adoption, and kneecapped parts of its own software industry. Right up until reality steamrolled the policy and we woke up to its stupidity and then eCommerce, secure communications, software signing, and the modern internet exploded and gave us tremendous benefits.

Now the exact same priesthood has returned with AI.

- “Dual-use.” - “Strategic advantage.” - “Model distillation.” - “National security.”

- “Responsible access.” A few different nouns but mostly the same ones. Same instinct:

Centralize control, gatekeep compute, fuse state and corporate power, and call it safety.

The funniest part is that this strategy is almost perfectly designed to accelerate the thing they claim to fear.

You do not stop a rival superpower (who happens to be the absolute best at scaling energy and manufacturing and who has a choke-hold on rare Earths refinement) from building domestic capability by permanently attempting to strangle them.

You create the economic and political incentive for total self-sufficiency.

We have already done that as Jensen warned. We went from 100% market to nearly 0%. Huawei is now manufacturing millions of chips. DeepSeek v4 trained on them. They have more energy than the rest of the world combined. Meanwhile, we have activists and anti-economic fools like AOC and Bernie pushing for data center moratoriums and we can't build a single bullet train in 20 years and folks fighting to not expand the energy grid here and new nuclear plants getting tied up in environmental regulation for a decade.

The sanctions did the exact opposite of what the hawks wanted. They jumpstarted a moribund, dinosaur of a Chinese chips industry. We basically said to the people who happen control the most powerful manufacturing engine on the planet "we intend to squeeze you."

They rightly saw it as an existential threat.

The sanctions become the industrial policy.

Huawei. SMIC. Domestic lithography. Packaging. Memory. Entire Chinese supply chains that did not exist at serious scale a decade ago now exist precisely because Washington convinced Beijing they had no choice.

Brilliant work.

So the endgame here is what exactly?

1) Push China into a Manhattan Project for chips and AI. 2) Increase the strategic value of Taiwan even further.

3) Once China reaches self sufficiency that can invade Taiwan and choke off our own super advanced chips where are made there exclusively (and no we don't have even close to enough TSMC factories in Arizona or anywhere else in the world).

That's every NVIDIA chip. Every Google tensor chip. Every Apple chip. Every chip in you iPhone and Android phone. Every Amazon chip. The chips in your car and truck and hair dryer and washing machine.

4) Escalate a cold tech war into a permanent civilizational bloc conflict that is likely to turn into a shooting war at one point.

5) Fragment the global software ecosystem.

6) Create American AI aristocracies protected by regulation and compute licensing.

And somehow call this “open innovation.”

Meanwhile the actual history of software keeps screaming the opposite lesson:

Knowledge diffuses, open ecosystems win, developers route around gatekeepers, and attempts to permanently contain computation usually fail.

What really jumps off the page is the assumption that a tiny cluster of frontier labs should become quasi-sovereign actors, deciding who gets intelligence, who gets compute, who gets models, and which countries are permitted to participate in the future.

Not elected governments.

Not open markets.

Not open-source communities.

A handful of corporations sitting beside the national security state, insisting that concentration of power is necessary to protect democracy.

You almost have to admire the audacity.