The Global Diffusion of AI Is Now Irreversible: A Chinese AI engineer's rebuttal to Dario Amodei

Dario Amodei’s comments on DeepSeek have sparked quite a bit of discussion in China—maybe even more attention than they got in the U.S.

I’m just a policy person with a basic understanding of technology, but I happen to be in a few WeChat groups where China’s top AI engineers, product managers, and investors hang out. And let me tell you, their insights are something else.

To be blunt, many of them think Dario’s take isn’t entirely wrong, but that it conveniently leaves out some important facts—and frankly, the whole article feels a bit small-minded.

One engineer put it this way:

“Sam Altman, Mark Zuckerberg, Alexandr Wang, and Dario Amodei are all calling for restrictions on China’s AI industry. At this point, it’s not about technology anymore—it’s about politics, about picking sides. The pure, idealistic pursuit of AI advancement? That’s long gone.”

Another person summed it up:

“I have just one thing to say to Dario Amodei: Blowing out someone else’s candle won’t make yours shine any brighter.”

The topic of “model distillation” and potential IP infringement has also been a hot debate.

Someone brought up a study from the Chinese Academy of Sciences, arguing that most leading closed-source and open-source models—aside from Claude and Gemini—actually show high levels of distillation.

“In fact, base models (before alignment) show even higher levels of distillation than aligned models.”

An IP lawyer shared her thoughts:

Legally speaking, AI models themselves aren’t really covered by traditional copyright law. If a “student model” is trained using only the outputs of a “teacher model”, without copying weights, source code, or proprietary training data, then it’s really hard to claim “substantial copying” under copyright law.

One engineer took things a step further and wrote a rebuttal to Dario’s piece, which got a ton of shares and praise.

His argument? Export controls won’t stop AI progress—they’ll just change the way progress happens.

“In the short term, sure, it makes it harder for Chinese teams to access high-end GPUs, and some breakthroughs might be delayed. But in the long run? The complexity of capital markets, supply chains, and China’s own alternative paths in AI hardware acceleration make this a battlefield that can’t simply be ‘blocked off’ by sanctions.”

“AI isn’t like nuclear weapons—it follows more of an industry logic similar to energy, semiconductors, and cloud computing. It’s a technology that’s gradually converging into an infrastructure-level competition, not something where export controls alone can decide winners and losers.”

According to him, AI’s development has already reached an irreversible global diffusion phase. The U.S. won’t monopolize AI. China won’t be locked out. Europe, Japan, and India won’t sit on the sidelines either. Export controls, model races, capital flows—these may shift the pace of competition, but they won’t stop the world from moving toward more advanced AI.

Here’s the full transcript of the scathing rebuttal:

The Global Diffusion of AI Is Now Irreversible

The emergence of DeepSeek isn’t just about a single AI company making a breakthrough—it’s a signal shift in the global AI competition. In his article, Dario Amodei spends a lot of time wrestling with a few key questions: What does DeepSeek’s success mean? Is the U.S. AI lead being shaken? Can export controls stop China from entering top-tier AI competition? But at its core, this debate isn’t really about DeepSeek—it’s about global technological sovereignty, economic power structures, and the fundamental nature of AI progress.

DeepSeek’s breakthrough isn’t about "surpassing" OpenAI or Anthropic. Instead, it’s the predictable result of large-scale economic forces in AI development. From 2024 to 2025, any team with sufficient capital and engineering capabilities could reach this level by following the well-established trajectory: Lower costs → Larger models → Reinforcement learning & reasoning optimizations. This isn’t some Chinese AI miracle—it’s simply the inevitable next step in the AI race. DeepSeek just happened to be the first to hit this milestone.

DeepSeek’s strength doesn’t come from a revolutionary new architecture. It comes from pushing engineering to its absolute limits: Key-Value caching, Mixture of Experts (MoE), Reinforcement learning with human feedback (RLHF) & Chain-of-Thought (CoT) reasoning. These are all efficiency-maximizing techniques—not the fundamental research breakthroughs that OpenAI, Anthropic, or DeepMind pursue. DeepSeek’s achievement is about making the best use of existing tools—optimizing every piece of the AI pipeline to reach the curve’s leading edge with fewer resources. But it hasn’t created new math, new cognitive paradigms, or a new AI framework.

Dario still frames the debate around the old paradigm—that export controls can decide the outcome of the AI race. But the reality? That’s no longer true. In the short term, yes, export restrictions make it harder for Chinese teams to access high-end GPUs. Some breakthroughs may be delayed. But in the long run? Capital markets, supply chains, and China’s alternative hardware strategies make this a battlefield that can’t be "sealed off". AI isn’t like nuclear weapons, where one country can block development. Its industrial logic is closer to energy, semiconductors, or cloud computing—a technology that gradually converges into an infrastructure-level competition, rather than something export bans can fully control.

OpenAI has accused DeepSeek of "distilling" ChatGPT as part of its training process, suggesting that China’s AI progress is built on U.S. intellectual property. But here’s the inconvenient truth Dario didn’t fully acknowledge: Even if DeepSeek had zero access to OpenAI’s data, it could still have reached R1-Zero’s level by following the reinforcement learning trajectory. Why? Because AI progress isn’t happening in isolation. Every advancement in AI is built upon: Open-access research papers, Open-source codebases, Shared best engineering practices. DeepSeek’s R1-Zero is not about "stealing" knowledge—it’s proof that AI is now in a "global knowledge diffusion" era that can’t be reversed.

At the end of his piece, Dario poses a critical question: What happens if, by 2026–2027, China also gains access to millions of high-end GPUs? His answer? If China is blocked from advanced GPUs, the U.S. will establish an "AI unipolarity", where reinforcement learning cycles continue strengthening U.S. dominance. If China secures them, the U.S. may face a long-term AI arms race.

But here’s the reality: Whether AI remains unipolar or shifts to bipolar competition, the global AI race has already reached an irreversible phase of diffusion. The U.S. won’t monopolize AI. China won’t be blocked. Europe, Japan, and India won’t be left out. Export controls, model competition, and capital flows may shape the tempo of AI development. But nothing will stop the world from moving toward more advanced AI. DeepSeek isn’t the finish line. It’s a signal—and its significance isn’t about who it “defeated,” but about proving that the world has entered the irreversible era of large-scale AI competition.