Is this a Sputnik moment? The world has reacted with astonishment to the release of a disruptive AI model from Chinese company DeepSeek, which appears to be able to perform as well as or, in some cases, better than ChatGPT and other cutting-edge models put out by U.S. companies. Americans had assumed their massive lead in funding, access to high-quality chips and innovation would keep them well ahead. That assumption now looks like hubris.
DeepSeek has created a 21st-century Sputnik moment
Export controls may have helped turbocharge Chinese innovation.
January 31, 2025 The Washington Post
The episode is in some ways a much bigger deal than Sputnik. Sputnik was about the Soviet Union’s space program competing with that of the United States. Few thought the Soviet economy in general was more technologically advanced than America’s. But DeepSeek is a private Chinese company that demonstrated its stunning prowess on the cheap in the most important technology for the future. It’s not exactly clear just how much DeepSeek’s model actually cost, to what extent it needed to use U.S. models for training and whether there was any closet Chinese government help. But given the enormous efforts that the U.S. government has made over the past few years to preserve its advantage — chip bans, export controls, etc. — DeepSeek has made a remarkable achievement. It suggests to me two lessons and two questions.
The first lesson is that, over time, open artificial-intelligence systems are likely to outperform closed systems. (An open system is like Lego blocks with instructions; a closed one is the built Lego structure with the instructions kept secret.) Many have pointed out that DeepSeek used Meta’s open-source Llama model to train. It also used Qwen, a family of AI models, also open-source, put out by the Chinese technology giant Alibaba. While DeepSeek is currently the best, China’s big technology companies have been releasing a number of AI models, mostly open-source, that are getting better and better. If the history of technology is any guide, the ability to see the innards of these models and understand their reasoning should lead to greater and faster technological innovation than using closed models that others cannot use for collaboration.
Second, constraints can be useful, as former Intel CEO Pat Gelsinger has noted. Just as art sometimes flourishes in repressive environments, in which restrictions force artists to be creative, so also engineers often operate best under constraints. Forced to use second-tier chips, Chinese engineers produced creative workarounds. This is not just true with DeepSeek. In 2023, Chinese telecommunications giant Huawei released a smartphone with a seven-nanometer chip, a kind that had been explicitly banned by U.S. export controls. There is some evidence that, after years of sclerosis, China’s chipmakers have responded to U.S. bans by becoming much more innovative.
In a fascinating interview last year, Liang Wenfeng, the CEO of DeepSeek, argued that his engineers were more motivated by doing research than making money, and appeared to contrast that attitude with the one prevalent in Silicon Valley, which is all about maximizing revenue, providing cloud services and generating cash flow. Demis Hassabis, who leads Google’s DeepMind and also shared the 2024 Nobel Prize in chemistry for AI-related scientific breakthroughs, is said to have fought to keep his team in London, far from Silicon Valley, so that it can focus on basic research.
The first question that DeepSeek raises is: Can the United States stop China from advancing along the technological frontier? Some argue that DeepSeek shows that export controls work: Its model needed many Nvidia chips, which it managed to procure before export bans were fully in place. Soon, China will not have access to the best chips and will suffer even more from the ban.
But as we have learned with the rounds and rounds of global sanctions against Russia, the world economy is large and porous. Stuff gets through. And China is not Russia. It is a vast, technologically sophisticated economy with millions of software developers and hundreds of high-quality firms in the technology space. Human talent on that scale will find ways to innovate, even if those measures keep China slightly behind.
The second question: What is the cost of this approach? If technology bans and export controls at best keep China behind a year — maybe just several months — is that gain worth the cost? That cost is Chinese retaliation, limiting the United States’ access to key materials that it needs for high technology. More important, a decoupled global economy also creates a closed ecosystem in which U.S. technology companies will not face competition from the best. Is Tesla going to innovate at the highest level if it is not facing its strongest Chinese rival?
A technology decoupling means that AI will become the central part of a new global arms race, totally unregulated and unconstrained, with the world’s two largest economies hurtling toward superintelligence no-holds-barred, and incorporating it into all military applications — including nuclear weapons. If artificial intelligence is as revolutionary a technology as predicted, having it unleashed in every realm of human life with absolutely no guardrails points to a scary future — one far more dangerous than anything people imagined because of the Sputnik satellite.
@FareedZakaria