American renowned artificial intelligence scholar Gary Marcus posted on his personal blog on July 20th, stating that Chinese artificial intelligence models have “almost caught up” with the most advanced levels in the United States. It is impossible for the United States to “win” this AI competition. Washington should stop viewing AI as a zero-sum game and instead explore paths of international cooperation and public goods.
Marcus is a renowned artificial intelligence scholar in the United States and an honorary professor at New York University. He has long been engaged in research on machine learning and cognitive science. In recent years, he has attracted much attention for advocating the development of more reliable general artificial intelligence.
Marcus mentioned in his blog that the Kimi K3 model recently released by Chinese artificial intelligence company Moonshot AI is now on par with top American models in terms of performance. Moreover, it uses an "open weight" model, allowing users to download it for free and run it locally. He believes that this news was one of the reasons for the decline in the stock market-related sectors in the United States last week, and it has once again shaken the business models of companies like OpenAI and Anthropic.
On July 20, 2026, spectators visited the Kimi booth featuring “The Dark Side of the Moon” at the 2026 World Artificial Intelligence Conference held in Shanghai. IC Photo
He pointed out that before the Kimi K3, the Zhipu GLM 5.2 and Alibaba’s new Tongyi Qianwen model also attracted widespread attention.
This is not an accidental event, but rather a trend.
Marcus stated that as early as the beginning of 2025, after DeepSeek’s release, he predicted that if the United States continued to focus its resources on large language models (LLMs), it would not be able to gain a decisive advantage over China; ultimately, both sides were more likely to end up in a draw.
He also predicted that due to the lack of a true technological moat, OpenAI would find it difficult to achieve stable profitability. NVIDIA would also face competitive pressure. The US CHIPS Act is limited in its ability to contain China. Models will become cheaper and more efficient, but issues related to hallucinations and reliability will remain long-term.
Now it appears that most of these things have happened.
In Marcus's view, the U.S. government has been too close to Silicon Valley, almost taking Silicon Valley's vision for generative AI as reality, leading to strategic misjudgments. He criticized that the U.S. "should not have placed all its bets on generative AI from the beginning," because this field has never shown strong technological barriers.
Regarding the current situation, he suggests that the U.S. Congress conduct an investigation into why the United States has lost its leading position, whether there was an over-reliance on a single technology, whether there is insufficient protection for intellectual property rights, and whether restrictions on immigration have led to a loss of talent. He specifically mentioned that some founders of Chinese AI companies studied in the United States and then returned to China to develop their businesses, which deserves reflection by the U.S.
More importantly, it's a strategic choice. Marcus listed seven possible options that the Trump administration might consider: not providing subsidies, banning open-source software, creating "moats" for American companies through regulation, supporting large AI laboratories, completely blocking Chinese models, and nationalizing OpenAI and Anthropic.
He is negative towards most of these proposals. In particular, he opposes the use of regulatory measures to exclude Chinese competitors, believing that this would drive up prices, suppress innovation, and harm American startups.
He also opposed the government funding ‘shelter’ for companies like OpenAI, stating that these companies are currently ‘mostly loss-making projects’, and the AI industry has not yet proven its ability to generate sustainable profits.
Marcus finally suggested that his preferred solution is to abandon the zero-sum mindset of 'winning the AI war', and to promote the establishment of an international AI cooperation mechanism similar to that of the European Organization for Nuclear Research (CERN), aiming to make artificial intelligence a global public good.
He recalled that as early as 2016, he proposed the idea of establishing an ‘AI version of CERN’. This would enable scientists from various countries to collaborate around clear goals such as medicine and science, and share their achievements with the world. It should not be monopolized by a few countries or companies.
The article specifically mentions that China recently reiterated its willingness to work with other countries to promote the beneficial and inclusive development of artificial intelligence at the Artificial Intelligence Conference, opening an important window for international discussions.
“Perhaps it’s finally time to seriously consider this matter,” he wrote. “If AI can be returned to the public domain and used for medicine and science through international cooperation, that might be the most promising direction at present.”
Marcus urged Trump that if he really wanted to win the Nobel Peace Prize, such an action might be an opportunity. Cooperation with China and using artificial intelligence for the benefit of the public would be a “generous gift given to humanity”.
