Walking into the office, a partner from a major American investment firm was surprised to find everyone using Chinese AI.

【By Observer Net, Yuan Jiaqi】

Walking into the office of the renowned Silicon Valley venture capital firm Andreessen Horowitz (a16z), partner Martin Casado found that the AI models used by the startups it invests in are likely all from China.

"I would say there's an 80% chance they're using Chinese open-source models," he added.

The change began in January this year. At that time, the Chinese startup DeepSeek opened a low-cost advanced AI model for free, making a name for itself and shaking the global stock market.

Following the "DeepSeek shockwave", Chinese AI models from tech giants like Alibaba and other companies also quietly gained more attention overseas.

The Economist published an article on the 21st analyzing that unlike American giants who are spending billions to crack competitors' proprietary models, Chinese companies are waging a completely different war. In the words of Stanford University AI expert and "father of Google Brain" Andrew Ng, this is a Darwinian survival struggle among Chinese large language model (LLM) developers over who is more open.

The Economist pointed out that in various intelligent tests released this year, the performance of Chinese open-source models exceeded similar models launched by American companies such as Meta, and their capabilities are continuously approaching top proprietary models. The article believes that the competitive enthusiasm of Chinese developers should sound a warning bell for the West.

Taking OpenAI, the developer of ChatGPT, as an example. Its CEO Altman recently admitted in an interview that the "open-source battle" in China has put significant pressure on OpenAI, forcing it to change its model release strategy.

Sam Altman, CEO of OpenAI, photo

During the mid-2010s, OpenAI had promoted the idea of "more openness" in the AI field, and the company's name was derived from this. However, after 2020, in order to maintain profitability, OpenAI stopped being "open" and instead sold its proprietary large language models, refusing to fully open-source its technology.

Until recently, Altman suddenly realized something was wrong—the usage rate of open-source models and open-weight models, including those from Chinese models, among his customers had significantly increased.

Envious and wanting to get a piece of the pie, OpenAI quickly launched two open-weight models, gpt-oss-120b and gpt-oss-20b, in August, which were freely available and supported developer customization.

This was the first time since OpenAI launched GPT-2 in 2019 that it had released open-weight models, and these were also the first such models to emerge in six years since it signed an exclusive cloud service agreement with Microsoft.

A US media outlet once hyped the move as a "major strategic shift" for the company, which had long managed its technology in a closed manner, indicating that OpenAI intended to bet on "increasing technological accessibility" to expand the developer ecosystem and strengthen its advantage against Chinese competitors.

The detail of naming these models with lowercase letters carries significant meaning. Not only are they relatively smaller in scale, but the market response after their launch has been mixed. Many voices have criticized them for lacking highlights, stripping away many of the core powerful functions of OpenAI's commercial products.

At the same time, OpenAI also released the highly anticipated new proprietary model GPT-5, but it faced a reputation disaster, forcing it to revert the default model of ChatGPT to the previous version due to the overwhelming criticism.

This defeat shocked Altman. But according to The Economist, this merely proves that OpenAI's so-called "embrace of openness" is insincere, and other American companies may be the same.

Not to mention some American companies that are going back in history.

While Altman realized that OpenAI's past approach of keeping models closed was "on the wrong side of history," Meta, which has gained widespread acclaim in the open-source community for its open AI model Llama, is now performing a "betrayal of its roots."

The CEO Zuckerberg declared that the company will be more cautious in choosing open content. He is also dedicated to building a "super intelligence," and is willing to hire top researchers from OpenAI, Google, Apple, and Anthropic with compensation as high as hundreds of millions of dollars, causing dissatisfaction among peers.

Alexandr Wang, founder and CEO of the US AI startup Scale AI, was also lured into the team to head a new laboratory.

This entrepreneur, who once vowed that "American AI must not lose to China," upon taking office, immediately planned to abandon Meta's most powerful open-source AI model Behemoth and develop a new closed-source model.

When American scholar Ethan Mollick saw this news in the media, he sighed. This AI expert, who has been deeply involved in the open-source field for over twenty years, started dealing with large language models early on, and was praised by US media as "the preferred AI expert for American policymakers and corporate leaders."

Even Meta is turning to closed-source, which made him sigh, "In this case, the United States has basically withdrawn from the competition in cutting-edge open-source large language models. Europe still has one competitor, while the rest of the market is almost entirely dominated by China."

On July 14, Ethan Mollick posted on social media

Ali Farhadi of the Allen Institute for AI also told The Economist that Chinese companies are doing everything they can to publicly release their best models; while American companies prefer to hold on to "shiny new things" as proprietary assets.

He spoke frankly, "Although it's hard to accept, I do believe we are now (in the open-source field) behind."

From a business perspective, the result of different path choices between Chinese and American companies is vastly different. Proprietary models created by American companies generate far more revenue than open-source models in China. Proprietary models are undoubtedly easier to make money, and the revenue can be used to support innovation research and development, which is an obvious advantage.

But compared to proprietary technology, open-source models can promote completely different application scenarios. Percy Liang, co-founder of the open-weight model platform "Together AI," explained that enterprises, governments, and researchers can more easily adapt open-source models to various "niche scenarios." These models can also help users run AI tools locally without relying on cloud services. Additionally, through providing auxiliary services such as customized support, developers can still achieve profitability.

In other words, while American laboratories are heavily betting on pushing the frontier of intelligence to earn huge profits, their Chinese competitors focus more on encouraging the widespread adoption of artificial intelligence.

This sounds a warning bell for The Economist, "If they (Chinese manufacturers) succeed, the impact of DeepSeek might just be the beginning."

This article is an exclusive contribution from Observer Net. Unauthorized reproduction is prohibited.

Original: https://www.toutiao.com/article/7541307958039249451/

Statement: The article represents the views of the author and is not necessarily endorsed by the publisher. Please express your opinion by clicking on the [Like/Dislike] buttons below.