"Our Watermark Is Right There in Chinese AI": The 'Kimi' Shock That Enraged the US

01. "Watermarks Discovered": US Treasury Secretary's Direct Offense

Treasury Secretary Scott Bessent delivered a direct blow during an interview with Fox Business on July 21, local time. "We have found watermarks of American Large Language Models (LLMs) in many Chinese AI models. This is unacceptable." He claimed that Chinese companies used the outputs of American models via a technique known as 'distillation' to train their own smaller models.

Distillation is an AI training method that uses the output of an existing, powerful model to build smaller, lighter models, which the US views as virtually 'technology theft.' While originally a legitimate technique for model compression, the core legal and political issue lies in acts that violate the commercial terms of use of American AI firms by employing fake accounts to extract massive amounts of output data—a practice known as a distillation attack.

Bessent warned, "Over the next few days or weeks, we will review this matter closely. If foreign models steal technology from our great companies, we have the capacity to sanction them." In a letter sent to the Senate Banking Committee last month, Anthropic claimed that China's Alibaba launched the "largest distillation attack in history" targeting its systems. Anthropic specified that Alibaba mobilized around 25,000 fake accounts between April 22 and June 5, 2026, to harvest over 28.8 million response data entries from its Claude model. Bessent's remarks are interpreted not merely as a warning, but as a clear signal pulling out bargaining chips publicly ahead of the September AI talks.

02. "Performance Gap Has Vanished": Kimi K3 Shock and Silicon Valley Unrest

At the center of the controversy is Kimi K3, released on July 16 by Chinese startup Moonshot AI. Claiming the ability to process 2.8 trillion parameters, the model demonstrated performance matching the latest models from Anthropic and OpenAI in several benchmark tests. According to Moonshot AI's announcement, while Kimi K3 fell short of Claude Fable 5 and GPT-5.6 Sol in overall comprehensive rankings, it outperformed Claude Opus 4.8 and GPT-5.5 in coding and agent evaluations. Kimi K3 maximized training efficiency by incorporating an ultra-long context processing capacity of 1 million input tokens, alongside a hybrid linear attention architecture (Kimi Delta Attention, KDA) and Attention Residuals (AttnRes).

Silicon Valley was particularly shocked when Kimi ranked first in front-end coding evaluations, surpassing both US AI labs. The bigger concern is that Kimi is an 'open-weight' model. While not entirely open-source, it is designed to allow the public download of its trained final parameters. Moonshot AI announced that it would release the full weights on July 27, and at the time of Secretary Bessent's remarks, downloads were not yet available. In contrast, OpenAI and Anthropic have stuck to closed models, restricting access.

Right after Kimi's release, the Nasdaq index stumbled. On July 17, the day after Kimi K3's debut, the Nasdaq fell 1.4%, the S&P 500 dropped 1%, and the Dow sank 407 points (0.77%). Taiwan's Taiex index tumbled by over 6%, while Japanese stocks fell 4%. Concerns spread that the rapid advance of Chinese models could undermine the very logic behind American AI infrastructure investment. Chinese state media declared that "the performance gap between Chinese and US AI models has virtually disappeared," a sentiment shared by some experts. Analysis suggests that the AI technology gap, once said to favor the US by more than three years, has now narrowed from 'years' to 'months.'

03. "September Talks Are a Watershed Moment": At the Crossroads of Sanctions or Coexistence

According to Reuters, the United States and China plan to hold a high-level meeting in September to discuss AI regulation. This is one of the key outcomes of the Trump-Xi summit in May and is likely to take place just before President Xi Jinping's visit to the US scheduled for September 24. Treasury Secretary Bessent is reportedly set to serve as the US representative.

The core agenda for the meeting is clear. The US is expected to push for potential restrictions on the domestic use of Chinese open-weight models, while China is likely to raise concerns about the hacking capabilities of advanced US models—such as Anthropic's unreleased Mythos model—and demand controls over international access. Anthropic's Claude Mythos is a frontier model with cyber capabilities powerful enough to detect software security vulnerabilities across major OS and browser platforms. However, US government controls are not ongoing. The Department of Commerce issued export control guidelines on Claude Fable 5 and Mythos 5 on June 12 citing national security, temporarily suspending global access, but lifted the controls on June 30, with Anthropic sequentially restoring access starting July 1.

Last year, the US Department of Commerce considered placing several Chinese AI labs on the 'Entity List' to effectively block their access, and the National Security Agency (NSA) along with the White House Office of the National Cyber Director prepared advisory recommendations regarding threats from Chinese AI labs. However, as open-access advocates such as former White House AI Tsar David Sacks and Sriram Krishnan departed the administration, protectionist voices have been gaining momentum.

Meanwhile, China is not sitting idly by. The Financial Times reported that China's Ministry of Commerce is considering measures to limit foreign companies' access to core AI and semiconductor data. Experts predict that "while it will be difficult to resolve major issues in the first meeting, they will attempt to reach consensus starting with basic terminology, such as the definition of frontier AI models." The September meeting will serve as the first test gauging the direction of the US-China technological order in the AI era.

---

At a time when assessments suggest America's AI lead—once estimated at more than three years—has shrunk to just months, the US is shifting the front line from a technology race to a battle over rules. The world is watching closely to see whether the September meeting will open the floodgates for cooperation or mark the dawn of a new Cold War.


Deep Dive Q&A
Q1. What is the 'Distillation' technique, and why does it ignite intellectual property (IP) infringement controversies?

A. Distillation is a technique that uses high-quality responses (outputs) obtained by querying high-performance Large Language Models (LLMs) as a training dataset to build smaller models. While widely used for model compression, it allows smaller models to 'free-ride' on and replicate the astronomical development costs and know-how poured into a competitor's massive model. This makes it a core issue regarding violations of commercial terms of service and unauthorized technology theft (IP theft).

Q2. Why did the release of Moonshot AI's 'Kimi K3' send shockwaves through Silicon Valley and financial markets?

A. Kimi K3, despite being an ultra-large model with 2.8 trillion parameters, was planned for release in an 'open-weight' format, allowing users to download and utilize its weights. This rattled markets by potentially destabilizing the proprietary monetization models and rationale for infrastructure investments of US AI firms that spent trillions of won to build closed models.

Q3. What are the core issues and key points to watch in the upcoming US-China high-level AI talks in September?

A. The US advocates blocking IP theft via distillation and regulating open-weight models, whereas China raises concerns about the national security hacking risks of ultra-high-performance US frontier AI models (such as Anthropic's 'Mythos') and demands limits on overseas access. The primary point to watch is whether the conflict has evolved from a race over technological speed into a 'rule-making competition' to establish global AI standards and security regulations.
※ Please note: This article was translated by AI and may contain errors.