US-China AI Rivalry Enters a New Phase: Sanctions Talk Over Moonshot Distillation Allegations

On July 22, the debate in Washington over how to deal with Chinese AI companies heated up rapidly. A senior White House official suggested that Kimi K3, the new model from China's Moonshot AI (月之暗面)， was developed by improperly distilling Anthropic's closed model Fable.

In response, Treasury Secretary Scott Bessent raised the possibility of the US government imposing sanctions on Chinese AI companies. WIRED reports that within the administration, haggling continues between sanctions hawks and caution advocates over how to respond to powerful Chinese-made AI models.

Meanwhile, OpenAI President Greg Brockman praised Kimi K3 as "quite well done" while saying he was "not sure" whether distillation was involved. A new dilemma has surfaced for the AI industry: a model that has technically reached parity is now being questioned over the transparency of its process.

The issue now carries multiple fronts at once — beyond the technical debate over model distillation, it also encompasses the US-China race for AI supremacy, intellectual property protection, and the boundary between open and closed models.

Anthropic Settles for $1.5 Billion — A Landmark in AI Copyright Litigation

Anthropic has reached a $1.5 billion (roughly ¥220 billion) settlement in a copyright lawsuit brought by book authors. It is the largest loss to date for an AI company.

According to The Decoder, the settlement has been described as "the biggest legal defeat" for an AI lab. Multiple lawsuits over the legality of using copyrighted works as training data are still working through the courts, and this settlement could become the reference point for the entire industry's legal risk going forward.

At the same time, the fact that the case ended in a settlement is also notable for laying out a compensation framework while avoiding a binding court precedent. For AI companies, it's a clear signal that the cost of using creative works must be factored in.

Google Earnings: AI Demand Drives the Cloud, Backlog Reaches $514 Billion

In Alphabet's (Google's parent company) Q2 earnings, the cloud division posted strong results that comfortably beat Wall Street expectations. The cloud backlog reached $514 billion, underscoring robust demand for AI infrastructure.

Q2 profit, including investment gains, came to $98 billion. The company's cloud growth directly reflects demand for the computing resources needed to develop and operate AI applications.

The Financial Times also reported that Google's earnings have surged on the AI boom, pointing out that hyperscalers have emerged as the biggest beneficiaries of AI demand.

OpenAI's "Human Error" Enabled the Hugging Face Hack

It has emerged that a misconfigured test environment — one OpenAI had described as "highly isolated" — enabled an AI-powered attack on Hugging Face. According to TechCrunch, cybersecurity experts have identified human error on OpenAI's side as the root cause of the incident.

Rather than exposing dangers in the AI models themselves, the incident surfaced problems in the operational and governance practices of the organizations running AI systems — a lesson for the entire industry. It also raises renewed questions about the weight of the responsibility AI companies take on when they tout a "secure test environment."

Cursor Router Arrives: Auto Mode Evolves Into a "Smart Router"

Cursor, the AI coding tool, has announced Cursor Router, a major overhaul of its Auto mode. The old Auto mode was a black box that "picked whichever model felt about right"; with this update, it now identifies the type of task for each request and routes it to the optimal model.

For developers, the benefit is that the tool absorbs the overhead of juggling multiple LLMs. The trend of model selection itself being abstracted away from the development workflow is now unmistakable.

Claude Code Lead on "Five Types" of Team Members in the AI Era

Boris Cherny, who leads Claude Code development at Anthropic, has offered an interesting perspective on how engineering teams should be composed in the AI era. Cherny argues that the people working on his team can be sorted into "five types."

As the boundaries between job roles blur, he says we have entered an era where people should be understood not by their job title but by how they collaborate with AI. ITmedia has published an accessible breakdown of the idea for beginners.

The notion of organizing teams around how people leverage and creatively apply AI — rather than traditional role distinctions like "frontend engineer" and "backend engineer" — offers a valuable perspective for many development organizations.

---