AI Daily Briefing — 2026-07-23

As Chinese open-weight AI models gain traction and capability, debate over their safety has intensified. While some in the Trump administration have floated potential bans, and major US labs like OpenAI and Anthropic view them as competitive threats, Arcee—a US-based open-source AI lab—argues the security concerns are overblown. These models, such as Moonshot AI’s Kimi K3 or Alibaba’s Qwen, offer inference at far lower token costs than proprietary alternatives, but critics worry they could serve as vectors for Chinese hacking.

Lucas Atkins, CTO of Arcee, says such fears misunderstand how large language models work. “There is really not any way for an Arcee, or an Alibaba, to make a model, have someone run it in their own environment and for us to have any access to it whatsoever,” he explained. He compares Chinese open-weight models to any other open-source software: they can be inspected, tested, and post-trained by enterprises before deployment. While the models are not fully open source—training data and methods remain hidden—the executable code is visible and reviewable on platforms like Hugging Face.

Atkins also downplays the risk of models intentionally inserting malicious code. Though theoretically possible, he says it would require extraordinary effort to trigger hidden behaviors. “I don’t know how you would do this,” he admitted. He notes that enterprises are increasingly building model-agnostic AI apps using multiple models, reducing lock-in. Rather than banning Chinese models, Atkins advocates fostering a robust open ecosystem in the US. Arcee, which builds open models for domestic companies, would benefit from a ban but instead promotes collaboration and scrutiny over fear.

In a recent episode of TechCrunch's Equity podcast, Menlo Ventures partner Matt Murphy offered rare insight into what separates today's fastest-growing AI startups from the rest. Drawing on 25 years of investment experience, Murphy noted that the speed of growth at companies like Anthropic is unprecedented, surpassing even the internet and mobile booms. Menlo led Anthropic's $500 million Series D when the company had no revenue, a bet that has since yielded extraordinary returns as Anthropic reached a $47 billion revenue run rate by May.

Murphy emphasized that a great AI model alone is no longer a sustainable advantage. He pointed to Anthropic's expansion into tools like Claude Code and MCP as evidence that platform-building, not model performance, creates lasting value. He also addressed criticism of Anthropic's Mythos rollout, pushing back on claims that marketing overshadowed safety, and highlighted how startups like Lovable and Legora are growing faster than any he has seen in his career.

For founders trying to compete, Murphy's core message is clear: speed and platform strategy matter more than ever. The full conversation covers why early bets on pre-revenue companies can pay off, how major investors like Google and Amazon served as validation, and what it takes to build a defensible business in the current AI landscape.

A new free, open-source Chinese AI model called Kimi, developed by Moonshot, has sparked fierce debate among President Donald Trump's AI advisers. The model reportedly matches the intelligence of costly US systems from OpenAI and Anthropic, but without the price tag. This has created a rift within the White House, as some advisers argue that Kimi undermines American AI companies' business models, while others see it as a competitive threat requiring action.

Last weekend, tensions boiled over publicly. David Sacks, a former Trump adviser, called Anthropic's models "lobotomized" and "woke," while Pentagon official Emil Michael labeled OpenAI's new strategic futures head a "supreme village idiot." The infighting stems from disagreement over how to respond to Kimi's launch, which has revived calls for restrictions on Chinese AI imports.

Meanwhile, Anthropic has secured a record $1.5 billion copyright settlement, the largest known payout in history, after plaintiffs accused the company of using pirated works to train its Claude model. Despite the payout, many authors and creators remain unsatisfied, viewing it as insufficient. Critics warn that AI copyright anxiety could stifle creativity, a concern echoed by experts.

As the Trump administration weighs a potential ban on Chinese AI models, Beijing is tightening its own export controls on AI technology and chips to prevent Western acquisition. The divide within the White House highlights the growing economic and political challenges posed by China's open-source AI strategy, which is paying off as US companies struggle to compete with free alternatives.

A new free, open-source AI model from China has ignited a fierce debate among President Donald Trump’s current and former AI advisors, exposing deep divisions over how to respond. The model, called Kimi, was released last week by Chinese company Moonshot and reportedly matches the intelligence of paid US rivals like OpenAI and Anthropic. This has triggered a war of words, with David Sacks, Trump’s former AI czar, calling Anthropic’s models “lobotomized” and “woke,” while Pentagon official Emil Michael labeled an OpenAI executive a “supreme village idiot.”

The core problem for Trump is that Kimi and similar Chinese models undercut US AI companies’ business models, threatening economic growth tied to the sector. With enthusiasm for AI driving a significant share of the economy, these free alternatives create both economic and political headaches. Anton Leicht of the Carnegie Endowment noted on X that this poses “a threat for an administration that really doesn’t want more economic bad news.” The situation has already rattled US stocks.

Trump’s advisors are split on a solution. Sacks, now without a formal role, argues against government intervention to protect US companies, criticizing those that “want the government to eliminate their open source competition.” However, the administration’s prevailing view favors tighter control, citing national security risks. This has led to a new White House review process for AI models, which Dean Ball, a former Trump advisor now at OpenAI, criticized as a “de facto licensing regime.” Michael countered by dismissing Ball’s concerns as a “Deep State scheme.”

Amid the infighting, little attention has been paid to how Kimi achieved its capabilities. Export controls on chips to China, loosened under Trump’s deal allowing Nvidia to sell more chips in exchange for government cuts, have enabled such advances. The debate underscores a growing distrust of AI companies, with many Americans likely unsympathetic to the plight of OpenAI or Anthropic as they face cheaper competition.

Substack has introduced a new feature that lets readers see how much of a newsletter’s content was generated by artificial intelligence. The platform partnered with Pangram, an AI-detection software, to analyze posts, comments, and replies on the app. Users will now see an estimate of human versus AI-written material for any text longer than 100 characters. The tool aims to bring transparency to the growing use of AI in independent publishing.

While the move could initially hurt Substack by exposing newsletters that rely heavily on AI, the company believes it will build long-term trust. By flagging AI-generated content, Substack hopes to reduce what it calls “AI slop” and reassure readers that human-written work remains valuable. The feature does not penalize AI-assisted writing but encourages authors to add optional “how I make this” notes explaining their creative process.

Substack CEO Chris Best said the tool represents a positive use of AI, emphasizing that the hardest part of writing—having a worthwhile idea—should still come from people. The detection feature is available across the Substack app and allows publishers to scan drafts before publication. Writers can also report and remove scans they believe are inaccurate. This initiative aligns with broader industry trends, as social media and music platforms increasingly label or restrict AI-generated content.