A major investigative report on AI surveillance towers along the US border, US–China consultations on AI security, the shifting balance of power in open models, and DeepSeek's plans for an extremely large-scale model — today brings noteworthy news on both fronts: surveillance technology and policy, and the model race.
US and China in Talks Over a Mutual Notification Mechanism for AI National Security Threats
According to WIRED, officials from the United States and China are in discussions about building a mechanism for mutually notifying each other of AI incidents that could threaten national security.
What's new here is that talks are progressing on a concrete notification mechanism, amid warnings that AI-related accidents or misuse could trigger miscalculation between the two countries. The details and timing of any agreement remain unknown at this point, but the effort can be seen as an attempt to extend to AI something akin to the hotlines that already exist in the nuclear weapons domain.
The US Border's "Virtual Wall" — Over 1,050 Deaths Within Range of AI Surveillance Towers
MIT Technology Review and Times of San Diego have published "Dying on Camera," a joint investigation spanning roughly 15 months. It is the first effort to comprehensively map deaths around the surveillance towers deployed along the US border (the so-called "virtual wall"), cross-referencing some 4,000 records of locations where bodies were found with information on roughly 600 towers identified by the EFF (Electronic Frontier Foundation), combined with terrain analysis.
The findings show that since 2015 alone, more than 1,050 people have died within the towers' surveillance range. Deaths occurred within range of nearly two-thirds of the towers analyzed, and the investigation also revealed terrain constraints: some towers can actually see as little as around 10% of their advertised surveillance range. Even the autonomous surveillance towers (ASTs) built by Anduril, regarded as among the most advanced systems, have seen more than 110 deaths within their range since 2021.
Drawing on interviews with more than 45 people, the reporting documents how these factors overlap: tower-side failures such as detection algorithms misfiring (including misclassifying groups of people as "livestock") and towers that had broken down, alongside agents failing to respond to alerts. According to a GAO (US Government Accountability Office) report, 803 towers are currently operational, and plans are underway to expand to 2,300 by 2034. Including the additions, this amounts to an investment of more than $6 billion, yet no framework is in place to verify effectiveness — prompting the investigators to call for a comprehensive audit of the deaths and improved data record-keeping.
Notably, the investigation also disclosed that Anthropic's Claude (via API) was used to process the vast trove of police records, with coordinate extraction performed in combination with human verification. It stands out as a real-world example of LLM use in journalism.
The Open-Model Balance of Power — "China Has a Clear Lead," Congressional Testimony Says
AI researcher Nathan Lambert has published the contents of a briefing he prepared for members of the US Congress. It frames open-weight models through the lens of US–China competition, and the current numbers look like this:
Chinese-origin models have racked up 3.2 billion cumulative downloads on Hugging Face, roughly twice the figure for US-origin models (the reversal came in July 2025)
The top three spots on the Artificial Analysis Intelligence Index are all held by Chinese models: GLM-5.3 (45 points), Kimi K3 (44 points), and GLM-5.3-Flash (42 points). The top US entry is Thinking Machines' Inkling (26 points)
Chinese open models are estimated to lag closed US frontier models by 2–5 months, while US open models are estimated to lag by 6–9 months
Penetration into academia is just as striking: a scan of papers across five major arXiv categories found Qwen mentioned in 30% of them, overtaking Llama (21%). Chinese models now account for over 80% of share on OpenRouter and more than 95% on the open-source coding agent OpenCode. What's more, a growing set of US companies — Harvey, Cursor, DoorDash, Airbnb — are building Chinese-made open models into their products.
Lambert estimates that even if distillation (training on the outputs of advanced models) were completely prevented, the gap with China would widen by only a month or two. Rather than cracking down through regulation, he argues, the United States should keep investing in its own open models and build ecosystem-level preparedness against the risks.
DeepSeek Is Training a 2-Trillion-Parameter Model, With an 8T Model Also Reportedly Planned
A post is drawing attention in Reddit's LocalLLaMA community claiming that DeepSeek is training a 2-trillion (2T) parameter model and also plans to build an 8T model down the line.
The current models are said to be 552 billion parameters for Flash and a total of 1.6T parameters for Pro (49B activated per token). Keep in mind that the post is a social media rumor, not an official announcement from DeepSeek. Even so, the community is debating it vigorously as a sign that parameter counts keep growing at the frontier of the open model race.
Amazon Blocks Meta's Muse Agent From Checkout
Amazon has reportedly blocked Muse, Meta's AI agent, from completing checkout (the purchase process) on its site.
With the arrival of an era in which AI agents buy products on users' behalf on the horizon, the move can be read as the world's largest retailer working to keep control of the "gateway to payment" in its own hands. It is drawing attention as one of the earliest cases of inter-platform tension in the agent economy surfacing as a concrete blocking measure.
Takeaways
The limits and social costs of surveillance AI, risk management between nations, structural shifts in the open model race, and the boundaries of the agent economy — all of these are changes happening at the points where "AI connects with the real world." The border surveillance investigation in particular exposes the precariousness of ever-expanding investment in AI systems whose effectiveness is never independently verified, and it looks set to become a reference point in future policy debates.