Daniel Remler

The UK AISI has never published anything on any Chinese model—would be nice if they did and settled this debate!

@ChrisPainterYup Very likely—Italy also had a G7 AI session for leaders (with Pope Francis) when they hosted in 2024. I doubt Sam and Dario were asked to stick around and talk about Ukraine, Iran, etc… But I agree the optics are striking.

@danielremler Interesting! Maybe the thing that’s striking me as odd is that this feels like “an official G7 meeting”, but I’m sure this is some kind of AI focused session?

I’m no expert on protocol, but in an earlier time would it have been odd to have the leaders of companies (and most of them American) sitting as interleaved co-equals with world leaders like this? Just feels really noticeable here

And for more on how adversarial distillation works read this recent paper by me and @BenHayum: cnas.org/publications/r…

Report by @eastland_maggie: bloomberg.com/news/articles/…

Alibaba now accused of conducting a massive campaign of adversarial distillation against Anthropic—unsurprising considering they face the same compute and expert data constraints as the rest of the PRC AI ecosystem. The question was always when, not if they would do this.

❗Tomorrow❗ @KhanSaifM, Emily Weinstein, @danielremler, and @ashleyrgold will outline potential U.S. responses to an increasingly critical challenge—the rapidly maturing Chinese AI ecosystem and its instrumentalization by the CCP.

Microsoft is reportedly considering a partnership with Chinese AI firm DeepSeek. A new report from @CNASdc says "DeepSeek-based agents are 12 times more likely to follow malicious instructions than their U.S. counterparts." x.com/CNASdc/status/…

AI is increasingly a core pillar of China's vision of global governance, nearly always framed in terms of the centrality of the UN and supporting developing countries. The "proposed" World AI Cooperation Organization sure sounds like it will be announced soon...

@dee_bosa And what kind of testing will MSFT do before they onboard the models—they haven’t disclosed doing any testing before hosting any Chinese model on Azure since DeepSeek-R1.

Was only a matter of time. Deepseek already hosted on hyperscaler clouds since last year… Microsoft moving closer to it for enterprise just normalizes DeepSeek and gives cover for others to embrace/adopt (which is already happening) Bigger q: does this give the Chinese stack a foot in the door in the US since DeepSeek is optimizing for Huawei chips?

@shashj Commerce/BIS isn't immune from judicial review—Anthropic could sue and IMHO has a pretty reasonable case (APA violation, exceeding statutory authority, etc.).

Given this gap, I expect China to produce a "Mythos-class" model by or around the end of 2026, probably with open weights. That and other capabilities will mark a meaningful threshold that puts Americans at serious risk (as I argue in my new report: cnas.org/publications/r…)

That puts DeepSeek (current Chinese lead) at least 6 months behind Anthropic (current American lead). I'd add another 2+ months given China's faster product cycle (less safety testing) and limited compute. Countering adversarial distillation could soon widen this gap further.

My best guess is that China is at least ~6-8 months from producing a model as cyber-capable as the (currently export controlled) Mythos and Fable 5, based on a comparison of the leading developers on the best agentic cyber benchmark plus some additional assumptions.

The report offers 6 recommendations—rapid assessments, cybersecurity advisories, testing standards, allied coordination, and semiannual ecosystem reporting. All doable with existing authorities. Full report here: cnas.org/publications/r…

Export controls have helped maintain the U.S. lead. But the United States, our allies, and the U.S. private sector have exerted near-zero pressure on Chinese AI systems specifically. Commerce has published just three, thin assessments of Chinese systems, each delayed weeks. /8

The Western AI community has spent years scrutinizing U.S. developers while giving China a free pass—I turn the tables in my new @CNASdc report, "Red Lines: Understanding the National Security Risks of China's Advanced AI": 🧵