Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Anthropic on USA vs China
TL;DR
Anthropic frames 2028 as a make-or-break year for U.S. AI leadership over China — Matthew Berman says the lab’s essay is unusually blunt, arguing democracies must stay ahead of the CCP on compute, model capability, and deployment before AI meaningfully shifts global power.
The essay’s core claim is that chips matter more than data or talent — Anthropic argues export controls on Nvidia-class hardware are the main reason China is still behind, while Berman agrees compute is crucial but thinks China’s algorithmic talent and workaround culture are being underestimated.
Berman’s biggest disagreement is the finish line: he thinks self-improving AI is real and could arrive by 2028 — while Anthropic says this is an ongoing competition with no final victory, he argues recursive self-improvement would create an uncrossable lead for whoever gets there first.
Anthropic warns a CCP AI lead would supercharge automated repression and military capability — the essay cites existing Chinese use of AI for censorship, surveillance, hacking, and PLA applications like DeepSeek-enabled unmanned swarms, and Berman says that dual-use risk is absolutely real.
The sharpest tension in the video is Anthropic vs. Nvidia on export controls — Anthropic says controls are working and Huawei may only reach 4% of Nvidia’s aggregate compute in 2026 and 2% in 2027, while Jensen Huang argues pushing China too hard could backfire by accelerating a rival stack optimized for Huawei chips.
Berman broadly buys Anthropic’s diagnosis but rejects parts of its prescription, especially on open source — he supports stopping distillation attacks and protecting U.S. innovation, but argues cheap, efficient open models will win enterprise adoption and that American AI needs open source to become the global default.
The Breakdown
Anthropic Comes Out Swinging on America vs. China
Berman opens by calling this the most concrete essay he’s seen from a major AI lab, and says Anthropic is basically panicking. The paper wastes no time: by 2028, either the U.S. and its allies stay ahead of the CCP in AI, or they risk handing the future’s rules and norms to an authoritarian regime.
Compute, Not Data, Is the Real Battleground
Anthropic’s central argument is that the most important ingredient in AI is access to chips, not data or researchers. Berman agrees that’s mostly right, but points out the irony: Anthropic and OpenAI don’t control silicon themselves, while China has world-class talent, a history of exploiting export-control loopholes, and enough ingenuity to stay close even under pressure.
Distillation, DeepSeek, and the Question of How China Keeps Up
The essay says Chinese labs remain competitive through loopholes and “large-scale distillation attacks” on U.S. models. Berman isn’t fully sold on that framing — he explains distillation simply as using a bigger model’s answers to train a smaller, cheaper one, but argues China’s proximity to the frontier likely owes as much to raw research talent and algorithmic innovation as to copying.
Two Futures for 2028 — and Why Berman Thinks Anthropic Is Underselling the Stakes
Anthropic lays out two paths: either America tightens controls, protects its lead, and negotiates from strength, or it fails to act and lets China catch up or overtake. Berman thinks the subtext is even bigger than the essay admits: 2028 matters because that may be when self-improving AI shows up, and whoever reaches recursive improvement first doesn’t just lead — they run away with it.
Why a CCP Lead Looks So Dangerous in Anthropic’s Telling
Berman says Anthropic is strongest when it describes what authoritarian AI leadership would actually mean: automated repression at a scale humans alone could never manage. He echoes the essay’s examples — facial recognition, biometric surveillance, censorship, hacking, and PLA use of DeepSeek-style models for cyber offense and unmanned swarms — and says the double-edged nature of AI is exactly the point.
Mythos, Cyber Capabilities, and the Race That Anthropic Says Isn’t a Race
Anthropic uses its unreleased Claude Mythos preview — reportedly a 10 trillion-parameter cyber-focused model from Project Glasswing — as evidence that frontier capability now has direct national-security implications. Berman notes the controversy around Mythos, with critics calling it fear-based marketing or a compute-serving problem, but agrees with the basic claim: if the CCP got a Mythos-level system first, it would almost certainly be used offensively.
Four Fronts of Competition — and Anthropic’s Contradiction on Intelligence
The essay breaks the contest into intelligence, domestic adoption, global distribution, and resilience, then says intelligence is the most important. Berman immediately pushes back, because Anthropic itself admits that if China spreads cheap, near-frontier AI globally, adoption can outweigh an intelligence deficit — and he thinks that’s exactly what open-source Chinese models are positioned to do.
Export Controls, Open Source, and Where Berman Parts Ways
Anthropic’s policy agenda is straightforward: close chip loopholes, stop distillation, and export American AI. Berman supports defending U.S. innovation, but he’s torn on export controls long term and flatly disagrees with Anthropic’s anti-open-source posture, arguing that if U.S. companies want the world to build on American AI instead of Chinese AI, open models will have to be part of the strategy.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
