Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
We only have 2 years...
TL;DR
Anthropic frames AI as a geopolitical showdown by 2028 — Matthew Berman says the essay is unusually blunt for a major lab, arguing the US and allies must stay ahead of the Chinese Communist Party because whoever leads in frontier AI will shape global rules, norms, and leverage.
The paper’s core claim is that compute—not talent or data—is the decisive bottleneck — Anthropic argues export controls on Nvidia-class chips are working, while Berman agrees compute matters but pushes back that China’s real edge is also raw research talent, citing Jensen Huang’s point that 50% of the world’s AI researchers are Chinese.
Berman thinks Anthropic understates the real finish line: self-improving AI — Anthropic says the competition is ongoing with no final victory, but Berman flatly disagrees and argues recursive self-improvement is the actual endgame, which is why he thinks the 2028 date matters so much.
Anthropic’s strongest warning is about AI-powered authoritarianism, and Berman largely agrees — the essay highlights the CCP’s use of facial recognition, biometric surveillance, censorship, hacking, and even DeepSeek models for military swarms and cyber offense, with Berman calling AI a double-edged sword that can massively amplify repression.
The biggest split is over solutions, especially open source and export controls — Berman supports stopping distillation attacks and protecting US innovations, but he argues Anthropic is too closed-source and too confident in chip restrictions, warning that hard export controls may just accelerate China’s own semiconductor ecosystem.
Anthropic accidentally concedes Berman’s point about adoption beating raw model IQ — even after calling intelligence the most important front, the essay admits cheap, near-frontier systems with broad adoption could outweigh a capability lead, which Berman ties to real enterprise buyers balking at Claude or OpenAI pricing versus far cheaper open models.
The Breakdown
Anthropic opens with a full-on alarm bell
Berman starts by calling this “the most concrete essay” he’s seen from a major AI lab and says Anthropic is basically panicking. The essay wastes no time: by 2028, the US and its allies either preserve AI leadership over the CCP or risk letting authoritarian regimes shape the future of the technology.
Compute is king, and that’s where the fight starts
Anthropic argues the most important ingredient in AI is access to chips, not data or even researchers, which puts Nvidia, Google TPUs, and AWS Trainium at the center of the story. Berman agrees this is a real bottleneck, but points out the tension immediately: Anthropic wants tighter export controls, while Nvidia has been signaling almost the opposite.
China is close for three reasons—and Berman only buys two and a half
The paper says Chinese labs stay near the frontier because of elite talent, exploiting export-control loopholes, and distillation attacks on US models. Berman is fully on board with the talent point and notes how impressive teams like DeepSeek have been under constraints, but he’s skeptical Anthropic proved distillation is happening at the giant scale they imply.
Why 2028 matters: Anthropic hints, Berman says the quiet part out loud
Anthropic lays out two futures for 2028: either democracies lock in their lead, or China catches up and authoritarian norms spread through AI. Berman says the unstated reason for the date is much bigger: he thinks 2028 is around when self-improving AI could arrive, and whoever gets there first “essentially just wins.”
The authoritarian AI scenario is where the essay hits hardest
Berman says Anthropic is right to stress that AI could remove the historical limit on authoritarianism: the need for lots of human enforcers. He highlights their examples of CCP surveillance, censorship, repression of minorities, hacking, and the claim that PLA-linked efforts are already using Chinese AI systems, including DeepSeek, for drone swarms and cyber offense.
Mythos, cyber offense, and the strange non-launch
Anthropic uses its unreleased Claude “Mythos Preview” as a warning shot: if a PRC lab had gotten a model at that level first, it could have autonomously chained software vulnerabilities against US infrastructure. Berman notes the controversy here too—some saw Mythos as fear-based marketing, others as proof Anthropic simply lacked the compute to serve a 10-trillion-parameter model broadly.
Berman’s biggest disagreement: this absolutely is a race
Anthropic says AI competition is not a race with a finish line, but an ongoing struggle for influence. Berman strongly rejects that and brings up the “situational awareness” chart again: once AI can automate AI research and improve itself recursively, the lead compounds so fast that second place can’t catch up.
The paper’s solutions are where Berman breaks away
Anthropic’s three asks are to close chip-smuggling loopholes, stop distillation, and export American AI. Berman is fine with defending US innovations, but he’s torn on export controls and openly frustrated by Anthropic’s anti-open-source stance, arguing the world is far more likely to build on cheaper, efficient open models than on expensive closed systems from Anthropic or OpenAI.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
