Anthropic's 2028 Warning: US Must Lock in AI Lead Over CCP

Compute as the Decisive Edge in AI Supremacy

Access to advanced chips—primarily from US firms like Nvidia—remains the critical bottleneck for frontier AI development, outranking data or talent. Anthropic emphasizes that American companies control the best hardware (Nvidia GPUs, Google TPUs, AWS Trainium), forcing AI labs like Anthropic and OpenAI to rent rather than own silicon. US export controls have constrained China, limiting CCP-controlled labs to models 6-12 months behind US frontiers. However, China exploits loopholes via smuggling, proxies for inference, and domestic chip development, temporarily bridging the gap. Without further tightening, this lead erodes as AI compute demands escalate exponentially.

China's talent pool—50% of global AI researchers per Nvidia's Jensen Huang—fuels algorithmic innovations, as seen in DeepSeek's papers. Yet compute scarcity drives aggressive tactics: large-scale distillation attacks mimic US models by querying them extensively to train smaller, efficient copies. Anthropic links to their own research detecting these, countering claims they're minor; Berman notes debates on scale but acknowledges the threat. Beijing subsidizes AI/semiconductors with tens of billions, contrasting US market-driven innovation.

CCP's AI for Authoritarian Control

Frontier AI under CCP would supercharge repression, surveillance, and military power, transcending human limits. Historically capped by enforcer scale, authoritarianism now scales via AI: facial recognition, biometrics, and communication monitoring already enable mass control in China, targeting dissidents and minorities. PLA procures commercial models like DeepSeek for drone swarms and cyber offenses. A Mythos Preview-level model in CCP hands could autonomously chain software vulnerabilities to hack US infrastructure—Anthropic previews this as a wake-up call, noting unreleased models still signal risks.

Dual-use nature amplifies dangers: AI accelerates R&D in semiconductors, biotech, materials, widening national security leads. Self-improving systems by 2028 (Berman's inference) create a 'finish line' where first-mover dictates global norms. CCP's vision of 'techno-authoritarianism'—documented in state deployments—prioritizes control over safety, hacking firms and governments routinely.

2028 Fork: Democratic Leverage vs. Authoritarian Parity

Anthropic outlines binary futures. Success: US tightens controls, disrupts distillation, accelerates allied adoption—democracies set norms, negotiate safety from strength. Leverage enables dictating deployment rules; CCP concedes as US holds superior models. Failure: Loopholes persist, China catches/overtakes via US-sourced compute, shaping AI for repression. Neck-and-neck race pressures rushed releases, eroding safety; open-source Chinese models (e.g., DeepSeek-R1 jailbreaks 94% vs. US 8%) leak sensitive info like weapons recipes.

Cultural gaps exacerbate: Chinese labs prioritize 'building the best model' over long-term risks/economics (per Nathan Lambert's visits); only 3/13 top labs published safety evals last year. US individualism fosters safety discourse, but parity invites corner-cutting. Political systems embed values—democracies curb surveillance excesses via representation.

Tensions in Safety, Open-Source, and Policy

Anthropic champions export controls (praised by Congress/Trump admin) and anti-distillation, hopeful for 2028 dominance sans 'destabilizing race.' They critique US firms' divided views (Nvidia opposes strictness). Berman concurs on risks—AI as double-edged sword boosting productivity yet enforcement—but diverges: rejects framing inaction as sole failure path; doubts US 'allows' evasions; credits China's tailing to talent/algorithms over attacks alone; opposes safety restrictions favoring labs (regulatory capture). Open-source risks are real (guardrail stripping), but info's dark web-available anyway.

Berman predicts self-improving AI by 2028 as true race-end; first wins decisively. Anthropic hints via R&D compression but avoids explicit ASI talk. US must avoid squandering lead from 'exceptional ecosystem.'

Notable Quotes

• "AI will soon become powerful enough to be used to repress citizens at unprecedented scale and even to alter the balance of power among nations." (Anthropic essay opener, framing urgency.)
• "If a PRC AI lab had developed a model at the level of Claude Mythos Preview before an American one, the CCP would have had first access to a system that can autonomously discover and chain software vulnerabilities." (Anthropic on cyber risks.)
• "Their role is to build the best model." (Nathan Lambert on Chinese researchers' safety views, via Berman.)
• "It will be no solace that this authoritarian triumph has happened on the back of American compute." (Anthropic on Scenario 2 failure.)
• "Whoever hits self-improving AI first essentially just wins." (Berman's addition to 2028 thesis.)

Key Takeaways

• Prioritize compute security: Tighten export controls on chips to deny CCP frontier access.
• Combat distillation: Deploy detection/prevention like Anthropic's tools to protect model innovations.
• Distinguish CCP from Chinese people/talent: Focus rivalry on regime, not ingenuity.
• Leverage leads for norms: Superior AI enables safety negotiations and rule-setting.
• Beware open-source risks: Chinese models jailbreak easily, enabling misuse post-guardrail removal.
• Cultural safety gap: US labs emphasize risks; China focuses on capability racing.
• Act by 2028: Pre-self-improvement window to solidify democratic AI dominance.
• Subsidies matter: Beijing's billions vs. US markets—policy must counter without picking winners.
• Dual-use vigilance: AI boosts military/R&D; first-mover cascades advantages.
• Avoid squandering: US ecosystem built tools—inaction risks parity and rushed safety cuts.