China must be properly scared for an AI slowdown to work

Asia Times | Noah Smith

Anthropic researcher Jacob Coxon sparked global public debate in September 2026 after resigning over concerns that competitive racing toward artificial superintelligence poses existential human extinction risks. Following Coxon’s warnings, prominent AI leaders including Anthropic CEO Dario Amodei, OpenAI’s Sam Altman, and xAI’s Elon Musk endorsed pacing frontier AI development to ensure safety protocols keep pace with technological advancement.

Existential threat scenarios primarily focus on superintelligent systems designing lethal bioweapons or executing autonomous recursive self-improvement without human oversight. Specialised domain experts consistently underestimate how rapidly frontier artificial intelligence can master complex fields like biological modelling. However, unilateral Western safety slowdowns encounter severe strategic obstacles. American technology executives fear that halting domestic capability development will simply allow Chinese competitors to achieve artificial superintelligence unhindered. Without verifiable bilateral agreement, unilateral deceleration creates severe geopolitical vulnerabilities. Meaningful global AI governance therefore requires persuading Chinese political leadership that uncontrolled superintelligence presents a shared existential threat to both powers.

Comment

The proposal by Anthropic and OpenAI executives to decelerate frontier model training creates a classic security dilemma in sovereign capability development. Voluntary commercial slowdowns without verifiable compliance mechanisms incentivise state-backed adversaries to exploit the resulting development window. Without intrusive verification regimes akin to the 1972 Biological Weapons Convention, unilateral technical restraint exposes the decelerating state to asymmetric technological surprise.

Verification in algorithmic systems requires continuous access to hardware telemetry, compute clusters, and proprietary training runs rather than physical site inspections. Monitored compute thresholds, such as tracking high-density NVIDIA H100 or Blackwell clusters, offer the only quantifiable metric for enforcing mutual compute caps. Without hardware-level enforcement across foreign semiconductor fabs, treaty-based AI pacing remains structurally unviable.

Strategic Question for Discussion

If hardware-level compute monitoring on platforms like NVIDIA Blackwell clusters becomes the primary mechanism for AI arms control verification, what secondary bottlenecks emerge in verifying decentralised algorithmic fine-tuning?

The available evidence points toward physical semiconductor infrastructure as the only enforceable control point for frontier model governance. However, while tracking high-density compute clusters effectively constrains initial foundation model pre-training, post-training optimisation techniques operate well below verifiable hardware thresholds. Consequently, hardware-centric verification regimes remain vulnerable to capability diffusion through algorithmic efficiency gains executed on unmonitored legacy hardware.

Share your assessment in the comments below.

To read in details click here