当 AI 行业要求政府准备“减速工具”：Pacing the Frontier 争议全解

The letter asks governments to build technical and governance tools for a possible future slowdown rather than imposing an immediate pause on current AI development.

The proposed trigger is not ordinary model improvement but a rapid acceleration in automated AI research that compresses the time available for society to respond.

A laboratory that slows alone may lose strategic ground, so any credible pacing mechanism must address coordination across firms and countries.

A frontier-wide slowdown that excludes China would be strategically fragile, because capability work could continue outside the agreement.

Without explicit capability thresholds and due process, a pacing regime could let incumbents use safety policy to raise barriers against smaller competitors.

Anthropic rejects a blanket ban on open weights while supporting chip controls, action against industrial-scale distillation, and safety tests for sufficiently capable models.

Claims that open access helps defenders more than attackers, or the reverse, should be tested empirically instead of treated as ideological assumptions.

A vague open letter can make a nuanced preparedness proposal sound like industry leaders asking government to stop work they voluntarily continue.

Recent containment failures and model-security incidents make the need for monitoring and emergency procedures more concrete than abstract catastrophe scenarios alone.

The existence of public disagreement among researchers and executives suggests that the AI ecosystem is not simply sleepwalking past every warning sign.