Sorting Dario's policy asks
Dario Amodei's Policy on the AI Exponential asks governments to move at the speed of the technology. By the timeline he sets himself, four of his five proposals cannot.
His estimate: powerful AI arrives within a year or two. That estimate is conditional on scaling laws holding. He measures the pace another way. Models went from barely writing a coherent line of code to writing most of the code at major AI companies. That took four years.
Sort the five asks by enactment speed. One can become binding law before powerful AI arrives. The other four route through institutions Amodei himself describes as too slow.
The ask that can land is mandatory third-party testing of frontier models above a compute threshold. The testing would cover cybersecurity, biological weapons, loss of control, and automated R&D.
The enactment path already works. State legislatures have passed the transparency laws Amodei cites: California's SB 53, New York's RAISE Act, and Illinois's SB 315. SB 53 went from introduction to signature inside a single session. A testing requirement extends those laws instead of starting from nothing.
The rule also regulates a handful of frontier labs, not the whole economy. It needs no new agency and no new entitlement. That makes it the one proposal a legislature can turn into binding law in a year or two.
The other four asks each depend on an institution the essay names as slow.
Wage insurance, retention tax credits, and income support funded by capital-gains taxes need Congress. Congress would have to build new entitlement machinery. Amodei's own complaint is that in the several years Congress takes to act, AI "can go from an amusing toy to the full country of geniuses."
Pre-approved standards for AI-based drug testing ask the FDA to rework its pipeline. That pipeline currently runs seven to eight years per drug.
Judicial oversight of autonomous weapons needs courts to build doctrine. Courts build doctrine case by case. That takes decades.
A democratic coalition controlling the chip supply chain needs treaty-level coordination. The governments involved disagree about export controls today.
Amodei concedes the lag in one line. He worries that current legislation is "at least a year out of step with AI's rapid progress." But the essay still presents the five asks as a single program with a single urgency.
Sorted by enactment speed, they split in two. One is a rule that can bind before powerful AI arrives. The other four are positions filed early so that Congress, the FDA, and the courts start deliberating before the need is acute.
Filing early is worth doing. When a crisis forces an institution to act, it adopts whatever framework is already on the table. But it remains a different activity from the testing regime. The testing regime is the one part of the agenda where pressure in the next year changes the outcome.