License to train

What it does

Before starting a very large training run, the lab submits a safety plan to a regulator and waits for approval. The regulator can say no.

Applies to: Training runs above a compute threshold (10^26 FLOP or a successor).

• Before beginning a training run above the threshold, submit a safety and security plan for the run and receive authorization.
• The regulator may impose conditions or withhold approval.
• Distinct from SB 1047, which required a self-certified protocol before training but no approval.

Where things stand

No U.S. law does this. SB 1047 required a written protocol and cybersecurity before training, self-certified. The EU AI Act requires notification, not approval. A true licensing gate on training has been proposed by think tanks and in early drafts of federal bills but not enacted.

Why it costs ~2.5 months

This comes before the model exists, so waiting for approval delays the next training run. The model estimates about four weeks of lost U.S. progress a year after the law takes effect.

Biggest unknown: Whether the regulator can say no, and how long it takes to say yes.

Why it lowers p(doom) by ~0.4%

Stops a lab from building a much more powerful model until it can show it can keep it under control.

This acts before a dangerous model exists, when prevention is easiest. It's among the stronger measures, though regulators may struggle to judge the cases.

The strongest case that it costs more

Licensing regimes in other industries take far longer than 30 days once they mature, and a regulator that can say no will eventually say no. The first refusal costs a generation, not a month. A lab also cannot plan a run it does not know will be approved, so the parallel-filing assumption breaks down as the regime tightens. Three months could be the floor rather than the ceiling.

The debate

For

• Sens. Blumenthal and Hawley proposed a licensing regime run by an independent oversight body, 2023.
• Sam Altman, OpenAI told the Senate the U.S. should consider licensing for models above a capability threshold, 2023.
• Brad Smith, Microsoft called for licensing so an agency is informed of testing before deployment, 2023.

Against

• Andrew Ng called AI licensing a massively, colossally dumb idea that would crush innovation, 2023.
• Yann LeCun, Meta accused leading lab CEOs of attempting regulatory capture, 2023.

Sources

• SB 1047 enrolled text, § 22603(a): pre-training duties (self-certified)
• EU AI Act, Article 52: notification obligations for GPAI with systemic risk