30-day training notice
Thirty-day pre-training notice. File a plan 30 days before a frontier run. No approval needed, but the clock runs.
What it does
The lab tells the government 30 days before starting a very large training run. No approval is needed; the run can start when the 30 days are up.
Applies to: Training runs above a compute threshold.
- Notify the regulator at least 30 days before starting a training run above the threshold, with a description of the run and its safety plan.
- No approval is required; the run may proceed when the period ends.
Where things stand
The 2023 U.S. executive order on AI required reporting on large training runs but was rescinded in January 2025. The EU AI Act requires notification within two weeks of a model meeting the systemic-risk threshold, which is after the fact. No current U.S. rule imposes a pre-run waiting period.
Why it costs ~9 days
Most runs are planned well ahead, but plans change and a fresh 30-day clock sometimes delays a run. That costs about 3 days over a year.
Biggest unknown: Whether labs actually know 30 days out that a run will happen.
Why it lowers p(doom) by ~0.05%
Gives the government a heads-up before the biggest AI models are built.
Advance notice stops nothing by itself. It helps the government prepare and respond.
The strongest case that it costs more
The frontier is increasingly made by many mid-sized runs, not one giant run per year. If the threshold catches routine runs, a 30-day notice on each is a standing tax on iteration speed. Labs would either batch experiments, which slows learning, or design around the threshold, which slows scale.
The debate
For
- President Biden required reports on planned training of the largest models, 2023.
- Commerce Department proposed quarterly reports on training runs above 10^26 operations, 2024.
- Sens. Romney, Reed, Moran and King proposed notice before frontier training runs begin, 2024.
Against
- President Trump revoked Biden's AI executive order on his first day, January 2025.
- Ben Horowitz, a16z criticized regulating math, FLOPs and R&D, 2024.
Sources
Rough starting points, not precise forecasts. Lead costs assume China doesn't depend on U.S. models, the case least favorable to safety laws, and count 3 years. On the menu you can change every assumption and put in your own numbers. Last priced 2026-09-26.