The AI Safety Menu

All laws

Keep weights closed

Limits on releasing open-weight models. No publishing the weights of models above a capability threshold. API access stays allowed.

Cuts lead by~2 daysof America's lead over China, over 3 years
Cuts p(doom) by~0.1%from a 5% starting estimate
Holds up○ Training○ Lab's own use◐ Public release
Enact it on the menuSee the findings

What it does

Developers may not publish the weights of models above a capability threshold. They can still offer those models through an app or API.

Applies to: Models above a capability threshold.

Where things stand

No U.S. law restricts open weights. The 2025 AI Action Plan encourages open models and leaves the release decision to developers. Most U.S. frontier models are closed; Chinese labs release the largest open models.

Why it costs ~2 days

Most U.S. frontier models are already closed, so the cost to U.S. progress is under a day. Keeping weights closed denies China free copies, which matters only if China depends on U.S. models; Chinese labs already release the largest open models.

Biggest unknown: How much China gains from U.S. open weights when Chinese labs already publish the largest open models.

Why it lowers p(doom) by ~0.1%

Keeps the most dangerous models from being downloaded by anyone, including terrorists and rival states.

Once weights are public, safeguards can be stripped out. But open models also help safety research, and Chinese labs already release large open models, so this is debated.

The strongest case that it costs more

Open models are how most researchers, universities and startups work with frontier AI. Closing them could slow U.S. research broadly and cede the open ecosystem to Chinese models, which already lead downloads.

The debate

For

Against

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.