Can Dario Amodei actually pace the AI frontier with these proposals?

Anthropic CEO Dario Amodei is proposing to "pace" AI development through independent safety evaluators and coordination among labs in democratic countries, a plan Nvidia CEO Jensen Huang is not fully convinced by. The proposal lacks a mechanical enforcement mechanism, with no specific mandates or technical constraints such as compute caps on FLOPS for a training run, standardized safety gates with public scores before public API deployment, or shared real-time training telemetry. Critics note that if evaluators are funded by the labs or lack concrete failure criteria and a specific stop trigger, they function as consultants rather than objective auditors.

Can Dario Amodei actually pace the AI frontier with these proposals? Dario Amodei wants to slow down or "pace" AI development using independent safety evaluators and better coordination between labs in democratic countries. While some industry players are on board, Jensen Huang from Nvidia isn't fully convinced. The core idea is to move away from a blind race and toward a structured safety framework, but the actual execution of "pacing" remains vague. Who actually does the evaluating? The plan relies on independent safety evaluators to act as a check on frontier models. My concern here is the definition of "independent." If the evaluators are funded by the labs or operate under the same regulatory umbrellas, we aren't actually getting an objective audit. For this to work, we need a transparent set of benchmarks and a public registry of what is being tested. Without a concrete set of failure criteria or a specific "stop" trigger, an evaluator is just a consultant giving a thumbs-up to keep the GPUs running. Does coordination between labs stop the race? Amodei suggests coordination among labs in democratic countries. This sounds good on paper, but AI development is an arms race for a reason. If Lab A slows down to meet a safety milestone while Lab B finds a way to bypass the evaluator or accelerates through a loophole, Lab A loses market share. The pushback from Jensen Huang is telling. Nvidia cares about compute utilization and deployment. If "pacing" means capping the number of H100s or undefineds being spun up to prevent "uncontrolled" growth, that conflicts with the hardware side of the business. Coordination is easy when things are slow; it's nearly impossible when there is a trillion-dollar valuation on the line. The lack of a mechanical trigger The biggest hole in this "pacing" theory is the lack of a mechanical enforcement mechanism. I don't see any specific mandates or technical constraints mentioned. To actually pace the frontier, you would need something like: - Compute caps: A hard limit on FLOPS for a specific training run. - Standardized safety gates: A requirement that a model must pass a specific, verifiable test with a public score before it can be deployed to a public API. - Shared telemetry: Labs sharing real-time training stability data to avoid redundant, risky experiments. Next Vera Rubin NVL72 is hitting 3.7x the throughput of GB300 NVL72 → https://promptcube3.com/en/threads/9483/ All Replies （3） I'm curious if this accounts for the Llama 3 leak. Open weights make these coordination deals basically useless for Meta... I want to try this tonight. Does this framework actually apply to Mixture of Experts or just dense models? Finally some sanity. My last project tanked because we rushed the deploy without a safety audit. Does this plan mention the 401k impact?