# Should AI Labs Slow Capability Advances?

> Published 2026-09-13 · https://www.promptzone.com/arif_lefevre/should-ai-labs-slow-capability-advances-3eif

Dario Amodei published an essay calling on frontier AI labs to deliberately slow capability gains until safety and evaluation infrastructure catches up. The post first appeared on Grok AI News.

Amodei outlined a three-step plan. Labs must grant permanent access to independent third-party evaluators. They must publish detailed capability reports before major releases. They must coordinate on shared safety thresholds that trigger pauses.

## The Three-Step Safety Framework

The first step requires ongoing evaluator presence inside each lab rather than one-off audits. The second step demands public release of model capability metrics on standardized benchmarks. The third step creates binding agreements among labs to halt training runs when predefined risk thresholds are crossed.

## Timeline of Agent Swarm Risks

Amodei warned that uncontrolled agent swarms could emerge within 6-12 months. These systems would combine planning, tool use, and long-horizon execution at scale. Current evaluation methods lack coverage for multi-agent coordination failures.

## Industry Response So Far

Sam Altman stated OpenAI would adopt comparable third-party evaluation practices. No other major lab has issued a matching public commitment. Early comments on the Grok AI News thread noted the absence of enforcement mechanisms.

## Comparison With Prior Safety Proposals

| Approach | Third-Party Access | Pre-Release Pause | Public Metrics | Enforcement |
|----------|---------------------|-------------------|----------------|-------------|
| Amodei 2025 plan | Permanent | Yes | Yes | Voluntary coordination |
| OpenAI Superalignment | Temporary | No | Partial | Internal only |
| EU AI Act | Audit-based | Conditional | Limited | Regulatory fines |

The Amodei proposal differs by requiring continuous evaluator presence and explicit pause triggers.

## Who Should Follow the Recommendations

Frontier labs training models above 10^26 FLOP should implement the three steps immediately. Smaller research groups and application developers face lower immediate risk and can continue at current pace. Regulators gain a concrete checklist they can reference in upcoming legislation.

## Practical Next Steps for Labs

- Publish current evaluator access policies on company blogs within 30 days.
- Adopt the same benchmark suite referenced in the essay for capability reporting.
- Join the existing voluntary coordination group already discussing shared thresholds.

> **Bottom line:** The proposal gives labs a concrete, testable framework instead of vague safety pledges.

Labs that adopt the plan early will face short-term speed trade-offs but reduce the chance of sudden regulatory shutdowns later.