Pacing the AI Frontier: What Amodei Proposed and What It Would Take
Anthropic CEO Dario Amodei has proposed to "pace the frontier" of AI development, laying out the idea in a blog post organized around three broad strategies. TechCrunch
The proposal centers on outside checks and coordination. It would rely on independent safety evaluators to test frontier systems, the most capable current models, and it calls for coordination between AI labs in democratic countries. TechCrunch
OpenAI CEO Sam Altman agreed with Amodei that it is time to "pace the frontier." TechCrunch The idea has drawn some wider industry support, and pushback from Jensen Huang. TechCrunch Huang is the chief executive of Nvidia, which makes the chips used to train AI systems. The New York Times
That public exchange runs alongside less visible talks between labs. OpenAI, Anthropic and Google have been in talks on AI safety for weeks. TechCrunch Google DeepMind has launched an institute to widen the debate over AGI, a term for highly capable future AI. TechCrunch
On TechCrunch Equity, hosts Kirsten Korosec, Anthony Ha and Sean O'Kane took up the practical question behind the slogan. They discussed whether companies can agree on what slowing down means and who gets to police it. TechCrunch
The broader context here is familiar from earlier work on standards and audits. Stating the goal is easy. Agreement is harder. Independent testing only works if outside testers have access, clear methods and time. Access means the model weights, the core settings learned in training, plus surrounding scaffolding, tool use and details of deployment, not just query access through an API for a few days. Methods means repeatable tests for abilities and safeguards that hold up when prompts or post-training, the extra tuning after initial training, change slightly. Time means testing before release plans lock in, while a poor result can still delay launch.
In my view, coordination is the harder problem. Labs in democratic countries still compete for talent, compute, business contracts and users. A voluntary agreement on pace would ask direct rivals to share safety findings, accept outside judgment and delay releases that could otherwise win customers. Worth flagging as an open issue is verification. Without a way to confirm what was trained, tested and shipped, coordination rests on trust between rivals.
In practical terms for people building with these models, the near-term stakes are about process rather than new technology. If independent review becomes a normal step, frontier teams will need test-ready builds, documentation and incident response plans as routine work. If lab-to-lab coordination holds, the safety talks reported among OpenAI, Anthropic and Google could become a standing forum rather than an ad hoc response. If it does not hold, each lab will keep setting its own bar and asking the public to trust internal testing.
My own expectation, shaped by watching my children grow up through the moves from desktop to mobile to cloud, is that adoption will not wait for rules to feel settled. Developers will build on whatever frontier models are available. That makes basic plumbing matter. Clear test criteria, published results where possible and shared language for risk would help downstream builders make better choices while the wider debate over pace continues. The proposal does not settle that debate. It gives it a structure to work with.
