In his September essay, We Must Pace the Frontier, Dario Amodei proposes slowing advanced AI development so safety work can keep up. He does not ask businesses to stop using AI. Read the essay.
That distinction matters when your decision is whether an assistant should organize inquiries, prepare estimates or update a customer record. The sensible response is to examine the responsibility you are handing over, and the evidence supporting that decision.
What he is proposing
Amodei proposes independent evaluators inside frontier labs, then government-supported coordination among democratic-country labs, followed by international agreements including China. Anthropic commits to the first step, with reviewers able to publish subject to limited redactions.
He argues that AI-assisted AI research and recent alignment incidents call for more time for safeguards, alignment research and testing. These are his assessments.
This is a proposal, not an industry agreement. It describes intended access, not an operating evaluator program. Amodei acknowledges verification and competitive pressures, and considers a comprehensive global pause unlikely soon. Proposal and implementation details.
The disagreement is about accountability
The Washington Post's editorial board argues that labs can slow themselves. Its critique.
A related proposal from Google DeepMind's Demis Hassabis, published in July, describes a standards body with independent technical and open-source representation. It would assess frontier models against updated benchmarks; non-frontier models would be exempt from that process. This is a separate proposal, not an implemented agreement among the companies. Hassabis's framework.
Whitmore's view is that the quality of oversight depends on what an evaluator can inspect, disclose and challenge. An impressive committee name is a poor substitute for those details. Buyers should also ask who remains responsible when an approved system behaves badly. A review process should make that answer clearer.
Build a business plan that survives a slower frontier
For buyers, the practical consequence is to separate improvements available today from capabilities promised tomorrow.
Consider a hypothetical firm choosing between two proposals. One organizes incoming requests using a model the team can test now. The other depends on a future agent handling an entire customer relationship. Evaluate the first against actual work. Ask the second provider to identify which promised outcomes depend on technology that has not arrived.
That distinction belongs in the budget. Fund a useful workflow with a measurable outcome. Treat future capabilities as an option to revisit, rather than an assumption needed to justify the purchase.
The same principle applies to upgrades. A new model should earn its place by improving the work: fewer corrections, better handling of exceptions, a lower operating cost or a capability the business actually needs. Decide those criteria before the release announcement lands.
Ask what will change the decision
Independent evaluation could become valuable purchasing evidence. Buyers should ask what a review covered, when it happened and whether its findings are available. An assessment of a model does not, by itself, assess the workflow your supplier assembled around it.
Whitmore's recommendation is to set a review date and explicit reasons to reconsider a provider. Those might include a material incident, a change in data handling, a price increase or a demonstrably better alternative.
A business should be able to explain its AI investment without forecasting the next frontier breakthrough.
