Contents
This article imagines future AI adoption based on Amodei’s proposal for external evaluation.
A new AI model is not always ready for immediate use in a company or laboratory. People need time to check its capabilities and safety, and decide which tasks to give it.
Dario Amodei proposed hosting external evaluators within AI developers. People from outside the company would examine its work and provide an independent view.
Shared standards could help adoption decisions
From here, imagine this approach spreading. If developers use shared evaluation standards, users can compare results more easily and judge which tasks an AI can handle.
External evaluation would not remove a company’s responsibility. It would still need to check suitability for its work and assign people to handle mistakes.
The gap between release and use could shrink
Imagine a laboratory using a new model to plan experiments in its release week, while a factory uses it for work instructions. Shared evaluation results have shortened part of their adoption process.
If AI also helps with evaluation, development, checking and adoption could follow each other more quickly. Capabilities demonstrated in a laboratory might reach work and daily life sooner.
More capable AI also needs clear decisions
If broad-capability AGI or AI far beyond human ability arrives, who decides it may be used, and what must be checked? The discussion of external evaluation could help establish that process.
