Proposed minimum for pacing: fix alignment before the next training run
The author argues labs should not train a new generation until alignment problems in the current one are resolved.
AGI & Safety#05
80%
Author's credence in the pause thesis
Image: AI Alignment Forum
80%Author's credence in the pause thesis
The short version
- The author proposes a minimum standard for pacing frontier AI development.
- Under this proposal, a lab cannot train the next generation of models until it has adequately resolved all significant alignment issues in the present generation.
- The author says that beating previous models on internal metrics is not enough, and that a user finding an alarming issue after deployment is a critical failure of both the model and the process.
- The author calls the proposal a heuristic rather than a concrete plan and says he still prefers a pause to pacing.
Why it matters
The proposal ties capability progress to alignment and interpretability progress, which could change how labs prioritize safety work. It is aimed at developers and researchers who differ on how easily current alignment issues can be fixed, so it offers a shared test that both optimists and pessimists could accept.
“You don't launch a jet that goes at Mach 3 before you've figured out why your Mach 2 jet keeps crashing”