Skip to content
Anton Leicht· Anton Leicht·Indexed by this site · Originally published· Governance →· Its place in the scan record →

Putting frontier AI pacing control into practice: three types of mechanisms and their respective costs

Putting Pacing Into Practice

AI summary

Anton Leicht argues for making frontier AI pacing control into a workable plan before political pressure arrives, and breaks it down into three types of mechanisms: compute input caps, safety thresholds, and speed limits. Input caps limit the compute used for R&D, but labs can circumvent the limits by offering inference to external companies and other means; under his assumed conditions of algorithmic efficiency improving threefold per year and R&D compute cut by 50%, the original research speed could be recovered in about eight months. Safety thresholds require a model to first pass designated evaluations or third-party or government assessments; the difficulty is that a defensible safety case cannot yet be produced, and the catch-up growth that follows each release would be harder to control than normal-paced progress. Speed limits set bounds on the rate of capability increase per unit time and require an aggregate capability benchmark that is hard to game, but compared with the first two types they are less likely to favor any one lab. He suggests advancing a combination in the order of input caps, speed limits, and safety thresholds, and calls on labs to take the lead in establishing compute accounting and to work with external evaluators to refine adversarial-robust capability benchmarks.

Read originalwriting.antonleicht.me