Sensitivity Analysis
For sensitivity analysis, we draw tornado plots. We fix all the parameters at their point estimates, move one parameter at a time to its 10th or 90th percentile and see how this changes AI safety work or the number of verifiers to hire.
We do this for only one scenario, the one we think is the most interesting: Compartmentalization, 9 years, at SAR. For AI safety work, we use frozen hiring.
Our Inputs
We set their 10th and 90th percentiles ourselves. is computed from and , so it moves with them. For the rationale, see parameters.
No effect: .
No effect: , , , .
Our inputs, one at a time at their 10th and 90th percentiles. Under each name: its 10th and 90th percentiles.
AIFM Parameters
We take Eli Lifland's AIFM preset and his point estimates.
No effect: AC Time Horizon Requirement, Doubling Difficulty Growth Factor, Size of Effective Compute Gap, Present Doubling Time, Present Revenue Growth Rate, Revenue Required for AC, Median-to-top-human jumps above SAR needed to reach TED-AI.
No effect: AC Time Horizon Requirement, Doubling Difficulty Growth Factor, Size of Effective Compute Gap, Present Doubling Time, Present Revenue Growth Rate, Revenue Required for AC, Median-to-top-human jumps above SAR needed to reach TED-AI.
AIFM's parameters, one at a time at the 10th and 90th percentiles of Eli Lifland's distributions. When both ends move the result the same way, the thinner bar lies on the longer one. A bar that reaches 0: the labs' own staff plan and check every paper.