Dispatch
MIT and Columbia's 'Racing to Ruin' finds transparency is double-edged and low trust makes disaster arrive with probability one
Covered in Import AI 468, the paper models AI racing dynamics and isolates two variables that determine whether a coordinated slowdown is reachable: transparency about technology development and how actors model each other's trustworthiness. The formal result is that when monitoring is sufficiently precise every equilibrium stops in finite time — but faster detection can first destroy early-stopping equilibria before restoring them, making transparency non-monotonic. The blunt conclusion: 'With low trust, every equilibrium races to ruin: the disaster arrives with probability one.'
Source
↳ Follow the thread