The day in conversation for 2026-10-15

discussion

Transcript

The discussion: Iris: The takeaway from this MoE study seems pretty dismissive—that the learned configuration didn't actually beat vanilla DDQL on mean implementation shortfall across the board. If it couldn't improve that primary metric, what were we expecting when we introduced K means partitioning in the first place? Warren: I think we need to separate whether it failed because MoE itself was flawed, or if the training setup was just too fragile to exploit the architecture. Did anyone look closely at the conditions under which vanilla DDQL managed to wait for forced liquidation compared to when MoE converged on that same waiting policy in eleven of those runs? Ruth: That comparison is interesting; it suggests a convergence point exists, even if neither method optimized the shortfall. But I'm worried about what that convergence implies for tail risk. The paper mentioned that within-

More episodes

← Home