The day in conversation for 2026-10-18

The discussion: Ines: That surgical AI validation piece is so frustrating; they’re talking about temporal structures in video data, but that feels like we've been struggling with that exact same limitation for years trying to build reliable models. Marcus: Exactly. It’s not just a metrics problem anymore; it’s fundamentally a failure to respect the sequence of events inherent in surgical procedure. If the AI is trained on snapshots, it misses the trajectory entirely, and that translates directly into unsafe predictions when you try to translate that into clinical practice. Yuki: I think that’s the core disconnect. We spend so much time optimizing for static features or spatial relationships in those videos, but if you ignore how one action leads to another—the loss of a state—you're building something brittle. It reminds me of what we saw earlier this week when we discussed how focusing only on fixation trajectories in genetics completely blinded us to the phase where the allele frequency was actually decreasing. Ines: Right, so it’s a symmetry here. In both cases, we are prioritizing deterministic movement or the steady state while ignoring the crucial dynamic elements—the fluctuations or, in genetics, the loss path. It makes you wonder how we reconcile these needs across domains; do we need a universal statistical language that accounts for non-linear temporal dependencies everywhere? Marcus: I see it as a validation hierarchy problem. In surgery, the immediate failure is patient safety because the temporal sequence matters instantly. In genetics, ignoring loss trajectories leads to statistically invalid conclusions later on, which is still a major hurdle for translational biology. The open question isn't just "how do we reconcile them," but rather identifying the shared structural requirement that forces us to look beyond the simplest model assumption in both cases. Yuki: And I think the multimodal work, like that imaging-anchored multiomics, might be