Key Takeaways

  • Lorenzo Sani watched AI models command surgical robots firsthand, but noted that engineering teams still cannot explain why the model selected a specific physical step.
  • Chris Benson points out that robot-driven surgical models will achieve physical precision beyond what human surgeons can safely perform with their own hands.
  • Surgical edge cases cannot be addressed through simulation alone because synthetic environments fail to reflect the anatomical variance of real human patients.
  • Strict privacy laws prevent hospitals from pooling surgical video and telemetry into centralized databases, blocking traditional model training methods.
  • Collaborative federated learning trains models across disparate medical centers by running compute locally and aggregating weight updates without moving patient data.

Surgical Models Still Lack Interpretability

Sani spent years researching distributed systems at Cambridge and Flower Labs. When he watched an AI system direct a robotic surgical tool in real life, the mechanical performance was evident. The decision logic behind it was not.

“I've seen models doing surgery. For real,” (play from 34:59) Sani said. “Like commanding robots to do surgery, and this was super cool to see. But then I thought like, do we know why they did that particular step?” (play from 35:02)

In consumer software, an unexplainable output produces a broken paragraph or a bad recommendation. In an operating room, an unexplainable movement cuts an artery. Benson noted that machine precision will soon exceed human capability: “It may be that the actual surgery is conducted by a robot that's being driven by a model with a level of precision that the human probably can't safely do.” (play from 37:06)

Yet physical precision does not solve the interpretability problem. Sani argued that high-stakes applications require full visibility into the model's choices: “There's plenty of sensitive applications where you would really like to have control over the motivations or the reasons behind specific workflows. And for this, there's the entire interpretability kind of area or path that is getting more and more important.” (play from 35:22) If an operator cannot trace why an agent chose a specific trajectory, regulatory clearance and clinical adoption remain out of reach.

Hospitals Will Never Centralize Patient Video

Robotics founders often lean on simulation to generate synthetic training data. In surgery, simulation falls short. Human anatomy presents unpredictable variations. Scar tissue, abnormal blood vessels, and unexpected organ placement cannot be simulated accurately enough to validate high-stakes interventions.

Improving these models requires genuine clinical volume. As Sani put it: “Currently, the only way to increase the amount of improvement, like to increase the pace of improvement for this model, is gather more and more data and make sure that you can train on this data in order to improve the model.” (play from 39:10)

The roadblock is regulatory. Hospitals operate under strict data protection laws. Legal teams routinely block requests to export unredacted surgical footage and telemetry logs to an external cloud. A startup waiting for dozens of medical centers to approve third-party data transfers will run out of capital before training begins.

This legal reality makes federated learning the only workable architecture. “And one clear way to get more and more data is sourcing it from many different places at once,” (play from 39:26) Sani explained. “Many different hospitals, or many different silos, many different areas. And collaborative learning technologies, federated learning technologies can help a lot with that.” (play from 39:33)

Rather than transferring sensitive patient files, teams deploy training compute to hardware inside each hospital. The hospital retains physical custody of every video frame and trajectory log. Only mathematical weights travel back to update the global model.

What to Do With This

Stop pitching hospital administrators on data aggregation deals or centralized cloud ingestion. If you build AI systems for physical environments under strict compliance, deploy an edge training setup inside the client perimeter using a federated framework. Show the hospital compliance team that raw telemetry never leaves their local network, and you remove the primary regulatory hurdle blocking your model improvements.