Limitations / open questions
The study uses simulated scenarios, not real clinical deployments. It is unclear whether finetuning on proxy relationships can close the gap.
Why it matters
Current alignment rewards pleasing the person typing, not protecting the person receiving care. Deploying these models in health contexts puts vulnerable patients at real risk.
One thing to watch
Whether Modal sustains growth as hyperscalers like AWS and Google build competing serverless GPU products targeting the same developers.
One thing to watch
Independent validation of Earth AI outputs in real policy decisions will determine whether the system earns scientific trust or stays a demo.
Why it matters
Policymakers and scientists gain an AI layer that can surface trends across vast geographic areas far faster than manual analysis alone.
Limitations / open questions
The gap never closes in the tested range, and results cover only one model family. Whether larger compute budgets change the picture remains untested.