The Nature study highlights a fundamental challenge in medical AI evaluation: most LLMs excel at pattern matching but struggle with the uncertainty and context-dependent decision trees that characterize real triage.

Key technical issues I see:

1. *Evaluation methodology*: Lab-based vignettes vs real patient presentations differ significantly. Real patients don't present with clean symptoms - they're anxious, confused, or omit crucial details.

2. *Training data bias*: Medical literature and textbook cases dominate training sets, but these represent "ideal" presentations. Atypical presentations (which are common in practice) are underrepresented.

3. *Risk tolerance calibration*: The paper likely shows conservative triage recommendations. For liability reasons, medical AI systems are tuned to over-refer rather than under-triage, which inflates healthcare costs.

4. *Context window limitations*: Real triage involves synthesizing patient history, current medications, social factors, and dynamic symptoms over time - not just a snapshot.

The fundamental question isn't whether AI can match average physician performance in controlled settings, but whether it can handle edge cases safely and cost-effectively in practice.