No Priors Ep. 17 | With Karan Singhal

Med-PaLM builds on PaLM/PaLM 2’s general reasoning and knowledge, then uses data-efficient techniques like instruction prompt tuning with doctor-written examples to adapt behavior for long-form medical QA without costly full retraining.

If you have only a handful of examples, use prompting; with tens of examples, consider prompt tuning; and once you have 100+ high-quality labeled examples, full fine-tuning tends to deliver the best performance relative to cost.

Existing medical AI work often relied on limited multiple‑choice benchmarks; Med-PaLM emphasized systematic automated evaluation plus detailed human evaluation by clinicians and laypeople, but Singhal stresses the need for workflow‑embedded, outcomes‑aware studies.

While stakes are high in medicine, patients and physicians already rely on imperfect web searches and time‑pressed clinicians; indecision or inaction is also a decision, so policy must weigh both risks and missed benefits of withholding capable systems.

LLMs are already being piloted for documentation and billing; Singhal expects next waves in radiology assistance, report QA, telemedicine augmentation, and clinical decision support—initially as physician co-pilots, not replacements.

As models like Med-PaLM 2 approach physician-level performance, even experts struggle to judge answers, making simple RLHF less reliable; this forces exploration of techniques like self-critique, debate, and AI‑assisted oversight that matter for broader AI safety.

To overcome institutional risk aversion, Singhal highlights tools that can query and cite trusted medical sources (e.g., Mayo Clinic) and architectures like automatic redaction or trusted execution environments that reduce exposure of sensitive patient data.