- This episode is a panel discussion from the Polymath Medical Salon at Florida Atlantic University’s Gruber AI Sandbox, moderated by Curt Jaimungal (host of Theories of Everything), exploring how AI—particularly large language models and multi-agent systems—could transform healthcare by making sophisticated, personalized medical expertise widely accessible. The panel includes AI researchers William Hahn, Elon Bernholtz, and Dan Elton, along with longevity entrepreneur Gil Blander.
Core ideas about AI and medicine
- The central vision discussed is “medical swarms”—large numbers of specialized AI agents (radiologist agent, nutritionist agent, psychologist agent, etc.) working in concert for a single patient, available 24/7 at near-zero cost, potentially giving every person the equivalent of hundreds of concierge physicians.
- This is framed not as replacing doctors but as democratizing access to high-quality medical reasoning, especially for people who currently lack it due to cost, geography, or system overload.
- A key near-term application is AI-assisted pre-diagnosis and clinical intake: an AI interviews the patient before the doctor visit, summarizes the history, and surfaces salient points—reducing the documentation burden and giving physicians more time for human interaction.
AI in medical imaging and complex chronic illness
- Dan Elton discusses using AI to extract precise biomarkers from medical images for risk prediction (e.g., cardiovascular disease) and to help manage complex chronic conditions like long COVID and chronic fatigue syndrome (which affects ~2% of people worldwide).
- He notes that the current medical system is poorly equipped for persistent, multi-factorial conditions because doctors are overtrained in biology and undertrained in psychology and sociology, and have only ~40 minutes per patient—gaps AI could help fill.
LLMs as models of cognition and memory
- Elon Bernholtz presents the view that LLMs and human cognition share core computational principles—specifically autoregressive next-token generation—and that this correspondence gives us a new tool for studying brain function, memory, and dementia.
- He proposes that dementia can be reinterpreted not as lost memories but as a shrinking “context window”—the brain losing the activation needed to generate the next token—and that experimentally reducing context windows in LLMs produces confabulation-like behavior similar to dementia patients.
- He also discusses confabulation in Wernicke-Korsakoff syndrome as an error-checking deficit and suggests LLMs let us model such breakdowns computationally.
Practical health markers and personalized intervention
- Gil Blander describes building a “digital twin” from ~50 blood biomarkers, DNA, wearable data, and food recognition to generate laser-focused lifestyle recommendations.
- Key biomarkers he highlights include ApoB (cardiovascular risk), hsCRP (inflammation), glucose/A1c (diabetic trajectory), and VO2max (fitness), while stressing that the right markers depend on the individual’s specific health situation.
Barriers to clinical deployment
- FDA has approved over 1,000 AI tools, but most are not commercially successful because hospitals lack budgets and insurance companies generally don’t reimburse for AI (only ~2 CPT codes for AI exist).
- Hospitals are increasingly reluctant to share data, recognizing its value, which slows research—though companies like Google and Microsoft are negotiating data-use agreements with hospital systems.
- Panelists expect AI “doctors at home” to become widespread before AI is fully integrated into clinical workflows.
The human element and emotional intelligence
- Recent benchmarks show the latest AI scores higher emotional intelligence than 62% of physicians in patient portal responses, though this is partly because overworked doctors give terse replies.
- Panelists agree that human touch, trust, and the feeling of being cared for remain essential—Robin Hanson’s research suggests people consume healthcare partly for emotional reassurance, not just medical outcomes.
- A medical student raises concern about whether AI can truly understand intimate human experiences like ecstasy or addiction, suggesting limits to AI’s emotional competence even as it improves.
Outlook and near-term predictions
- In the next three months, realistic AI gains include chart summarization, pre-visit patient interviewing, and distilling large medical literature into actionable bullet points for patients and clinicians.
- Longer term, the panel anticipates AI-driven drug discovery, scientific hypothesis generation (tools like Google’s Co-Scientist are mentioned), and universal access to polymathic health advisors that simultaneously serve as physician, therapist, dietitian, and coach.
- For the younger generation, the advice is to learn to code and engage with AI tools, since software creation is becoming dramatically cheaper and more accessible—comparable to how literacy moved from a specialized profession to a universal skill.