The Brain and the Bot Podcast

Daniel Brandt Vegas, Steven Wong
  • 0.0 (0)
  • Medicine
  • Monthly

Two clinician educators learning and discussing Artificial Intelligence related topics from a Medical Education perspective.

Episodes

  1. Oct 6

    #4 Learning, in the Mind of the Beholder

    At end-of-week feedback, two learners said the same thing: good week, but "I only saw one or two disease states — I wish I'd seen more." That comment kicks off this episode. Is the scavenger-hunt view of a clinical rotation the right one, and what does it miss when the pneumonia patient in bed four is there because of a catastrophic stroke two months ago? We work through what competency-based medical education has actually taught us after a dozen years: that a checklist of observed EPAs can't capture competence, because competence is the ability to adapt to a shifting mix of contextual variables — the wheezing patient with no edema and no history, who doesn't speak your language. Then we borrow two ideas from machine learning to reframe it. The long tail: you can build a system (or a learner) that handles the textbook 95% fairly easily; it's the endless stream of edge cases that's hard and high-stakes. And zero-shot learning: a toddler needs two picture books to tell a cow from a bird, while older algorithms needed thousands of ICU patients to recognize sepsis — so what does it take to generalize from one heart-failure patient rather than fifteen? That leads somewhere unexpected: maybe junior learners should move faster through the core, and senior learners should be allowed to slow down and dig into context? Chapters (00:00) "I only saw two diseases this week" (01:30) The stroke patient who was there all along (03:15) Learning is in the mind of the beholder (05:00) What CBME taught us: EPAs aren't competence (06:30) Context is the variable that matters (09:00) Holistic assessment and the ITER (10:15) The long tail, from Amazon to self-driving cars (14:15) 95% textbook, 5% everything else (15:45) Zero-shot learning: the toddler vs. the sepsis algorithm (19:30) Do these two ideas contradict each other? (20:30) Recognizing a disease vs. managing it (21:30) Spend more time with fewer patients? (23:45) The fast-paced ward and the four-hour consult (25:30) Should juniors go faster and seniors go slower? (29:30) Signposting the long tail for learners Mentioned in this episode Competency-based medical education (CBME), EPAs, and ITERs The Royal College of Physicians and Surgeons of Canada's revised CBME implementation Long-tail theory (business origins; autonomous-vehicle edge cases) Zero-shot / few-shot learning vs. data-hungry ICU deterioration models AI use: - Nano banana for cover art - Descript for transcription editing

  2. Oct 6

    #3 Should AI summarize the morning handover??

    What Is Clinical Practice? Morning handover is the most expensive hour of the day: an exhausted post-call team presenting overnight admissions to an incoming team that hasn't started yet. It's non-negotiable, it's where a huge amount of teaching happens, and it's a prime target for AI summarization. In this episode we ask whether it should be. We start by pulling handover apart into its two jobs — the clinical transfer of care and the educational exercise of synthesizing a case and pitching it — then dig into what AI summarization actually is under the hood. Why large language models confabulate rather than "hallucinate," why probabilistic output is a different kind of computer than the deterministic calculators clinicians are used to, and what it means when a scribe tool writes a normal abdominal exam that never happened. From there: should the learner's presentation be replaced by an AI summary, or compared against one? What gets lost if trainees never learn to tell the story of a case? And borrowing from self-driving cars and aviation, what error rate is medicine actually willing to accept from a machine versus a human? Chapters (00:00) Why morning handover is prime real estate (01:45) Two jobs in one: clinical handover vs. the oral case presentation (04:00) Summarization as a core clinical skill (05:30) Could AI do the summary? Where that question leads (07:45) Hallucination or confabulation? What LLMs actually do (12:30) Probabilistic vs. deterministic: why the computer doesn't give the same answer twice (15:00) The scribe that documented an exam that never happened (17:15) Changing our mental model: it's not Ctrl+F (18:30) Replace the presentation or compare against it? (21:30) De-skilling and the risk of outsourcing the story (23:15) An emergency-department summarization study, scored against physicians (26:30) What a new morning-review workflow could look like (29:00) Self-driving cars: how much error is acceptable? (32:45) Error tolerance in medicine vs. aviation and engineering (34:15) Human plus machine — and who's in charge? Mentioned in this episode Melvin et al. (Toronto, 2019) — qualitative study of the oral case presentation in internal medicine handover Confabulation vs. hallucination as framings for LLM errors Our earlier episode on de-skilling AI use: - Nano banana for cover art - Descript for transcription editing

  3. Oct 6

    #1 Beyond Medical Expertise, AI Imposters

    What do you actually offer a senior resident who's months away from being your colleague? They've passed their boards, their knowledge is solid, and they can clearly do the job. In this episode we work through that question from a recent clinic supervision and land on two practical tools: deliberate role modeling (borrowing from Northouse's behavioral leadership theory) and intellectual candor — showing learners where you still struggle, not just what you know. From an AI corollary, the things that have made language models dramatically more useful over the past two years — context engineering, "soul" and user documents in agent frameworks like OpenClaw, built-in self-reflection loops — look a lot like what we're trying to instill in advanced learners: values, systems thinking, and a sense of the culture they're entering, not more facts. We also ask whether AI models suffer from imposter syndrome or overconfidence, why sycophancy is a self-efficacy problem, and what a model "breakdown" after failed attempts tells us about the cultures we build for humans and machines alike. Chapters (00:00) What can you teach a learner who already knows the textbook? (02:55) Imposter syndrome and self-efficacy in new attendings (05:45) Role modeling as behavioral leadership: task-based and relational (10:20) Intellectual candor: showing your own struggles on purpose (14:50) Growth mindset and the confidence it takes to be vulnerable (15:30) The AI parallel: prompt engineering, context, and OpenClaw's "soul" document (19:00) It's not how smart, it's what values you're infusing (23:30) Correcting practice patterns and reinforcing group culture (25:45) Do AI models have imposter syndrome? Confidence thresholds and sycophancy (29:00) The old "you must know the answer" culture, and why it's shifting (30:30) Wrap-up: advanced learners Mentioned in this episode Northouse, Leadership: Theory and Practice (behavioral approach) Elizabeth Molloy on intellectual candor in medical education (2019) CanMEDS roles beyond Medical Expert OpenClaw and agent "soul"/user documents AI use: - Nano banana for cover art - Descript for transcription editing

  4. Jul 23

    #2 Deskilling

    Looking at the topic of possible "Deskilling" when it comes to using AI tools in medicine. What is "deskilling" What's the history of "deskilling" in medicine Newest evidence of "deskilling" in medicine.

  • 4 Episodes

About

Two clinician educators learning and discussing Artificial Intelligence related topics from a Medical Education perspective.

Information