Episode Summary
Executive Summary: Peter Attia interviews Harvard biomedical informatics chair Zach Cohen about AI’s evolution from rule-based expert systems to today’s transformer-based models, and how that shift is already reshaping medicine. They argue current AI can augment clinicians, improve diagnosis, streamline workflows, and empower patients now—while also raising concerns about misuse, regulation, and the future of medical work.
Main Topics: The three waves of AI (Priority: 5/5): Cohen traces AI from postwar rule-based systems, to data-driven deep learning, to transformer-based large language models, emphasizing why earlier approaches failed and why the current wave scaled. Why modern AI works: data, deep nets, GPUs (Priority: 5/5): He argues the breakthrough came from large labeled datasets, multi-layer neural networks, and GPU parallelism, which together enabled practical machine learning at scale. AI in medicine today (Priority: 5/5): The discussion covers where AI already performs well in medicine—especially imaging, EKG interpretation, administrative tasks, and patient-facing diagnostic support. Radiology, pathology, dermatology, and multimodal models (Priority: 4/5): They explore how image-based specialties are most immediately augmentable, especially as transformers combine images with clinical text and history. Primary care shortages and workflow augmentation (Priority: 4/5): Cohen argues AI can help nurse practitioners and physician assistants fill gaps in primary care, since the U.S. already lacks enough doctors in many settings. Risks, regulation, and misuse (Priority: 4/5): The conversation addresses harmful uses of AI, including misinformation, fraud, and biosecurity risks, as well as concerns about regulatory capture and overregulation. Future possibilities: surgery, mental health, and digital legacy (Priority: 3/5): They discuss robotic surgery, AI-assisted psychotherapy, and the possibility of recording enough personal data to create a posthumous digital approximation of a person.
Key Arguments: AI’s progress is best understood as a series of waves: rule-based expert systems failed because they were brittle, labor-intensive, and unable to scale; modern systems succeed because they learn from data. The real enablers of current AI were not just algorithms but also massive datasets and GPU hardware, which made deep learning practical. In medicine, AI is already strong in narrow, high-volume, pattern-recognition tasks such as EKGs, radiology, pathology, dermatology, and administrative documentation. Transformers and multimodal models extend AI beyond image-only recognition by combining text, images, and clinical context, making them more useful for real-world medical decision-making. AI can help address the shortage of clinicians by augmenting nurse practitioners, physician assistants, and lower-volume doctors rather than simply replacing abundant physicians. Patients are already using AI to help solve diagnostic mysteries, and this may become a major force in correcting misdiagnosis and improving access to expertise. The biggest near-term danger is not Skynet-style autonomy but human misuse: misinformation, fraud, biosecurity, and social manipulation. Regulation is necessary, but there is a risk that large companies will shape rules in ways that entrench their own market position. The most likely medical transformation is not a futuristic breakthrough but better deployment of existing models, better workflows, and reimbursement structures. AI may eventually support earlier detection of chronic disease, including neurodegeneration, by integrating voice, gait, eye movement, imaging, and longitudinal data.
Data Points: AI generations: 3 - Cohen describes the current era as the third generation of AI. Medical school slots unfilled in some specialties: ~50% - He says about half of certain specialty training slots, including pediatric endocrinology and developmental disorders, are not being filled. Primary care shortage by 2035: ~50,000 doctors - He cites AAMC estimates that the U.S. will be short roughly 50,000 primary care doctors by 2035. ImageNet scale: millions of images - Used as an example of the large labeled datasets that enabled modern deep learning. Medical text corpora: 1–6 terabytes - He estimates the scale of text used to train large language models from human-generated text. 2012 image recognition breakthrough: year 2012 - He identifies 2012 as the moment deep neural networks on GPUs clearly outperformed competitors in image recognition. Transformer paper: 2017 - He references the 'Attention Is All You Need' paper as the key transformer breakthrough. GPT-3.5 public tipping point: December 2022 - He says chatbots became mainstream around the release of GPT-3.5. Echocardiogram training set: 1 million echocardiograms - He cites a recent study training on a million echo studies and a million reports. Harvard undergraduates receiving mental health support: 60% - Attia notes Harvard’s mental health demand is very high, with 60% of undergraduates receiving some support. Autonomous vehicle safety claim: 0 fatalities - He relays a company claim that if every vehicle matched its autonomy level, fatalities would disappear. Rewind AI storage: gigabytes - He says his personal recording system stores surprisingly little data because it compresses audio and snapshots efficiently.
Pivotal Quotes: "The goalposts around the Turing test keep getting moved." — Zach Cohen: He explains why debates over whether AI is truly intelligent are less useful than evaluating what it can actually do. "There are three things that have taken the relative failures of first and second attempts at AI and got us to where we are today." — Zach Cohen: He summarizes the key enablers of modern AI: data, deep neural networks, and GPUs. "I think that unlike my 10 year prostatectomy by robot prediction, I'm not as certain." — Zach Cohen: He contrasts his confidence in near-term medical AI augmentation with his uncertainty about more speculative long-term predictions.
Implications: AI is already useful in medicine, especially for imaging, triage, documentation, and patient self-advocacy. The next decade will likely be shaped less by new breakthroughs than by workflow integration, data access, regulation, and whether AI augments clinicians or entrenches incumbents.
About Peter Attia Drive
Expert insight on health, performance, longevity, critical thinking, and pursuing excellence. Dr. Peter Attia (Stanford/Hopkins/NIH-trained MD) talks with leaders in their fields.