OpenAI’s Big Reset + A.I. in the Doctor’s Office + Talkie, a pre-1930s LLM

Episode Summary

Executive Summary: This episode covers OpenAI’s strategic pivot—loosened Microsoft ties, a major Amazon deal, Stargate’s recalibration, and the Elon Musk lawsuit—then shifts to AI in medicine, where doctors are rapidly adopting AI scribes and decision tools while warning about privacy and deskilling. It ends with a look at Taki, a vintage LLM trained on pre-1930 text, as a tool for forecasting and historical simulation.

Main Topics: OpenAI’s strategic reset (Priority: 5/5): The hosts break down OpenAI’s rewritten Microsoft partnership, which gives OpenAI more flexibility to work with other cloud providers and reshapes AGI-related revenue terms. Amazon partnership and cloud competition (Priority: 5/5): OpenAI’s expanded deal with Amazon lets it sell through Bedrock, add Codex, and develop custom models, signaling a broader push to reach enterprise customers beyond Azure. Stargate and infrastructure reality (Priority: 5/5): OpenAI is scaling back some data center ambitions and leaning more on third-party capacity, reflecting tensions between grand compute plans and practical financial constraints. Elon Musk vs. OpenAI trial (Priority: 4/5): The federal trial focuses on Musk’s claim that OpenAI betrayed its nonprofit mission; the hosts discuss legal stakes, internal emails, and the odds of meaningful remedies. AI in medicine adoption (Priority: 5/5): Dr. Adam Rodman explains how AI scribes, evidence tools like Open Evidence, and patient-facing chatbots have become routine in medicine, while clinicians adapt their workflows. Risks and limits of medical AI (Priority: 4/5): The segment emphasizes privacy concerns, hallucinations, overreliance by patients, and the risk that AI may erode training and diagnostic skills in younger clinicians. Taki and vintage language models (Priority: 3/5): The final interview explores a pre-1930-trained LLM designed to test forecasting, historical reconstruction, and the effects of training cutoffs on model behavior.

Key Arguments: OpenAI’s new deals show it is no longer structurally locked into Microsoft, which should help it distribute models more widely and grow revenue. The Stargate slowdowns do not necessarily mean retreat; they may indicate a shift toward capital efficiency and preparing for a public offering. The Musk lawsuit is more likely to distract and complicate OpenAI than to destroy it, though it could create serious headaches if Musk were to win. AI adoption in medicine is already mainstream because doctors are bringing in tools that make their jobs easier, especially scribes and decision support systems. Medical AI should be used for preparation, reference, and general questions, but not as a substitute for clinical judgment on high-stakes decisions. The biggest short-to-medium-term concern in medicine is not that AI will replace doctors, but that it will deskill the next generation. Taki demonstrates how training data cutoffs shape model behavior and why historical LLMs are useful for studying forecasting, leakage, and cultural attitudes over time.

Data Points: Microsoft stake in OpenAI: about $135 billion - Described as Microsoft’s valued stake in OpenAI during the partnership discussion. OpenAI revenue-sharing arrangement with Microsoft: continues until 2030 - The rewritten deal removes the AGI-triggered cutoff and extends revenue sharing to 2030. Amazon investment in OpenAI: $50 billion - Announced as part of the expanded OpenAI-Amazon deal. Stargate infrastructure project: $500 billion - The joint OpenAI infrastructure project referenced in Financial Times reporting. Reported OpenAI subscription projection: ChatGPT Go to 112 million users - The Information report said OpenAI projected massive growth for its lower-cost subscription tier. Reported OpenAI premium subscription projection: high-end subscriptions to fall 80% to about 9 million - The Information report described a shift away from premium subs toward cheaper tiers. Doctors using AI professionally: more than 80% - American Medical Association survey cited during the medical AI segment. Doctors using Open Evidence in a day: 1 million consultations in 24 hours - Open Evidence reported this usage figure in March. Doctors using Open Evidence: approximately 40%+ - Hosts cited recent data suggesting broad uptake among U.S. physicians. Prescription refill trial: almost 200 routine drugs - Utah trial of an AI agent to autonomously renew prescriptions for routine medications. CT scan detection lead time: up to 3 years early - Mayo Clinic’s Red Mod identified pancreatic cancer-related changes years before diagnosis. Polyp detection skill drop after AI use: 6 percentage points - Rodman cited a study where doctors’ detection performance declined after using AI assistance. Taki dataset size: 240 billion tokens - The pre-1930 training corpus size for the vintage LLM project. Model size comparison: smaller than GPT-3, bigger than GPT-2 - Rodman described Taki’s scale relative to earlier OpenAI models.

Pivotal Quotes: "We awakened an ancient evil." — Kevin Roose: Joking about OpenAI’s Codex model randomly mentioning goblins and other creatures. "We are here because Musk didn’t get his way at OpenAI." — OpenAI lead counsel William Savitt: Summarizing OpenAI’s defense in the Elon Musk trial. "The biggest sort of normal doctor technology ... are what are called AI scribes." — Dr. Adam Rodman: Explaining the most common AI tools now used by physicians.

Implications: OpenAI is moving toward a more flexible, commercial, and capital-efficient structure, while medicine is becoming an early proving ground for AI utility and risk. The next big questions are whether AI can scale profitably, avoid harming training, and deliver real-world benefits without overclaiming.