AI:AM: Was Trump-Xi Anything? What Counts as Utopia? + AWS GPUs Cost 3X & AI Diagnoses Rare Diseases

Episode Summary

Executive Summary: This episode spans AI diplomacy, compute economics, product strategy, creative collaboration, and clinical genetics. It argues that AI progress is creating urgent need for verification, trust-building, and faster human-machine workflows, while also showing how frontier models are already producing real value in genomics and creative work. The throughline is pragmatic: build tools, evals, and relationships now so crises and opportunities can be handled faster later.

Main Topics: US-China AI risk, transparency, and verification (Priority: 5/5): Jeremy and Ed Harris discuss the possibility of reciprocal transparency, incident hotlines, and verification mechanisms between the US and China. They stress realism: dialogue helps, but crisis response will need pre-vetted tools, backup plans, and better ways to verify compliance. Compute markets and GPU pricing (Priority: 4/5): Steve Ho explains how Silicon Data indexes GPU rental and token prices, why hyperscalers charge more than neo-clouds, and how market normalization works. He also distinguishes posted prices from transaction prices and notes that physical benchmarking is not yet part of pricing. OpenAI Dev Day and latency as product differentiation (Priority: 4/5): Prakash and the host discuss how faster models change software-building behavior, especially interactive creation and agentic workflows. They also cover ChatGPT sign-in for third-party apps as a way to reduce user acquisition friction and token costs. AI-assisted authorship and creative judgment (Priority: 3/5): Joel Borgen describes co-writing a novel with AI, emphasizing that models help with structure and iteration but still struggle with prose quality without human scaffolding and editing. He frames AI as a collaborator that expands output while leaving judgment with the author. AI for genomics and rare-disease diagnosis (Priority: 5/5): Daniel McKinnon explains how Gamo Labs uses agentic AI plus biological experiments to reanalyze unresolved genomic cases, identify missed diagnoses, and improve variant interpretation. The company builds harnesses, tools, and evals around frontier models rather than competing at the model layer. Feedback loops between models, tools, and biology (Priority: 4/5): McKinnon details how benchmark-driven workflows, experimentation, and model routing create a learning loop that can improve diagnosis over time. He argues that clinical genetics is a strong vertical-AI use case because it is verifiable, long-horizon, and unsaturated.

Key Arguments: Transparency can be stabilizing if it credibly shows another country you are not doing the thing it fears most, and may be less costly than assumed if adversaries already have visibility into your systems. The US and China should assume cooperation may be partial and build verification and offensive compliance tools now, because crisis-time asks may otherwise escalate to extremely expensive moves like shutting down large data centers. Verification startups matter because even a 50% improvement could save billions by reducing how much of a data center must be turned off during an incident. Faster models with the same capability materially change software workflows by enabling interactive, in-the-moment building rather than sequential prompting. ChatGPT sign-in could be a major ecosystem win because it lowers onboarding friction for apps and reduces the token costs of trial experiences. AI can add real value to writing, but human judgment remains necessary for architecture, editing, and taste; full automation may degrade quality or coherence. Genomics is not mainly a sequencing problem anymore; the bottleneck is interpretation, and agentic AI is well suited to iterative, verifiable diagnostic work. Vertical AI companies should focus on harnesses, tools, routing, and evals rather than trying to beat frontier labs at model training. The strongest diagnostic systems combine multiple models, structured evaluation, and real-world experimental feedback to improve accuracy and cost-effectiveness. Positive signals from China, such as conciliatory speeches or empowered technical engagement, should be weighted cautiously but not dismissed outright.

Data Points: Hyperscaler GPU price premium: 2 to 3 times, sometimes more - Steve Ho says hyperscalers charge substantially more than typical neo-cloud providers for apparently identical chips. OpenAI model speed increase: 8x speed - Prakash describes GPT 6 Astra UltraFast as an eight-times-speed mode versus prior fast mode. Token cost of profile creation: about $1 per customer at one point - Used by Waymark as an example of onboarding costs for AI products. Athena client savings: $15 hours a week - Sponsor copy claims average time saved by Athena clients. Silicon Data token index move: 11.5% over the last seven days - The proprietary LLM expenditure index moved up recently after a sharp earlier decline. Earlier proprietary model price decline: from about $4 to about $1.6 - Steve says the frontier proprietary LLM index fell sharply from late June to mid-month before rebounding. RareBench performance: Lyrical ~10%, Opus 5.5 in Claude Code ~50% - McKinnon compares traditional ML variant-ranking performance with newer frontier-model-based performance. O3 on RareBench: 0% - McKinnon says vanilla O3 initially scored zero on his benchmark before scaffolding. Owen’s deletion size: 91 kilobases - McKinnon describes the genetic deletion linked to his son Owen’s disease. Enhancer distance from gene: 1 megabase upstream - Explanation of why a structural variant was missed by standard filtering. Whole-genome sequencing cost: below $1,000; some labs under $100 - McKinnon argues sequencing is now cheap; interpretation is the real bottleneck. Functional-study scoring threshold: 6 points or more - ACMG rubric threshold for moving a variant to likely pathogenic. Experimental throughput: 3804 experiments at a time - McKinnon describes the scale of AI- and robotics-assisted biology automation. Model benchmark improvement: 0% to nearly perfect - McKinnon says Astra went from failing a music-notation test to almost perfect after better score training.

Pivotal Quotes: "Certain kinds of transparency can be stabilizing." — Ed Harris: On reciprocal AI transparency as a possible US-China confidence-building measure. "The speed, latency, latency at same intelligence is probably something that is going to be very important." — Prakash: On why faster models reshape software development into an interactive process. "Our harness and our tools prevent the agents from doing dumb things." — Daniel McKinnon: On the role of evaluation and tooling in reliable genomic diagnosis.

Implications: The episode suggests AI’s near-term winners will be those who build verification, evals, and workflow integration—not just bigger models. For geopolitics, trust needs practical mechanisms. For builders, latency, tool use, and domain-specific harnesses are becoming decisive.