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

What It Covers

Four segments cover US-China AI verification diplomacy with Gladstone AI's Harris brothers, GPU pricing disparities between hyperscalers and neo-clouds via Silicon Data's Steve Ho, AI-assisted novel writing with author Joel Borgen, and Gamo Labs founder Daniel McKinnon's use of AI agents to diagnose rare genetic diseases in previously unresolved pediatric cases.

Key Questions Answered

• •US-China AI Verification Gap: Without pre-established verification infrastructure, the only available response to a major AI incident would be demanding China shut down all data centers above a certain compute threshold — an astronomically costly ask for both sides. Verification startups should engage intelligence community contacts now, before a crisis, because vetting new national technical means historically takes years of serial bureaucratic time that cannot be compressed.
• •Reciprocal Transparency as Low-Cost Signal: Offering China limited visibility into US frontier lab activities may cost little in practice, since Chinese actors already have deep access to US systems. Structured transparency — showing enough to confirm the US is not pursuing decisive strategic dominance — could stabilize relations without meaningful intelligence loss, and could be piloted first between OpenAI and Anthropic as a domestic test bed.
• •Hyperscaler GPU Markup: Renting H100s from AWS, Azure, or Google Cloud consistently costs two to three times more than equivalent neo-cloud providers. The premium reflects bundled software, compliance tooling, long-term enterprise relationships, and product differentiation — not raw compute performance. Silicon Data's index separates these categories, treating hyperscaler pricing as a distinct product class rather than a comparable unit of compute.
• •Ultrafast Inference Changes Software Development: OpenAI's GPT-6 Astra Ultrafast runs at eight times the speed of prior fast modes, enabling real-time interactive co-creation — building a game collaboratively while staying in flow, rather than submitting prompts and waiting. Speed at equivalent intelligence levels is the next major capability threshold, shifting AI from an asynchronous tool to a synchronous creative partner during active development sessions.
• •AI Novel Co-authorship Requires Human Architecture: AI models handle scene execution well but consistently produce unreadable prose without detailed human scaffolding — a chapter plan, beat structure, and post-generation editing pass. Author Joel Borgen pre-architected the first half of his novel before engaging models, then edited AI output down roughly 14%, removing over-explanation and generic phrasing. Context length expansion was the single largest unlock enabling full-novel coherence across sessions.
• •Genomic Interpretation Bottleneck, Not Sequencing: Whole genome sequencing now costs under $100 at high-throughput labs, but interpretation remains the unsolved problem. Vanilla Claude Opus scores approximately 50% on Gamo Labs' RareBench variant prioritization benchmark, versus roughly 10% for the leading traditional ML tool Lyrical. Gamo's pipeline uses model ensembling with case-specific routing — Claude, Grok, and Gemini each perform better on distinct case clusters — plus a harness that prevents agents from hallucinating gene name substitutions.