The AI Revolution Isn’t What Silicon Valley Promised — with Josh Tyrangiel

What It Covers

Journalist Josh Tyrangiel, author of *AI for Good*, argues that meaningful AI progress happens outside Silicon Valley — in hospitals, government agencies, and battlefields — where practitioners apply narrow, disciplined AI to specific problems rather than chasing utopian promises. He examines sepsis detection, IRS modernization, autonomous warfare, Chinese AI competition, and journalism's survival.

Key Questions Answered

• •Narrow AI deployment: Cleveland Clinic reduced sepsis mortality by 41% — saving over 1,000 lives annually — by applying the Bayesian Health prediction model through Epic Health systems with relentless refinement over one year. The key was accepting 90% accuracy as sufficient rather than waiting for perfection. Leaders should define two specific parameters before deploying AI: quality improvement and cost reduction only.
• •Government AI modernization: Operation Warp Speed's COVID vaccine distribution succeeded using a $16 million Palantir digital twin of the supply chain — not a perfect end-to-end system, but functional enough to run logistics like a video game dashboard. The IRS was similarly modernizing its 1960s-era individual master file using machine learning before DOGE disrupted progress, demonstrating that political will, not technology, is the primary barrier.
• •Chinese AI dumping threat: Beijing is executing a deliberate strategy of subsidizing cheap LLMs — potentially one-thirtieth the cost of US frontier models — to convert enterprise customers globally, mirroring its steel and infrastructure playbook in Africa and Latin America. Combined with alleged model distillation (IP theft), this threatens Silicon Valley's revenue model before AI companies have achieved profitability on compute-heavy products.
• •AI and warfare transformation: Ukraine's military demonstrates that AI-powered drone warfare has compressed decades of military evolution into roughly one year — Russian soldiers now survive only hours on the battlefield. Future US conflicts will likely be managed remotely from command centers like CENTCOM Tampa using small autonomous drones, making expensive legacy platforms like B-1 bombers strategically vulnerable and potentially obsolete against asymmetric AI-enabled opponents.
• •AI regulation framework: The International Atomic Energy Commission offers a viable regulatory model for AI — requiring declared model sizes, mandatory inspections, and international safety certification above certain capability thresholds. The primary obstacle is the unexamined assumption that China would never cooperate, despite evidence that both the US and China prioritize stability and have already held preliminary AI discussions at the presidential level.
• •AI writing utility: Use AI as a structural diagnostic tool rather than a content generator — paste struggling paragraphs and request logically coherent bullet points to identify whether the underlying argument actually holds together. This approach works particularly well for explaining complex technical concepts through metaphor. AI-generated prose mimicking established writers produces recognizable distortions, making it unsuitable as a direct substitute for developed voice.