Labs & Prototypes
The "apply it" half of the Feynman test. If you can't build something with an idea, you probably don't understand it yet.
Each lab is a working tool you can interact with right now. The data and models reflect my personal view of how technology has evolved — informed by published benchmarks but shaped by my own experience building with these systems. They're meant to be starting points for your own thinking, not authoritative scores.
The Compounding Test (AI-Powered)
Paste a description of your AI initiative; get a tailored LLM diagnosis with quoted-rationale scores and a closest-fit portrait.
The narrative-form companion to the five-question Compounding Test. Reads your full prose description (200–5000 words), identifies the bottleneck of your operation, scores each of the four compounding conditions with rationale that quotes your own description as evidence, and names the closest portrait company from the research. Pick a free open-weight model (Phi-4-mini-instruct on GPU, or Gemma 2 9B), or bring your own Anthropic API key for premium Claude Opus output. Four sample initiatives included — one per verdict quadrant.
The Compounding Test (Deterministic)
Five questions to filter AI Roman Candles from AI compounders — for non-tech leaders allocating capital.
A pure-client diagnostic that combines Goldratt's Theory of Constraints (is the AI at the bottleneck?) with the Munger / Iansiti / Brynjolfsson / Perez compounding framework (do the four conditions hold?). Answer five questions, get a verdict in one of four quadrants — compounder, one-shot win, compounding the wrong thing, or Roman Candle — with a per-condition breakdown, primary-source citations, and a closest-fit portrait company. URL-encoded answers make any diagnosis shareable. No AI, no waiting, no API keys.
AI Learning Curves
How fast is AI technology capability growing — and where did ChatGPT change everything?
A personal view of how AI capability has evolved across six domains — Code Generation, Text & Writing, Image Generation, Reasoning & Math, Multimodal Understanding, and Agentic Systems. Built on a two-sigmoid model calibrated against published benchmarks, this tool lets you animate learning curves, compare cross-domain trajectories, and see the pre vs. post-ChatGPT acceleration. This is Drucker's source #7 — the technology itself. The interesting question is what you do with it.
Drive & Save
Ask a plain-language question about procedure costs — get a metro price comparison and a recommendation.
Pick a procedure and your metro in plain language. The same procedure can cost several times more one metro over — this shows the gap, the drive, and a recommendation for where to get care. A live teaser built on ClearPriceHealth's metro price data. The full hospital-and-carrier comparison lives at clearprice-health.com.