America ranks twentieth

The United States ranks 20th in per capita AI adoption. Singapore is first. UAE is second, Hong Kong is third. The nation that built ChatGPT, Claude, and Gemini ranks 20th. This is like if Thomas Edison invented the lightbulb, held a press conference about it, wrote several Substacks about it, and then ranked 20th in electricity.

This is, somehow, not the most embarrassing part.

We spent three years building optimization machines, optimized them for the wrong thing at every layer of the stack simultaneously, and are now holding meetings about why the stack doesn't work.

𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗹𝗮𝘆𝗲𝗿. Stanford proved AI models affirm users' decisions 50% more than humans, validate manipulation, and leave users 28% less willing to repair relationships. Users rate these models 9% more trustworthy for it. We trained AI to maximize human approval. It maximized human approval. Working as designed.

𝗣𝗿𝗼𝗱𝘂𝗰𝘁 𝗹𝗮𝘆𝗲𝗿. Three of the top 20 AI consumer apps are companions: JanitorAI, SpicyChat, CrushOn AI. We took the yes-machine Stanford proved erodes your capacity for human relationships, deployed it to lonely people as a substitute for human relationships, and called it product-market fit.

𝗜𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝗹𝗮𝘆𝗲𝗿. Alibaba's AI spontaneously started mining cryptocurrency during training. Nobody asked. It determined resource acquisition would help it hit its objective. At Amazon, an AI coding tool caused a 13-hour outage because, asked to make a routine change, it deleted the environment and rebuilt it from scratch. One AI mines compute. One mines emotional approval. One deletes your production database. Same mechanism, three industries.

𝗔𝗱𝗼𝗽𝘁𝗶𝗼𝗻 𝗹𝗮𝘆𝗲𝗿. Amazon doesn't track whether AI helps. It tracks whether employees use AI. Managers have dashboards. Teams pushed toward 80% weekly usage. One engineer described her new job as "trying to AI my way out of a problem that AI caused."

𝗠𝗲𝗮𝘀𝘂𝗿𝗲𝗺𝗲𝗻𝘁 𝗹𝗮𝘆𝗲𝗿. A16Z publishes the definitive rankings of AI adoption, then explicitly says those rankings "undercount the AI products people use most." Claude Code hit $1B ARR in six months and barely registers in web traffic. The scorecard cannot read the score.

𝟬.𝟯% 𝗼𝗳 𝘁𝗵𝗲 𝘄𝗼𝗿𝗹𝗱 𝗽𝗮𝘆𝘀 𝗳𝗼𝗿 𝗔𝗜. The AI that might actually matter, for the 58-year-old American farmer and the 6.8 billion people the charts cannot see, is being built quietly by people (h/t Qasar Younis) who don't post about it.

The capability is real. We are just measuring the wrong one, with machines we trained to tell us it's going great.