Old wiring → new wiring
Old wiring: I’m worth what I know. New wiring: A price tells you what’s scarce. It was never a verdict on a life.
The story
On May 9, 2015, I stood on a red dot at Northwestern and told a room full of people to evolve faster.
The talk did well. About 600,000 people have watched it since. For most of the next decade I carried around a quiet confidence that I understood how a person keeps up with a world that’s speeding up. The talk was called Evolve Faster, after all.
Upon hearing this overconfidence, the universe smirked and cracked its knuckles.
A few years ago I sat down with an AI model and decided to test it on something I knew cold. I described the ideas from that talk, the ones I’d turned over for years until they felt earned, and asked what it knew.
Under the hood: the price of a thought
In November 2022, a million tokens from a model as capable as the one behind the first ChatGPT cost about $20. By October 2024, the same level of capability cost 7 cents. That’s from Stanford’s AI Index, and it works out to roughly 280 times cheaper in under two years.
One caution, because I’d rather be accurate than impressive. That’s one capability level on one test. The research group Epoch AI looked across several and found the price of a fixed level of performance falling somewhere between 9 and 900 times a year, depending on the task.
Take the slow end. Nine times cheaper per year means whatever a machine does today for a dollar, it does next year for eleven cents. I can’t think of anything else I buy that behaves like that.
Intelligence has entered its Costco era. Shelves to the ceiling, the unit price dropping, everybody taking home more than they meant to.
The rewire
It was good wiring for a long time. Answers were scarce, so having them was a living and an identity. When something you spent years being proud of gets cheap to buy, borrow or generate, it feels like somebody marked you down personally. But a price is information about what’s scarce. It was never a verdict on a life. I had just built a fair amount of my life as if it were, and I didn’t notice until the price moved.
The machine can hand me three better angles in four seconds. It can’t make me willing to take them, and it can’t make me stop defending the old version because the old version had my name on it. That work is still mine, and it turns out to be most of the work.
This week: the repricing audit
Twenty minutes, one sheet of paper.
- Write down five things you’re paid or praised for. Be specific. “Strategy” doesn’t count. “I can read a term sheet and find the clause that will hurt us” does.
- Give the model each one cold. Paste this first: I’m going to describe something I’m known for being good at. Don’t flatter me. Do it as well as you can. Then tell me which parts of a top professional’s version you could not have produced, and why.
- Grade it the way you’d grade a new hire. For anything it got 70% right or better, write down what the missing 30% was. Judgment? Taste? Knowing the people involved? Being the one who signs?
- Answer “so what do you do?” using only the 30% list.
Sources
- Stanford HAI, “AI Index 2025: State of AI in 10 Charts” ($20 to $0.07 per million tokens, November 2022 to October 2024): https://hai.stanford.edu/news/ai-index-2025-state-of-ai-in-10-charts
- Epoch AI, data insight on LLM inference price trends, March 12, 2025 (9 to 900 times per year): https://epoch.ai/data-insights/ai-inference-price-trends