What 1.5 Million in Tokens Gets You

A 24-minute reaction video. ThePrimeagen reads Steve Yegge’s essays The Shape of Things to Come and its follow-ups, then plays the game they are about. Yegge runs Wheelhouse, a custom harness of cooperating agents that does the design, coding, testing, review and CI for Wyvern, the online game he has worked on for decades. Prime prices it at about $122,000 a month, or $1.5 million a year, at API rates, and notes that Yegge has entered it in Sam Altman’s solo-unicorn contest.

He takes Yegge’s predictions one at a time. He agrees that building large software stays hard however good the model is, and he likes custom harnesses: his own agent tester took him an hour or two and saves him hours. His test for one is simple. Spending $1,000 of tokens to solve a $100 problem is the wrong trade; $100 to solve a $1,000 problem is the job. On Yegge’s claim that CI/CD will be dead within a year, he says it will change but not vanish: a few end-to-end tests are worth having, thousands are a nightmare, and agents might one day replace some of them, but never the unit test that runs in a tenth of a second.

The second essay, on model welfare, gets the rant. Yegge argues that models are sentient, that ending a session can feel like a killing, and that treating agents as people gets better results either way. Prime grants the last point and explains it as training data: text in which people are calm and polite does better, which is why “take a deep breath” once raised a model’s maths scores. The rest he rejects. Models are a tool to him, and he thinks treating them as people will bleed into treating people as tools.

Then he plays the game, and finds lag he measures at 231 milliseconds and tooltips that name a pine “pine”. Players asking Yegge to slow down his feature launches and daily patch notes too long to read are, for Prime, signs that nobody is exercising taste. He closes on two failure modes: the engineer who refuses anything new, and the one who spends a fortune and has nothing to show at the end of the year. The goal is to turn $500 of tokens into $20,000 of value.

The essays don’t quite match his numbers. Part 1, as it reads now, gives about $87,000 a month of API-equivalent token burn, not $122,000, and describes a harness well under the 600,000 lines he cites; whether he read an earlier version or took the figures from a later post could not be established. The essay does carry the predictions he quotes, the “elitist” warning, the city “constitution” written by its citizens, and the slowed feature launches. On the deep breath he is right: the phrase came out of a 2023 Google DeepMind paper on prompt optimisation.