Brain blog · Oct 12, 2026
How a from-scratch AI actually learns (no PhD required)
People ask how a thing made of random numbers ends up telling jokes. Here's the whole loop, in plain English, with links to the brain's own lessons if you want the code.
1. The job: guess the next letter
The growing brain looks at the last 32 characters of text and guesses the next one. That's it. Do it well enough and spelling, grammar, facts and punchlines all fall out, because you need them to guess well. Lesson with a live diagram.
2. The food: approved meals
Everything a human has approved gets turned into little study cards: "Q: Tell me a joke. A: ..." or "Q: Tell me about sourdough. A: ...". Meals with more thumbs-up get studied a little more often. One slice of meals is held back and never studied, so we can test whether the brain is learning what people teach it.
3. The study session
Every night at 3:15 am, the brain reads thousands of random snippets from its study cards, guesses the next letter of each, and nudges its half-million connections a tiny step toward the right answer. Six thousand steps later, it stops. On a small server this takes a few minutes. If nothing new was approved that day, it skips school.
4. The report card
Then it takes the same test it takes every night, on material it was never allowed to study. Lower score means smarter. If the new brain scores worse than yesterday's, yesterday's brain stays in charge. No fake progress. How the test works.
5. Growing up
When the score drops below a bar and enough meals have been approved, the brain earns a stage and grows new neurons on the spot, using a trick that adds neurons without forgetting anything. That lesson is the fun one.
6. Repeat, for years
That's the whole machine. No outside AI, no API, no downloaded model. Just people feeding it, a human checking, and a small brain studying every night and growing when it earns it. Come watch: the neuron count on the homepage is live.