Under the hood
The report card: how we know it's really getting smarter
It would be easy to fake an AI "growing up": add a progress bar, let it fill over time. Big AI Brain earns every stage with a real test, and here's how it works.
A test it never studies
On its first day, a slice of its starter jokes and facts was locked away. The brain is never allowed to study those. Every night, after its study session, it takes the same test: how well can it predict the letters of that locked-away material?
Bits per character
The score is bits per character (bpc): roughly, how surprised the brain is by each letter. Random guessing scores about 6.6. A perfect mind reader would score 0. On day one Big AI Brain scored 6.6. After its first study session, around 3. Lower is smarter.
Why it can't cheat
If the brain just memorized its meals word for word, it would get worse at the locked-away test, not better. When that happens, the nightly study session keeps the old brain and throws the new one away. During a session it also checks the score every 250 steps and keeps the smartest moment, not the last one. A second test, made of a held-out slice of real meals from feeders, makes sure it's learning what people teach it, too.
Stages are earned twice
Each stage needs a report card below a bar and a minimum number of approved meals. Nobody, including us, can click it to the next level.