A ChatGPT co-creator just shipped an AI that cannot write a single word. Diogo Almeida spent years at OpenAI on the research that made ChatGPT talk like a person. Two years in stealth, $40M raised, and last week he shipped Jev. It does not generate text. At all. You hand it messy state and a list of allowed answers. It returns one of them, plus how sure it is. It can't hallucinate. There is no string to invent. It can still be wrong, but not creative about it, and it tells you when it's guessing. Output tokens are free. 🤯 Unmetered. $0.042 per million in, zero out. The structure saves more than the price does. Forty questions through a chat model is forty calls, each re-sending the whole document and everything said so far. Caching softens that. It doesn't fix it. Jev reads the state once and answers all forty at once. I couldn't resist testing it. The first thing I ran was a rent roll. Two real ones, two different PM systems. One stacks 1,600 rows of charges under 320 units, writes status as "C," and prints the "Unit" header one column off from the actual unit numbers. I used to write a custom script per format. Last one was 76 lines. One tool this time. No new code. 39 to 49 questions per roll. Which column is the unit. What "NTV" means. Whether a pet rent line with the dog's name in it is rent or a fee. Plain code did the math. Both tied to the seller's own totals. 17 of 17 checks on one, 16 of 16 on the other, three runs each. 1.3 seconds. A tenth of a cent. Same questions through Claude. Sonnet 5 ran 7 to 10 times slower, cost 59 times more, and missed a unit in 2 of 3 runs. Haiku missed it all three. That unit had a resident and no status. Jev came back under 50% confident and kicked it to a person. A confident wrong answer on a rent roll is how you overpay for a building. It's running on T12 extraction now, and on deal screening against my buy box. At these speeds people aren't calling this thing. They're polling it. Someone has a bot playing Doom at ten decisions a second for about $7 an hour. The launch demos ran driving and drone agents in simulation. Every AI product you use today is request and response. You ask, you wait, you read. That shape exists because generation is slow and metered, so you ration the asks. Remove the wait and the meter and AI stops being something you consult. It becomes something that runs, deciding as it goes. Make anything cheap enough and people don't spend less on it. They use far more of it. Cheaper steel didn't shrink steel bills. It put steel in everything. Same here. You check every row instead of a sample. Every deal in the funnel instead of the three you had time for. Two rent rolls is two rent rolls. Jev flipped on a couple of low-confidence calls between runs. One Haiku miss was my own code's fault, fixed across all three before I counted. Been a long time since a launch made me cancel my afternoon.