TypeSafe AI's manifesto is titled Composable AI: Build Prod, Not God. They say their job is to make intelligence something you can call from code, so more software can use it. In an appendix they define an economic revolution as world productivity growth hitting 3 percent a year within five years and staying there for ten. They mean total factor productivity: how much extra the whole economy produces after you account for more workers and more machines. They call that unprecedented. They say it is possible if the gains spread through the real economy.
They are not racing AGI. Today's models already know enough to create a lot of economic value. The models know a lot, reason across fields, and do things that looked impossible a few years ago. After years and trillions of dollars, most software still is not smart, and daily life is mostly the same. That should tell us something, they write: the bottleneck is not intelligence. It is that today's intelligence is hard to build on.
Early cars were horseless carriages. Inventors put a motor in and kept the high seats, the springs, and even the whip socket in some models. New tools get forced into the shape of the thing they replace. Current AI is trained to be a helpful, clear, pleasant assistant. That is reasonable if a person is on the other side. The result is AI that needs a person in the loop instead of running in the background. Software has never worked that way. Even complex software is built from simple logic, stacked in layers, with every branch something you can check. They want a programmer to call AI for a judgment, and keep code for exact computation. Computers already branch on bits. Imagine if they could also branch on common sense, understanding, and intent. An appendix note calls that neuro-symbolic AI: smart if-statements.
The people who built databases did not imagine Google. The people who built internet protocols did not imagine Stripe. They made a lower layer dependable enough to run in the background and to stack on. Intelligence today is like databases before SQL: powerful, but every use is custom. Once a smart decision is as dependable as a database query, builders will stack them the same way.
Safety comes first if you want to stack. You let a part run unattended if it is reliable. You only build on it if you can trust it. It takes trust to bury a part five layers down. People will only do that if they can inspect it, test it, and constrain it piece by piece.
They name three steps. Ship that shape of AI at the best intelligence for the money. Make it reliable enough for real automation. Then give the world higher-level pieces that are stable enough to stack. The last line is: We're building prod, not God.
Their docs describe Jev as TypeSafe's main model, and the first System One model: a model that returns answers code can use, not chat. You send the current state and typed questions. You get structured answers your code can use. No generated text. No parsing. Three question types: Choice, which picks among options; Score, which grades against a rubric; and Noul, the probability a yes or no is true, from 0 to 1. You can mix all three in one call. Each question is evaluated on its own, in parallel, against the same state. Adding questions, they write, barely changes how long you wait.
Editorial
Chat is the whip socket. What you still need is a yes or no your program can check, and throw away if it does not trust it. Same leftover as a database that can refuse a bad write.
They are right that stacking is the bottleneck, if the decision is as easy to call as a query. I have not run Jev. I do not know if its confidence matches how often it is right, or if it is only a cheap classifier. The 3 percent productivity path is a hope in an appendix, not a result of the model.