The most talked-about AI model of the past week is not a bigger, smarter chatbot. It is a model that refuses to chat at all. Jev, from a startup called TypeSafe AI, does not write you a paragraph. It returns a calibrated decision: a yes or a no, a pick from a list you defined, a score, each with a confidence number attached. It was built by Diogo Almeida, one of the researchers who worked on ChatGPT and RLHF, and it came out of stealth to a wave of developer enthusiasm, a reported forty million dollars in funding, and ten-billion-dollar valuation chatter. TypeSafe claims it is radically cheaper than a conversational model, on the order of hundreds of times, though that number is still self-reported and worth treating as the ceiling rather than the expectation.
I read the coverage with a finance hat on, and the thing that struck me is how badly the industry has been fitting the wrong shape of tool to our work. Finance almost never needs a paragraph. It needs a decision. Is this invoice a duplicate. Should this payment run today or wait. Does this remittance match that open receivable. Is this vendor bank-detail change a fraud flag. Which of these four hundred disputes gets worked first. What GL code does this line belong to.
Every one of those is a decision with a defined set of possible answers, made thousands of times a day. A model that produces eloquent prose is both the wrong shape and, priced by the million tokens, the wrong economics for running that on every transaction.
Read the full article on Substack