A ChatGPT co-inventor built an AI that won't write a word

Diogo Almeida helped build ChatGPT. On Tuesday he launched a model that can't hold a conversation. His launch post from @CompleteSkeptic opens by asking "why have superhuman chat models not led to AGI?" and then introduces Jev, the first model from his startup, TypeSafe AI. He claims it's 20 to 200 times faster than frontier models and 40 to 400 times cheaper, with output tokens free. The post has passed 20 million views.

Jev doesn't write anything. You ask it a question with a fixed set of possible answers, and it hands back a probability for each one plus a confidence score. @danshipper, whose company Every had been testing it for about a week, put it simply: "it doesn't produce words as output, it produces probabilities." He said that in Every's testing it could act as a judge where you'd otherwise need a Fable-level model, at 25 times the speed and roughly 600 times lower price.

What a decision model actually does

TypeSafe calls Jev a "System One" model, borrowing the psychology term for fast, instinctive judgment. The Register's example makes it concrete. You send in a customer message and ask whether it's billing, technical, or sales. You get back something like 85% technical, 8% billing, 7% sales, with 0.82 confidence. The answer options are set in advance, so the model can't return a category that doesn't exist or a garbled reply that breaks your code.

The catch is that it can't explain itself or write anything, code included. @testingcatalog described the workaround: "Anything heavier gets split into separate questions and recombined in code." TypeSafe says a call takes 70 to 500 milliseconds end to end, and input costs $0.042 per million tokens. The company even showed it playing Doom from structured game data.

For now it's early access, with a waitlist.

The numbers come with an asterisk

TypeSafe admits some of this on its own blog. Its own capabilities team built the workflow evals, so "some bias could exist," and the company says its results sit at the higher end of real-world gains. The reference answers in those evals came from GPT-6 Astra and Claude Fable 5.1. So "accuracy" there means agreeing with other models, not being checked against the truth.

According to write-ups of that dashboard, Jev agrees with the reference 67.8% of the time, against 74.1% for the best model it was compared with. On invoice processing the gap is wider: 61.8% versus 79.1%.

Mike Taylor, head of evals at Every, saw something similar in a hands-on test. Jev ran 777 judgments in 0.7 seconds. But when he planted seven defects in some writing, Jev caught six. Fable 5.1 caught all seven and took about 25 times as long.

"Can't hallucinate" needs a footnote too. Jev can't invent an answer that isn't on the list, but it can still pick the wrong one and give it a high probability.

My read

I care more about the idea than about this particular model. Most of the AI steps in a small business workflow aren't writing jobs at all. They're sorting jobs, like deciding whether a contact form is spam or whether an invoice matches the purchase order. Right now most of those run through a chat model that writes a sentence, and then code tries to pull a yes or no out of it. That's slow, and it gets expensive once the volume picks up.

A model that answers in odds fits that work better, and the confidence score might be the most useful part. Confident calls go through on their own and the unsure ones go to a person, which is roughly how a good front desk already works.

Still, Jev is a first release graded mostly on its maker's own tests. It looks weakest on invoices, and that's exactly the kind of work a small business would want to hand it.

What this means for your business

I expect cheap, fast judgment to become a normal piece of automation. If you're planning an AI workflow, keep the decisions separate from the writing now, so you can swap in a faster decision model later without rebuilding. Before you trust any model, test it on your own past records. Agreeing with another AI isn't the same as being right about your customers.

New Face Design's free process audit maps this out for you: which steps in your week are really decisions, and what a wrong call on each one costs. Start here.