TypeSafe AI, the startup founded by former OpenAI researcher Diogo Almeida, released a transformer model called Jev this week. It’s not a large language model. It doesn’t write text at all. Instead, it returns probabilities, or what the company calls “calibrated decisions,” for outcomes the developer defines up front. TechCrunch AI reports that developer demand was so heavy the company briefly couldn’t serve its API.
Almeida isn’t a random founder. He helped build ChatGPT and co-invented reinforcement learning from human feedback (RLHF), the training method behind most of the chatbots you use today. His pitch for Jev starts with a confession: “We have lightning in a bottle, and yet it is not useful.” His diagnosis, as told to TechCrunch AI, is that the industry has spent four years optimizing for human language, while “computers speak a different language.”
Here’s what Jev actually does, and why developers are excited.
It outputs numbers, not sentences
You define the possible outputs ahead of time. Jev hands back a probability for each one. That’s the whole trick. Because the model can’t invent an answer outside the set you gave it, it cannot hallucinate in the usual sense. TypeSafe describes it as a “System One model”: intuition over reasoning, built for one narrow task at a time.
It’s absurdly cheap
Output tokens are free. Input tokens are priced per billion, not per million. That’s a different pricing tier from anything the frontier labs sell, and it changes which jobs are worth handing to a model at all.
Early users report big speed and accuracy wins
Pranit Sharma, a software engineer at Vercel, told TechCrunch AI his team had used an OpenAI model to run a safety classifier on commands. Swapping in Jev made results five to 18 times faster, with better accuracy. Nikhil Mudholkar, CTO of Bryo AI, tested Jev against Gemini for sorting business emails. Gemini was slightly more accurate, but 10 to 20 times more expensive. What sold him was the confidence score: “it is the only one that hands back a real probability which makes it ideal for automating workflows.”
It works as a watchdog for other AI
Using an LLM agent to monitor another LLM agent gets expensive fast. Almeida sees Jev as the cheap guard at the door: scoring agent traces, flagging jailbreaks, deciding whether a step looks safe. Armin Ronacher, CTO of Earendil (the company behind the open source model harness Pi), also pointed to model routing. Predicting which workload needs which model is useful but too pricey to do with an LLM. Jev’s cost makes that real-time sorting practical.
Everything it learned came from synthetic data
Almeida says Jev trained exclusively on synthetic data using a method he calls “reinforcement learning from calibrated decisions.” He called the decision to generate all their own data “one of the best bets I’ve ever made in my life,” better than RLHF. Half the company is a lab dedicated to that subfield.
The caveats
Ronacher put the trade-off plainly: Jev “delegates the hallucination problem a little bit to the user.” A 50% score is a coin toss you should ignore. A 95% score you can act on. That judgment now lives in your code, not in the model.
Almeida also won’t discuss the architecture. Outside observers suspect Jev sits on top of an open-weight LLM. And in at least one head-to-head test, a full LLM still edged it on accuracy. Jev wins on cost and speed, not on raw intelligence.
Why this matters
The name is the thesis. William Stanley Jevons was the economist who noticed that when a resource gets cheaper, people use far more of it. Almeida expects the same for intelligence: “smart software all over the place in a way that’s emergent and distributed,” closer to the early internet than to the mega-apps being built today.
What stands out to me is how obvious this looks in hindsight. Most production AI calls are classification jobs wearing a chatbot costume. Is this email spam? Is this command dangerous? Which model should handle this request? None of those need prose. Ronacher’s read is that nobody bothered building this earlier because subsidized LLM pricing meant “you often don’t have to be creative yet.”
That subsidy won’t last forever. Expect competitors to follow, and expect TypeSafe to ship more modalities. Almeida, asked whether TypeSafe is a frontier lab, said the main product of frontier labs is “fear or hype,” and he’d rather sell intelligence. The full interview is worth reading at TechCrunch AI.