Written by Admin Alex · Fact-Checked by M.Ali · Info Verified September 2026
We review and update this article regularly as new information becomes available.
TL;DR: Diogo Almeida, the OpenAI researcher who helped invent RLHF and build ChatGPT, left to found TypeSafe AI two years ago. His new model, called Jev, doesn’t generate text at all. It outputs calibrated probabilities instead, which means it structurally can’t hallucinate the way conversational models do. Developers testing it are reporting results five to eighteen times faster than competing approaches, and demand got high enough that TypeSafe briefly ran out of serving capacity.
Every AI model you’ve used generates text one token at a time, guessing at the next most likely word based on everything that came before it. That’s true whether it’s ChatGPT, Gemini, or Claude. It’s also the root cause of hallucination: nothing structurally stops the model from confidently guessing wrong.
Diogo Almeida thinks the industry has been solving the wrong problem. Almeida was an OpenAI researcher who helped invent reinforcement learning from human feedback, the training technique that made ChatGPT usable in the first place. Two years ago he left to start TypeSafe AI, and this month the company came out of stealth with a model called Jev that works nothing like the LLMs everyone’s gotten used to.
Jev doesn’t write sentences. It outputs probabilities and what TypeSafe calls “calibrated decisions,” a structure built for software automation tasks rather than conversation. Because it isn’t generating open-ended text, it can’t hallucinate in the traditional sense, since its outputs are constrained to a defined set of possibilities from the start. TypeSafe describes it as a “System One” model, built for fast intuition rather than the slower step-by-step reasoning newer LLMs lean on. It’s trained entirely on synthetic data, and its pricing structure is unusual too: output tokens are free, while input tokens get metered by the billion.
The early developer reaction has been loud. A Vercel engineer testing Jev reported results five to eighteen times faster than their existing setup, with better accuracy on top. A team at Bryo AI found it ran 10 to 20 times cheaper than the alternatives they’d been using, with more reliable confidence scoring alongside the savings. Demand got intense enough after launch that TypeSafe temporarily lost the ability to serve API requests, a problem most two-year-old startups would happily trade for almost anything else in the world.
Almeida’s own description of the gap he’s trying to close is blunt. “We have lightning in a bottle, and yet it is not useful,” he said. “Computers speak a different language.” The idea is that raw model capability has outpaced the interfaces built to use it, and that a narrower, non-conversational model built specifically for structured decisions closes that gap better than another general-purpose chatbot ever could.
Why this matters beyond one product launch
Almeida built RLHF. If anyone has standing to say the field over-indexed on making models talk instead of making them reliably decide, it’s probably him. The framing here is a direct challenge to the dominant approach of the last three years: bigger context windows, more reasoning steps, more conversational polish, all stacked on top of an architecture that’s still fundamentally guessing the next word.
Whether Jev generalizes beyond the narrow automation tasks it’s been tested on is the open question. Non-autoregressive models built for calibrated decisions instead of open text generation have shown up before without displacing the conversational giants. But the specific developer numbers here, five to eighteen times faster with fewer errors, are hard to wave off as marketing spin.
Bottom Line: TypeSafe just shipped a serious challenge to the assumption that every useful AI model has to talk like a chatbot. Expect competitors to move fast now that the demand is proven in public. The real test isn’t this launch week, it’s whether Jev’s advantages hold up once the task set gets messier than the clean automation cases it’s winning right now.
The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.
