Former OpenAI researcher's AI system evaluating options on a screen, not generating text.

Editorial illustration for Former OpenAI Researcher Builds AI That Judges Options, Not Writes Text

Former OpenAI Researcher Builds AI That Judges Options,...

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

September 16, 2026 • 2 min read

Diogo Almeida spent years training the language model that became ChatGPT's backbone, listed as a co-author on the InstructGPT paper that shaped how OpenAI built instruction-following systems. Now he's betting the next useful AI product won't write anything at all.

Almeida's startup, TypeSafe AI, launched a model called Jev that skips text generation entirely. No emails, no code, no chat replies. Instead, Jev sits inside other software and does one narrow job: given a set of possible answers defined by a developer, it scores which one fits best and attaches a probability to it.

The pitch is that most software doesn't need a model that can write a sonnet or debug Python. It needs something fast and cheap that can make a call between a handful of fixed options, over and over, at scale. TypeSafe's documentation walks through how this might work for something as ordinary as a customer service inbox, where a single message has to get routed correctly before anyone can act on it.

Startup TypeSafe AI has introduced an AI model called Jev that takes a different approach than typical chatbots. Instead of generating text, emails, or code, Jev is built to deliver narrow judgments and probabilities inside other programs.

Former OpenAI researcher builds an AI model that judges options instead of writing text - THE DECODER

Why this matters

Jev is a bet that not every AI problem needs a chatbot bolted onto it. For developers building support desks, fraud checks, or routing systems, a model that outputs a label and a probability instead of three paragraphs of hedged prose is faster to run and cheaper to serve at scale. That's a real advantage if your product just needs a yes, no, or "flag for review," not a conversation.

But TypeSafe AI's own example shows the ceiling clearly: Jev tells the software a request looks like a refund, the software still has to apply fixed rules, and a human still has to step in when the label is unclear. The judgment layer is narrow by design, which means the surrounding logic and staffing still carry the weight. Founders should read this as a tool for specific classification bottlenecks, not a replacement for reasoning systems. Researchers should watch whether narrow judgment models like this generalize past the categories they're trained on, or whether every new use case means retraining from scratch.