AI
The ChatGPT Pioneer Who Quit Talking, and the Model That Answers in Pure Probability
Diogo Almeida helped invent the technique that made ChatGPT possible. Reinforcement learning from human feedback, the training method perhaps most responsible for the current age of AI, was co created by Almeida during his years as an OpenAI researcher. Two years ago he walked away, convinced that chat was the wrong target for the work software actually needs done. This week his startup TypeSafe AI opened early access to Jev, a model built on a striking premise, that computers speak a different language than humans do, so stop forcing the model to speak human.
Jev belongs to a different species than large language models. It generates zero text. Instead it returns what TypeSafe calls calibrated decisions, a choice from up to 255 predefined options, a numerical score, or a binary probability, each carrying a confidence value between zero and one. The model runs end to end in 70 to 500 milliseconds. Input is priced at $0.042 per million tokens, and output tokens are free. Because every output shape is fixed in advance by the developer, the model is structurally incapable of emitting a malformed answer, which is the engineering version of hallucination proof output.
Developers are responding with unusual speed. Demand was so strong at launch that TypeSafe briefly lost the ability to serve users from its API. Pranit Sharma, a software engineer at Vercel, reported that his company replaced OpenAI's Luna 5.6 in a safety classifier with Jev and got results five to 18 times more quickly, with greater accuracy. Bryo AI chief technology officer Nikhil Mudholkar tested Jev against Gemini on business email classification. Gemini was slightly more accurate and 10 to 20 times more expensive, and Jev was the only model handing back a real probability, which Mudholkar called ideal for automating workflows.
Under the hood, Almeida trained Jev exclusively on synthetic data using a technique he calls reinforcement learning from calibrated decisions. He named the model after William Stanley Jevons, the 19th century economist behind the paradox that cheaper resources get used more. The bet is explicit. As intelligence gets cheaper, it spreads everywhere. Almeida put it bluntly to TechCrunch, saying the main product of frontier labs is fear or hype, and he would like his main product to be intelligence.
For readers, the takeaway is architectural. The chatbot era taught machines to talk. The automation era needs machines that decide, millions of times per second, inside code that runs beyond human eyes. Jev points at a future where the smartest model in your stack is the one that stays silent. That future just got a price tag that software teams can actually pay.
Quick answers
What is this story about?
Diogo Almeida helped invent the technique that made ChatGPT possible. Reinforcement learning from human feedback, the training method perhaps most responsible for the current age of AI, was co created by Almeida during his years as an OpenAI researcher. Two years ago he walked away, convinced that chat was the wrong target for the work software actually needs done. This week his startup TypeSafe AI opened early access to Jev, a model built on a striking premise, that computers speak a different language than humans do, so stop forcing the model to speak human.
Why does this story matter?
For readers, the takeaway is architectural. The chatbot era taught machines to talk. The automation era needs machines that decide, millions of times per second, inside code that runs beyond human eyes. Jev points at a future where the smartest model in your stack is the one that stays silent. That future just got a price tag that software teams can actually pay.
Sources
- TechCrunch: A new kind of AI model from a ChatGPT inventor is thrilling developers
- Glonce: TypeSafe launches Jev, a model for structured decisions, not chat
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.