On Monday, October 15th, Diogo Almeida, a former OpenAI researcher, launched TypeSafe AI, a startup that emerged from stealth mode with $40 million in seed funding and the unveiling of its first product, Jev. Unlike traditional chatbots, Jev is designed to respond to structured queries—such as choices, scores, or probabilities—rather than generating free-flowing text. This approach enables Jev to operate within 70 to 500 milliseconds and eliminates hallucinations, a common issue in conventional chatbots.
The engineer behind ChatGPT’s charm (and flaws)
Almeida, a graduate of the Georgia Institute of Technology, worked at Google Brain before joining OpenAI, where he spent a decade researching artificial intelligence. He is listed as a co-author in the GPT-4 technical report for his contributions to human feedback reinforcement learning (RLHF) and InstructGPT, technologies that enable chatbots like ChatGPT and Claude to engage in fluid conversations with users.
Almeida criticizes RLHF, arguing that it trains models to prioritize answers that humans rate highly, even when those answers are unreliable. This has led systems to favor plausible but potentially fabricated responses over admitting uncertainty, as exemplified by ChatGPT’s praise of an audio recording of a fart as if it were music.
Two years of silence to ‘fix his own mistake’
After leaving OpenAI in 2024, Almeida spent two years building TypeSafe AI alongside co-founders Erik Gafni and Sasha Sheng, a former researcher at Meta/FAIR. The company remained unknown to the public until Monday, October 15th, when it announced a $40 million investment led by DCVC and the launch of Jev.
In his launch statement, Almeida wrote, “After co-inventing ChatGPT, I kept wondering: why haven’t superhuman chat models led to AGI [general artificial intelligence]?” He believes that models trained to please humans in conversations are not necessarily suited for making autonomous decisions within software systems, and Jev was created to address this gap.
A model that deliberately refuses to chat
Jev reverses the logic of the large language models that Almeida helped popularize. Instead of generating text token by token sequentially, it processes structured questions simultaneously and returns answers in JSON format, such as {“revenue”: 0.08, “technical”: 0.85, “sales”: 0.07}, which can be directly used by a system to route a customer to a specific queue.
TypeSafe AI describes Jev as a “System One model,” inspired by Daniel Kahneman’s theory of two human thinking systems: one fast and intuitive, and the other slower and more rational. The goal is for the model to function as a software’s “automatic thinking,” providing short, immediate answers without simulating a conversation.
Why doesn’t Jev ever ‘hallucinate’ an answer?
The model is trained using a technique called Calibrated Decision Reinforcement (RLCD). This method teaches the AI to accurately estimate the probability of correctness for each answer. When Jev indicates 85% confidence, this reflects the actual likelihood that the decision is correct, unlike common chatbots, which can appear confident even when they are wrong.
The name Jev honors economist William Stanley Jevons, referencing the idea that efficiency gains tend to increase demand for a technology rather than reduce it.
Almeida describes the proposal as a “call for a frontier intelligence function,” where unstructured input generates a probabilistic and typed decision. It remains to be seen whether the market will accept the idea that the next wave of automation depends less on fluent conversations and more on models that deliberately do not know—and do not want to—chat.
Frequently Asked Questions
- What is TypeSafe AI?
TypeSafe AI is a startup founded by Diogo Almeida in 2024 after he left OpenAI, together with Erik Gafni and Sasha Sheng. On Monday, October 15th, the company announced a $40 million seed round led by DCVC.
- Who is Diogo Almeida?
Diogo Almeida is an AI researcher who has worked at Google Brain and OpenAI. He co-authored the RLHF and InstructGPT techniques that underpin chatbots like ChatGPT and Claude.
- What is Diogo Almeida’s criticism of RLHF?
He argues that RLHF taught models to prioritize answers that humans rated highly, even when those answers were unreliable, leading them to hallucinate plausible responses.
- How does the Jev model work?
Jev processes structured questions in parallel, returning choices, scores, or probabilities in JSON format without generating continuous text.
- What is the Calibrated Decision Reinforcement used in Jev?
It is a training technique where the AI learns to accurately estimate the probability of correctness for each answer, ensuring that the declared confidence matches the actual likelihood.
- Why doesn’t Jev hallucinate answers?
Because it does not generate free text and limits its outputs to structured formats, operating in milliseconds and eliminating the hallucination problem found in traditional models.
