TypeSafe AI launches Jev, an AI model for structured decision-making, and criticizes traditional chatbots

Diogo Almeida, a former OpenAI researcher who helped develop the technology behind ChatGPT, launched his new startup, TypeSafe AI, on Monday, May 15. The company emerged from stealth mode with an initial $40 million funding round led by DCVC and unveiled its first public product: Jev, a model designed to make structured decisions without generating free text.

Almeida, a Georgia Institute of Technology graduate and former Google Brain engineer, spent a decade at OpenAI. He is listed as a co-author in the GPT-4 technical paper for his fundamental contributions to human feedback reinforcement learning (RLHF) and InstructGPT, techniques that now underpin most large language models, including ChatGPT and Claude.

Why a former OpenAI engineer now criticizes his own creation

In recent interviews, Almeida argues that RLHF, by training models to please humans, has led them to prioritize plausible answers even when they are unreliable, resulting in “hallucinations.” He illustrates this with an example where ChatGPT praised an audio recording of a fart as if it were music.

After leaving OpenAI in 2024, Almeida spent two years building TypeSafe AI alongside co-founders Erik Gafni and Sasha Sheng, a former Meta/FAIR researcher. Their goal was to develop an AI that, instead of conversing, would make autonomous and reliable decisions within software systems.

Jev: a model that doesn’t converse but decides

Jev operates differently from traditional large language models. Instead of generating text token by token, it responds to structured queries—such as choices, scores, or probabilities—processed simultaneously. A sample response might be a JSON object with fields like {“revenue”: 0.08, “technical”: 0.85, “sales”: 0.07}, directly indicating which service queue the case should be routed to.

The company calls this approach a “System One model,” inspired by Daniel Kahneman’s theory of the two systems of human thought. The model functions as a software’s “automatic thinking,” providing quick, concise answers without the prolonged reasoning typical of a chatbot.

No hallucinations and with calibrated confidence

Jev is trained using a technique called Calibrated Decision Reinforcement (RLCD). This method enables the AI to accurately estimate the probability of correctness for each answer. When the model claims 85% confidence, that figure reflects the actual likelihood of being correct, unlike common chatbots that may appear confident even when wrong.

Response time ranges from 70 to 500 milliseconds, making it dozens to hundreds of times faster than traditional language models in equivalent tasks. Because it does not generate free text, Jev cannot “hallucinate” answers, eliminating a problem that has plagued Almeida since his time at OpenAI.

The model’s name honors economist William Stanley Jevons, who observed that efficiency gains tend to increase demand for a technology rather than reduce it.

A call for the next wave of automation

Almeida describes Jev as a “frontier intelligence function call,” where an unstructured input generates a probabilistic and typed decision. He questions whether the market will accept the idea that future AI automation depends less on fluent conversations and more on models that deliberately don’t want to chat.

Frequently Asked Questions

What is TypeSafe AI? TypeSafe AI is a startup founded by Diogo Almeida, Erik Gafni, and Sasha Sheng after they left OpenAI in 2024. The company announced an initial $40 million funding round led by DCVC and launched the Jev model.

Who is Diogo Almeida? An AI researcher with experience at Google Brain and OpenAI, he co-authored technical contributions to GPT-4, RLHF, and InstructGPT, which are fundamental to chatbots like ChatGPT and Claude.

What is Almeida’s criticism of RLHF? He argues that RLHF taught models to prioritize answers well-rated by humans, even when they are unreliable, leading them to prefer plausible and invented answers over admitting ignorance.

How does Jev work? The model responds to structured queries with choices, scores, or probabilities processed in parallel, without generating continuous text, and is classified as a “System One model.”

What is Calibrated Decision Reinforcement? RLCD is a training technique that enables the AI to accurately estimate the probability of correctness for each answer, ensuring that the declared confidence matches the actual likelihood.

Why doesn’t Jev hallucinate? Because it does not generate free text and limits its outputs to structured formats, in addition to having a fast response time, the model avoids hallucinations.