A new kind of AI model from a ChatGPT inventor is thrilling developers
News Source
β’Fri, 18 Sep 2026 18:49:30 +0000
π° What Happened
A former OpenAI researcher has launched a new kind of AI model. Diogo Almeida helped build ChatGPT and invent a key training method called RLHF. He left OpenAI two years ago to start a company called TypeSafe AI.
This week TypeSafe AI released a model called Jev. Unlike ChatGPT, Jev does not write text. It gives probabilities, which the company calls calibrated decisions. The company says this makes the model very cheap and fast.
Users set the possible answers in advance, so Jev cannot make things up. Its output tokens are free, and input tokens are charged by the billion. Demand was so high that the company briefly could not serve users through its API.
π The Backstory
RLHF is a method that trains AI to answer in ways humans prefer. It is one of the main reasons chatbots like ChatGPT feel helpful today. Even so, Almeida said the model felt like 'lightning in a bottle' that was not useful for automation.
He argued that AI has focused on human language, but computers speak a different language. Jev is still based on a transformer, the same basic design as large language models. Developers see it as a cheaper way to power software tasks.
π― Why It Matters
Cheaper, non-text AI could cut the cost of automating everyday software jobs. It also shows the AI race is moving beyond chatbots that only write words.
A former OpenAI researcher has launched a new kind of AI model. Diogo Almeida helped build ChatGPT and invent a key training method called RLHF. He left OpenAI two years ago to start a company called TypeSafe AI.
This week TypeSafe AI released a model called Jev. Unlike ChatGPT, Jev does not write text. It gives probabilities, which the company calls calibrated decisions. The company says this makes the model very cheap and fast.
Users set the possible answers in advance, so Jev cannot make things up. Its output tokens are free, and input tokens are charged by the billion. Demand was so high that the company briefly could not serve users through its API.
RLHF is a method that trains AI to answer in ways humans prefer. It is one of the main reasons chatbots like ChatGPT feel helpful today. Even so, Almeida said the model felt like 'lightning in a bottle' that was not useful for automation.
He argued that AI has focused on human language, but computers speak a different language. Jev is still based on a transformer, the same basic design as large language models. Developers see it as a cheaper way to power software tasks.
Cheaper, non-text AI could cut the cost of automating everyday software jobs. It also shows the AI race is moving beyond chatbots that only write words.