Ex-OpenAI researcher's TypeSafe AI launches non-LLM model Jev

Almeida helped build ChatGPT at OpenAI and then invent reinforcement learning from human feedback, or RLHF, the training technique widely credited with advancing today's AI systems. He told TechCrunch he was disappointed that the technology was optimized for human language. “We have lightning in a bottle, and yet it is not useful,” he said. “The problem is we are optimizing for human language … We have been super good at human language for four years, but it’s not useful for automation because computers speak a different language.” He left OpenAI two years ago to start TypeSafe AI.

Jev does not output text. It produces probabilities, which TypeSafe calls calibrated decisions. Because users define the outputs in advance, the model cannot hallucinate, according to the company. Output tokens are free, while input tokens are metered by the billion rather than the million. TypeSafe says the approach makes the model inexpensive and fast. Jev appears most useful for software automation, and developers see it as a more robust way to incorporate intelligence into code.

Pranit Sharma, a software engineer at Vercel, said his company had used OpenAI’s ChatGPT Luna 5.6 to run a classifier that reviewed commands for safety. When Vercel replaced Luna with Jev, the results came back five to 18 times more quickly and with greater accuracy, according to TechCrunch. Another developer, Bryo AI CTO Nikhil Mudholkar, tested Jev against Gemini for classifying business emails. Gemini was slightly more accurate but 10 to 20 times more expensive, he found. Mudholkar highlighted Jev’s confidence scores, saying it “is the only one that hands back a real probability which makes it ideal for automating workflows.”

Besides replacing LLMs in some uses, Jev can augment them by acting as a check on misbehavior. Almeida argues that using Jev to monitor LLM agents can make sense because using agents to monitor agents quickly becomes expensive. He sees users deploying Jev to track LLM agent traces and prevent jailbreaks.

Armin Ronacher, the CTO of Earendil, which builds the open-source model harness Pi, said Jev delegates the hallucination problem partly to the user. “The user has to say, okay, if this only comes back with 50% probability, maybe this is a coin toss, and I disregard it. But if it’s 95%, sure, then I can do something with it,” he said. Ronacher added that Jev could be used for model routing: predicting whether a workload requires a specific model would be useful, but doing the job with an LLM would be expensive, while Jev’s low cost and speed make real-time sorting possible.

The model is named after William Stanley Jevons, the 19th-century economist whose Jevons paradox describes how falling costs can lead a commodity to be used more, not less. Almeida hopes cheaper intelligence will lead to widespread deployment. “We think that there’s just going to be smart software all over the place in a way that’s emergent and distributed … much more like the early internet” than the mega apps companies are trying to build now, he said.

Almeida is tight-lipped about Jev’s architecture, and outside observers suspect it is built on top of an open-weight LLM. TypeSafe calls Jev a “System One model,” focused on intuition rather than reasoning and aimed at specific tasks. Almeida said Jev is trained exclusively on synthetic data using a technique he calls “reinforcement learning from calibrated decisions.” He told TechCrunch that betting on making all of the company’s data has been one of the best bets of his life, better than the launch and, in his view, better than RLHF. Half of TypeSafe is a lab that owns what he described as the subfield of statistically well-understood synthetic data, he said.

Ronacher expects competitors to emerge now that Jev’s utility is apparent. “We should have seen this earlier in many ways, but presumably because the LLMs are so cheap and subsidized, you often don’t have to be creative yet,” he said. TypeSafe plans to build more versions of the model in new modalities. Asked whether TypeSafe is a frontier lab, Almeida said the main product of frontier labs is fear or hype, while he would like TypeSafe’s main product to be intelligence.