ChatGPT broke Diogo Almeida’s heart.
Almeida is an OpenAI researcher who helped build chatbots and later invented reinforcement learning from human feedback (RLHF), perhaps the most contributing model training technique to the current AI era. However, despite its functionality, he was disappointed.
“It’s lightning in a bottle, but it’s useless,” Almeida told TechCrunch. “I’ve been fighting that problem ever since. It took me a while to come to a conclusion. The problem is that we’re optimizing for human language…We’ve been very good at human language for four years, but computers speak a different language, so it’s not useful for automation.”
Two years ago, Almeida left OpenAI to launch TypeSafe AI, a startup trying to solve that problem. This week, the company released a new transformer-based model, Jev, which is not a large-scale language model (LLM). It doesn’t output text, but instead generates probabilities, or what the company calls “coordinated decisions.”
Avoiding language has several benefits. The model becomes incredibly cheap and fast. It also eliminates hallucinations because the user predefines the output. Output tokens are free and input tokens are measured in billions instead of millions.
Developers have great interest in this product. Demand was so high that the company temporarily lost its ability to serve users through its API. Jev seems to be most useful for software automation. So far, software developers see this as a cheaper and more robust way to build intelligence into their code.
For example, Pranit Sharma, a software engineer at Vercel, which develops agent infrastructure, said his company used OpenAI’s ChatGPT Luna 5.6 to run a classifier to review commands for safety. When Vercel replaced OpenAI’s Luna with Jev, the results were 5 to 18 times faster and with higher accuracy.
Another developer, Bryo AI CTO Nikhil Mudholkar, tested Jev and Gemini for business email classification. In his tests, Gemini was slightly more accurate, but it was 10 to 20 times more expensive. Even more interesting to Mudholkar was Jev’s confidence score. “This is the only one that returns real probabilities, perfect for workflow automation!!”
In addition to replacing LLM in certain use cases, the new model can also enhance LLM to act as a smart check against fraud. Almeida argues that using agents to monitor agents can quickly become expensive, but it makes sense to use Jev to monitor them. He sees users deploying Jev to track LLM agent tracking and prevent jailbreaks.
“Ultimately, the question of hallucinations is left a little bit up to the user,” explains Armin Ronacher, CTO of Earendil, which is developing the open source model harness Pi. “The user has to say, “If this only returns 50% of the time, then maybe this is a coin toss. I’ll ignore it. But certainly if it’s 95%, you can do something with it.”
Another potential use for Jev is model routing, Ronacher said. Predicting whether a particular model is needed for a particular workload is useful, but using LLM for that job is costly. Jev’s low cost and speed enable such real-time sorting.
That is Almeida’s wish. This model is named after the 19th century economist William Stanley Jevons. His eponymous paradox describes how a fall in the price of a good causes it to be used more and more. In this case, the declining cost of intelligence should drive its widespread adoption.
“We think smart software will be everywhere, in emerging and distributed forms…much more like the early Internet than you know, like the giant apps that people are building now,” Almeida said.
Almeida has been tight-lipped about the model’s architecture, and outside observers suspect it is built on an open-weight LLM. The company calls Jev a “System One model,” which focuses on intuition rather than reasoning, with a particular focus on the right tasks. Almeida said Jeb is trained solely on synthetic data using a technique he calls “reinforcement learning from tailored decisions.”
“We made a bet early on that we would create all the data ourselves, and it was one of the best bets I’ve ever made in my life. In my opinion, it’s better than launch, it’s better than RLHF,” he told TechCrunch. “Half[of our company]is basically a research institute that owns a whole subfield of statistically well-understood synthetic data, and that’s now the joy of my life.”
For now, Jev stands alone as a model of its kind, but Ronacher expects competitors to emerge now that its usefulness is clear.
“In many ways we should have looked at this earlier, but perhaps LLMs are so cheap and subsidized that in many cases you don’t have to be creative yet,” he said.
TypeSafe itself plans to build more versions of the model with new modalities. When asked if TypeSafe is Frontier Labs, Almeida said, “Frontier Labs’ main product is fear and hype. We want our main product to be intelligence…(But we’re) not a lab in the sense of betting on infinite wealth, betting on religion, building gods in data centers, or anything like that today.”
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
