TechCrunch says ChatGPT has disappointed Diogo Almeida, the OpenAI researcher who helped create the chatbot and invented reinforcement learning from human feedback (RLHF). Despite his capabilities, Almeida was not satisfied. “We have lightning in a bottle, but it’s not useful,” he said. "I've been struggling with this problem ever since. It took me a while to come to the conclusion that the problem is that we're optimizing for human language... We did a great job with human language for four years, but it's not useful for automation because computers speak a different language." Two years ago, Almeida left OpenAI to found TypeSafe AI, a startup that is trying to solve this problem. This week the company released a new Transformers-based model, Jev, which is a non-large language model (LLM). It doesn't output text, but instead produces probabilities, or what the company calls "calibrated solutions." Eliminating language makes the model incredibly cheap and fast, and since users pre-define the output, it can't "hallucinate". Output tokens are free, but input tokens are measured in billions, not millions
lions. The screenshot shows a comparison of Jev and the OpenAI model answering the same queries. Developers are showing great interest in the product; the company lost the ability to serve users through its API for some time due to high demand. Jev seems to be most useful for software automation. At the moment, programmers see it as a cheaper and more reliable way to implement intelligence into their code. For example, Pranid Sharma, a software engineer at Vercel, an agent infrastructure company, said his company used OpenAI's ChatGPT Luna 5.6 to run a classifier that checks commands for safety. When Vercel replaced Luna with Jev, it obtained results 5 to 18 times faster and with greater accuracy. Another developer, Nikhil Mudholkar, CTO of Bryo AI, tested Jev against Gemini for business email classification. In his test, the Gemini was slightly more accurate, but 10 to 20 times more expensive. More interesting to Mudholkar were Jev's confidence scores - "it's the only model that returns real-world probability, making it ideal for workflow automation!" Besides
Oh, the new model can not only replace LLM in certain cases, but also complement them, acting as a smart check for incorrect behavior. Using agents to monitor other agents can quickly become expensive, but using Jev for this purpose makes sense, according to Almeida. He sees users using Jev to track traces of LLM agents and prevent them from being hacked. “At the end of the day, this delegates the hallucination problem a little bit to the user,” explained Armin Ronacher, CTO of Earendil, which is developing the open-source Pi model. "The user should say: if it only returns 50% probability, it's probably just a coin flip and I'm ignoring it. But if it's 95%, then I can do something about it." Another potential application for Jev is model routing, Ronacher added. Predicting whether a given load requires a particular model would be useful, but using LLM for this task would be expensive. The low cost and speed of Jev make such real-time sorting possible. And that is Almeida's hope. The model is named after William Stanley Jevons, the 19th century economist whose parade
The oxal concept describes how a decrease in the cost of a product can lead to its increasing use. In this case, reducing the cost of intelligence should lead to its widespread distribution. “We believe that smart software will be everywhere, in ways that will emerge and be distributed...much more than the early internet, than you know, like the mega-apps that people are trying to build right now,” Almeida said. Almeida is coy about the model's architecture, which outside observers suspect is built on top of an open LLM. The company calls Jev a "system one model," focused on intuition rather than reasoning, and specifically targeted to the right task. Almeida says Jev is trained entirely on synthetic data using a technique he calls “calibrated decision-based reinforcement learning.” “We made an early bet that we would build all of our data, and it was one of the best bets I've ever made—better than our startup, better than RLHF in my opinion,” he told TechCrunch. "Half [of our company] is a laboratory that actually owns
this whole sub-industry of statistically well-understood synthetic data, and that's now my joy in life." For now, Jev stands alone as such a model, but Ronacher expects competitors to emerge, now that its usefulness is clear. "We should have seen this sooner in many ways, but presumably because LLMs are so cheap and subsidized, you often don't need to be creative yet," he said. TypeSafe itself will develop more versions of the model in new modalities. Asked if TypeSafe Frontier Labs, Almeida responded: “Frontier Labs' core product is fear or hype. I would like our core product to be intelligence... [but we are not] a laboratory in the sense that, you know, betting on infinite wealth, or religion, or building God in a data center, or whatever is the topic of today." Material prepared by neural network based on the original publication. AI may be wrong in details (numbers, names, dates, quotes) - for important information, check with the original source. Virginia Governor creates AI task force and limits data centers PrismML introduced Ternary Bonsai 2 27B with high performance
Kyutai has released a Voice of Reason model for verbally solving math problems. Nokia opens access to AnyJev, a library for decision-making without training. OpenAI introduced GPT-6 Sol and Luna with new API pricing, reducing the cost by 50%.
