ChatGPT Broke Diogo Almeida’s Heart: Former OpenAI Researcher Launches TypeSafe AI with ‘Jev,’ a Novel Model Eschewing Language for Pure Automation

Diogo Almeida, a pivotal figure in the early development of modern artificial intelligence, found himself deeply disillusioned by the very technology he helped create. As an OpenAI researcher, Almeida played a crucial role in building the foundational architecture of the now-ubiquitous ChatGPT and is credited with inventing Reinforcement Learning from Human Feedback (RLHF), the sophisticated model-training technique widely considered responsible for the current explosion in AI capabilities. Yet, despite the remarkable prowess of these large language models (LLMs), Almeida was profoundly disappointed by their practical utility for automation.

His frustration crystallized into a stark observation: "We have lightning in a bottle, and yet it is not useful," Almeida told TechCrunch. This sentiment underscores a fundamental disconnect he identified between the impressive ability of LLMs to generate human-like language and their actual applicability in robust, automated systems. "I’ve been battling that problem since then," he explained, reflecting on his journey. "It took me a while to come to the conclusion: 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."

This realization spurred Almeida to action. Two years ago, he departed from OpenAI, the company where he had made such significant contributions, to establish TypeSafe AI. His new startup was founded with a singular mission: to bridge this perceived gap between sophisticated AI and practical, reliable automation. This week, TypeSafe AI unveiled its flagship product, Jev, a new transformer-based model designed specifically to address the limitations Almeida identified in LLMs.

Crucially, Jev is explicitly not a large language model. Unlike LLMs, which are engineered to produce coherent, natural language text, Jev operates on an entirely different principle. Instead of generating sentences or paragraphs, it outputs "probabilities," or what TypeSafe AI terms "calibrated decisions." This fundamental design choice allows Jev to communicate directly in a language more akin to what computers understand and require for automated processes.

Eschewing the complexities of human language offers Jev several distinct and compelling advantages. Foremost among these are its remarkable cost-effectiveness and blistering speed. Because users define the potential outputs in advance, Jev functions as a highly efficient classifier and decision-maker, rather than a creative text generator. This architecture inherently prevents the model from "hallucinating" – a common and persistent challenge with LLMs, where models generate factually incorrect or nonsensical information. Furthermore, its economic model is revolutionary: Jev’s output tokens are free, while its input tokens are metered by the billion, a dramatic reduction in cost compared to typical LLM pricing structures that meter by the million.

The market’s immediate response to Jev has been overwhelmingly positive. The company briefly experienced an inability to serve all users from its API due to unprecedented demand, a clear indicator of the perceived value and utility of their new offering. Jev appears to be most potent in applications related to software automation, where developers are eager to integrate intelligence into their code in a more robust and cost-efficient manner than previously possible with LLMs.

Early adopters have already reported significant improvements. Pranit Sharma, a software engineer at Vercel, a company specializing in agentic infrastructure, shared a compelling case study. Vercel had been utilizing OpenAI’s ChatGPT Luna 5.6 to power a classifier responsible for reviewing commands for safety. Upon replacing OpenAI’s Luna with Jev, Vercel observed a dramatic acceleration in performance, achieving results 5 to 18 times more quickly, coupled with notably greater accuracy in its classifications.

Another developer, Nikhil Mudholkar, CTO of Bryo AI, conducted his own comparative tests, pitting Jev against Google’s Gemini for the task of classifying business emails. While Gemini demonstrated a marginal edge in accuracy, Jev proved to be an astounding 10 to 20 times more cost-effective. Mudholkar highlighted a particularly valuable feature of Jev: its confidence scores. He noted that Jev "is the only one that hands back a real probability which makes it ideal for automating workflows!!" This capability allows developers to programmatically determine the reliability of a decision, enabling more sophisticated and trustworthy automated systems.

Beyond directly replacing LLMs in specific use cases, Jev also presents a powerful opportunity to augment existing LLM deployments. Almeida argues that Jev can serve as an intelligent, cost-effective safeguard against misbehavior. While using LLM agents to monitor other LLM agents can quickly become prohibitively expensive, deploying Jev for such tasks offers an economical solution. Almeida envisions users leveraging Jev to track LLM agent traces, identify anomalies, and prevent "jailbreaks" – instances where users manipulate LLMs to bypass safety protocols.

Armin Ronacher, CTO of Earendil, a company that develops the open-source model harness Pi, elaborated on this aspect. He explained that Jev effectively "delegates the hallucination problem a little bit to the user." With Jev, the model provides a probability, and the user (or the automated system) then decides how to act. "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," Ronacher elucidated. This empowers developers with fine-grained control over how AI outputs are integrated into their applications.

Ronacher also pointed to another potential transformative application for Jev: model routing. The ability to predict efficiently whether a given computational workload requires a specific, specialized AI model could significantly optimize resource allocation. However, employing a complex and expensive LLM for this routing task would negate much of the cost benefit. Jev’s inherent low cost and rapid processing speed make this kind of real-time, dynamic sorting and model selection a practical and economically viable possibility.

This widespread, distributed intelligence is precisely Almeida’s hope and vision. The model itself is named after William Stanley Jevons, the influential 19th-century economist whose eponymous paradox describes how the falling cost of a commodity often leads to a dramatic increase in its consumption and utility. In this context, Almeida believes that the falling cost of intelligence, enabled by models like Jev, should lead to its pervasive deployment across countless applications and systems. "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 you know like the mega apps that people are trying to build right now," Almeida articulated, painting a picture of a future where intelligence is embedded everywhere, rather than centralized in monolithic applications.

While Almeida remains tight-lipped regarding the precise architectural details of Jev, outside observers speculate that it may be built upon an open-weight LLM. TypeSafe AI refers to Jev as a "System One model," drawing a parallel to cognitive psychology’s concept of fast, intuitive thinking, as opposed to slower, more deliberate reasoning. The company emphasizes that Jev is "specifically focused on the right task," underscoring its specialized, efficient nature. Almeida revealed that Jev is trained exclusively on synthetic data, utilizing a novel technique he calls "reinforcement learning from calibrated decisions."

Almeida expressed profound confidence in this innovative training methodology. "We made an early bet that we will be making all of our data, and that has been one of the best bets I’ve ever made in my life — better than our launch, in my opinion, better than RLHF," he shared with TechCrunch, highlighting the significance he places on this approach. He further explained that a substantial portion of TypeSafe AI operates as a dedicated lab "that basically owns this entire subfield of statistically well-understood synthetic data, and that is now my life joy."

For the moment, Jev stands largely alone in its specific niche as a model designed for calibrated, probabilistic decision-making rather than language generation. However, Armin Ronacher anticipates that competitors will inevitably emerge now that the clear utility and market demand for such a model have been unequivocally demonstrated. "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," Ronacher commented, suggesting that the prevailing focus on LLMs may have inadvertently delayed innovation in this crucial area.

TypeSafe AI itself has ambitious plans for the future, intending to develop more versions of Jev and explore new modalities for its application. When questioned about whether TypeSafe views itself as a "frontier lab" – a term often associated with high-stakes, often speculative AI development – Almeida offered a nuanced and grounded perspective. He stated, "the main product of Frontier Labs is fear or hype. I would like our main product to be intelligence… [but we are] not a lab in the sense of, you know, like bet on infinite wealth, or a religion, or building God in a data center, or whatever is the thing of today." Almeida’s vision for TypeSafe AI is clear: to deliver practical, accessible intelligence that solves real-world automation problems, moving beyond the sensationalism and grand pronouncements that often accompany the AI frontier.