New AI Model by ChatGPT Creator Inspires Excitement Among Developers

Diogo Almeida, a former researcher at OpenAI, has shared his discontent regarding the capabilities of chatbots like ChatGPT. Despite contributing to its development and innovations like reinforcement learning from human feedback (RLHF), Almeida believes these advancements are not translating effectively into practical applications.

“We have lightning in a bottle, and yet it is not useful,” Almeida lamented in an interview. He explained that while human language optimization is impressive, it falls short in automation since computers require a different mode of communication.

After departing OpenAI two years ago, Almeida established TypeSafe AI, aiming to address the shortcomings he identified. Recently, the startup unveiled a new model called Jev, which is not classified as a large language model (LLM). Instead of generating text, Jev focuses on delivering probabilities—termed “calibrated decisions.”

This innovative approach brings several advantages: it enhances speed and cost-efficiency, while also eliminating the common issue of models “hallucinating” responses. Unlike traditional LLMs, Jev’s outputs are predetermined by users, making it a reliable tool in automation.

With strong interest from developers, demand surged to the point where TypeSafe temporarily faced challenges in handling API requests. Jev is particularly attractive for software automation projects, with many developers viewing it as a more economical and effective solution to enhance their coding practices.

For instance, Pranit Sharma, a software engineer at Vercel, reported that replacing OpenAI’s ChatGPT Luna with Jev resulted in processing speeds 5 to 18 times faster and greater accuracy.

Another developer, Nikhil Mudholkar, the CTO of Bryo AI, compared Jev against Gemini for classifying business emails. While Gemini demonstrated slightly higher accuracy, it came at a cost 10 to 20 times greater. Mudholkar noted Jev’s confidence scores as a significant advantage, stating, “it is the only one that hands back a real probability which makes it ideal for automating workflows!”

In addition to potentially replacing LLMs, Jev can complement them by monitoring their behavior. Almeida suggests that Jev could be employed to track LLM actions and prevent unauthorized manipulations. By shifting some responsibility for assessing risk to users, Almeida sees a practical approach to the hallucination challenge.

Moreover, Jev holds promise for model routing, allowing for predictive decisions on which workload requires a specific model—yet another area where LLMs might not be cost-effective. Jev’s affordability and rapid response could facilitate effective real-time sorting, presenting new opportunities for businesses.

Jev, named after the 19th-century economist William Stanley Jevons, aims to increase the utilization of intelligence as its cost decreases. Almeida envisions a future where intelligent software proliferates in a manner reminiscent of the early internet, rather than focusing solely on developing mega-apps.

While Almeida remains tight-lipped about the specifics of Jev’s architecture, insiders speculate it builds upon an open-weight LLM foundation. TypeSafe describes Jev as a “System One model,” prioritizing intuitive responses tailored to specific tasks. Almeida emphasized that Jev’s training relies solely on synthetic data through his innovative method of “reinforcement learning from calibrated decisions.”

As TypeSafe continues its development, Almeida reflects on the company’s commitment to synthetic data, claiming it has been a pivotal decision in their journey. “Half of [our company] is a lab that basically owns this entire subfield of statistically well-understood synthetic data,” he stated.

Though Jev currently stands alone as a model of its kind, Ronacher anticipates competitors will emerge as its benefits become evident. “We should have seen this earlier… but presumably because the LLMs are so cheap and subsidized, you often don’t have to be creative yet,” he noted.

TypeSafe also has plans to develop additional versions of Jev across various modalities, with Almeida indicating that their primary goal is to advance intelligence rather than engage in hype-driven narratives.