Diogo Almeida, an OpenAI researcher who helped develop ChatGPT and invented the reinforcement learning from human feedback (RLHF) technique, left the company two years ago to address what he saw as a fundamental flaw in modern AI: its excessive focus on human language. Almeida told TechCrunch that while their achievements were remarkable, the reliance on text made these systems impractical for automation because “computers speak a different language.”
His startup, TypeSafe AI, launched this week with Jev, a transformer-based model that functions differently from large language models (LLMs). Rather than generating text, Jev outputs probabilities, described by the company as “calibrated decisions.” By eliminating natural language generation, the model significantly reduces costs and increases speed, while user-defined output schemas prevent hallucinations entirely. Input processing is metered by the billion rather than the million.
Developer interest has been immediate, causing API demand to spike so sharply that TypeSafe temporarily lost the ability to serve users. Early adopters have found Jev particularly effective for software automation. Pranit Sharma, a software engineer at Vercel, reported that switching from OpenAI’s Luna 5.6 to Jev for safety command classification resulted in processing times 5 to 18 times faster with improved accuracy.
Nikhil Mudholkar, CTO of Bryo AI, compared Jev against Google’s Gemini for classifying business emails. While Gemini showed a marginal edge in accuracy, it was 10 to 20 times more expensive. Mudholkar highlighted Jev’s confidence scores as a key advantage, noting it provides real probabilities ideal for automating workflows.
The model also offers potential for augmenting existing LLMs. Armin Ronacher, CTO of Earendil, which builds the open-source model harness Pi, explained that Jev can monitor LLM agents to prevent jailbreaks and detect misbehavior at a lower cost than using another LLM for supervision. He added that the system allows users to filter results based on probability thresholds, disregarding low-confidence outputs while acting on high-certainty ones.
Ronacher also identified “model routing” as a viable use case, where Jev’s low latency and cost enable real-time sorting of workloads to determine if a specific model is required. The model is named after 19th-century economist William Stanley Jevons, referencing his paradox that falling commodity costs lead to increased usage. Almeida envisions a future of distributed, emergent smart software akin to the early internet.
TypeSafe AI has not disclosed Jev’s full architecture, though observers suspect it is built upon an open-weight LLM. The company describes it as a “System One model” focused on intuition rather than reasoning, trained exclusively on synthetic data via a method Almeida calls “reinforcement learning from calibrated decisions.” Almeida termed his investment in synthetic data production one of the best decisions of his career, surpassing even the launch of RLHF.
While Jev currently stands alone in this category, Ronacher predicts competitors will emerge now that its utility is clear. TypeSafe plans to expand Jev into new modalities, with Almeida distinguishing his company from frontier labs driven by hype or fear, aiming instead to make intelligence its primary product.
Named after Jevons paradox, of course. Cheap intelligence will just consume more compute. I’ll believe it’s practical when I see enterprise adoption hit triple digits.
Interesting that Vercel uses it for safety classification. But does this mean the open-weight LLM base limits its reasoning capabilities compared to true frontier models?
Wait, so it outputs probabilities instead of text? I assumed all AI had to generate tokens. This feels like a fundamental shift in how we build agents.
Finally, someone admits LLMs are terrible at actual automation. The speed and cost savings sound like a game changer for engineers.