TypeSafe AI's Jev Model Chooses Actions, Not Words

A new AI system from TypeSafe AI sidesteps text generation entirely, instead returning probability scores for predefined actions—a design choice that founder Diogo Almeida believes better matches how software automation actually works.

Almeida, who contributed to ChatGPT's development at OpenAI before leaving two years ago, argues that language models excel at processing human language but remain inefficient for automation tasks. "We have lightning in our hands, but it is of no use to us," he told TechCrunch. "We have become excellent at working with human language, but that does not work for automation because computers communicate in a different language."

How Jev differs from language models

Jev is a transformer-based system, but TypeSafe AI does not classify it as a large language model. Rather than generating arbitrary text, Jev calculates probabilities across a developer-defined set of possible responses. The company calls these outputs "calibrated decisions."

Because developers specify the response options in advance, Jev avoids the hallucination risks associated with open-ended generation. The architecture also eliminates output token billing—only input tokens count—and the system can process billions of tokens rather than the millions typical of LLMs. TypeSafe AI reports that demand after this week's launch temporarily overwhelmed its API capacity.

Early performance benchmarks

Vercel engineer Pranit Sharma reported that switching from ChatGPT Luna 5.6 to Jev for command safety checks delivered responses five to eighteen times faster while improving classification accuracy. Nikhil Mudholkar, CTO of Bryo AI, tested Jev against Gemini for business correspondence classification and found Gemini slightly more accurate but ten to twenty times more expensive. Mudholkar noted that Jev's calibrated confidence scores prove especially valuable for triggering automated workflows. "This is the only system that returns a true probability, which makes it ideally suited to workflow automation," he said.

Armin Ronacher, CTO of Earendil, highlighted how probability scores help allocate responsibility between system and user. A fifty-percent confidence result can be flagged for human review, while a ninety-five-percent score may justify autonomous action.

Potential applications beyond classification

Jev can both replace language models for specific tasks and validate their outputs. Use cases include analyzing software agent logs, detecting risky behavior, and blocking attempts to circumvent guardrails. Another application is model routing—automatically selecting the best system for a given task. Using a large language model for routing decisions is expensive; Jev's speed and low cost could enable real-time routing at scale.

Why it matters

If Jev maintains its reported speed, accuracy, and cost advantages in production environments, it could establish a new category of AI tooling optimized for decision-making rather than conversation. The approach directly addresses a gap in current automation stacks: systems that need to choose an action quickly and reliably without the overhead of natural language generation. For enterprises running high-volume workflows, the economics could be compelling—particularly if probability calibration proves reliable enough to reduce human oversight.

Architecture and future development

TypeSafe AI has not disclosed Jev's architecture, though observers speculate it may be built on an open-weights language model. The company describes Jev as a "System One" model focused on intuitive decision-making for specific tasks. Training relies exclusively on synthetic data using what TypeSafe AI calls reinforcement learning based on calibrated decisions. The team plans to release versions supporting additional data types.

Almeida named the system after British economist William Stanley Jevons, whose paradox holds that making a resource cheaper often increases consumption. Almeida hopes affordable "intelligence" will lead to distributed, autonomous software that resembles the early internet more than today's monolithic applications. Ronacher expects competing solutions to emerge once Jev proves its market value.

These details were first reported by TechCrunch.