The Shift from Language to Decision-Making

Traditional Large Language Models (LLMs) are optimized for human language, which often makes them inefficient or unreliable for machine-to-machine automation. Diogo Almeida, co-inventor of RLHF, argues that the current industry obsession with text generation is a bottleneck. His new model, Jev, abandons text output entirely in favor of "calibrated decisions." By outputting probabilities rather than natural language, Jev eliminates hallucinations, as the possible outputs are predefined by the developer.

Performance and Economic Advantages

Jev is designed for high-throughput software automation, where speed and cost are critical. Because it does not generate complex text, it is significantly faster and cheaper than standard LLMs.

  • Speed: In testing, Vercel reported that Jev performed classification tasks 5 to 18 times faster than previous LLM implementations.
  • Cost: Input tokens are metered by the billion rather than the million, and output tokens are free, making it viable for high-frequency tasks where LLMs would be cost-prohibitive.
  • Reliability: Unlike LLMs that provide probabilistic text, Jev returns explicit confidence scores. This allows developers to build deterministic logic around the model's output—for example, automatically executing a task only if the confidence score exceeds 95%.

Practical Applications in Agentic Workflows

Jev is currently being deployed as a specialized tool for "System One" tasks—those requiring intuition rather than deep reasoning. Key use cases include:

  • Safety and Classification: Replacing heavy LLMs for command review or email classification.
  • Model Routing: Acting as a low-cost, real-time router to determine which model should handle a specific workload.
  • Guardrails: Monitoring LLM agent traces to detect jailbreaks or misbehavior, effectively using a smaller, cheaper model to supervise more expensive, complex ones.

Synthetic Data and Architecture

Almeida emphasizes that Jev is trained exclusively on synthetic data using a technique termed "reinforcement learning from calibrated decisions." By controlling the data generation process, the team has created a model that is statistically well-understood. This approach represents a departure from the "frontier lab" model of scaling compute to build general-purpose intelligence, focusing instead on delivering specific, reliable utility for software developers.