2026-11-04 · Fonte：InfoWorld
Diogo Almeida's TypeSafe AI says its new Jev model returns decisions with probabilities in 70–500 milliseconds, sidestepping verbose token generation that inflates costs in automated workflows.
TypeSafe AI, a startup founded by former OpenAI researcher and RLHF co-inventor Diogo Almeida, has introduced a large language model built to talk to software rather than to people. The model, called Jev, takes the current state of a task or workflow as input and returns a single defined decision together with the probability attached to that decision. According to Almeida's blog post announcing the system, those decisions cover cases such as which tool to invoke, what action to take next, whether to approve a request, or when to hand a task off to another model. The output is designed to be consumed directly by software applications or other AI models inside automated workflows.
The premise is that verbosity has become a tax on AI budgets. Today's large language models consume or generate a high volume of tokens even when the underlying request calls for a simple recommendation, driving up usage costs. Automated workflows sharpen the problem: because pipelines of software agents make far more decisions than human-mediated processes, the token bill scales with the number of decisions. And where a machine is often looking only for a categorical answer, the natural-language padding around that answer is spend with little return.
Jev's speed case rests on skipping sequential text generation. Because it does not have to produce text tokens one after another, Almeida said, the model responds faster than conventional systems: latency ranges from 70 milliseconds to 500 milliseconds, against several seconds for the LLMs TypeSafe tested. He added that the concise response format also makes the model cheaper to run.
For enterprise teams embedding AI into automated workflows, the pitch is twofold: fewer tokens consumed per decision, and answers delivered at speeds compatible with machine-paced operations. TypeSafe is targeting the share of AI usage where the reader is another program rather than a person — a segment where, per Almeida, verbose natural-language output adds cost without value. The report also signals savings could extend beyond token costs.
Tags: TypeSafe AI · Jev · Diogo Almeida · LLM · Automated Workflows · Inference Latency
萬安算交所 AIXX · AIXX · Exchange verificado · Ir ao mercado