TypeSafeAI's Jev targets automation with fixed decision outputs

A different approach to generative AI

The Wall Street Journal reported on Thursday that Jev, an AI model launched three weeks ago in Silicon Valley, is drawing attention as a possible alternative to large language models. Startup TypeSafeAI unveiled Jev on Sept. 15. Unlike a chatbot, Jev does not generate text. It uses machine learning to classify inputs into predetermined outputs, including yes-or-no answers, scores or answers selected from a list.

TypeSafeAI calls the approach Reinforcement Learning for Calibrated Decisions and positions Jev as a model for automation software. CEO Diogo Almeida previously helped develop ChatGPT at OpenAI before leaving in 2024. He said Jev is fast, inexpensive and consistent, and can address limitations associated with LLMs, which he described as hard to predict.

Adoption and investment

Almeida said about 25 percent of Fortune 500 companies are already using Jev. He also said daily throughput reached 1 trillion tokens a week earlier and described growth as exponential. TypeSafeAI received a $40 million investment from venture capital firm DCVC. The Information reported last week that the startup is in talks to raise more than $1 billion, with some investors citing a valuation above $10 billion. Almeida declined to comment.

He said major LLMs are well suited to tasks involving human intervention but are "unbelievably useless" for automation. His argument is that trusted software should be built on a dependable foundation, with features added one at a time, rather than assembled around the variable behavior of chatbots.

Decision-focused products grow

OpenAI unveiled a "Decisions API" at the end of September that uses its "Luna" model to provide fixed answers to user-defined questions. Data analytics startup Databricks also released an "ai_decide" function. Almeida said he expected copycat products but was confident Jev would perform better.