started · updated
Decision-making AI models emerge as industry shifts toward automation
A new category of artificial intelligence focused on decision-making models is emerging, moving away from traditional text-generating large language models (LLMs) toward specialized systems designed for automation.
TypeSafe AI recently released Jev, a model that uses reinforcement learning for calibrated decision-making. Rather than generating text, Jev classifies inputs into predefined outputs such as yes/no answers, numerical scores, or selections from a list. This approach aims to provide faster, lower-cost, and more predictable results for software automation. Diogo Almeida, a former OpenAI researcher, noted that Jev processes approximately one trillion tokens daily and is already utilized by roughly 25% of Fortune 500 companies. The success of Jev has prompted major tech players to follow suit, including OpenAI with its Decisions API, Databricks with ai_decide, and Cloudflare with its open-source Clef models.
In competition with this trend, the Shanghai-based AI company StartLux has released StartLux-Decision, an open-source suite of decision models ranging from 0.8B to 27B parameters. According to company reports, the 27B version outperformed Jev in 31 out of 38 benchmarks on the Decision Index 0.2.1. StartLux-Decision is designed to handle multiple judgments in a single request and supports continuous operations, such as navigating web interfaces or managing business workflows, by outputting direct selections and probabilities without the need for autoregressive text generation.
Entities
Diogo Almeida · Jev · OpenAI · StartLux · TypeSafe AI