Semantic Compute: From Interpreters to Compilers (seldon-ai.com)

TypeSafe AI's Jev has gained remarkable traction in the developer community, amassing nearly 13% adoption among AI Gateway's paid teams within just 24 hours of its launch. This shift suggests a readiness for a new programming paradigm in AI, moving from merely calling models to compiling semantic functions that enhance decision-making processes in software. Jev processes context and questions to provide typed decisions, integrating with existing code to classify actions efficiently; early tests showed median act latency dramatically reduced from 1.97 seconds to 0.46 seconds. This reflects a broader trend toward Semantic Compute, which aims to optimize the execution of semantic behavior independently from the models that perform these tasks. The concept of Semantic Functions is central to this discussion, as they represent explicit mappings between inputs and acceptable behaviors, allowing for flexible execution plans. This shift facilitates the separation of logical behavior from execution, enabling diverse implementations—ranging from LLMs to deterministic code—to optimize how decisions are made. By proposing a kernel of operation families such as PREDICATE, CHOICE, and SCORE, the architecture promotes an efficient and managed approach to executing complex semantic tasks. Overall, Jev's emergence signals a significant evolution in the AI programming landscape, paving the way for more robust and adaptable systems that can streamline decision-making processes across applications.