Today's AI trends
OpenTelemetry is separating its GenAI semantic conventions into a dedicated repository, with documented vocabulary for agent and workflow operations, models, token data, retrieval, and message-related fields. The conventions are active rather than fully settled, so teams should standardize the fields they need while isolating version-sensitive mappings. OpenTelemetry
AWS published a reference architecture for AgentCore memory lifecycle policies using nightly Step Functions workflows for pruning, scoring, consolidation, metrics, and audit output. Its pattern combines TTL expiry, relevance decay based on recency and access, and LLM-based consolidation; the practical implication is to design retention and deletion alongside retrieval, rather than letting agent memory accumulate indefinitely. AWS Machine Learning Blog
SGLang 0.5.19 reports 786 pull requests from 214 contributors and adds beam search alongside deployment guidance for new model families and accelerator configurations, including selected FP8 KV-cache and host-memory-tier guidance. For platform teams, serving is increasingly a compatibility and memory-hierarchy problem, not merely an HTTP endpoint in front of a model. SGLang release notes