Reading up on SGLang

Thinking Machines argues current turn-based AI voice systems have a scalability ceiling and proposes interaction models using time-aligned micro-turns and a two-model architecture for true real-time collaboration.

Baidu releases Unlimited-OCR, a VLM-based OCR that processes entire documents in a single pass using reference sliding window attention to avoid memory overflow.

MiniMax researcher Olive Song reveals how they debug RL training, handle model hacking, and prioritize alignment and engineering discipline in open-weight model development.

The article explains how prompt caching works through vLLM's paged attention and automatic prefix caching, detailing KV-cache reuse and practical tips for improving cache hits.