Optimizing RLHF and llamafile 0.9.2

🔥 What's hot right now
RLHFless is optimizing RLHF training with serverless computing, achieving up to 1.35x speedup and 44.8% cost reduction. By pre-computing shared prefixes, it tackles resource variability and idle time effectively.

🚀 Just shipped
llamafile v0.9.2 is a major update adding support for DeepSeek Distil R1, Gemma 3, and IBM Granite. It also introduces LocalScore, a benchmarking utility to evaluate LLM performance on your specific hardware.

🛠 Useful for the array
AMA-Bench is a new evaluation framework for long-horizon memory in LLM agents, focusing on continuous interaction. It proposes AMA-Agent using causality graphs to identify gaps in current memory systems.