Tag
Hugging Face turned real coding agent harnesses like Claude Code, Codex, Hermes, Pi, and OpenCode into RL training environments via a capture proxy, requiring no changes to the harnesses or training code, with all models, tasks, data, and code fully open sourced.
Ai2 发布 Olmo-core 3，一款面向大规模 MoE 训练的开源框架升级，采用 DDP-based 路由设计与多种并行技术，可在保持约 3.2B 激活参数的情况下将总参数扩展至万亿级别，吞吐量较旧实现提升约 2.7 倍。
AllenAI released Olmo-core 3, an open, scalable training framework redesigned for large mixture-of-experts LLMs, benchmarked at over one trillion total parameters with minimal throughput loss as the expert pool grows.
MiMo-V2.6 introduces Groupwise Advantage Redistribution to enhance reinforcement learning for AI agents by comparing sibling attempts and using graded feedback, showing steady performance improvements across multiple task domains.
The article details the infrastructure and challenges in training the Dyna-2 AI model at a million-hour scale, focusing on data lifecycle and GPU cluster optimizations.
KaichouYou discusses challenges in scaling concurrent rollouts for RL training infrastructure, highlighting a stress test of sandbox scaling on Qwen3 235B with SemiAnalysis, including a writeup of errors and fixes.
Discusses how sandbox startup latency and scaling in RL training infrastructure can significantly impact training performance, referencing a detailed analysis by SemiAnalysis on matching trainer and generator throughput.
The article argues that entry-level work serves as training infrastructure for developing judgment and skills, and that AI adoption must account for this apprenticeship function to avoid weakening the path to senior expertise.
The tweet discusses Microsoft AI's use of Ray actors for training the MAI-Thinking-1 model, enabling finer granularity for heterogeneous compute and better CPU resource utilization in GPU clusters.