FLOPS、算力、卡时、集群规模混着说，是看懂算力新闻的第一道坎。本文厘清 FLOPS 与 FLOPs 一字之差的两个概念，给出 6ND 训练算力估算经验与卡时换算方法，解释峰值与有效算力之间的 MFU 差距，最后附一张看懂算力新闻的换算清单。

2017 年的 Transformer 之后，架构研究沿三条主线展开：把注意力做便宜、把注意力换掉、把 FFN 做稀疏。本文梳理稀疏注意力、FlashAttention、SSM、混合架构与 MoE 的脉络，并给出读新架构论文的三问。

Looped Transformers achieve parameter efficiency by repeatedly executing a shared block across recurrent loops. Each loo…

Most modern multimodal large language models (MLLMs) build on a pretrained visual encoder that provides a strong visual …

In this tutorial, we work with **MSEB**, the Massive Sound Embedding Benchmark from Google Research, and approach it fro…

Comparisons in video self-supervised learning often evaluate complete training recipes rather than isolating the method …

IT之家 9 月 23 日消息，小米 MiMo 大模型负责人罗福莉今日发文，宣布 MiMo-V3 即将采用全新架构。其核心 HySparse 2 今日发布，带来更少的预填充、更小的 KV 缓存、更出色的长上下文检索。 在 1M（100 万）…

World-action models (WAMs) transfer visual and motion priors from pretrained video generators to robot control by jointl…

Modern Transformer design and compression both reduce to allocating capacity under a budget. The standard scalars for th…

Learn how to leverage NVIDIA’s cuDNN Frontend Graph API to build custom kernel fusions, autotuning engine configurations…

World-action models (WAMs) jointly model future observations and actions, typically predicting the future as RGB images.…

]

把 2026 年可核查的公开证据整理成一张六层能力路线图——预训练、后训练 RL、推理时计算、上下文与记忆、智能体与 Harness、世界模型。含 Meta ScaleRL 40 万 GPU 小时实验结论、RL 预算占比 10%–30% 口径、Chinchilla 对比、Meta-Harness 6x 差距等数据锚点，并给出优先级表与算法工程师/产品经理的行动建议。