高分作者动态瀑布流

手撕 DeepSeek-V4 (1) : 模型架构

小冬瓜AIGC | X-R1开源框架 | 现高校LLM对齐研究 原创课程帮助学员拿下OpenAI, Meta, 字节SEED等手撕 DeepSeek-V4 手撕 DeepSeek-V4 系列，含代码小冬瓜AIGC：手撕 DeepSeek-V4 (1) : 模型架构DeepSeek-V4 前瞻，含代码小冬瓜AIGC：【手撕NSA】DeepSeek新作…

deep dive into DeepSeek-V4 (1) : model architecture

小冬瓜AIGC published a Zhihu article relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

如何Day 0快速部署开源大模型？

前言本周五（4.24）上午DeepSeek V4发布，没有五一发布，没有晚上发布，是DeepSeek少有的仁慈。我给梁文锋磕一个。V3 发布时间：2024年12月26日 北京时间 19 时 R1 发布时间：2025年1月20日 北京时间 20 时这次发布时间完全照顾国人的，估计这个发布时间的…

Chinese technical note: 如何Day 0fast deploymentopen-source LLMs？

果冻虾仁 published a Zhihu article relevant to inference and AI systems, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

SkillReducer：为Skills瘦身，破解Token低效难题

目前，很多公开的Skill存在严重冗余，这既浪费成本，也稀释了模型的注意力。因此，SkillReducer 则专注于解决该问题。标题：SkillReducer: Optimizing LLM Agent Skills for Token Efficiency论文地址：https://arxiv.org/pdf/2603.29919背景论文通过对 55,315 个公开技…

Chinese technical note: SkillReducer：为Skills瘦身，破解Token低效难题

吃果冻不吐果冻皮 published a Zhihu article relevant to AI agents and coding workflows, inference and AI systems, frontier and open model development, research signals. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

Claude Code 模型 RL训练中的Reward Hacking

作者: Jiacai Liu总结随着RL infra发展，通过大规模的强化学习提升大模型的能力已成为各家训练的共识。RL训练的目标为最大化模型在环境交互中的累积奖励。但RL训练远非简单的通过看reward，entropy，test accuracy 等曲线指标这么简单。其根本原因在于，即…

Chinese technical note: Claude Code 模型 RLtraining中的Reward Hacking

skydownacai published a Zhihu article relevant to post-training and RL, AI agents and coding workflows, frontier and open model development, evaluation and reliability. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

北向agent化，将重新定义AI Infra的易用性和生态

未来的一些写作计划过年那会儿，我提到北向agent化的一些想法，后面我们团队相继发布了，MindSpore AKG Agent和Model Agent：AKG kernel Agent：利用multi-agent进行kernel的生成和迁移https://mp.weixin.qq.com/s/EAJCL2vtH3AmRWd3FeegSQ 月底的时候，我们还将发布一个Agent，具体内容到…

Chinese technical note: 北向agents化，将重新定义AI Infra的易用性和生态

金雪锋 published a Zhihu article relevant to AI agents and coding workflows. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

超节点亲和AI框架

分享一下与香港科技大学广州陈雷老师团队一起合作的成果 《HyperParallel: A Supernode-Affinity AI Framework》 arxiv： HyperParallel: A Supernode-Affinity AI Framework 当前AI框架面临的挑战AI框架是位于模型算法和硬件的中间层，是支持AI领域快速发…

Chinese technical note: 超节点亲和AI框架

金雪锋 published a Zhihu article relevant to research signals. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

【招人帖-实习】腾讯微信基础-大模型算法研究员

微信基础-大模型算法研究员-Agent方向(北京) 【岗位职责】 1. 负责 Agent Harness 框架的设计与开发，包括 Agent 调度、任务编排与执行上下文管理； 2. 负责 AI Agent 记忆系统的构建与优化，覆盖记忆的写入、组织、检索与生命周期管理，探索跨会话持续记忆…

Chinese technical note: 【招人帖-实习】腾讯微信基础-LLMs算法研究员

紫气东来 published a Zhihu article relevant to AI agents and coding workflows, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

Vibe Engineering 进化史：三重门背后的思维跃迁

最近在公司内外部做了几次关于 Coding Agent 的分享，收到了一些好评。整理了下内容，分享给大家，相信应该有些价值。 内部分享中有比较多具体实践，涉及到内部项目的信息就略过了。这篇文章主要关注于 AI coding 的思维转变方面。 Coding Agent 产品推荐对…

Chinese technical note: Vibe Engineering 进化史：三重门背后的思维跃迁

字节 published a Zhihu article relevant to AI agents and coding workflows. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

Kimi K2.5 干货有点多啊

笔者现在也在搞Agentic相关的训练，趁着春节假期，想来补补课，看看这个方向上一些最近比较有名的工作。目前听来，像Minimax M2.5、Kimi K2.5、Qwen3.5，都在Agentic上下了很大的力。但Writing的时候，Qwen3.5的技术报告还没出来；M2.5似乎并没有论文，只有…

Chinese technical note: Kimi K2.5 technical notes有点多啊

兽族机枪兵 published a Zhihu article relevant to AI agents and coding workflows, frontier and open model development, research signals. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

管中窥豹nano-vllm（三）：ModelRunner补充

本文需要辅助本系列第一篇文章：《管中窥豹nano-vllm（一）：从样例到主流程》中的 ModelRunner 一节。在那篇文章中，介绍了 ModelRunner 的__init__和 run 方法。不过对里面的一些细节进行了省略。本文进行补充。allocate_kv_cachealocate_kv_cache 在 Mod…

Chinese technical note: 管中窥豹nano-vllm（三）：ModelRunner补充

果冻虾仁 published a Zhihu article relevant to inference and AI systems, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

投机采样方案里，EAGLE-3是如何做到一块显卡同时部署目标模型和草稿模型的？

笔者最近发现，有些投机采样方案，能够做到target model和draft model放在同一张显卡里去部署。觉得非常不可思议，于是来看了下这块内容。 一块显卡同时部署俩模型的第一个工作我实在找不到。但发现EAGLE-3是目前比较流行，甚至可以说是工业应用级别最流行…

Chinese technical note: 投机采样方案里，EAGLE-3是如何做到一块显卡同时deployment目标模型和草稿模型的？

兽族机枪兵 published a Zhihu article relevant to inference and AI systems. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

终于有时间好好看看DeepSeek-OCR了。。

上班之后，笔者感觉精力一直非常有限，看到很多新工作出来却没工夫好好研读。终于能乘着春节假期，好好充充电了。之前一直看网上说DeepSeek-OCR好像和一般的OCR不太一样，好像提出了新的长上下文压缩方案，非常感兴趣，所以春节假期的第一天就赶紧来看了一…

Chinese technical note: 终于有时间好好看看DeepSeek-OCR了。。

兽族机枪兵 published a Zhihu article relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

vllm使用plugin集成外部模型结构

模型结构概念介绍 vllm是最流行的LLM 推理服务的解决方案之一，它除了提供LLM 推理的吞吐优化、性能优化的能力以外，也内置了各种各样的模型结构，在这个代码目录中：https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/modelsvllm 的模型结构一直对标着huggingface 的 transformers中的模型结…

Chinese technical note: vllm使用plugin集成外部模型结构

果冻虾仁 published a Zhihu article relevant to inference and AI systems, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

LLM(34)：Batch size 与 Learning rate 的理论与实践

在之前的文章中，笔者比较简略讨论过 scaling law 对 Batch size 与Learning rate 超参设置的指导，当时的主要结论如下：LLM(26)：从信息论的角度解释 scaling law当计算预算增加时，应增大 Batch size；当计算预算增加时，应减小 Learning rate;当计算预算…

Chinese technical note: LLM(34)：Batch size 与 Learning rate 的理论与实践

紫气东来 published a Zhihu article relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

【手撕Engram】DeepSeek 的 Conditional Memory 能取代 Attention 吗？（超长文、附代码）

小冬瓜AIGC | X-R1开源框架 | 现高校LLM对齐研究 原创课程帮助学员拿下OpenAI, Meta, 字节SEED等没有取代，Engram 聚焦短距离序列建模。手撕 DeepSeek-V4 系列：小冬瓜AIGC：手撕 DeepSeek-V4 (1) : 模型架构DeepSeek-V4 技术前瞻系列，都附上本人手撕代码 …

Chinese technical note: 【deep dive intoEngram】DeepSeek 的 Conditional Memory 能取代 Attention 吗？（超长文、附代码）

小冬瓜AIGC published a Zhihu article relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

AKG kernel Agent：利用multi-agent进行kernel的生成和迁移

分享一下最近团队和湖南大学一起合作的成果，如何通过multi-agent技术进行kernel的自动生成和迁移。背景：算子开发领域主要存在四个挑战：一、算子开发效率低，门槛高，依赖专业知识从gpu cuda/ascend aclnn，到cutlass模版编程，再到Triton/TileLang的Pyth…

Chinese technical note: AKG kernel agents：利用multi-agents进行kernel的生成和迁移

金雪锋 published a Zhihu article relevant to AI agents and coding workflows, inference and AI systems, research signals. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

深读 Manus：寻找 Agent 产品的“品味”与“定力”

一晃有 3 个月没有更新了，今天正好是 26 年的第一天，听了个非常精彩的播客，终于又激发起一些分享和表达的欲望。 这个播客是来自于张小珺商业访谈录的 对 Manus 首席科学家 Peak 的访谈，我对其评价为 25 年最棒的一期播客，Peak 的技术深度，对产品与商…

Chinese technical note: 深读 Manus：寻找 agents 产品的“品味”与“定力”

字节 published a Zhihu article relevant to AI agents and coding workflows. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

【手撕 mHC】详解DeepSeek残差链接mHC进化之路（超长文、附代码）

小冬瓜AIGC | X-R1开源框架 | 现高校LLM对齐研究 原创课程帮助学员拿下OpenAI, Meta, 字节SEED等手撕 DeepSeek-V4 系列： 小冬瓜AIGC：手撕 DeepSeek-V4 (1) : 模型架构DeepSeek-V4 技术前瞻系列，都附上本人手撕代码 notebook 小冬瓜AIGC：【手撕NSA】Deep…

Chinese technical note: 【deep dive into mHC】详解DeepSeek残差链接mHC进化之路（超长文、附代码）

小冬瓜AIGC published a Zhihu article relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

做正确且难的事——小冬瓜AIGC的25年终总结

小冬瓜AIGC | X-R1开源框架 | 现高校LLM对齐研究 原创课程帮助学员拿下OpenAI, Meta, 字节SEED等在今年的最后一天，聊聊 2025 LLM 的发展和博主的年终总结。2025 LLM 简记25年的 RL 和基础架构飞速发展RL强化学习（RL）迎来跨越式发展。年初 DeepSeek-R1 引…

Chinese technical note: 做正确且难的事——小冬瓜AIGC的25年终总结

小冬瓜AIGC published a Zhihu article relevant to post-training and RL, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

你真的搞懂了LLM性能压测的各项指标吗？

之前发现不同框架的性能差异有出入，当时并没有太在意。最近，对社区开源的LLM性能压测工具做了一个深度体验。发现各个框架对于各指标（如：ILT/TPOT）的定义真的是五花八门，稍微不注意可能就会用错里面的指标。本文梳理了各性能压测工具的计算公式，试图…

Chinese technical note: 你真的搞懂了LLM性能压测的各项指标吗？

吃果冻不吐果冻皮 published a Zhihu article relevant to frontier and open model development, evaluation and reliability. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

LLM Agent RL的一些实践感悟

最近几个月在各种场景上做了大量的agent RL训练，例如search agent， 数据分析agent 等等。既有小的dense model，也有大的moe。有单一数据场景，有多源合板数据场景。有失败的经历，也有成功的经历。这里抽空分享下一些感悟，比较乱，随便看看就行: 稳定性…

Chinese technical note: LLM agents RL的一些实践感悟

skydownacai published a Zhihu article relevant to post-training and RL, AI agents and coding workflows, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

DeepSeek 视觉语言大模型技术演进（从DeepSeek VL/VL2到DeepSeek OCR）

随着 ChatGPT 迅速走红，这两年大家在日常工作中使用 LLM 进行的场景越来越多。本系列将针对主流算法架构进行讲解。 大模型算法演进大模型算法架构：QWen技术演进及剖析大模型算法架构：DeepSeek技术演进及剖析大模型算法架构：LLaMA3 技术剖析大模型算法架…

Chinese technical note: DeepSeek 视觉语言LLMs技术演进（从DeepSeek VL/VL2到DeepSeek OCR）

吃果冻不吐果冻皮 published a Zhihu article relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

多模态大模型入门：举一个例子说明SW-MSA

例子：8x8 特征图上的 SW-MSA 假设我们有一个 8×8 的特征图，并且每个窗口的大小为 4×4（即 M=4）。我们将分两步来展示 W-MSA 和 SW-MSA 的工作原理。第一步：W-MSA1、常规窗口划分 将 8×8 的特征图划分为4个不重叠的 4×4 窗口。 假设这四个窗口分别为 W…

Chinese technical note: multimodalLLMs入门：举一个例子说明SW-MSA

xhchen published a Zhihu article relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

谈谈LLM生成文本的惩罚参数

之前在 一文搞懂大模型生成文本的解码策略 中简单介绍过惩罚参数（重复惩罚、频率惩罚、存在惩罚），本文来详谈一下这三者的区别以及代码实现。重复惩罚（repetition_penalty）工作机制： 直接针对当前上下文中（包括输入+已生成的token）已经出现过的 tok…

Chinese technical note: 谈谈LLM生成文本的惩罚参数

吃果冻不吐果冻皮 published a Zhihu article relevant to inference and AI systems, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

DeepSeek团队发布视觉压缩OCR模型，哪些信息和技术亮点值得关注？

Chinese technical note: DeepSeek团队发布视觉压缩OCR模型，哪些信息和技术亮点值得关注？

vibe memo published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

DeepSeek-OCR与Glyph: 视觉压缩会重塑LLM的输入范式？

DeepSeek-OCR与Glyph不约而同地证明，将文本渲染为图像输入模型，能在保持高精度的同时实现3-10倍的压缩效率提升，从根本上解决了长上下文处理的算力瓶颈。这种“视觉-文本压缩”范式不仅让模型天然支持富文本格式，还避免了传统分词器的历史包袱和安全风险…

Chinese technical note: DeepSeek-OCR与Glyph: 视觉压缩会重塑LLM的输入范式？

vibe memo published a Zhihu article relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

如何评价 Thinking Machines Lab发表的首篇长文《克服 LLM 推理中的不确定性》？

Chinese technical note: How to read Thinking Machines Lab发表的首篇长文《克服 LLM inference中的不确定性》？

吃果冻不吐果冻皮 published a Zhihu answer relevant to inference and AI systems, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

RL崩溃背后的训推mismatch

本文最初发在 notion blog 中 ： When Speed Kills Stability: Demystifying RL Collapse from the Training-Inference Mismatch，作者: Jiacai Liu @skydownacai, Yingru Li @Yingru Li, Yuqian Fu @storm, Jiawei Wang, Qian Liu, Yu Shen为方便观看，在…

Chinese technical note: RL崩溃背后的训推mismatch

skydownacai published a Zhihu article relevant to AI engineering. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

Tongyi DeepResearch series: WebWeaver, WebResearcher, ReSum, WebSailor-V2,AgentScaler

概述 Paper, WebWeaver解决问题: Open-Ended Deep Research现有方法受限于静态研究流程(规划与证据获取使用固定的worfklow，不是迭代可反馈的)和一次性生成范式（易受“长上下文遗漏”和幻觉问题影响）。“搜索-然后-生成”方法：智能体在生成报告前收集所…

Tongyi DeepResearch series: WebWeaver, WebResearcher, ReSum, WebSailor-V2,agentsScaler

vibe memo published a Zhihu article relevant to AI agents and coding workflows, evaluation and reliability. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

让Agent“记得对、做得准、跑得快”：Context Engineering实战

最近看了一个非常不错的 关于 Context Engineering 的分享，里面综合讨论了 Anthropic，Cognition，Manus，Chroma，LangChain 等公司在这个领域方面的最新思考和实践，有点像当年分享过的 Applied LLMs，非常值得一看。AI Agent 开发这个领域还处在比较早期…

Chinese technical note: 让agents“记得对、做得准、跑得快”：Context Engineering实战

字节 published a Zhihu article relevant to AI agents and coding workflows, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

Writing effective tools for agents — with agents

如何让智能体调用的数百种工具发挥最大效用？本文将介绍在各种智能体AI系统中提升性能的最有效技术，帮助你在现实世界任务中取得出色表现。 翻译文章: https://www.anthropic.com/engineering/writing-tools-for-agents概述模型上下文协议（MCP）可以为大型语言模型（LLM）智能体赋能，使其能够调用数…

Chinese technical note: Writing effective tools for agentss — with agentss

vibe memo published a Zhihu article relevant to AI agents and coding workflows, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

RL训练中为什么熵减往往意味着训练收敛?

最近半年以来，有关于RL+Entropy的研究非常的多。对于离散的动作空间 \mathcal{A} , 策略 \pi 在状态 s 处的entropy为 \mathcal{H} \left( \pi \left( \cdot |s \right) \right) :=\mathbb{E} _{a\sim \pi \left( \cdot |s \right)}\left[ -\log \pi \left(…

Chinese technical note: RLtraining中Why熵减往往意味着training收敛?

skydownacai published a Zhihu article relevant to post-training and RL. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

【招人帖】腾讯微信基础大模型算法

招募范围：社招、校招、暑期实习、27届及以后实习（有青云计划名额）【岗位职责】1.负责社交大模型方向的记忆检索、Agent函数调用、风格化基座模型等方向的算法突破2.紧密贴合业务，通过后训练（SFT&RL）提升模型的专项问题解决能力3.基于微信场景数据提供…

Chinese technical note: 【招人帖】腾讯微信基础LLMs算法

紫气东来 published a Zhihu article relevant to post-training and RL, AI agents and coding workflows, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

深入 vLLM：高吞吐量 LLM 推理系统的剖析

深入了解 vLLM：高吞吐量 LLM 推理系统的剖析，从分页注意力、连续批处理到多 GPU 动态扩展，逐步揭示现代 LLM 引擎的核心组件和高级功能。 译文: https://www.aleksagordic.com/blog/vllm VLLM相关组件设计目的概述一、LLM 引擎与引擎核心分页注意力 (PagedAttention) 和 KV …

Chinese technical note: 深入 vLLM：高吞吐量 LLM inference系统的剖析

vibe memo published a Zhihu article relevant to inference and AI systems, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

在SFT训练阶段，有大量短文本数据与少量的长文本数据，从训练效率和模型性能的角度，应该如何选择批处理策略？

在SFT训练阶段，有大量短文本数据与少量的长文本数据，从训练效率和模型性能的角度，应该如何选择批处理策略？1、sorted batching 在传统的批处理（batching）中，由于序列长度不一，通常需要通过填充（padding）将同一批次内的所有序列对齐到最大长度，这…

Chinese technical note: 在SFTtraining阶段，有大量短文本数据与少量的长文本数据，从training效率和模型性能的角度，应该如何选择批处理策略？

xhchen published a Zhihu article relevant to post-training and RL. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

主流大模型中的“混合结构”

Hybrid MoE MoE可以应用于所有FFN层，但也有一些大模型使用MoE-Dense混合结构，在某些层将FFN层替换为MoE，而在某些层还是使用FFN层。Hybrid Attention位置编码 代表作：Llama4，Command R+ RoPE已经是主流大模型使用的位置编码，但现有的基于RoPE的方法在…

Chinese technical note: 主流LLMs中的“混合结构”

xhchen published a Zhihu article relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

Linear Attention vs Self Attention

相关论文Transformers are RNNs: Fast Autoregressive Transformers with Linear AttentionTransNormerLLM: A Faster and Better Large Language Model with Improved TransNormerMiniMax-01: Scaling Foundation Models with Lightning AttentionLinear At…

Linear Attention vs Self Attention

xhchen published a Zhihu article relevant to frontier and open model development, research signals. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

烧掉上亿 Token 后，我总结的 Coding Agent 高阶玩法

今天来继续聊聊如何用好 Claude Code，Codex，Cursor CLI 这类产品的一些经验和思考。 技术栈选择我们在公司推广 AI Coding 类工具时很早就发现，大模型对于不同编程语言、框架等方面的能力程度差异还是挺大的。比如我们之前很多项目用了 Scala 语言以及一…

Chinese technical note: 烧掉上亿 Token 后，我总结的 Coding agents 高阶玩法

字节 published a Zhihu article relevant to AI agents and coding workflows, inference and AI systems, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

如何评价 OpenAI 8 月 6 日推出的开源模型 gpt-oss？和国产开源模型相比怎么样？

Chinese technical note: How to read OpenAI 8 月 6 日推出的开源模型 gpt-oss？和国产开源模型相比怎么样？

小冬瓜AIGC published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

Reasoning LLM（六）：内在奖励

众所周知，在强化学习训练中的关键环节就是奖励信号的获取，准确的奖励信号对于训练的效果至关重要。在经典RL 中，奖励信号可以看作环境的一部分 —— 即行动后环境的真实反馈，而在 RL 训练 LLM 中，奖励值的来源主要有两种方式：批判式：即 RLHF 中的 RM…

Chinese technical note: Reasoning LLM（六）：内在reward

紫气东来 published a Zhihu article relevant to post-training and RL, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

Meta 推荐芯片MTIAV2分析（ISCA 25）

本文分析仅个人专业论文解读，不代表任何组织观点。 论文，Meta’s Second Generation AI Chip: Model-Chip Co-Design and Productionization Experiences， 在大模型时代，坚持做推荐芯片的不多，可以见到做wafer scale的大推理芯片，做推荐小模型芯片的的…

Chinese technical note: Meta 推荐芯片MTIAV2分析（ISCA 25）

南爸安爸 published a Zhihu article relevant to inference and AI systems, frontier and open model development, research signals. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

大模型中目前最快最好的解码策略是什么呢？还有哪些问题值得研究？

Chinese technical note: LLMs中目前最快最好的解码策略What is呢？还有哪些问题值得研究？

吃果冻不吐果冻皮 published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

OpenAI 上线新一代编程神器 Codex，有哪些技术亮点？程序员的工作会被彻底颠覆吗？

Chinese technical note: OpenAI 上线新一代coding神器 Codex，有哪些技术亮点？程序员的工作会被彻底颠覆吗？

字节 published a Zhihu answer relevant to AI agents and coding workflows. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

MoE训练中的Top-K运算不会导致不可导（不连续）吗？

Chinese technical note: MoEtraining中的Top-K运算不会导致不可导（不连续）吗？

小冬瓜AIGC published a Zhihu answer relevant to AI engineering. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

重读 Google 旧文 Pathways，寻找 veRL 中 Single-controller 思想源头

在研究强化学习训练框架 veRL 时，初学者首先会被灌输 single controller 和 multi-controller 这两个基础的概念。此概念最早可以追溯到 Google 在 2022 年发布的一篇旧文 Pathways，veRL 在设计中也深受 Pathways 文章的影响。今日重读此文，一些死去的记…

Chinese technical note: 重读 Google 旧文 Pathways，寻找 veRL 中 Single-controller 思想源头

方佳瑞 published a Zhihu article relevant to post-training and RL. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

ChatGPT最实用的提示（Prompts）写法有哪些？

Chinese technical note: ChatGPT最实用的提示（Prompts）写法有哪些？

字节 published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

字节豆包大模型团队与香港大学发布全新 RLHF 框架提升 1.5~20 倍吞吐量，这意味着什么？

Chinese technical note: 字节豆包LLMs团队与香港大学发布全新 RLHF 框架提升 1.5~20 倍吞吐量，这意味着什么？

方佳瑞 published a Zhihu answer relevant to post-training and RL, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

veRL：All in RL元年的必修课

如果为 AI Infra 每年选择一个主旋律关键词，2023 年是 Pretrain，2024 年是 Serving，2025 年我认为应该是 RL。 RL 本一个老生常谈的话题，在 AlphaGo 时代已经经历过研究热潮。ChatGPT 问世时采用了基于人类反馈的强化学习（RLHF）技术，将 PPO 应用于人…

Chinese technical note: veRL：All in RL元年的必修课

方佳瑞 published a Zhihu article relevant to post-training and RL, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

SageAttention：即插即用的8-bit Attention 最佳实践

LLM 的量化加速是近两年的热点话题，老方法过江之鲫，新文章层出不穷。然而，过往的研究焦点大多集中于对模型中的Linear 层进行量化。尤其是在 LLM 推理的decode阶段，如 GPTQ 和 AWQ 等仅权重量化（weight-only quantization），SmoothQuant 等 激活权重协…

Chinese technical note: SageAttention：即插即用的8-bit Attention 最佳实践

方佳瑞 published a Zhihu article relevant to inference and AI systems, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

如何评价Nvidia发布的大模型推理PD分离架构Dynamo？

Chinese technical note: How to readNvidia发布的LLMsinferencePD分离architectureDynamo？

南爸安爸 published a Zhihu answer relevant to inference and AI systems, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

百度旗下 AI 芯片昆仑芯支持单机部署 DeepSeek 满血版大模型，此举有何深意？

Chinese technical note: 百度旗下 AI 芯片昆仑芯支持单机deployment DeepSeek 满血版LLMs，此举有何深意？

吃果冻不吐果冻皮 published a Zhihu answer relevant to inference and AI systems, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

DeepSeek-R1 API定价为什么这么便宜，在目前全量模型部署非常困难的情况下，是否注定亏损？

Chinese technical note: DeepSeek-R1 API定价Why这么便宜，在目前全量模型deployment非常困难的情况下，是否注定亏损？

方佳瑞 published a Zhihu answer relevant to inference and AI systems, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

吃瓜DeepSeek 推理成本需要的相关概念：Throughput，TPOT，TTFT

近期，DeepSeek 推理成本问题备受关注，引发了广泛讨论。针对各种 DeepSeek R1 的 API 服务展开了对比。围绕 MaaS 是盈利还是亏损也展开了激烈的争论。此外，微信全量接入 DeepSeek R1 后能否承受得住也成为了焦点话题。大家讨论DeepSeek部署成本时候常说的…

Chinese technical note: 吃瓜DeepSeek inference成本需要的相关概念：Throughput，TPOT，TTFT

方佳瑞 published a Zhihu article relevant to inference and AI systems, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

为何腾讯、阿里、华为等大厂难出DeepSeek级颠覆产品？是组织僵化还是创新无能？

Chinese technical note: 为何腾讯、阿里、华为等大厂难出DeepSeek级颠覆产品？是组织僵化还是创新无能？

南爸安爸 published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

如何评价 DeepSeek 的 R1 与 R1-Zero 模型？

Chinese technical note: How to read DeepSeek 的 R1 与 R1-Zero 模型？

紫气东来 published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

如何评价deepseek预发布的deepseek-R1？

Chinese technical note: How to readdeepseek预发布的deepseek-R1？

小冬瓜AIGC published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

如何评价 DeepSeek 的 R1 与 R1-Zero 模型？

Chinese technical note: How to read DeepSeek 的 R1 与 R1-Zero 模型？

skydownacai published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

如何评价 MiniMax 于 2025 年 1 月 15 日发布的 MiniMax-01 系列模型？

Chinese technical note: How to read MiniMax 于 2025 年 1 月 15 日发布的 MiniMax-01 系列模型？

紫气东来 published a Zhihu answer relevant to AI engineering. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

可否给一个实际例子，结合GPU硬件架构和Transformer架构讲一讲，如何提高训练和推理速度？

Chinese technical note: 可否给一个实际例子，结合GPU硬件architecture和Transformerarchitecture讲一讲，如何提高training和inference速度？

紫气东来 published a Zhihu answer relevant to inference and AI systems. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

阿里云Qwen2两小时登顶HuggingFace开源大模型榜首，你怎么看？

Chinese technical note: 阿里云Qwen2两小时登顶HuggingFaceopen-source LLMs榜首，你怎么看？

vibe memo published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

单卡可Million-context推理TTFT 10倍加速 - MInference 1.0

本文介绍 @灰墙 与我的最新工作 - MInference 1.0！MInference 1.0是基于dynamic sprase attention的pre-filling推理加速算法，其可以scale至单卡1M的context length，并在TTFT（time-to-first-token）上取得10倍的加速效果。MInference 1.0在众多tasks上 …

Chinese technical note: 单卡可Million-contextinferenceTTFT 10倍加速 - MInference 1.0

九号 published a Zhihu article relevant to inference and AI systems. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

为什么加速LLM推断有KV Cache而没有Q Cache？

Chinese technical note: Why加速LLM推断有KV Cache而没有Q Cache？

小冬瓜AIGC published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

大模型推理优化哪些公司或者团队做的好？

Chinese technical note: LLMsinference优化哪些公司或者团队做的好？

vibe memo published a Zhihu answer relevant to inference and AI systems, frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

训练多模态模型有哪些trick？

Chinese technical note: trainingmultimodal模型有哪些trick？

vibe memo published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

请问现在有哪些研究和数据集可以评测大语言模型llm的长文本理解能力？

Chinese technical note: 请问现在有哪些研究和数据集可以evaluation大语言模型llm的长文本理解能力？

九号 published a Zhihu answer relevant to frontier and open model development, evaluation and reliability. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

全面评测LLMs长文本能力的Benchmarks

长文本是目前各家LLMs的主要竞争场景。 长文本处理能力的对LLMs的重要性是显而易见的：其赋予了LLMs区别于任何其他技术的解决问题能力，比如一个完整项目的代码生成，根据原始产业数据生成研报，整本书籍的摘要问答，网文和剧本等超长文本的生成等。 目前大…

Chinese technical note: 全面evaluationLLMs长文本能力的Benchmarks

九号 published a Zhihu article relevant to frontier and open model development, evaluation and reliability. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

读 Mora 论文笔记

今天来喷一下 Mora 这篇文章。 https://arxiv.org/pdf/2403.13248.pdf群里面有很多不明就里的朋友在转发，然后我看了一遍，有点想评论的欲望。disclaimer, 我现在也在做生成模型，mora 这文章里评测的一些产品中有我司的产品。我的结论是这文章完全是蹭热度的文章，文笔不错…

Chinese technical note: 读 Mora 论文笔记

ZZZZJ published a Zhihu article relevant to evaluation and reliability, research signals. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

如何看待Anthropic公司在ChatGPT4.5推出前，宣布推出Claude 3？

Chinese technical note: 如何看待Anthropic公司在ChatGPT4.5推出前，宣布推出Claude 3？

九号 published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

谷歌正式推出开源大语言模型 Gemma，声称超越 Meta Llama-2 竞品，将带来哪些影响？

Chinese technical note: 谷歌正式推出开源大语言模型 Gemma，声称超越 Meta Llama-2 竞品，将带来哪些影响？

九号 published a Zhihu answer relevant to AI engineering. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

评测Gemma - 对比Mistral，Qwen1.5

简要评测谷歌Gemma，对比基线为Qwen1.5，Mistral，与Llama-2.总的来说：代码能力：与Qwen1.5同属第一梯队，超越Mistral，离CodeLlama还有一些距离。数学能力：未超越Qwen1.5，与Mistral并列排第二。文本建模能力：与Mistral有较大差距，但鲁棒性显著超过Lla…

Chinese technical note: evaluationGemma - 对比Mistral，Qwen1.5

九号 published a Zhihu article relevant to frontier and open model development, evaluation and reliability. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

可以用大语言模型评估大语言模型的结果吗？

Chinese technical note: 可以用大语言模型评估大语言模型的结果吗？

九号 published a Zhihu answer relevant to AI engineering. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

出「月考卷」来杜绝LLMs在评测中的作弊

本工作发表在AAAI‘24 Main Track，在此查看会议版论文：LatestEval: Addressing Data Contamination through Dynamic and Time-Sensitive Test Construction在基于leaderboard的LLMs评测体系中，刷榜问题是几乎无法杜绝的。首先，出于PR的需求，学术界和工…

Chinese technical note: 出「月考卷」来杜绝LLMs在evaluation中的作弊

九号 published a Zhihu article relevant to frontier and open model development, evaluation and reliability, research signals. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

万亿LLM/多模态 MoE和万亿Embedding大模型系统设计上的共性和差异的一点体会

适合5～10分钟阅读，思路不十分严谨，文后写了一些思考，适合对MoE和embedding计算都有一定基础的同学，欢迎评论和讨论。核心观点： 万亿embedding和MoE在系统设计思想上有非常多共性，也有一些局部差异，两个研究方向可以互相借鉴。Embedding主要聚焦加速…

Chinese technical note: 万亿LLM/multimodal MoE和万亿EmbeddingLLMs系统设计上的共性和差异的一点体会

南爸安爸 published a Zhihu article relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

为什么NLP模型通常使用AdamW作为优化器，而不是SGD？

Chinese technical note: WhyNLP模型通常使用AdamW作为优化器，而不是SGD？

兽族机枪兵 published a Zhihu answer relevant to AI engineering. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

你会花3w学chatgpt和mid journey，然后就业吗？

Chinese technical note: 你会花3w学chatgpt和mid journey，然后就业吗？

兽族机枪兵 published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

模型推理CUDA100% GPU0%是什么情况？

Chinese technical note: 模型inferenceCUDA100% GPU0%What is情况？

ZZZZJ published a Zhihu answer relevant to inference and AI systems. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

ChatGPT的出现对哲学会产生哪些挑战?

Chinese technical note: ChatGPT的出现对哲学会产生哪些挑战?

金雪锋 published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

求：复旦强化学习老师推荐?

Chinese technical note: 求：复旦reinforcement learning老师推荐?

skydownacai published a Zhihu answer relevant to post-training and RL. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接

如何评价 OpenAI 的超级对话模型 ChatGPT ？

Chinese technical note: How to read OpenAI 的超级对话模型 ChatGPT ？

金雪锋 published a Zhihu answer relevant to frontier and open model development. The original Chinese excerpt is included below so the feed can preserve the raw source while giving English readers enough context to triage the item.

原文链接