CrazyTensor

[Tanishq Mathew Abraham, Ph.D. @iScienceLuvr]

DeepSeek releases DeepSeek-V4.1-Flash!! 🔥

This model is wild... the benchmarks are showcasing it's GPT-5.6 Sol level, yet it's only 552B params?! This feels like it shouldn't be possible, what's the catch? Perhaps benchmarkmaxxed?

Let's look into the architecture and training of the model:

The key focus seems to be on more aggressive KV cache compression:

"We adopt a Causal Encoder-Decoder (CED) architecture, in which decoder global KV is projected from the final encoder hidden states. This design enables the model to activate 8B parameters per token during prefill and 16B during decode, which is particularly cost-effective for input-heavy agentic scenarios."

"DeepSeek-V4 can be viewed as an SWA-based local-processing backbone augmented with compressed global context."

The idea behind CED is that the decoder's KV cache is constructed from the encoder output, bypassing full decoder computation.

They also introduce Compressed Sparse Attention 2 (CSA2) that has three operating modes that differ in how they obtain main KV, indexer K, and Top-K indices. (frankly I don't understand this part very well 😭)

DeepSeek-V4.1-Flash uses a variant of mHC called single-pass mHC, and they also incorporate a 196B Engram module to decouple memorization from computation.

The model is natively multimodal: a vision embeddings generated from "DeepSeek-ViT" (a pretty standard ViT arch) are passed jointly with the text tokens into the model. This is trained first with SigLIP loss then with autoregressive loss for the combined vision encoder+LLM.

"we train DeepSeek-V4.1-Flash on a large-scale multimodal corpus comprising 45T tokens."

They use FP4 KV cache with quantization-aware training to further save storage.

Regarding post-training:

"In this release, we refrain from introducing novel post-training algorithms."

"at the current stage, the marginal return of engineering the data and environment pipeline substantially exceeds that of algorithmic novelty in post-training."

They utilize the model itself to construct its own training environments, based on data they are getting from model use internally.

"As we transitioned from DeepSeek-V3 to V4, the rapidly growing number and diversity of agentic training environments motivated us to build DeepSeek Elastic Compute (DSec), a production-grade sandbox platform for large-scale agentic training and evaluation."

"We therefore introduce a scalar effort level 𝑏 as an explicit conditioning signal during reinforcement-learning."

max --> b=100, high --> b=75, low --> b=50.

"As the last stage of post-training, the final full-vocabulary OPD task is trained on datasets from all domains using over 40 teacher models."

Damn, this is a dense report, I've barely touched the surface tbh, very interesting!!

model: https://t.co/CXQzLhSk8h

paper: https://t.co/ULZRPXiarS

[meng shao @shao__meng]

最近分享的 5 门来自 Stanford、CMU 和 MIT 三所大学的 AI 课程，我再整合分享一次，视频内容很多，可以先收藏慢慢学习 🔖

1. 斯坦福大学 CS329A Self-Improving AI Agents
人定义意图、Agent 执行的新软件工程
https://t.co/qkXLsdlDmT

2. 斯坦福大学 CS 329Z: Engineering AI Agents
从模型到系统：流水线 → 复合 AI → 自主 Agent
https://t.co/q7whzZSWub

3. 斯坦福大学 CS146S: The Modern Software Developer
人定义意图、Agent 执行的新软件工程
https://t.co/AhFJW6xl8V

4. MIT 多模态 AI 课程 2026 版
YouTube 全集上线，融合对齐 → 推理智能体
https://t.co/GhRjioday0

5. CMU 秋季新课 AI Agents
OpenHands 作者亲授，从零搭 Harness、建评测、用 RL 训智能体，讲义视频已公开
https://t.co/8WMWTkDwSN

[山景城小路 @Zen_with_AI]

斯坦福免费机器人入门课！
CS223A《机器人学导论》，16讲全公开。

讲师 Oussama Khatib 是斯坦福计算机科学教授���机器人领域传奇人物。

课程从空间变换、正逆运动学、雅可比、轨迹生成、动力学到运动规划与力控制，打下扎实数学基础，特别适合想系统入门的人。

免费观看：https://t.co/L1izbvkxKv
YouTube 播放列表也有完整讲座。

[Paul Liang @pliang279]

Sharing lecture videos for the **How to AI (Almost) Anything/Multimodal AI** course I taught at MIT in spring 2026.

This course became quite a hit the last time I shared it in spring 2025. Spring 2026's updated version contains updated topics on multimodal agents, reasoning, self-evolving AI, and new modalities like touch and smell. Also includes slides (not videos) of guest lectures on multimodal AI for health, design, manufacturing, cities, & transportation.

Youtube playlist: https://t.co/awhFUqHGBt

Course website and materials: https://t.co/m7LZaCQ4fw

Today's AI can be applied to almost anything - from language to vision, audio, sensors, medical data, music, art, smell, and taste. This course covers the principles of AI (focusing on deep learning and foundation models), how we can apply AI to novel real-world data modalities, and multimodal AI that can process many modalities at once, such as connecting language and multimedia, music and art, sensing and actuation, and more.

[Graham Neubig @gneubig]

We started to post videos for CMU 11-768, AI Agents!

All of the videos will be posted to this playlist, so please bookmark/follow it if you want to be notified of new ones! I'll also try to post them on this thread too.

https://t.co/HAJgKXve0x

[astaxie @astaxie]

真的很感慨，斯坦福的课程更新速度太快了。

CS329Z 已经在教 AI Agent：RAG、Tool Use、MCP、Memory、Multi-Agent、Eval、Coding Agent……

再看看很多大学生还在学习上一个时代、甚至已经被淘汰的东西。

强烈建议身边的大学生都去学学这门课，真的很值得。

https://t.co/7CWzlKMzlu

[ClaudeDevs @ClaudeDevs]

Retailers running shopping agents on Claude have seen carts up to 35% larger and shoppers 60% more likely to complete a purchase.

See our blog to learn more about the architecture, latency & cost techniques, and eval practices:
https://t.co/lAgsVD5JSL

[ClaudeDevs @ClaudeDevs]

We're open-sourcing Claude Commerce Agents.

This is a blueprint for building shopping and merchant agents, with reference implementations across retail, travel, telecom, and entertainment.

[Mihail Eric @mihail_eric]

I’m excited to finally announce the newest edition my Stanford course 𝗧𝗵𝗲 𝗠𝗼𝗱𝗲𝗿𝗻 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿. It has been 9 months in the making.

Last November, with the release of Claude Opus 4.5, coding agents experienced a step function improvement in capability. We all felt it. The LLMs were more powerful, could reason for longer, solve harder tasks.

This year’s iteration of my course reflects the 2026 metamorphosis of software engineering.

My core belief is simple: AI-native developers of the LLM era are going to become the most important members of any software organization. I have designed my course to train this next generation of engineers.

𝗪𝗵𝗮𝘁’𝘀 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝘁𝗵𝗶𝘀 𝘁𝗶𝗺𝗲 𝗮𝗿𝗼𝘂𝗻𝗱

First, 85% of my Fall 2025 class material is being thrown out. The Fall 2026 syllabus reflects the core capabilities AI-native engineers must have: agent skills, advanced context engineering, MCP portals, agent-ready codebase principles, agentic code review, security, parallelizing background agents, software factories, and more.

Second, I am going to teach my students how to have software taste. Every student will be required to ship pull requests to production-grade, real-world codebases. The course is collaborating with the top open-source AI repos who will offer support and mentorship to students on how to meaningfully contribute to their projects.

This has never been done before in any university course so I am incredibly grateful to our OSS Partners: @browserbase, @HeyGen, @CopilotKit, @semgrep, @OpenHandsDev, @milvusio, @marimo_io, Pi, @crewAIInc, @warpdotdev, @vercel, @cmux, @arizeai, @UnslothAI, and @anyscalecompute.

𝗪𝗵𝗮𝘁’𝘀 𝘀𝘁𝗮𝘆𝗶𝗻𝗴 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲

I’m fortunate to again have AI software engineering leaders and founders as guest speakers to share their learnings from building top coding agent products. Thank you to @leerob from @cursor_ai, @bcherny of @claudeai code, @EnoReyes of @FactoryAI, @silasalberti of @cognition, @0xine of @semgrep, Rajesh Bhatia of @Cloudflare , @amasad of @Replit, and @eladgil.

All resources will be available online. All classes will be available to the public.

9/22 on Stanford campus. See you in class. 

https://t.co/wTokHyUMsz

[lifcc @mylifcc]

Netflix 研究员 Cameron Wolfe 8 月 24 日发了一篇 LLM 强化学习总览。
从策略和轨迹写到 PPO、GRPO，再写到 R1 之后那一串补丁。
很多人把后训练理解成再刷一遍 SFT。这篇把真正在烧算力的那一段拆开了。

[DAIR.AI @dair_ai]

Nice benchmark to measure agentic e-commerce capabilities.

They ran an agent for one simulated year of e-commerce operations and it ends up with 27.3% of the money a human makes.

MerchantBench is a 365-day order-level simulation grounded in 98,843 real product records with 26 tools for agent interaction. Agents handle product sourcing, listing and pricing control, cash-flow management, and feedback that arrives at wildly different delays.

Scoring runs on cumulative net assets, so incoherence compounds rather than averaging out. Eight LLMs across two agent frameworks, 48 runs of 365 simulated days each, and the best configuration still lands far under the human baseline.

Bounded tasks with immediate success criteria have been flattering agents that cannot hold a plan for a month.

Paper: https://t.co/wEWNbzXUjb

Track more trending AI papers in our academy: https://t.co/LRnpZN7L4c

[Artificial Intelligence Papers @SciFi]

MerchantBench: Benchmarking LLM Agents for Long-Term Coherence in E-Commerce Operations

Qiming Shi, Yulong Tao, Linbo Jin, Zhaolu Kang, Yibo Dou, Jiawen Zhu, Tianjun Pan, Shaokang Fu, Chengyu Wang, Siyue Li, Yaping Cheng, Di Weng, Chengfu Huo
https://t.co/fqCyyy1LJL [𝚌𝚜.𝙰𝙸]

[Movez @0xMovez]

Anthropic Head of Product:

“Fable 5 - is our best model for self-improving agentic systems. It can run for days on a single /goal.

add /loops, dynamic workflows, dreaming and you become unstoppable.”

in 11 minutes, the Anthropic team shows how to build long-running systems with Fable 5 from scratch.

Worth more than a $500 agent-building course.

Live from Anthropic’s latest stage in Japan. Unpublished.

[Lance Martin @RLanceMartin]

https://t.co/es0JQM4MS9

[高军 @GoJun315]

Anthropic 黑客松获奖者，把自己 10 个月重度使用 Claude Code 沉淀下来的整套打法，毫无保留地开源了出来。

ECC，一个面向 Claude Code、Cursor、Codex 等主流 AI 编程工具的 Agent 工程化框架，目前在 GitHub 上狂揽了 18 万 Star。

它不是简单一份���置模板，而是一整套把 AI 编程工具当成基础设施来用的完整系统：

- 60 个专用 Agent，覆盖规划、架构、TDD、代码审查、安全审计、构建排错等场景
- 232 个按需加载的 Skill，写代码、写文章、做投资材料、跑营销都有现成模板
- 内置 AgentShield 安全扫描，专门审计 CLAUDE.md、MCP 配置、hook、skill 里的安全风险
- 三个 Opus 4.6 跑红蓝对抗，攻击方找漏洞、防守方评估、审计方给出最终风险评级
- 连续学习层会自动把每次会话里有效的模式，沉淀成可复用的 skill
- 覆盖 TypeScript、Python、Go、Rust、Java、Swift 等 12 个语言生态

GitHub：https://t.co/UG9LsDURSX

可作为 Claude Code 插件一键安装，建议重度用 Claude Code 写代码的朋友使用。

也推荐研究一下这套配置思路，能学到很多把 AI 编程工具真正用出生产力的玩法。

[Yanhua @yanhua1010]

https://t.co/qIWuePdel2

[Max For AI @MaxForAI]

来自微软研究院的研究员@DimitrisPapail和@VaishShrivas

刚刚发布了一个全新的Agent训练方法，叫ECHO。

先说结论：
这可能是2026年Agent RL里最便宜、最暴力、也最会被抄进所有训练框架的方法之一。

它干的事非常反直觉：

让Agent学习世界给他的回声。

以前CLI agent训练，大概是这样：
模型敲命令，终端返回结果。
模型看一眼结果，继续敲下一条命令。

但训练的时候，GRPO主要只更新action token，也就是「模型自己敲了什么」。

至于终端返回的stdout、stderr、traceback、测试失败、文件列表、错误日志，很多时候只是上下文，用完就过去了。

Dimitris和Vaish觉得这很荒谬。

因为这些东西其实就是世界给agent��反馈。

你敲错命令，世界告诉你错在哪。
你跑测试失败，世界把traceback吐给你。
你ls一下目录，世界告诉你当前状态。
你改完代码再跑一遍，世界告诉你这次有没有修好。

这不就是程序员成长的全过程吗？

所以ECHO做的事简单到离谱：
别把这些terminal output浪费掉。

在原来的GRPO action loss之外，再加一个environment token prediction loss。

也就是让模型不光学习「下一步该怎么做」，还学习「我做完之后，世界会怎么回应」。

结果很夸张。

TerminalBench-2.0 pass@1：

8B模型：2.7→5.2
14B模型：5.2→10.8

训练到同等性能的速度最高快2.3倍。

最关键的是，这个提升几乎不加成本。

没有额外rollout、teacher model、新verifier。

只是把本来已经在轨迹里的终端反馈，重新变成训练信号。

过去agent跑错命令、测试失败、卡在报错里，可能只是一堆垃圾rollout。

但在ECHO里，这些报错、日志、traceback、文件状态，全部都能训练模型理解：

我做了这个动作，世界会给我什么后果。

这会改变agent训练的一个基本直觉。

过去我们总觉得agent要变强，需要更强teacher、更贵verifier、更多专家轨迹。

但其实只需要让agent在环境里行动，然后把环境的回声当老师就行了。

更离谱的是，他们还做了一个无reward实验。

关掉verifier reward，只让模型继续预测终端输出。

结果在一些任务上，模型居然还能继续涨。

这说明一件很恐怖的事：

只要环境反馈足够密，agent可能不需要每一步都有人告诉它对错。

它可以通过「行动→观察后果→预测后果」这条链路，慢慢校准自己的世界模型。

这就是为什么我觉得ECHO不是一个普通训练trick。

它更像是agent时代的一个底层信号发现：
世界每一次回应，都是训练数据。

所有报错、日志、用户反馈、网页变化、工具返回、API response，过去都只是agent的上下文。

以后它们都可能变成agent自我改进的燃料。

如果这个思路成立，未来最强的agent不一定只是prompt写得好、工具接得多、模型参数更大。

真正的差距会来自：

谁能更好地把「世界的反馈」变成训练信号。

我最喜欢篇blog里面的这句话：
The world is a loss function.

这句话我感觉会被反复引用。

[思维怪怪 @0xLogicrw]

OpenAI 后训练核心成员翁家翌（Jiayi Weng）以个人名义提出了一种名为「启发式学习」的强化学习新范式，并开源了全部实验代码。他用 Codex（GPT-5.4）反复玩 Atari 打砖块游戏，但 GPT-5.4 自始至终没有被重新训练过。真正在进步的，是 GPT-5.4 写出来的那套游戏策略代码。

流程是这样的：GPT-5.4 先写一版打砖块的 Python 策略，跑一局，���录像，找出哪里打丢了球，然后自己改代码再跑。经过几轮迭代，策略代码从 387 分涨到了 864 满分。全程没有任何神经网���被训练，纯靠 AI 反复修改 if-else 规则、调落点预测、加死循环检测。最终那套代码包含球路预测器、卡球检测器、回归测试和实验日志，已经长成了一个完整的软件系统。

这和传统强化学习的核心区别在于「学到的东西存在哪」。传统做法把知识压进神经网络参数里，人看不懂，学新任务还容易把旧的覆盖掉（即灾难性遗忘）。翁家翌的做法反过来：知识就是代码，人能读、能改、能加测试锁住，不会因为学新东西就丢了旧本领。

除了打砖块满分，他还在 MuJoCo Ant（模拟机器蚂蚁走路）上跑出超 6000 分的深度强化学习级成绩，在 Atari57 全套 57 个游戏上逼近了 PPO 基准。但翁家翌也明确画了边界：纯代码搞不定复杂感知任务，比如用 Python 写 if-else 去认图片。

他设想的终局是混合架构：底层用轻量神经网络负责视觉等感知，中层用启发式学习处理实时逻辑和安全规则，顶层由大模型审查日志、改代码，再周期性地用底层积累的高质量数据更新自身。过去手写规则之所以被淘汰，不是因为规则没用，而是人类维护不起。现在 AI 写代码够快够好，这条老路重新走得通了。

[CLS @ChengleiSi]

学LLM必看Tatsu

[梭哈.AI @SUOHA_AI]

非常值得刷的视频课--麻省理工《How to AI (Almost) Anything》

这门MIT公开课教你如何用AI（几乎）做任何事情——共12节课，覆盖大量专业领域 https://t.co/iqUHxw9NLI

大部分内容容易忘，建议把这份中英双语版讲义+图文笔记（社区整理）保存到NotebookLM / Claude / ChatGPT构建你自己的知识库，方便随时对话检索👇
https://t.co/vQ5FmN9Jde