wang @lv_roc

微软联合交大、同济、复旦发了一个框架叫 SkillOpt，思路很有意思：像训练神经网络一样训练 AI Agent 的技能文件 它不动模型权重，训练的是 Skill——就是你给 Claude Code 或 Codex 写的那些 prompt 和指导文档 做法是把神经网络训练的那套搬过来：轮次、批量大小、学习率、验证门控，全部套在自然语言上。执行任务 → 记录过程 → 复盘 → 修改 Skill → 验证效果，自动闭环迭代 结果很夸张。7个大模型、6类任务、3种 Agent 环境，总共52组测试，全部第一或并列第一 但最有价值的不是跑分，是它的迁移能力： 在 GPT-5.4 上优化好的技能文件，直接给 GPT-5.4-nano 用，性能提升 5.6 分 在 Claude Code 里优化好的技能，直接搬到 Codex 里用，提升 29.4 分 甚至能从一个数学基准迁移到另一个 一个 best_skill.md 文件，到处能用 对那些天天手动调 prompt 调到头秃的人来说，这个东西等于把调参这件事自动化了

学无止境，当你有了 AI 编程作为工具之后…

Yihui @yihui_indie

4 days ago

【完整爆肝教程】Codex + TanStarter 出海模板快速上线 AI 出海网站 ！

30 3 104 23K 142

You can now run Kimi K2.7 Code locally! 🌘 We shrank the 1T model to 325GB (-48%) via Dynamic 2-bit where important layers are upcasted. Run at >40 tok/s on 330GB RAM/VRAM setups. Run full precision on 610 GB. Guide: unsloth.ai/docs/models/ki… GGUF: huggingface.co/unsloth/Kimi-K…

Kimi.ai @Kimi_Moonshot

6 days ago

🌘 Kimi-K2.7-Code, our latest coding model, is now released and open-sourced! 🔷 Improved coding & agent performance over K2.6: +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite. 🔷 Reasoning efficiency: Less overthinking, with 30% lower

keyboard_arrow_left Previous keyboard_arrow_right Next

631 2K 14K 2.5M 3K

原来还能这么做：把 OpenAI、Anthropic、Google 等十几家 LLM 提供商的接口统一成一个，切换模型只改一个字符串就行。核心就靠 provider:model 路由加适配器，没有黑魔法，但开发体验直接从“翻文档”变成了“改个前缀”。这种抽象层思路比工具本身更值得琢磨。 github.com/andrewyng/aisu…

震惊了，抖音上祁博士一天卖 50w 的数字人 agent，我2 分钟就开发完成了。 用的就是Pixelle-Video这个项目，已经22k stars。包括数字人口播、动作迁移、图生视频全支持。 支持ComfyUI，输入主题，从写脚本到加 BGM 到出片，一条龙自动跑视频。 老王部署到本地，做了个简短的视频，属于插图式的视频，如果你需要更复杂的视频，需要自己配置云端模型比如 seedance2，kling 等等。 Pixelle-Video 最厉害的地方，是它把视频生产完全做成了可配置，支持本地部署模型和云端大模型。 文案、画面、配音、剪辑，它拆成四个可替换的模块，每块后面都能换模型，可以自由切换模型，使用非常方便 >文案层：LLM 读主题，吐出带时间戳的结构化脚本，每>句对应一段画面。 >画面层：脚本每句转生图提示词，扔给 ComfyUI 或直连 DashScope 出图，图生视频和数字人口播也走这一层。 >语音层：脚本原文走 TTS 合成，多语言加音色克隆，不用自己录音。 >合成层：画面对齐语音时间轴，叠上 BGM，输出 MP4。 仓库：github.com/AIDC-AI/Pixell… P.S. 想到了就能出片，懂一点 AI 编程，这个项目就能自己做成适合各行业的数字人 agent。

老王霸 AI Lab @laowangbabababa

2 weeks ago

x.com/i/article/2061…

41 198 682 681K 2K

Claude Code fully dissected! Researchers from UCL reverse-engineered the leaked Claude source. What they found changes how you should think about agent design. Only 1.6% of the codebase is AI decision logic. The other 98.4% is operational infrastructure. Permission gates, tool routing, context compaction, recovery logic, session persistence. The model reasons. The harness does everything else. This is the opposite of what most agent frameworks do today. LangGraph routes model outputs through explicit state machines. Devin bolts heavy planners onto operational scaffolding. Claude Code gives the model maximum decision latitude inside a rich deterministic harness, and invests all its engineering effort in that harness. The core loop is a simple while-true. Call model, run tools, repeat. But the systems around that loop are where the real design lives: A permission system with 7 modes and an ML classifier. Users approve 93% of prompts anyway, so the architecture compensates with automated layers instead of adding more warnings. A 5-layer context compaction pipeline. Each layer runs only when cheaper ones fail. Budget reduction, snip, microcompact, context collapse, auto-compact. Four extension mechanisms ordered by context cost. Hooks (zero), skills (low), plugins (medium), MCP (high). Each answers a different integration problem. Subagents return only summary text to the parent. Their full transcripts live in sidechain files. Agent teams still cost roughly 7x the tokens of a standard session. Resume does not restore session-scoped permissions. Trust is re-established every session. That friction is the point. The bet behind all of this is simple. As frontier models converge on raw coding ability, the quality of the harness becomes the differentiator, not the model. Paper: Dive into Claude Code (arXiv:2604.14228) We've shared an article on Agent Harness and what every big company is building. Read it below.

Akshay 🚀 @akshay_pachaar

2 months ago

x.com/i/article/2040…

74 522 3K 2.0M 8K

Kimi 2.7 ranked 2nd after Fable 5 and before GPT-5 xhigh We have re-run our ErdosBench smoke test on 14 problems with Kimi 2.7, Qwen 3.7 Max, Grok 4.3 and compared it with the top performers from previous runs. Kimi 2.7 is amazingly good. More below.

国内大厂手撕八股又领先五年

Xudong Han @Xudong07452910

7 days ago

现在顶级 AI 实验室的入场券，早就不只是有学术光环了！ 最近看到一篇很硬核的 ML 面试复盘文章，作者拿到了 DeepMind 等多家顶级 AI 公司的 offer，文章里面有个很现实的观察： 哪怕你手里有多篇 AI 顶会的一作，简历也只是把你送进面试间。

35 165 1K 132K 2K

x.com/i/article/2065…

苹果官方出的github库：apple/container 用Swift开发 ,专门给Apple芯片优化。 干一件事: 在Mac上用轻量虚拟机跑Linux容器 ,不再依赖Docker Desktop那一套笨重的东西。 本地起开发环境更快、更省内存 ,M系列芯片的Mac终于有个原生顺手的容器方案。

How do you give a code LLM knowledge of an entire repository without paying for it at every single query? We introduce Code2LoRA: a hypernetwork that turns a repository into its own LoRA adapter. Repo knowledge baked into weights → zero inference-time token overhead.

Code2LoRA seems an incredibly interesting idea. Qwen2.5-Coder-1.5B is not the most powerful LLM around, but it's enough to validate the concept. Instead of stuffing repository context into the prompt at every query, distill it into a LoRA adapter. One forward pass over the repo snapshot, one adapter, zero extra inference tokens. For evolving codebases, a single layer GRU tracks commit history on top of that snapshot. Each git diff updates the hidden state in <10ms. You get a fresh adapter at every commit without need for a full retraining. Great job Liliana! I bet this will lead to something cool in the near future 🙌

Liliana Hotsko @liliana_hotsko

a week ago

How do you give a code LLM knowledge of an entire repository without paying for it at every single query? We introduce Code2LoRA: a hypernetwork that turns a repository into its own LoRA adapter. Repo knowledge baked into weights → zero inference-time token overhead.

40 125 1K 164K 2K

Train your own LLM from scratch! A step-by-step repo that walks you through building and training a transformer model from scratch using PyTorch. From downloading training data all the way to generating text. The architecture is built from the ground up following the original "Attention is All You Need" paper. MLP, single head attention, multi-head attention, transformer blocks, and the full transformer model - all coded and explained with detailed diagrams at each step. Training data comes from The Pile - a diverse 825GB open-source dataset covering books, articles, code, websites, and more. The repo includes scripts to download it, preprocess and tokenize it using tiktoken, store it in HDF5 format, and feed it into training batches. You can train a 13M parameter model on a single Colab T4 GPU. At 13M parameters the model starts generating proper grammar and coherent short sentences. For billion-parameter training you need at least an A100 or RTX 4090. The repo includes a full GPU compatibility table so you know exactly what's possible on your hardware. Includes a complete SFT and RLHF guide as a separate notebook for taking your trained model further. Key capabilities: • End-to-end pipeline: data download → preprocessing → training → text generation • Full transformer implementation from scratch with PyTorch • Trains models from 13M to 2B+ parameters on a single GPU • Training data from The Pile (825GB, 22 diverse datasets) • Tokenization via tiktoken (r50k_base) • SFT and RLHF guide included 100% open source. I've shared the link in the replies!

🛠️ 开源框架推荐：《Agent Skills》—— Addy Osmani 出品，让 AI Coding Agent 真正像 Google 高级工程师一样写代码。 大多数 AI 编程工具最大的问题，不是「不够聪明」，而是太爱走捷径：跳过规格文档直接写代码、不写测试、不做安全审查、也不知道代码能不能直接 ship。结果就是「能跑，但不敢上线」。 Agent Skills 正是为了解决这个问题而生。它把 Google 软件工程文化（来自《Software Engineering at Google》的工程实践）直接编码成了 AI Agent 的行为约束，让 AI 在每个开发阶段都自动激活结构化工作流，而不是凭感觉随机应对。 核心特性 1. 7 个阶段性斜杠命令，覆盖完整开发生命周期： /spec（规格）→ /plan（规划）→ /build（构建）→ /test（测试）→ /review（审查）→ /code-simplify（简化）→ /ship（发布），还支持 /build auto 一键自主完成规划与实现。 2. 23 个结构化技能：每个技能都融入了 Google 高级工程师处理同类问题的系统方法、质量门控和「反捷径」机制。 3. 防走捷径设计：专门针对 AI 常见的「不写 spec 直接写代码」「不测试就提交」等系统性坏习惯进行约束。 4. 广泛兼容：Claude Code（强烈推荐）、Cursor、GitHub Copilot、Gemini CLI、Windsurf、OpenCode、Kiro IDE 等主流 AI 编程工具均支持。 5. MIT 开源：可自由扩展为团队专属的工程规范技能库。 特别适合：正在用 AI 工具做中大型项目、希望 AI 生成的代码达到生产级质量的工程师和团队 Lead。让你的 AI Agent 也拥有 Google 级别的工程素养，从现在开始。🚀 github.com/addyosmani/age… （已获 51.5k ⭐） #AIAgent #ClaudeCode #Codex #AIEngineering #Cursor #VibeCoding

哈哈哈哈哈没想到真的有人做了这个！

@

57 years ago

0 0 0 0 0

Here's a teaser of our Mac-1 model. > 6.6B model > runs locally (on any Mac) > requires 7GB RAM (12GB ideal) > can use 487 MacOS native tools > perform multi-tool chained tasks > reasoning: ON > output: ~65 tok/s We built a robust application layer around the model to make UI/UX MacOS native. The "model-focused" SaaS era is here. Stay tuned for more.

CJ Zafir @cjzafir

2 weeks ago

Our first model Mac-1 6.6B beating 3 giant models. - Haiku 4.5 - GPT 5.4 mini - Gemini 3 flash Running this model on my Macbook M3 24GB. (model takes only 7GB RAM) It searches web, call tools, ask follow-ups, tell jokes, find contacts, search files, write emails, book events,

162 108 2K 776K 1K

正式开源 html-video 🚀 html版剪映来了！ 你的 Agent 现在可以通过写 html轻松做出世界级水准的产品宣传、知识解说视频，成本极低！🔥 历时 3 天，3 万行代码！支持20多套顶尖视频风格模板，分页编辑，mp4 导出，支持包括Claude Code、Codex、hermes、cursor等主流 Agent接入即用💥地址见评论区

Shoutout to the open source projects behind this: • Serve-sim powers the streaming simulator by @Baconbrix github.com/EvanBacon/serv… • SnapshotPreviews extracts SwiftUI previews by @sentry github.com/getsentry/Snap…

Meet Gemma 4 12B! A unified, encoder-free multimodal model designed to bring high-performance intelligence directly to your laptop, and released under an Apache 2.0 license. Bridging the gap between edge efficiency and advanced reasoning. Here is what’s new with Gemma 4 12B: 👇

INSTEAD OF WATCHING AN HOUR OF NETFLIX TONIGHT. This 1 hour Stanford lecture by Joel Peterson will teach you more about negotiation and getting what you want than most people learn in years. Bookmark it and give it an hour, no matter what.

Rahul @sairahul1

3 weeks ago

x.com/i/article/2060…

13 34 185 237K 605