breezedeus

Wow, this tweet went very viral! I wanted share a possibly slightly improved version of the tweet in an "idea file". The idea of the idea file is that in this era of LLM agents, there is less of a point/need of sharing the specific code/app, you just share the idea, then the other person's agent customizes & builds it for your specific needs. So here's the idea in a gist format: gist.github.com/karpathy/442a6… You can give this to your agent and it can build you your own LLM wiki and guide you on how to use it etc. It's intentionally kept a little bit abstract/vague because there are so many directions to take this in. And ofc, people can adjust the idea or contribute their own in the Discussion which is cool.

x.com/i/article/2034…

🧠💥今天英语科技圈彻底炸锅。 @eonsys 公布了一个极具历史节点意义的 demo： 他们用 12.5 万个真实神经元连线（Connectome），在虚拟世界里“复活”了一只果蝇。 一条截然不同的智能路线正在成型。 视频不是动画模拟👇 1️⃣ 到底发生了什么？ 研究人员把真实果蝇的大脑完整连接图，接入了基于 MuJoCo 物理引擎的虚拟躯体。 结果，这只“数字果蝇”开始自主走路、整理身体（grooming）、甚至进食。 形成了一个真正的闭环：感觉输入 -> 神经元网络放电 -> 动作输出。 2️⃣ 为什么这比大模型更让人震撼？ 过去几年，主流 AI 的核心路径是“拟合行为输出”（学它看起来像什么 / what it does），本质上是在拿海量数据训练一个像果蝇的策略网络。 但这次的逻辑变了。 研究者是从生物本体出发（重建它本身是什么 / what it is）。 没有任何行为数据的强化学习训练。 当真实的生物脑结构在代码中被精确跑起来之后，那些复杂的生物行为，就直接从神经连线里自然涌现出来了。 3️⃣ 它的终极意义 脑仿真真正跨过门槛的一刻，“技术奇点”不再是深度学习的专属。 它向我们证明了一件极其重要的事情： 一旦神经结构足够完整，行为本身就是可计算、可复现、可嵌入身体的。

x.com/i/article/2027…

Baidu just released a new VLM : PaddleOCR-VL-1.5🔥 @PaddlePaddle huggingface.co/PaddlePaddle/P… ✨ 0.9B - Apache 2.0 ✨ 94.5% OmniDocBench v1.5 ✨ Multilingual OCR: strong on rare characters & ancient texts

AI phones - large models or small models? We recently open-sourced OpenPhone📱— a 3B parameter mobile agent foundation model! After a year of trial and error, here's what we learned about AI phones ✨ Open-Sourced AI Phone Agents: github.com/HKUDS/OpenPhone 🤔 How do AI phones actually work? Simple: AI helps you operate your phone. But how does AI communicate with different apps? Option 1: API Calls 🔌 Ideally, we'd just call app APIs directly. Reality check — there are basically none! Big tech won't open their APIs because apps ARE their traffic moat. Building individual MCPs for each app? Engineering nightmare 💥 Option 2: GUI Interaction 🖱️ Since no APIs, let's do what humans do — look at screens and tap stuff. This approach is super generalizable, should work with any app. That's why most AI phones go the GUI Agent route now. GUI Agents are basically multi-modal models: - Input: screenshot + task description - Output: coordinates for next tap - Capability: screen understanding + task reasoning 📱 Three technical approaches for Phone Agents - Pure cloud ☁️ What most AI phones do currently — heavily rely on cloud-based large models. Performance is definitely better than small models, but privacy🔒 and cost💰 concerns are real. - Pure on-device models 📱 This is the direction OpenPhone is exploring. 3B parameters strikes a good balance — runs on phones, fast, private, and cost-effective. The trade-off is limited performance on complex tasks, given it's only 3B parameters. - Hybrid edge-cloud 🤝 Probably the most practical route. Simple stuff and anything privacy-sensitive stays on-device, complex reasoning hits the cloud. The trick is the routing strategy — when to make the switch? Interesting part is teaching the on-device model to recognize its own capability boundaries. 🔮 Some Random thoughts 1. GUI Agents still have plenty of issues: slow, error-prone, multi-app accuracy sucks. Rich MCP ecosystem would make life easier, but don't hold your breath. 2. Right now everyone's just collecting data, then SFT+RL to optimize models. Basically throwing data at the problem — hopefully we get smarter ways to do this. 3. AI phone ceiling isn't just tech — it's ecosystem. Future apps might go dual mode: APIs for agents, GUI for humans🚀 4. Computer-Use Agents are shifting toward coding — writing code instead of just clicking around💻, because code execution is way more accurate and efficient. Works great on desktop, mobile's still challenging. 5. Future Digital Agents might need to pack everything into one model: coding + multimodal + tool-use.

强烈推荐 Lenny @lennysan 的这个硅谷产品圈 Podcast 的原始音频文稿资料！！！含金量非常高，一共 320 份 .txt 的字幕原稿。 里面收录了非常多 AI 领域的名人访谈。例如： ✨ 李飞飞 Dr. Fei-Fei Li - AI 教母下个赌注：空间智能 ✨ Alexander Embiricos - OpenAI 内部视角 ✨ Anton Osika - 欧洲增长最快的 AI 公司，实战策略 （这几期链接我放在下方 threads 里面了） 我拿自己本地的 AI 学习工具，分析了原稿内容，生成了详细的调研报告（包含判断信息真伪、补充相关资料等）。文件资料包里包含了： 1. 嘉宾内容分析报告（.md .pdf） 2. 采访稿中英文双语原稿（.md .pdf） Markdown 的版本可以配合 AI 使用，PDF 则比较方便阅读。GitHub 两个格式都有 😎（贴心如我） GitHub 链接：github.com/Penny777btc/le…

EvoCUA 🔥 multimodal agent that can actually operate a computer, released by @Meituan_LongCat huggingface.co/collections/me… ✨ 8B & 32B - Apache2.0 ✨ Multi-turn control of Chrome, Excel, PowerPoint, VSCode ✨ 32B reaches 56.7% on OSWorld