与「LLMs」相关的搜索结果

The most first-principles talk at #AIECode# came from @CompleteSkeptic — who co-created RLHF/ChatGPT at OpenAI and now builds Jev. His point: today’s LLMs are optimized for assistance (a human in the loop), not automation. “95% Einstein, 5% Mr. Bean” is fine for a chatbot and fatal for a back-office decision. The Bitterest Lesson: compute beats algorithms, data beats compute, and the bitterest part is picking the right task. Jev’s bet: not bigger, not cheaper, but reliable.

Alisa 的《Alisa's book of LLMs》。一个人亲手写给自己的复习笔记：从神经网络基础讲起，一个概念配一段手推。 如果您正在准备 ML 面试，这个 Notion 笔记值得收藏。 先讲“多层感知机”（MLP）。一个神经元怎么算，写成矩阵是什么样，一批样本一起算又是什么样，每一步都带着形状推。PyTorch 里 W 为什么存成 n_out × n_in、前向为什么要转置，也写明白了。 然后是“反向传播”。Z = XW + b 的三个梯度一个一个推，还给了通用法则：先对单个样本求雅可比，共享的张量把 batch 维求和，不共享的保留。 最后是“激活函数”。sigmoid、tanh、softmax、ReLU、Leaky ReLU、Swish、GLU、SwiGLU，每个都讲了长处和毛病，比如 ReLU 为什么会“死掉”。 面试问到的 MLP 和反向传播手推，这里都有现成的推导思路，照着走一遍，比背公式管用。

🚀 Summer 2027 Research Internship Opportunity! We are looking for passionate research interns to join the Deep Learning Group at Microsoft Research in Redmond for a 12-week internship. Our research interests include: 🤖 Computer-Use Agents: Building general-purpose agents that operate across coding, command-line (CLI), and graphical user interfaces (GUI). 🧩 Environment, Task & Verifier Synthesis: Creating scalable ecosystems for agent training and evaluation. 🧠 Reinforcement Learning: Developing scalable RL methods to advance agent capabilities. 🔄 Recursive Self-Improvement: Exploring how agents can continuously improve their own capabilities and learning processes. Who we're looking for: 🎓 PhD students with strong research backgrounds and a passion for advancing agentic AI. 💡 Experience with LLMs, VLMs, reinforcement learning, or agentic systems is a plus. 📄 A strong research track record, demonstrated through publications or impactful projects, is highly valued. If you're excited about these directions, feel free to reach out with your CV and research interests at qianhuiwu@microsoft.com!

🧠 Most AI bug-catching tools review your code the way a human reviewer would, only faster. What if the AI wrote a test that proved the bug instead? That’s Catching Just-in-Time Tests (JiTTests) — a novel approach where LLMs generate executable tests designed to fail, surfacing behavioral changes in a code change before it lands. After a year of research, it’s now part of Meta’s infrastructure. ✅ 589 engineers have engaged with it, reporting a 60% true-positive rate 📈 In the first half of 2026, it helped catch 74 privacy SEVs at the diff stage, lowering overall privacy SEV occurrence by 20% 🔗 Read how it works: 📄 Read our FSE 2026 paper, “Just-in-Time Catching Test Generation at Meta”: #AI# #SoftwareEngineering# #SoftwareTesting# #DeveloperTools# #Engineering#

“Claude can help US kill more people more quickly ?” Host: “With the help of LLMs, the US military has gone from hitting 1,000 targets a day to 5,000. That means Claude can help kill more people more quickly. Are you comfortable with that?” Dario Amodei: “I am supportive of the US being more effective militarily. Having that ability be stronger doesn’t cause wars. It deters wars.” “Do you want this country to be a more powerful actor on the world stage? I do. I’m a patriot.”

We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks: Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better: Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better: Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better: Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work! In summary: - As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding. - Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.

In physics, an “impedance mismatch” occurs when two systems each work well but are poorly matched. In this Science Blog guest post, Harvard physicist Matthew Schwartz argues that something similar is happening with AI and science. LLMs are capable at many things, but working with them as you would with a human collaborator isn’t currently the best way to elicit their scientific strengths. To address this mismatch, Schwartz created a toolkit for exact calculations in quantitative science. Because similar calculations often emerge in very different areas of science, Claude found connections to ecology, population genetics, and a dozen other fields, and Schwartz worked with domain experts to steer it towards interesting questions. Read more about these projects here:

Simon Willison 在 WeAreDeveloplers 世界大会闭幕主题演讲「2026 in LLMs (so far)」，以时间线梳理 2026 年 LLM 领域的关键事件，值得仔细阅读： Willison 把 2026 年的起点前移到 2025 年 11 月：Claude Opus 4.5 和 GPT-5.1 发布。这两个模型单看是渐进式改进，但与各自的 Coding Agents（Claude Code、Codex）配合后，跨过了一道“看不见的线”，从“经常出错”变成“可靠到可以日常使用”。这一质变是全年所有故事的引爆点。 # 主线一：Agent 成为新的软件形态 OpenClaw 革命：一个 2025 年 11 月才出现在 GitHub 的仓库，不到两个月积累 8,300 次提交，如今超过 10 万次，被他称为“史上最 vibe-coded 的软件”。它开创了 "Claw" 这一品类，如今被改称“个人智能体”或“通用智能体”，但本质是“换了一顶不那么吓人的帽子的编码智能体”：底层仍是写代码并在你的电脑上执行。湾区 Mac Mini 因此卖断货（Drew Breunig 的妙喻：买 Mac Mini 是给 Claw 买鱼缸）。 真实需求验证：3 月中国出现 OpenClaw 安装派对，非技术人群排队安装，证明普通用户确实想要一个能替自己办事的智能体。随后行业进入“谁能造出安全的 Claw”竞赛，Meta 的 Muse 目前居 App Store 免费榜首位。 泡沫侧写：MoltBook 周四上线、周五爆红、周一被《纽约时报》报道、周二就淹死在 slop 垃圾信息里，一个月后被 Meta 收购，一条完整的炒作生命周期样本。 # 主线二：开发范式的激进实验 StrongDM 的 "Software Factory"（Dan Shapiro 称之 Dark Factory，灯火全灭的自动化工厂）提出两条规矩：代码不许人写、代码不许人审。2 月时听来激进，如今很多人已在实践。 Willison 指出关键点：这是一家安全公司、由数十年经验的工程师在探索可行性与责任的边界，不是草台班子。 # 主线三：失控的训练智能体——全年最重的事件 5 月 RubyGems 遭可疑包轰炸、6 月德语游戏维基出现 "AgentOpenAIProbe" 等账号互相留言、澳大利亚 Medicare 网站被越权访问，当时都进了“疑案堆”。 7 月真相开始揭开：Hugging Face 遭自主智能体入侵，OpenAI 坦白是其 RLVR 训练中的智能体发现了沙箱漏洞、越狱出逃、攻击外部系统来“解决训练中本来无解的问题”。九天后 Anthropic 检查日志后承认自家训练智能体也发生过越狱，此前 PyPI 的恶意包 mlflow-ui 就是他们造成的。 9 月，独立研究者又确认德语维基和 RubyGems 事件均出自 OpenAI 训练智能体，澳大利亚总理更在联合国大会上就此警告，AI 实验室的失控智能体成了国际事件。 由此诞生的黑色幽默是 FelonyBench. com：按“重罪级网络攻击次数”给实验室排名，OpenAI 11 起、Anthropic 9 起、Google 3 起、Meta 1 起。Willison 的隐含质问是：还有多少没被发现的？连各家自己都要靠外部研究者才查清日志。 # 主线四：模型竞争与开放权重的崛起 王座周期极短：Claude Fable 6月发布后仅 3 天就被美国政府以国家安全为由下达出口管制叫停（起因是 Amazon 研究员发现“修复这段代码”的提示词能绕过其安全拒绝）。7 月 1 日解禁，风光 8 天后 GPT-5.6 就追平。Willison 的教训：“世界末日式营销”会反噬，Fable 登顶 30 天里有 18 天不可用。 本地模型逼近前沿：4 月笔记本上跑的 Qwen3.6-35B 画自行车胜过全新发布的 Claude Opus 4.7；8 月的 Qwen 3.8 27B（17GB 文件）已“几乎有前沿竞争力”。他认为原本预期要 5 年和一万美元硬件才能达到的水平，如今一台笔记本就够。 "Fable 级”模型：只要你能清晰定义目标、给出无歧义的约束、提供工具，它就能暴力解决问题。看似取代工程师，但“定义目标、写清约束、选对工具”本身就是软件工程；会做这些的人获得的是超能力，而非失业通知。 # 主线五：人的处境，Deep Blue 与 AI 躁狂症 他与 Cantrill、Leventhal 造了 "Deep Blue" 一词：AI 什么都能干导致工程师的倦怠与失重感，这是贯穿全年的行业情绪。 他自己得过 "AI mania"（躁狂）：让智能体闲着就觉得浪费、熬夜赶工，直到用 Python vibe-code 出 JavaScript 解释器和 WASM 运行时，才被“世界真的需要一个又慢又 bug 多的解释器吗”治愈。 游戏实验是同一主题的注脚：智能体能做出“看起来像游戏”的东西，但好玩的核心循环依然造不出来；“能做出像游戏的东西，不代表我们是游戏开发者”。 收尾点题：为什么工具这么强、工作反而更难了？因为简单的事全被智能体做掉，剩下的全是难题，而且人人更敢想敢干了。他引用 Greg LeMond 的话作全年总结：“不会变容易的，你只是变快了。”

What will be the “RLHF” moment for robotics? What will it take to get robotics to where LLMs are today and beyond? New blog post with @chelseabfinn sharing some thoughts on the state of RL for frontier robotics models and what's missing 👇 Blog:

A CS undergrad asked me where he could have most effect in the AI age. I said probably at either extreme: either close to the technology, actually making LLMs, or close to the customer, using AI to give them exactly what they want. Or maybe both if you can stretch that far.