与「llm」相关的搜索结果

X AI ALGORITHM UPDATE (BIG ONE) 🚨 X updated their open source algorithm yesterday and the biggest number in it is gone: copy link went from 20 to 0 points. Small accounts should focus on: posts people like, reply to and stay on (so stop counting on link copies) tl;dr: - on Nov 13 X's open source code set the score for copying a post's link to 0 - copy link: 20 → 0 - it was the top score in the whole system, 4x a reply - the code also names a "copy link bait" spam check - likes (0.708), replies (5) and DM shares (5) didn't move full details: - The number that's gone Since I started reading X's open source code in August, one number sat at the top: 20, for someone tapping share and copying your post's link. Nothing else came close. As of the Nov 13 release, it's 0. X still guesses who'll copy your link. It just multiplies that guess by 0 now. X didn't explain it. It's just what the code says. (I check the open source algo daily with Claude and share updates with you) - Cross it off your list My own checklist called copy link the biggest number. Last week's post said it too. It was true then. It isn't now. If you saved that list, cross that line out. - The "copy link bait" check X's code now mentions a spam check called "copy link bait". In the code it adds a spam label to the post, and that label already keeps a post out of strangers' feeds. It only shows up in a test, so I can't see if it's running or how strict it is. My read: X saw people baiting link copies, so it turned the score off and built something to catch the bait. That's a read, the code doesn't say why. - Why that matters if you're small If your plan was posts people copy the link to, that plan now earns 0 in X's ranking. A copied link can still bring people in from outside X. It just doesn't push your post up anymore. What still pays: likes at 0.708 (the highest ever), people staying on your post, replies at 5 and back and forth with mutuals at 20. - Also new: photo and video tags If a post has a spam label, people you tag in its photo or video who don't follow you won't get notified. Same idea as last week's rule for replies. - What didn't change Every other score. Your new posts still get their extra push in the first 2 hours and first 200 views, and early likes still pick which one gets it. Here's a recap of what you should be doing: - aim for likes, replies and time spent, the copy link score is gone - a DM share still earns 5, same as a reply or a quote - likes sit at 0.708, never higher - time on your post keeps counting at its best rate yet - 10+ seconds after someone taps in adds 0.385 - write your replies yourself, under big accounts most of all - give people a reason to hang around your profile - in videos the opening seconds carry it, an open alone scores 0 - posts read to the end help X find you strangers - smaller accounts (below 50K) get a shot at a spot near 16 for a fresh original - early likes decide which new post gets that spot - post when your followers are awake, the push lasts 2 hours and 200 views - mutuals going back and forth earn 20, so follow back the regulars - only your originals ever get shown to strangers - like 1 and like 32 each unlock another way to reach strangers - one strong post per feed, your second only counts 62% - clips longer than 10 seconds for the video lists - a post likely to win you a follow gets +4 - give your pinned post a look, it can affect the whole account What you should NOT be doing: - counting on copy link shares, they score 0 now - baiting link copies, the code names a spam check for exactly that - a spam label also stops your photo tags reaching non-followers - AI-written replies under big accounts, "llm slop" is a flag X uses - copy-paste replies, labeled ones skip non-followers' notifications - chasing taps, 0.274 and dropping - drifting off your topic, each probable "not interested" takes 53.07 away - videos people open and leave, the open itself earns 0 - posting at 3am, it wastes the 2 hour push - expecting replies and reposts to reach strangers - an insult label, it hides that post from non-followers - earning a misinfo or low quality label, which adds a warning and locks strangers out of liking or sharing - 3 adult photos or videos among your latest 5 posts, that puts a label on you - forgetting videos get rechecked at 64 and 512 likes - one account label and strangers stop seeing you - pinning a link to a site X rates badly, that can mean a label for a week - posting your own old words again risks the copypasta flag - a low quality link plus tagged strangers reads as spam - bait for engagement, big or small it gets a label - several posts back to back, they split the same feed - hashtag piles, they do nothing - bookmarks for reach, they score 0 - hoping a 3 day old post still lands in For You, it stops at 2 days

很多人翻墙之后，只会刷推特 其实真正的信息差，藏在海外这些网站里。 下面这 10 个地方，建议全部收藏。每天花半小时逛一圈，信息获取能力可能超过身边 90% 的人！ 1、Reddit 互联网最大的兴趣社区之一。AI、投资、编程、创业、科技，几乎任何领域都有专属 Subreddit。真人经验、深度讨论、行业内幕，挖宝圣地！ 2、GitHub 全球最大的开源代码社区之一。AI Agent、LLM、大模型、自动化工具、开发框架，想找宝藏开源项目，这里简直挖不完！ 3、Hacker News 硅谷程序员的情报站。AI、开源项目、技术趋势、创业故事，每天都有高质量讨论，技术信息密度爆表！ 4、Product Hunt AI 工具和 SaaS 产品的发射场。每天都有新产品上线，很多爆火的 AI 项目都能在这里提前发现。 5、LinkedIn 千万别只把它当招聘网站。全球 CEO、投资人、创业者、科技公司高管都在这里分享一手行业动态。 6、Telegram 海外信息集散中心。AI、加密货币、投资、开源工具，各种频道和社群层出不穷，不少一手消息传播速度非常快。 7、Threads Meta 旗下社交平台。海外创作者、开发者、创业者聚集地，氛围比 X 更轻松，经常能看到有意思的新观点。 8、Discord 海外互联网社区的大本营。AI 项目、Web3、游戏、开发者社区都喜欢在这里建群，很多项目的核心讨论和早期动态都在 Discord。 9、Medium 海外技术和商业长文聚集地。AI 技术教程、创业复盘、投资思考、产品分析，适合系统性补充知识。 10、Substack 独立作者和行业专家的内容订阅平台。大量投资人、科技创业者、研究员在这里发布深度报告，有些内容比新闻媒体更有洞察力。 最后说一句： 不要把翻墙后的互联网，只当成一个刷 X 的娱乐工具。 Reddit 找讨论，GitHub 挖项目，Hacker News 看趋势，Product Hunt 找工具，Substack 追踪大佬观点。 这才是打开海外互联网的正确姿势！ 建议收藏，也转发给还在只会刷 X 的朋友。 你还有哪些值得推荐的海外网站？评论区补充！

一篇论文了解决策模型的训练《Towards Foundation Decision Models: Drex》 决策模型在很多场景比如分类、路由等任务上很有用，没有当时 Jev 刚出来吹的那么神，但是可以落地应用。 和传统的 BERT 比起来就是用 LLM 做backbone，知识容量和能力大幅提升。

一篇论文了解决策模型的训练《Towards Foundation Decision Models: Drex》 决策模型在很多场景比如分类、路由等任务上很有用，没有当时 Jev 刚出来吹的那么神，但是可以落地应用。 和传统的 BERT 比起来就是用 LLM 做backbone，知识容量和能力大幅提升。 话说这个论文的插图，我很怀疑是gpt生成的。

I think the best case is to make the "monitor" as un-agentic as possible. We don't mind being "monitored" by various random tiny insects we don't even notice, even guard dogs are fine, but we really do mind if it's a globally connected networked system. And the intermediate case, human security guards with no recording equipment, we mind only a medium amount. The best idea I've come up with so far is to make these monitoring devices out of open-source verifiable hardware that has etched on it an open-weight LLM that classifies for a limited number of activities (basically, just violence; sorry, no scanning for jaywalking or not wearing a mask or drugs or unauthorized manufacturing or illegal immigrants, really just violence) and only records and forwards if that's what it sees. The device has no storage, no ability to load new algorithms, this is all it does. Also, a hash of every recording is published onchain (and a device can't record the next one until it sees confirmation of the previous one) so we can all count how much recording is done. Finally, everyone has a legal right to take any such device they see and grab it and dissassemble and verify the hardware themselves. Explicitly fund university clubs, hacker groups, etc to regularly do this.

大家都知道LLM，那SLM呢？ 小语言模型的三种主流构建方法 1. 知识蒸馏（Knowledge Distillation） 基于“教师-学生”框架迁移大模型能力， 代表模型有 DistilGPT、TinyLLaMA。 2. 从头轻量化训练（Training from Scratch） 设计轻量架构并在中等规模语料上独立训练， 代表模型有 MiniBERT、ALBERT. 3. 模型压缩技术（Model Compression） 在现有架构上通过剪枝、量化 或低秩分解缩减体积并保留性能

Agent 最缺的不是更大的上下文窗口，而是理解整个代码库的能力：当一个项目有数万文件、跨模块调用和海量文档时，为什么还要让 AI 一遍遍 Grep、Read、重新理解整个工程？ 📦 项目档案 项目：Graphify-Labs/graphify GitHub Stars：⭐ 147299 技术栈：Python + Tree-sitter + NetworkX + Leiden 核心定位：将代码、文档、SQL Schema、PDF、图片及音视频转换成可查询、可追溯的知识图谱，让 AI Agent 从「搜索文件」升级为「理解工程关系」。 GitHub： 🛠️ 一、底层架构：不是传统 RAG，而是图结构知识引擎 Graphify 最值得研究的地方，是它没有走传统向量检索路线。 传统 RAG 通常需要经历： 文件切分 → Embedding → 向量入库 → 相似度检索 → 上下文拼接。 Graphify 直接构建 Graph。 底层采用 NetworkX 图数据结构，以 Node 和 Edge 表达工程实体及其关系。 整个处理链路可以概括为： detect() → extract() → build() → cluster() → analyze() → report() → export() 每个阶段职责独立，通过 Python Dict 和 NetworkX Graph 传递数据。 这意味着它不只是索引代码，而是在构建整个项目的结构化关系网络。 🛠️ 二、Tree-sitter AST：让代码解析摆脱 LLM 依赖 这是最有工程价值的设计。 Graphify 使用 Tree-sitter AST 解析约40种编程语言，提取： • Classes：类定义及结构。 • Functions：函数与方法。 • Imports：模块依赖关系。 • Call Graph：跨文件函数调用。 • SQL：表、视图、外键及 JOIN 关系。 其中33种语言使用 Tree-sitter Grammar，另外7种使用内置解析器。 关键在于：代码结构提取完全在本地完成，不需要调用 LLM，也不需要把代码上传到云端。 对于大型代码库，这种确定性解析比单纯依赖模型阅读文件更容易复现和追踪。 🛠️ 三、EXTRACTED / INFERRED：让每条知识关系都有证据 很多知识图谱最大的隐患是：模型推断出来的关系，很容易被误认为真实事实。 Graphify 给关系增加了置信度标签： • EXTRACTED：源码中明确存在的关系，例如 import、函数调用。 • INFERRED：通过上下文、命名和调用关系推断出来的连接。 • AMBIGUOUS：存在不确定性的关系，需要人工复核。 INFERRED 关系还会附带 0.0—1.0 的 confidence_score。 这套机制非常重要。 它实际上把知识图谱从一个简单的关联网络，变成了一个能够区分事实、推断与不确定性的工程知识系统。 🛠️ 四、Leiden + SHA256 Cache：解决规模与重复计算 Graphify 使用 Leiden 社区发现算法，根据节点之间的连接密度，将复杂工程拆分成不同功能子系统。 例如，一个大型后端项目，可以自动识别认证、数据库、路由、权限等不同社区。 同时，SHA256 内容哈希缓存会记录文件指纹。 文件没有变化，就跳过重复提取；只有新增或修改的文件才重新处理。 音视频还可以通过本地 faster-whisper 转录，再进入统一知识图谱。 这让增量更新不必每次重新扫描整个工程。 五、真正解决了什么工程问题？ 想象一个场景： 你接手一个陌生的微服务项目，需要搞清楚： 用户认证最终连接了哪些数据库模块？ 修改某个核心函数会影响哪些下游组件？ 某个架构决策为什么会存在？ 两个看似无关的模块究竟有什么隐藏依赖？ 传统方式：Grep、Read、跳转、反复询问 Agent，消耗大量上下文。 Graphify：直接执行 query、path、explain。 它还能生成三个核心产物： • graph.json：完整知识图谱，可供后续查询。 • GRAPH_REPORT.md：核心节点、社区及关键关系分析。 • graph.html：可交互的可视化图谱。 甚至支持 Mermaid 调用链导出、Markdown Wiki 和 Obsidian 知识库。 最重要的设计思想是： 不要让 AI 每次都重新阅读整个世界，而是提前把世界组织成它能够理解和遍历的结构。 这才是 AI Coding 从上下文堆砌走向工程知识管理的重要一步。 对于大型 Monorepo、遗留系统维护、跨模块依赖分析，以及需要长期理解代码库的 Agent，这套架构尤其值得研究。 🔗 GitHub 开源地址：

目前本地部署的 AI 服务栈基本凑齐了，链路如下： 🎙️ ASR：QwenASR + Whisper 🗣️ TTS：IndexTTS + BreezyTTS 🎵 音乐：Yue2 + MiniMax Music 🖼️ 图像：QwenImage 2.1 🧠 LLM：Qwen3.8 27B 🎬 视频：MiniMax H3 + LTX 2.5 从语音识别、语音合成、音乐、图像、大模型到视频生成，本地自建服务一条龙

党哥的调查结果： 1. blackwell series总共卖出300万片，其中100万片的算力在中国，中国当之无愧就是nvidia第二大客户； 2. 中国头部大模型7~8家，每家都有3~10万的GPU算力，外加十几家二线LLM厂商，neocloud，上百个电网和运营商自建data center，各地私有化部署，100万片只是一个非常保守的估计。

Multi-agent workflows have a scaling problem! The agents inside the same workflow rarely need the same amount of capacity at the same time. Say you have a planner, multiple research agents, and a final synthesis agent. The planner might finish quickly while requests start piling up behind one of the research agents. Scaling the entire application would add capacity everywhere, even though only one part of the workflow actually needs it. CanyonOS is an open-source, workflow-aware platform for multi-agent Python applications and the harness between your agents. Each agent runs as a separate service with its own queue and replica limits, so the busy research agent can get additional capacity while the planner and synthesis agents stay where they are. It also handles communication between the agents and gives you visibility into what happens across the workflow after a request comes in. CanyonOS records the path a request took, including the agents it called, parallel branches, model usage, execution time, and estimated model costs. And you don't need to add tracing decorators or a separate instrumentation SDK throughout your agent code. You still integrate the workflow with CanyonOS, but tracing is handled at the platform layer. I also ran a multi-agent portfolio analysis workflow through it with separate agents working across three stocks. A single request produced 10 traceable steps across the workflow, and I could see which agents ran, where the LLM calls happened, how long each stage took, and the estimated model cost. For multi-agent applications, being able to operate each agent independently while still seeing the complete execution path becomes much more useful as the workflow grows. CanyonOS is open source. Link to the GitHub repo in the comments!