Blum
Alexander Embiricos, product lead for dots at OpenAI: "That's one key idea there, persistence, and it's probably the most important idea." In this interview he explains why a dot gets its own cloud computer and keeps one context across Slack and voice calls. 45 mins with the person in charge of dots. No recap gives you more. Watch the interview first, then read the guide below to set up your own dot in 10 steps.
🎬
视频
#585 🔗 原推
2026-10-04 20:58:48
Blum
Stop asking AI to design your whole app in one prompt. That's how you end up with the same app as everyone else. Paul Bakaus (ex-Google, creator of a 75K-star design skill) explains what to do instead in a 16-min talk. 05:58 - why you can't one-shot design 07:16 - 4 questions to answer before designing 08:10 - steering the agent with words like "bolder" 12:02 - the design workflow, from shaping to polish 14:09 - why taste can't be automated People pay for design courses to learn this. This talk is free. Watch it first, then read the article below and get rid of the AI look in your products.
中文: 停止要求人工智能在一个提示下设计你的整个应用。 这就是你和其他人一样使用同一个应用程序的方式。 保罗·巴考斯(前谷歌，75K星设计技能的创造者)在一段16分钟的演讲中解释了该怎么做。 05:58 - 为什么你无法单枪一击 07:16 - 设计前需要回答的4个问题 08:10 - 用“bolder”等词引导代理 12:02 - 从塑形到抛光的设计工作流程 14:09 - 味觉无法自动化的原因 人们付费学习这门设计课程。这个谈话是免费的。 先观看，然后阅读下面的文章，并清除产品中的AI外观。
🎬
视频
#584 🔗 原推
2026-10-04 16:05:00
Blum
Harrison Chase: "Create your private evals, because eval defines what good looks like inside the organization." In 24 mins he walks through how LangChain builds the private evals behind its own agents. You will learn more from that single idea than from most paid courses on agent evaluation. Watch the talk, then read how that same thinking cut LangChain's median thread cost by 64%.
中文: 哈里森·蔡斯: 创建你的私人椭圆形，因为 eval 定义了组织内部的好看外观。 24分钟后，他介绍了朗查因如何在自身代理人身后建造私人椭圆形雪道。 你将从这一理念中学到比大多数付费课程在代理评估方面更多。 观看演讲，然后阅读同样的想法如何使LangChain的中位线程成本降低了64%。
🎬
视频
#583 🔗 原推
2026-10-04 13:00:10
Blum
The two Anthropic engineers who created Agent Skills say they stopped building AI agents. In this 16-min talk, they explain why and what they build instead. 01:20 - why each field doesn't need its own agent 02:56 - a skill is just a folder of files 08:08 - MCP for connections, skills for expertise 10:37 - plans for skill tests, versions and dependencies 12:02 - a skill library the whole team shares I've seen paid AI courses that teach less than this. Watch it first, then read the article below on setting this up for your own research or business.
中文: 两名创建Agent Skills的人类工程师表示，他们停止了人工智能的构建。 在这场16分钟的演讲中，他们解释了原因以及他们构建的内容。 01:20 - 为何每个字段不需要自己的代理 02:56 - 一项技能只是文件的文件夹 08:08 - 人际关系、专业技能的MCP 10:37 - 技能测试、版本和依赖项的计划 12:02 - 整个团队共享的技能库 我见过付费的人工智能课程，教授的比这多。 先观看，然后阅读以下关于为自身研究或业务设置此内容的文章。
🎬
视频
#582 🔗 原推
2026-10-03 18:56:44
Blum
@OpenAIDevs But does it actually keep context across apps, or is that just the pitch?
中文: @OpenAIDevs 但它真的会在应用程序中保留上下文，还是只是宣传?
#581 🔗 原推
2026-10-02 22:41:32
Blum
Sam Altman: "Dots are remarkably capable, always-on agents built to handle really anything you can think of." The DevDay keynote explains how each dot gets its own cloud computer, its own browser, and access to more than 4,000 connected apps. On its own, it beats a $500 AI course covering the same ground. Watch the keynote first, then read the guide below to build your own agent team.
中文: 萨姆·阿尔特曼: 圆点功能非常强大，始终能随时使用，能够真正满足任何你能想到的事物。 DevDay 主题演讲解释了每个点如何拥有自己的云计算机、自己的浏览器，以及超过 4000 个已连接的应用程序。 它独自击败了覆盖同一领域的500美元人工智能课程。 先观看主题演讲，然后阅读以下指南，打造自己的代理团队。
🎬
视频
#580 🔗 原推
2026-10-02 13:59:29
Blum
RT @Blum_OG: Stop handing your AI bot full access. In a 31-min workshop, a SpaceXAI engineer shows how to stop babysitting Grok Bot without risking your customers or your money. 04:46 - from reading tickets to answering customers 09:19 - start with one bot and one workflow 14:27 - when the bot isn't sure: handoff to a human 16:42 - refunds in Stripe with a 14-day rule 28:07 - read-only first, then writes with approval I've seen paid AI courses that teach less than this. Watch it first, then read the article below on hiring your first AI employee.
中文: RT @Blum_OG:停止将你的AI机器人完全访问。 在一个31分钟的研讨会上，一位SpaceXAI工程师展示了如何在不冒客户或资金风险的情况下停止照顾婴儿机器人。 04:46 - 从看票到回答顾客 09:19 - 从一个机器人和一个工作流程开始 14:27 - 当机器人不确定时:交给人类 16:42 - 在Stripe退款，规定14天 28:07 - 先读，然后经批准写作 我见过付费的人工智能课程，教授的比这多。 先看，然后阅读下面关于聘用你第一位人工智能员工的文章。
🎬
视频
#579 🔗 原推
2026-10-01 12:31:16
Blum
Stop handing your AI bot full access. In a 31-min workshop, a SpaceXAI engineer shows how to stop babysitting Grok Bot without risking your customers or your money. 04:46 - from reading tickets to answering customers 09:19 - start with one bot and one workflow 14:27 - when the bot isn't sure: handoff to a human 16:42 - refunds in Stripe with a 14-day rule 28:07 - read-only first, then writes with approval I've seen paid AI courses that teach less than this. Watch it first, then read the article below on hiring your first AI employee.
🎬
视频
#578 🔗 原推
2026-10-01 08:48:16
Blum
@GoogleDeepMind https://twitter.com/Blum_OG/status/2105553572549316891/photo/1
#577 🔗 原推
2026-10-01 07:01:05
Blum
OpenAI engineers no longer fix dozens of bugs a day. Their AI agents do. Sam Altman and the OpenAI team show exactly how in 54 minutes. 05:55 - dots taking on bug fixes and API migrations 10:16 - a dot moving a launch review and reading test feedback 12:43 - tagging a dot in a Space page to redo a chart 19:39 - GPT-6.1 Sol: near-Astra intelligence at a fifth of the price 21:04 - Ultrafast and regular Astra building the same rocket 38:51 - Codex clicking through an app on every screen size This free video explains more than most paid AI courses. Watch it first, then read the article below with ready-to-use prompts for your own dot.
🎬
视频
#576 🔗 原推
2026-09-30 20:48:50
Blum
@DeRonin_ @MiaAI_lab @TheAhmadOsman @jun_song @ggerganov @danielhanchen already follow a few of these and they're seriously good, gonna follow the rest now too
中文: @DeRonin_ @MiaAI_lab @TheAhmadOsman @jun_song @gerganov @danielhanchen 已经关注其中一些，他们非常优秀，现在也会关注其余部分
#575 🔗 原推
2026-09-30 20:10:08
Blum
@ollama @bespokelabsai @madiator https://twitter.com/Blum_OG/status/2105389174874616182/photo/1
#574 🔗 原推
2026-09-30 20:07:50
Blum
@milesdeutscher https://twitter.com/Blum_OG/status/2105383153926103097/photo/1
#573 🔗 原推
2026-09-30 19:43:54
Blum
Harrison Chase, LangChain co-founder: "The main job of a harness is to bring context to the model at the right point in time." In under 24 mins, he explains agent loops and how to test them. This free video is worth more than most paid AI agent courses. Watch the talk, then follow the guide below to give your agent tools, set spending limits and test its results.
中文: 哈里森·蔡斯，朗查恩联合创始人: 安全带的主要工作是在正确的时间点为模型带来背景。 在不到24分钟的时间里，他讲解了代理循环以及如何测试它们。 这个免费视频比大多数付费AI代理课程更值得。 观看演讲，然后按照以下指南提供您的代理工具，设定支出限额并测试其结果。
🎬
视频
#572 🔗 原推
2026-09-29 20:40:32
Blum
@OpenAIDevs Looks like my evening’s booked testing this beast https://twitter.com/Blum_OG/status/2105012837659930720/photo/1
#571 🔗 原推
2026-09-29 19:12:24
Blum
@claudeai can’t wait to test it on my tasks https://twitter.com/Blum_OG/status/2104645587501187147/photo/1
中文: @claudeai 迫不及待想在我的任务中进行测试
#570 🔗 原推
2026-09-28 18:53:05
Blum
Stop paying for AI automation courses until you've watched this. TypeSafe's CEO spends 18 mins explaining why AI learns to give answers people like, even when those answers are wrong. 04:29 - AI assistants vs automation 06:15 - what RLHF rewards 07:19 - when ChatGPT gives flattering feedback 11:41 - automating small, repetitive tasks 16:19 - training for calibrated decisions This video teaches more than courses people routinely pay hundreds of dollars for. Watch first, then read article below on making Claude Code loops cheaper with Jev.
🎬
视频
#569 🔗 原推
2026-09-28 18:32:04
Blum
Diogo Almeida, TypeSafe CEO: "Don't have AI make decisions with stakes." His talk explains why LLMs trained on human preferences can sound right while getting decisions wrong. You'll get more practical knowledge from this talk than from a paid AI course. Watch the talk, then use the Jev guide below to set confidence thresholds and route uncertain cases to a human.
🎬
视频
#568 🔗 原推
2026-09-27 16:17:23
Blum
If you have to approve every tiny step, your AI agent has become another job. LangChain's CEO lays out a better way to delegate work to agents in about 21 mins. 05:51 - fixed step order in code 10:36 - undoing an agent's code changes 13:34 - agents triggered by events 15:50 - approving and correcting tool calls 17:25 - Agent Inbox for human review I'd put this ahead of plenty of paid agent tutorials. Watch the video first, then read the article below for the architecture of Company Brain.
中文: 如果你必须批准每一个微小的步骤，你的人工智能代理已经成为另一项工作。 朗任首席执行官提出了一种更好的方式，可在大约21分钟内将工作委托给经纪人。 05:51 - 代码中的固定步骤顺序 10:36 - 撤销代理代码的更改 13:34 - 事件引发的特工 15:50 - 审批和更正工具调用 17:25 - 代理收件箱供人工审核 我会把这个放在大量付费代理教程之前。 先观看视频，然后阅读以下关于公司大脑架构的文章。
🎬
视频
#567 🔗 原推
2026-09-27 10:19:23
Blum
Diogo Almeida, TypeSafe CEO: "I care about intelligence per dollar." Jev's creator explains why cost can matter more than speed when analyzing large datasets. In just over 2h, he shows how to split AI tasks into small decisions you can test before putting them into a product. Watch the interview, then use the article below to turn an idea into a working product with AI.
中文: 迪奥戈·阿尔梅达，TypeSafe 首席执行官: 我每美元都关心情报。 杰夫的创造者解释了为何在分析大型数据集时，成本比速度更重要。 在超过2小时时，他展示了如何将人工智能任务划分为可以测试的小决策，然后再将其放入产品中。 观看采访，然后使用下面的文章将一个想法转化为人工智能的工作产品。
🎬
视频
#566 🔗 原推
2026-09-26 10:21:06
Blum
Stop guessing whether your agent is getting better. LangChain's co-founder shows how to compare agent setups and use recorded runs to find problems. 03:44 - hooks around model and tool calls 06:55 - when to build a custom harness 12:29 - agent accuracy, latency, and cost 13:04 - debugging failed agent runs 17:29 - suggested fixes from agent traces This 24-min video is worth watching before you spend another week fixing the wrong part of your agent. Watch this first, then read the guide below on building a harness with Jev.
中文: 别再猜到你的经纪人是否正在好转。 朗尚的联合创始人展示了如何比较代理设置，并利用录制的运行记录来发现问题。 03:44 - 围绕模型和工具调用的钩子 06:55 - 何时制作定制线束 12:29 - 代理准确性、延迟和成本 13:04 - 调试失败的代理运行 17:29 - 建议从代理痕迹中修复 这段24分钟的视频值得一看，之后再花一周时间修复你代理的错误部分。 先看这个，然后阅读下面关于与杰夫一起建造安全带的指南。
🎬
视频
#565 🔗 原推
2026-09-25 19:26:31
Blum
Boris Cherny: "If Claude has a target to iterate against, it can do much better." He shows how tests and screenshots let Claude inspect its own output. In 18 mins, Cherny covers tools, planning, tests, and project instructions. Spend your time on this before spending a f*cking fortune on another course. Watch his explanation, then use the article below to set up your weekly review.
中文: 鲍里斯·切尔尼: 如果克劳德有目标可以抵御，它可能会做得更好。 他展示了测试和截图如何让克劳德自行检查其输出结果。 18分钟内，切尔尼将涵盖工具、规划、测试和项目说明。 花点时间，再花点钱去另一条赛道。 请观看他的解释，然后使用以下文章来设置你的每周评价。
🎬
视频
#564 🔗 原推
2026-09-25 15:13:20
Blum
@atoms_res @hifiHQ @andocorporation Ando raising $20M for messaging between humans and agents is interesting But $72M across three projects seems like a lot for mostly early access forms
中文: @atoms_res @hifiHQ @andocorporation Ando 为人类与特工之间的通讯筹款筹集了2000万美元，这很有趣 但在三个项目中，7200万美元似乎对早期访问形式来说似乎很多
#563 🔗 原推
2026-09-25 11:59:02
Blum
Diogo Almeida, TypeSafe CEO: "You get what you optimize for." His 36-minute talk traces the gap between helpful chat and reliable automation. This gives you a clearer understanding of training goals than many paid AI courses. Watch the talk, then use Together's guide to train your own classifier for about $17 in training costs.
中文: 迪奥戈·阿尔梅达，TypeSafe 首席执行官: 你得到你优化的所有内容。 他长达36分钟的演讲，记录了实用聊天与可靠自动化之间的差距。 这让你比许多付费人工智能课程更清楚地了解培训目标。 观看演讲，然后使用“携手”指南，以约17美元的培训费用训练自己的分类器。
🎬
视频
#562 🔗 原推
2026-09-24 19:08:22
Blum
Stop paying premium prices for every coding task. In under 16 mins, DigitalOcean engineers show how to route coding tasks to different models and check what happens to quality, speed, and cost. 02:48 - small models vs frontier models 04:51 - a separate model for routing 06:57 - models for bug fixes, code, and tests 09:07 - router vs Opus: quality and speed 09:41 - OpenCode demo: app, tests, and README I've seen $500 courses that teach less than this. Watch this first, then read article below on experimenting with cheaper models in Codex and choosing the right model for each task.
中文: 停止为每一项编程任务支付高价。 在16分钟内，DigitalOcean的工程师将编程任务路由到不同的模型，并检查质量、速度和成本的情况。 02:48 - 小型模型与前沿模型 04:51 - 用于路由的独立模型 06:57 - 用于错误修复、代码和测试的模型 09:07 - 路由器与Opus:质量和速度 09:41 - OpenCode 演示:应用程序、测试和 README 我见过500美元的课程，教的比这还少。 先观看，然后阅读以下文章，了解在Codex中尝试使用更便宜的型号，并为每项任务选择合适的型号。
🎬
视频
#561 🔗 原推
2026-09-24 13:30:00
Blum
Greg Brockman (OpenAI president): "Astra has really hit something that I'm like, okay, I think this is pretty reasonable to call it AGI." He explains what Astra can do and why the right tools and context matter for getting useful work done. This helps you decide what to hand over to AI and what you still need to check yourself. Watch the video, then use the article below to split a project into clear tasks with checks for completion.
🎬
视频
#560 🔗 原推
2026-09-24 09:22:00
Blum
Diogo Almeida (Jev creator): "System messages are like disgusting global variables." He explains how to split AI workflows into decisions you can test separately. This one video is more useful than many AI courses people pay for. Watch the breakdown, then use the article's prompts to test Jev's relevance checks on videos from your niche.
中文: 迪奥戈·阿尔梅达(杰夫 创作者): 系统消息就像令人厌恶的全局变量。 他解释了如何将人工智能工作流程划分为可单独测试的决策。 这个视频比人们付费的许多人工智能课程更有用。 观看细分，然后使用文章的提示来测试Jev对来自您细分市场的视频的相关性检查。
🎬
视频
#559 🔗 原推
2026-09-24 05:20:49
Blum
LangChain co-founder Harrison Chase: "There's a harness that orchestrates a model and some context." In under 24 minutes, he explains the loop behind tool use and context. That helps you understand what changes when you switch runtimes. I'd put this free talk ahead of many paid courses on agent basics. Watch the talk, then read the article to build products that work with different agent harnesses.
中文: 朗查恩联合创始人哈里森·蔡斯: 有一种线束可以编排模型和一些背景内容。 在不到24分钟内，他解释了工具使用和上下文背后的循环。这有助于你了解在切换运行时哪些变化。 我会把这篇免费演讲提前到许多关于经纪人基础课程的付费课程之前。 观看演讲，然后阅读文章，构建出可与不同代理设备同工的产品。
🎬
视频
#558 🔗 原推
2026-09-22 18:32:37
Blum
Stop explaining your preferences to your AI agent over and over. LangChain's co-founder shows how to turn user feedback into saved instructions an AI agent can use next time. 06:52 - immediate and background memory updates 31:27 - tools for saving and searching memories 46:12 - correcting an outdated memory 53:00 - sorting emails using past examples 1:11:54 - updating agent instructions from feedback A workshop like this shouldn't be available for free. Watch the workshop first, then read the article below for a closer look at Instinct's memory.
中文: 停止向你的人工智能代理解释你的偏好。 LangChain 的联合创始人展示了如何将用户反馈转化为人工智能代理下次可以使用的保存说明。 06:52 - 即时和后台内存更新 31:27 - 保存和搜索记忆的工具 46:12 - 纠正过时的记忆 53:00 - 使用过前实例整理邮件 1:11:54 - 根据反馈更新代理说明 这样的研讨会不应该免费提供。 先观看研讨会，然后阅读下面的文章，以更仔细地了解Instinct的记忆力。
🎬
视频
#557 🔗 原推
2026-09-21 20:35:00
Blum
Stop paying for AI automation courses until you've watched this. The founder of TypeSafe explains why AI gets things wrong and what it needs to work reliably. 04:29 - assistance and automation 06:15 - what RLHF actually rewards 08:22 - why wrong answers sound convincing 11:54 - automating small, repetitive tasks 16:19 - calibrated decisions for software I've seen $500 courses that teach less than this. Watch this, then read the article below on how to use Jev correctly.
中文: 停止支付人工智能自动化课程的费用，直到你看过这个为止。 TypeSafe 的创始人解释了为什么人工智能会出错，以及它需要什么才能可靠地工作。 04:29 - 协助与自动化 06:15 - 什么 RLHF 实际奖励 08:22 - 为什么错误的答案听起来令人信服 11:54 - 自动化小型重复性任务 16:19 - 软件经过校准决策 我见过500美元的课程，教得比这少。 观看本文，然后阅读下面关于如何正确使用Jev的文章。
🎬
视频
#556 🔗 原推
2026-09-21 17:30:00
Blum
Cormac Brick (Tech Lead at Google AI Edge): “So you can get really robust and reliable function calling using this fine-tuning workflow.” He explains how fine-tuning helped a small AI model reach over 90% success on specific app tasks. Watch the talk, then use the article below to train on your own reviewed examples.
中文: 科马克·布里克(谷歌AI Edge科技领先) 因此，通过这种微调工作流程，你可以获得非常稳健且可靠的功能调用。 他解释了微调如何帮助小型人工智能模型在特定应用任务上取得超过90%的成功。 观看演讲，然后使用以下文章，根据您自己评过的示例进行训练。
🎬
视频
#555 🔗 原推
2026-09-21 12:42:40
Blum
GPT-4 co-author Diogo Almeida: "We are doing a third thing that is optimized for calibrated decision-making" In 18 minutes, he explains why models trained for human approval can be overconfident. You’ll get more out of this video than you would from a $700 course. Watch the talk, then read the guide below to learn what Jev is good at and how to use it.
中文: GPT-4 的合著者迪奥戈·阿尔梅达: 我们正在做第三件针对经过校准决策而优化的事情 在18分钟内，他解释了为何经过人类训练的模型可能过于自信。 你从这个视频中得到的比从700美元课程中得到的更多。 观看演讲，然后阅读下面的指南，了解捷文夫擅长什么以及如何使用。
🎬
视频
#554 🔗 原推
2026-09-20 17:56:43
Blum
@atoms_res @BITFOOTS_ 6x on ZEC is a solid carry while the rest of the book bleeds At least one position is doing the heavy lifting
中文: @atoms_res @BITFOOTS_ 在 ZEC 上是一次可靠的携带，而书的其余部分则在流血 至少有一个职位正在承担繁重的工作
#553 🔗 原推
2026-09-20 15:55:31
Blum
Andrej Karpathy: "The hottest new programming language is English." In under 40 minutes, he explains how LLMs have revolutionized what we can do in the digital environment and how to use them effectively. This talk contains much more useful info than the $500 course. Watch the full talk, then read the article below to set up a working harness for your agents.
🎬
视频
#552 🔗 原推
2026-09-20 13:32:38
Blum
@atoms_res @Starknet @playcambria @opensea 7 chances to grab something Checking Starknet first
中文: @atoms_res @Starknet @playcambria @opensea 7 获取机会 先检查Starknet
#551 🔗 原推
2026-09-20 09:01:22
Blum
a plain yes-or-no decision inside agent can cost up to 400x more than it has to Harrison Chase, LangChain's CEO, spent 47 mins on what's actually inside an agent's harness 08:27 - what actually counts as a harness 18:14 - the building blocks of an agent 29:51 - why agents need a sandbox 32:37 - sandboxes against prompt injection 41:51 - the mess of evaluating agents watch this first, then read the article below on the model built to handle that decision for a fraction of the price
中文: 一个普通的“是”或“否”决定，其价格可能比实际成本高400倍 朗查因公司首席执行官哈里森·蔡斯花了47分钟，在经纪人的安全带内实际花费了大量资金 08:27 - 真正算作线束的东西 18:14 - 代理人的积木 29:51 - 为什么代理需要一个沙盒 32:37 - 沙盒与及时注射 41:51 - 评估代理人的混乱 先看这个，然后阅读下面关于该模型的文章，该模型旨在以一小部分价格来应对这一决定
🎬
视频
#550 🔗 原推
2026-09-19 20:00:00
Blum
a plain yes-or-no decision inside agent can cost up to 400x more than it has to Harrison Chase, LangChain's CEO, spent 47 mins on what's actually inside an agent's harness 08:27 - what actually counts as a harness 18:14 - the building blocks of an agent 29:51 - why agents need a sandbox 32:37 - sandboxes against prompt injection 41:51 - the mess of evaluating agents watch this first, then read the article below on the model built to handle that decision for a fraction of the price
中文: 一个普通的“是”或“否”决定，其价格可能比实际成本高400倍 朗查因公司首席执行官哈里森·蔡斯花了47分钟，在经纪人的安全带内实际花费了大量资金 08:27 - 真正算作线束的东西 18:14 - 代理人的积木 29:51 - 为什么代理需要一个沙盒 32:37 - 沙盒与及时注射 41:51 - 评估代理人的混乱 先看这个，然后阅读下面关于该模型的文章，该模型旨在以一小部分价格来应对这一决定
🎬
视频
#549 🔗 原推
2026-09-19 20:00:00
Blum
a plain yes-or-no decision inside agent can cost up to 400x more than it has to Harrison Chase, LangChain's CEO, spent 47 mins on what's actually inside an agent's harness 08:27 - what actually counts as a harness 18:14 - the building blocks of an agent 29:51 - why agents need a sandbox 32:37 - sandboxes against prompt injection 41:51 - the mess of evaluating agents watch this first, then read the article below on the model built to handle that decision for a fraction of the price
中文: 一个普通的“是”或“否”决定，其价格可能比实际成本高400倍 朗查因公司首席执行官哈里森·蔡斯花了47分钟，在经纪人的安全带内实际花费了大量资金 08:27 - 真正算作线束的东西 18:14 - 代理人的积木 29:51 - 为什么代理需要一个沙盒 32:37 - 沙盒与及时注射 41:51 - 评估代理人的混乱 先看这个，然后阅读下面关于该模型的文章，该模型旨在以一小部分价格来应对这一决定
🎬
视频
#548 🔗 原推
2026-09-19 19:55:31
Blum
Andrej Karpathy: "You can outsource your thinking, but you can't outsource your understanding." In this 29-minute talk, he explains why an agent still needs a person directing it. It reveals more useful tips than most paid courses about agents. Watch the talk first, then read the article below for the method that puts it into practice.
中文: 安德烈·卡帕蒂: 你可以将思维外包出去，但无法将理解外包。 在这场29分钟的演讲中，他解释了为何经纪人仍然需要一个人来指导。 它比大多数付费课程都揭示了更多关于代理的实用建议。 先观看演讲，然后阅读下面的文章，了解将其付诸实施的方法。
🎬
视频
#547 🔗 原推
2026-09-19 16:14:03
Blum
@atoms_res @ChumpCoinX @thinkingcatRH @schiffygld $BONER at $58.6M is the clear outlier here That gap to the next one is huge
中文: @amos_res @ChumpCoinX @thinkingcatRH @schiffygld $BONER 售价为 5860 万美元，是这里显而易见的例外 与下一个之间的差距非常大
#546 🔗 原推
2026-09-18 16:02:37
Blum
@atoms_res @CryptoRubic @zodl_app @zec_bit @TachyonZcash @zilkroad_ Zcash ecosystem finally getting the attention it deserves Free mints and shielded bonding curves are a solid combo
中文: @atoms_res @CryptoRuby @zodl_app @zec_bit @TachyonZcash @zilkroad_Zcash 生态系统终于得到了应有的关注 自由铸币厂和带屏蔽的粘接曲线是一个可靠的组合
#545 🔗 原推
2026-09-18 12:03:28
Blum
You can feed your AI everything you know and it will still ignore most of it. An Anthropic engineer spent 2h live building an agent, showing exactly why a full folder means nothing without context engineering. 05:15 - what a harness is made of: tools, prompts, skills 25:20 - the agent loop: act, observe, repeat 49:30 - read-only vs read-write file permissions 58:15 - context engineering with ls and cat 1:21:20 - building a multi-step research agent This workshop shouldn't have been given to the public for free. Watch this first, then read the article below on how to feed your AI the info so it actually uses it.
中文: 你可以为你所知道的一切人工智能提供食物，而它仍然会忽略其中的大部分内容。 一位人类工程师花了2小时时间在现场构建一个代理，展示了为什么没有上下文工程，一个完整的文件夹毫无意义。 05:15 - 什么是工具、提示、技能 25:20 - 代理循环:行动、观察、重复 49:30 - 只读文件与读写文件权限 58:15 - 与猫一起进行上下文工程 1:21:20 - 构建多步骤研究代理 本研讨会本不应免费向公众提供。 先看这个，然后阅读下面的文章，了解如何为人工智能提供信息，以便它真正地使用它。
🎬
视频
#544 🔗 原推
2026-09-16 18:54:49
Blum
@atoms_res @arcat_meme @minarafun 8 tickers but half of them are just cats Arc really has a type
中文: @amats_res @arcat_meme @minarafun 8 个勾选点，但其中一半只是猫 Arc 确实有一种类型
#543 🔗 原推
2026-09-16 16:04:41
Blum
the founder of @videoclawapp quit his developer job and moved back in with his parents to build his own products. then came the videos needed to market them. he bought recording gear, paid editors and tried a bunch of AI tools. vibe coding was fast. video creation was still slow. eventually, he built Videoclaw. its desktop video creation agent is now launching in public beta. i’d start by asking it to cut a long product demo into something people might actually finish watching. then work through the changes in chat, much like vibe coding. the AI editor, video generator, avatar maker and voice tools are all in the same app. AI automates complex production work, and simple requests let you make precise edits. you can ask it to fix a caption without hunting through menus yourself. that’s useful for developers and founders trying to explain what they’ve built. marketers have plenty to make with it, too. the rest of us get a way to speak video without needing the time, skills or team to handle production ourselves. acrylic paints, Photoshop, Canva, AI image generation. easier tools keep bringing more people into creating. Videoclaw lowers that barrier for video, and i’m curious what people who’ve never made one will do with it. maybe the next generation of creators ends up calling themselves content engineers. seems plausible when so much of the work happens through instructions. the public beta is free to use, with $10 in AI generation credits included. the first 300 people get $50. it’s a Mac app for now, with more platforms coming soon. download it at https://t.co/a1vUsu5oQC, connect your ChatGPT or Claude, and start prompting.
中文: @videoclawapp 的创始人辞去了开发者的工作，搬回父母身边，为自己生产产品。 然后，他们制作了营销所需的视频。他购买了录音设备、付费编辑，并尝试了大量人工智能工具。视频编码速度很快。视频制作仍然很慢。 最终，他创建了Videoclaw。其桌面视频创作代理现已在公开测试版中推出。 首先，我要求它将一个长的产品演示内容剪成人们可能真正看完的内容。然后通过聊天来调整，就像氛围编码一样。 人工智能编辑器、视频生成器、头像制作器和语音工具都在同一应用中。人工智能可自动完成复杂的制作工作，而简单的请求可让您进行精确编辑。您可以要求它自行修复说明，而无需通过菜单进行搜索。 这对开发者和创始人来说很有用，他们试图解释自己构建的内容。营销人员也有很多内容可以尝试。我们其他人在无需时间、技能或团队的情况下，就能通过视频来表达自己。 丙烯颜料、Photoshop、画布、AI图像生成。更便捷的工具不断吸引更多人来创作。Videoclaw 降低了视频的门槛，我好奇那些从未制作过视频的人会用它做什么。 也许下一代创作者最终会称自己为内容工程师。当如此多的工作通过指令进行时，似乎就是合理的。 公共测试版免费使用，包含10美元的AI生成学分。前300人获得50美元。 目前这是一款Mac应用，即将推出更多平台。请在 下载，连接 ChatGPT 或 Claude，然后开始提示。
#542 🔗 原推
2026-09-16 15:25:04
Blum
@atoms_res @arc Free mint with a deadline Guess I'll set an alarm for Sep 17
中文: @amos_res @arc 免费薄荷，截止日期 猜猜我会为9月17日设置闹钟
#541 🔗 原推
2026-09-16 12:28:28
Blum
Stop routing every agent call through your most expensive model. Engineer at Uber explains how to route agent calls to the right model. 00:00 - 70% of PRs now come from agents 01:26 - one gateway routes every model call 03:51 - cutting the token tax by 40% 08:39 - one context graph instead of 20 scattered systems 17:29 - the bottleneck is now deciding what to build I've seen $500 courses that teach less than this. Watch this first, then read the article below on running your own agents for less, without reducing your results.
🎬
视频
#540 🔗 原推
2026-09-16 07:46:52
Blum
Your code can route support tickets and send low-confidence cases to a human Jev is @typesafeai's first public System One Model, built for fast, cheap decisions inside software For a support ticket, Jev can assess urgency, check whether it concerns billing and identify whether a human is needed, all in parallel in one request It returns typed answers with probabilities, so you can combine decisions in code and set confidence thresholds for automatic action or review At $0.042 per million input tokens, with free output, I'd use the same approach for form follow-ups and deciding which notifications are worth sending Read @CompleteSkeptic's launch post and join Jev's newly opened developer waitlist
中文: 您的代码可以路由支持票，并将低置信度的案例发送给人类 Jev 是 @typesafeai 的首个公共系统一型，用于软件内部快速、低成本的决策 对于支持票，Jev 可以评估其紧迫性，检查是否涉及账单问题，并确定是否需要人员，同时同时提出一项请求 它返回输入的带概率的答案，因此您可以在代码中组合决策，并设置置信阈值以自动操作或审核 每百万个输入代币0.042美元，免费输出，我会采用相同的方法进行表单跟进，并决定哪些通知值得发送 阅读@CompleteSkeptic的发布帖子，加入Jev新开的开发者候补名单
#539 🔗 原推
2026-09-15 21:19:57
Blum
Odyssey-3 genuinely surprised me. one world model that powers a humanoid, controls robot arms, drives a car, pilots a drone, plays video games and trains other AIs. the team behind @odysseyml built self-driving cars at Cruise and Wayve and has been on world models since 2023. so when they call task-by-task robot training brute force, I listen. their answer is a physics agent. it learns how the world works by watching it: physics, motion, cause and effect, how people behave. a new machine plugs into that knowledge. they say a few hours of its own data gets it going, with an action head turning the model's understanding into controls. it's basically what LLMs did for knowledge work. one broad model underneath, and every upgrade to it makes everything built on top better. what stuck with me most: it also generates worlds where AI agents can practice, so their mistakes happen in a simulation first. they think world models will end up understanding physics better than we do. big claim. I want to see it hold up when it's public in the coming weeks.
中文: 奥德赛-3 真的让我感到意外。 一种为人形驱动的世界模型，控制机器人手臂、驾驶汽车、驾驶无人机、玩电子游戏以及训练其他人工智能。 @odysseyml 团队在 Cruise 和 Wayve 公司生产自动驾驶汽车，自 2023 年起就已加入世界车型。因此，当他们称任务型机器人训练为暴力力量时，我会倾听。 他们的答案是物理代理，通过观察它来了解世界的运作方式:物理、运动、因果，以及人们的行为方式。 一台新机器会插入这种知识。他们表示，几秒钟的自身数据就能发挥作用，而一个操作负责人将模型的理解转化为控制。 这基本上就是LLM为知识工作所做的。下面有一个广泛的模型，每一次升级都让所有内容都更为美好。 最让我印象深刻的是:它还能在人工智能智能中产生能够练习的世界，因此他们的错误首先在模拟中发生。 他们认为，世界模型最终会比我们更了解物理。我希望在接下来的几周内公开它时会挺身而出。
#538 🔗 原推
2026-09-15 17:37:03
Blum
RT @Blum_OG: I gave the same prompt to GLM 5.3 Flash, Kimi K3 and DeepSeek v4 Pro and got 3 working apps back. Built all 3 inside @boltdotnew and shipped the best one to a live link. On a normal plan I'd never burn prompts on a comparison like that. Forge is an experimental mode in Bolt that runs open models with up to 50x more usage. You can build and iterate on web apps without worrying about usage limits. One monthly bar, no daily cap. Draft 10 versions if you need to, then switch to Standard or Max when it's time to ship. On Bolt's own build benchmark, the Forge models score 91% of the top paid model. Close enough for drafting. The deal behind it: opt in, and your builds help train open models. It asks every time you switch in. The research preview runs Sept 14 to Oct 14, 2026. Forge is free on Pro plans until Oct 14, and there's a $9/month early-access plan, Bolt Lite, with a waitlist. It's experimental, so duplicate your project before bringing anything serious in. Watch the 3 builds, then try Forge with the link below.
🎬
视频
#537 🔗 原推
2026-09-14 19:31:25
Blum
I gave the same prompt to GLM 5.3 Flash, Kimi K3 and DeepSeek v4 Pro and got 3 working apps back. Built all 3 inside @boltdotnew and shipped the best one to a live link. On a normal plan I'd never burn prompts on a comparison like that. Forge is an experimental mode in Bolt that runs open models with up to 50x more usage. You can build and iterate on web apps without worrying about usage limits. One monthly bar, no daily cap. Draft 10 versions if you need to, then switch to Standard or Max when it's time to ship. On Bolt's own build benchmark, the Forge models score 91% of the top paid model. Close enough for drafting. The deal behind it: opt in, and your builds help train open models. It asks every time you switch in. The research preview runs Sept 14 to Oct 14, 2026. Forge is free on Pro plans until Oct 14, and there's a $9/month early-access plan, Bolt Lite, with a waitlist. It's experimental, so duplicate your project before bringing anything serious in. Watch the 3 builds, then try Forge with the link below.
🎬
视频
#536 🔗 原推
2026-09-14 17:09:51
Blum
@EMostaque Which model are you referring to that was the UK's top frontier model?
中文: @EMostaque 你指的是英国顶级前沿车型的哪个型号?
#535 🔗 原推
2026-09-14 12:06:01
Blum
@EMostaque World class cybersecurity is a huge ask for an org whose core mission is evals But youre right, the skill sets barely overlap
#534 🔗 原推
2026-09-14 07:07:32
Blum
@MiniMax_AI Love seeing FastH3 hit Apple Silicon That 4-step distillation is huge for local workflows
中文: @MiniMax_AI 喜欢看到 FastH3 冲击苹果硅 这种四步蒸馏对本地工作流程来说非常重要
#533 🔗 原推
2026-09-14 06:10:39
Blum
Harrison Chase: "a big part of memory is reflecting on traces and then updating something." the LangChain cofounder explains how agents use past runs to revise instructions i'd choose this free interview over a paid course on agent memory set aside 40 minutes to watch, then follow the guide below to test the method on a recurring bug
🎬
视频
#532 🔗 原推
2026-09-13 18:05:00
Blum
@garrytan What kind of advice do you give 1:1 that cant be shared publicly?
#531 🔗 原推
2026-09-13 17:03:23
Blum
@EMostaque Are you asking how the speed limit would actually be enforced technically?
中文: @EMostaque 你在问，限速实际上会如何从技术上来执行吗?
#530 🔗 原推
2026-09-13 09:45:06
Blum
@EMostaque Open models handling most tasks feels inevitable The swarm risk is real though, coordination changes everything
中文: @EMostaque 开放模型处理大多数任务感觉不可避免 群体风险是真实存在的，协调会改变一切
#529 🔗 原推
2026-09-13 09:23:34
Blum
@garrytan Big companies shipping fun tools for tinkerers is a good sign
中文: @garrytan 大公司为修补工运送有趣工具是个好兆头
#528 🔗 原推
2026-09-12 20:01:18
Blum
Stop spending crazy amounts on tokens. Co-founders of Factory AI and LangChain explain why the best agent doesn't mean the most expensive agent. 20:59 - the case against a model-independent harness 44:20 - why memory might be AI's most overused word 55:35 - agent readiness, the deterministic feedback agents need 1:04:41 - what missions cost, and how routing cuts the bill This conversation shouldn't have been distributed for free. After watching the video, read the article below to understand why there are always not enough tokens and how to get rid of this problem.
中文: 停止在代币上花费大量资金。 Factory AI 和 LangChain 的联合创始人解释了为何最佳代理并不代表最昂贵的代理。 20:59 - 针对模型无关线束的案例 44:20 - 为什么记忆可能是人工智能中使用最多的词汇 55:35 - 代理准备，需要确定性反馈 1:04:41 - 任务成本如何，以及路线如何降低费用 这个对话本不该免费分发。 观看视频后，阅读下面的文章，了解为何代币总是不够多，以及如何摆脱这一问题。
🎬
视频
#527 🔗 原推
2026-09-12 16:30:00
Blum
@EMostaque Lower settings usually still work for math and physics, just with less detail in the steps
中文: @EMostaque 较低设置通常仍适用于数学和物理领域，但步骤中的细节较少
#526 🔗 原推
2026-09-12 13:04:41
Blum
Microsoft engineer: “the average spend on tokens goes down by almost 78%" she explained how to properly manage tokens and stop complaining about limits paid courses charge $700 for this info, but here it's free watch this 21-min video, and then read the article below to learn how to use tokens wisely and avoid overspending on them
中文: 微软工程师: 代币的平均支出下降了近78% 她解释了如何正确管理代币，并停止抱怨限制 付费课程为此信息收取700美元，但此处免费 观看这段21分钟的视频，然后阅读以下文章，了解如何明智地使用代币，并避免过度消费
🎬
视频
#525 🔗 原推
2026-09-12 12:54:07
Blum
@garrytan Do you think this is a natural evolution or a sign that the market is getting crowded?
#524 🔗 原推
2026-09-12 06:56:23
Blum
@EMostaque Love without brains is just vibes and chaos
中文: @EMostaque 没有大脑的爱就是氛围和混乱
#523 🔗 原推
2026-09-12 05:58:38
Blum
@OpenAIDevs 20 million requests per second before the rewrite is wild Rust must have felt like a cheat code
中文: @OpenAIDevs 在重写之前每秒有 2000 万次请求 生锈一定感觉像是作弊码
#522 🔗 原推
2026-09-11 20:01:57
Blum
@OpenAIDevs GPT-6 Astra already needs skill reviews to behave Sounds like my last three teammates
中文: @OpenAIDevs GPT-6 Astra 已经需要技能评价才能表现 听起来像是我最后三个队友
#521 🔗 原推
2026-09-11 18:35:22
Blum
@EMostaque Finally, a startup that understands the real enemy: warm toes
中文: @EMostaque 终于，一家了解真正敌人的初创公司:温暖的脚趾
#520 🔗 原推
2026-09-11 11:29:33
Blum
@EMostaque I don’t disagree on the risk But going fully in silico isn’t a switch you flip overnight Physical labs still do things computers can’t model yet
中文: @EMostaque 我在风险问题上并不存在分歧 但完全采用硅不是一夜之间就能翻转的开关 物理实验室仍然做着计算机无法模拟的事情
#519 🔗 原推
2026-09-11 08:39:01
Blum
stop spending your evenings fixing the same mistakes your AI agent keeps making a LangChain engineer shows how to make AI check its own work before handing it back to you 08:06 - state, nodes, and edges 10:47 - checking documents and generated answers 11:48 - retrying code after failed tests 37:12 - stopping loops with a step limit 54:46 - tracking the steps an agent takes they shouldn't have given this workshop away for free watch this first, then read my guide below to build a research workflow that checks claims
中文: 停止在晚上解决你的人工智能代理持续犯的同样错误 一位朗查恩工程师展示了如何让人工智能在将其交还给你之前检查自己的工作 08:06 - 状态、节点和边缘 10:47 - 检查文件并生成答案 11:48 - 测试失败后重试代码 37:12 - 使用步骤限制停止循环 54:46 - 跟踪代理人采取的步骤 他们本不该免费赠送这个工作坊 先看这个，然后阅读我下面的指南，构建一个检查理赔的研究流程
🎬
视频
#518 🔗 原推
2026-09-11 07:44:42
Blum
@mattshumer_ Subsidized tokens were never sustainable long-term The shift to paid access was inevitable
中文: @mattshumer_ 补贴代币从来就不是长期可持续的 转向付费获取是不可避免的
#517 🔗 原推
2026-09-10 21:05:05
Blum
@rasbt encoder-decoder is a bold shift for them V5 naming would have been cleaner honestly
中文: @rasbt 编码器解码器对他们来说是一个大胆的转变 V5 命名老实说会更干净
#516 🔗 原推
2026-09-10 20:14:32
Blum
@rasbt Single GPU and eight days of patience Nice work
中文: @rasbt 单 GPU 和八天耐心 不错的工作
#515 🔗 原推
2026-09-10 19:24:03
Blum
stop paying for AI courses until you've watched this an Anthropic engineer shows how to make AI agents check and improve their own work in under 40 mins 13:28 - a basic agent that makes slides 19:15 - the prompt for grading slides 28:02 - a QA loop for slide generation 33:08 - when AI judges give wrong scores 35:27 - explanations first, scores second i've seen $500 courses that teach less than this watch this first, then read the article below on using AI to check your agent's work
中文: 停止支付人工智能课程费用，直到你看过这个 一位人类工程师展示了如何在40分钟内让人工智能代理检查并改进自身的工作 13:28 - 一种制作幻灯片的基本代理 19:15 - 分级幻灯片的提示 28:02 - 用于幻灯片生成的QA循环 33:08 - 当人工智能评委给出错误的分数时 35:27 - 解释第一，得分第二 我见过500美元的课程，教书比这少 先看这个，然后阅读下面关于使用人工智能检查你的代理工作的文章
🎬
视频
#514 🔗 原推
2026-09-10 17:34:40
Blum
@EMostaque @DarioAmodei Is 25% his current public number or has it moved since the Navier-Stokes result?
中文: @EMostaque @DarioAmodei 是他当前公开号码的25%，还是自Naviare-Stokes结果以来就已移动的?
#513 🔗 原推
2026-09-10 09:54:21
Blum
Boris Cherny: "every model is very different" his 36-minute interview covers why prompts and tools change with each model release learning to adapt to new models beats many paid courses watch the interview, then use Anthropic's article to audit your Claude prompts https://twitter.com/Blum_OG/status/2097796948095053917/video/1
🎬
视频
#512 🔗 原推
2026-09-09 21:19:02
Blum
@EMostaque Moonshot mate is a solid promotion Now you get to chomp at the bit with extra authority
中文: @Mostaque Moonshot mate 是一个可靠的推广 现在你可以凭借额外的权限来扼杀这一点
#511 🔗 原推
2026-09-09 21:13:55
Blum
@OpenAI 250+ people on defense is a serious operation The playbook part is smart, most orgs never share the architecture
中文: @OpenAI 250 以上国防人员是一项严肃行动 剧本部分很聪明，大多数人从未分享过这些架构
#510 🔗 原推
2026-09-09 20:39:43
Blum
@mattshumer_ what do you think is the most realistic first step people can take after reading it?
中文: @mattshumer_ 你认为人们在读完后能迈出最现实的第一步是什么?
#509 🔗 原推
2026-09-09 20:25:01
Blum
@mattshumer_ The scale of coordination this would take is hard to even picture
中文: @mattshumer_ 这需要协调的程度甚至难以想象
#508 🔗 原推
2026-09-09 15:51:33
Blum
LangChain co-founder Harrison Chase: "use the file system as much as possible." his 32-minute talk explains how agents store notes and separate working contexts there's more useful advice here than in many paid courses watch the talk, then use the article below to decide what belongs in your Grok Bot's description, memory, and shared workspace
中文: 朗查恩联合创始人哈里森·蔡斯: 尽可能使用文件系统。 他32分钟的演讲解释了代理人如何存储笔记和独立的工作情境 这里比许多付费课程更有用的建议 观看演讲，然后使用以下文章来确定您的Grok Boot的描述、内存和共享工作空间
🎬
视频
#507 🔗 原推
2026-09-09 07:31:22
Blum
@EMostaque Humanoids dont perspire but they do depreciate
#506 🔗 原推
2026-09-09 05:33:58
Blum
@garrytan Sober zone with mandatory recovery could actually give people a real path out But it only works if the city enforces it
#505 🔗 原推
2026-09-09 04:19:12
Blum
@EMostaque Anthropic finally finishing what Valve started They might actually release it before we get AGI
中文: @EMostaque Anthropic 终于完成了 Valve 的启动 他们可能会在我们收到AGI之前发布它
#504 🔗 原推
2026-09-08 22:52:29
Blum
@MiniMax_AI @higgsfield @krea_ai @runwayml @HeyGen Yasushi Akimoto for sure His track record with Japanese IP is unmatched
中文: @MiniMax_AI @higgsfield @krea_ai @runwayml @HeyGen 秋本邂邂邂邂茂 他使用日本IP的记录无与伦比
#503 🔗 原推
2026-09-08 18:52:28
Blum
Andrew Ng: "even though people talk about evals, for some reason people don't do it." in under 27 mins, he explains how evals help locate a failing step in an AI workflow a lesson you'd pay for in an AI course, shared for free watch the video, then read the article below for practical ways to test AI answers and catch mistakes
中文: 安德鲁·吴: 尽管人们谈论薖一入，但出于某种原因，人们并不这样做。 在不到27分钟的时间里，他解释了埃瓦尔斯如何帮助定位人工智能工作流程中的一个失败步骤 在人工智能课程中你需要付出的一课，免费分享 观看视频，然后阅读以下文章，了解检验人工智能答案和发现错误的实用方法
🎬
视频
#502 🔗 原推
2026-09-08 16:28:29
Blum
@EMostaque Yang-Mills feels right because it has a clear physical target, unlike the more abstract open problems
#501 🔗 原推
2026-09-08 07:27:27
Blum
@EMostaque @__alpoge__ The part about the prompt itself being written by prompting Codex is a wild detail It makes the 'very little human input' claim look even thinner
#500 🔗 原推
2026-09-08 05:40:30
Blum
@EMostaque enlightenment as the only real alignment is a take but it sidesteps that even clear minds can disagree on what the right path is
中文: @EMostaque 启蒙作为唯一真正的对齐是一持 但它回避的是，即使是清醒的头脑，也可能在正确的道路上存在分歧
#499 🔗 原推
2026-09-07 18:48:18
Blum
@garrytan Categorization is doing a lot of work here Founders still pick what to build, but the RFS shapes what gets a closer look
中文: @garrytan Categorization 正在这里做大量工作 创始人仍然选择要构建什么，但RFS塑造了更近的面貌
#498 🔗 原推
2026-09-07 15:36:15
Blum
@EMostaque HVAC techs are about to out-earn the AI models that predicted their job growth
#497 🔗 原推
2026-09-06 22:15:35
Blum
@mattshumer_ What kind of project would you hand off for days without checking in?
中文: @mattshumer_ 哪些项目在不办理入住的情况下会推迟几天?
#496 🔗 原推
2026-09-06 16:29:13
Blum
@EMostaque The jump from instantly generating code to people running out of ideas is doing a lot of work
中文: @EMostaque 从即时生成代码到人们想法的耗尽，正在做大量工作
#495 🔗 原推
2026-09-06 16:03:28
Blum
smart models repeat old mistakes when runs start without evidence from earlier work Anthropic's Lamis Mukta: "The intelligence alone is not going to compound because they need this context that helps them perform the specific tasks that you need them to." 32 minutes on why smarter models still repeat your workflow mistakes worth more than a $500 course on agent memory the video explains why agents need memory. the article below shows how to build it
中文: 智能模型在运行开始时会重复旧错误，而无法从早期工作中提供证据 蚁皮的拉米斯·穆克塔: 仅凭智力无法复合，因为他们需要这种环境，帮助他们完成所需的具体任务。 32分钟介绍为何智能模型仍会重蹈你的工作流程覆辙 价值超过500美元的代理记忆课程 该视频解释了代理人员为何需要内存。下文展示了如何构建
🎬
视频
#494 🔗 原推
2026-09-06 07:15:03
Blum
@EMostaque What did it actually make?
中文: @EMostaque 它究竟做了什么?
#493 🔗 原推
2026-09-05 17:56:16
Blum
@mattshumer_ What kind of updates are you planning?
中文: @mattshumer_ 你正在计划哪些类型的更新?
#492 🔗 原推
2026-09-05 16:22:54
Blum
@garrytan @AsideAI The jump from 2 hours to under 3 minutes is the part that hits hardest Setup friction is usually where these tools lose people
中文: @garrytan @AsideAI 从2小时到3分钟以下是影响最重的部分 设置摩擦通常是这些工具会流失人的地方
#491 🔗 原推
2026-09-04 20:29:59
Blum
OPENAI JUST DROPPED GPT-6 ASTRA 🚨 the strongest update for builders is how Astra handles long, messy work Terminal-Bench Science 0.1: Astra 64.6%, Fable 5.1 52.6% AutomationBench: Astra 41.4%, Fable 5.1 31.4% Terminal-Bench 4.0: Astra 57.9%, Fable 5.1 55.8% Codex also gets notes that survive context-window changes, plus search across old messages and tool output Astra just crushed the newly released Fable 5.1 across the board that's a clear sign Anthropic needs to step it up to stay in the race
中文: 奥普纳伊刚刚放弃了GPT-6 ASTRA 🚨 对建筑商而言，最有力的更新是Astra处理漫长而混乱的工作方式 终端级科学 0.1:Astra 64.6%，Fable 5.1 52.6% 自动化板头:Astra 41.4%，Fable 5.1 31.4% 终端-台式机4.0:Astra 57.9%，Fable 5.15.8% Codex 还会收到在上下文窗口变化中存活下来的笔记，并搜索旧消息和工具输出 阿斯特拉刚刚全面粉碎了新发布的Fable 5.1 这明显表明人类需要挺身而出才能继续参加比赛
media 0 共 1 项
#490 🔗 原推
2026-09-04 19:28:08
Blum
@MistralAI @aiDotEngineer @bfl_ai @cognition @huggingface Paris in September sounds perfect for a sold-out crowd Just make sure the coffee breaks live up to last year's hype
中文: @MistralAI @aiDotEngineer @bfl_ai @cognition @huggingface 巴黎九月的音乐听起来非常适合售罄的人群 确保咖啡与去年的炒作不尽同
#489 🔗 原推
2026-09-04 19:07:37
Blum
@MiniMax_AI @togethercompute This is the kind of event where the real savings numbers finally get discussed out loud
中文: @MiniMax_AI @togethercompute 这是真正节省开支数据最终被大声讨论的事件
#488 🔗 原推
2026-09-04 19:01:43
Blum
@OpenAIDevs What are these timestamps for?
中文: @OpenAIDevs 这些时间戳是什么?
#487 🔗 原推
2026-09-04 18:37:04
Blum
@EMostaque @cognition A zillion is the only unit that makes sense for that crew Token budgets probably look like research grants
中文: @EMostaque @cognition A 亿是唯一适合该团队的单位 代币预算可能看起来像是研究资助
#486 🔗 原推
2026-09-04 11:38:19
Blum
O*NET's Robotics Engineers page lists a $122,930 median wage Stanford professor Chelsea Finn: “you don’t need a PhD necessarily to do a lot of that engineering work.” her 58-minute talk explains what Robotics Engineers actually build and why their work is becoming more valuable watch it, then read the article below for a six-month roadmap into that work
中文: O*NET的机器人工程师页面列出了122930美元的中位工资 斯坦福大学教授切尔西·芬恩: 你不需要博士学位来完成大量的工程工作。 她58分钟的演讲解释了机器人工程师究竟如何构建，以及他们的工作为何变得越来越有价值 观看后，阅读以下文章，了解该工作为期六个月的规划
🎬
视频
#485 🔗 原推
2026-09-04 11:06:31
Blum
@OpenAI https://twitter.com/Blum_OG/status/2095787222037671965/photo/1
#484 🔗 原推
2026-09-04 08:13:06
Blum
@perplexity_ai 13.5% higher score at 6.1% lower cost is a pretty clean tradeoff Most models only win on one axis
中文: @perplexity_ai 以低6.1%的价格高得分，是一次相当彻底的权衡 大多数模型仅在一个轴上获胜
#483 🔗 原推
2026-09-03 21:11:52
Blum
@garrytan the editor pipeline was just a group chat hyping each other up
中文: @garrytan 编辑器的流程只是一个互相炒作的群聊
#482 🔗 原推
2026-09-03 21:08:13
Blum
@mattshumer_ Deleting your entire Mac is a wild way to beta test But I get it, Astra must be something else to win you back
中文: @mattshumer_ 删除整个 Mac 是测试版的一种疯狂方式 但我了解到，阿斯特拉一定是别的东西才能赢回你
#481 🔗 原推
2026-09-03 19:35:18
Blum
@cohere @AstonMartinF1 The email draft idea is a fun touch Keeping it light while the team stays locked in
中文: @cohere @AstonMartinF1 邮件草稿创意很有趣 保持轻盈，团队保持紧闭
#480 🔗 原推
2026-09-03 18:02:46
Blum
@EMostaque Quoting scripture to justify a bill feels like a strange way to make tech policy
中文: @EMostaque 引用经文来证明一项法案的合理性，似乎是一种制定技术政策的奇怪方式
#479 🔗 原推
2026-09-03 16:22:19
Blum
@EMostaque meta isnt really a frontier ai lab theyre a social media company that happens to have an ai division
中文: @EMostaque 元实际上并无前沿的 ia 实验室 他们是一家恰好拥有一家业务部门的社交媒体公司
#478 🔗 原推
2026-09-02 21:43:56
Blum
Harrison Chase (CEO of LangChain): “The key is finding these UX patterns where the agent does a ton of work, but the human’s still in the loop at key points.” his 21-minute talk shows how to build agents that do useful work and ask for help when judgment matters it damn well beats any paid course on the topic watch it first. then follow the article step by step and build your first working workflow
中文: 哈里森·蔡斯(朗瑞公司首席执行官) 关键是找到这些用户体验模式，即中介在关键点上进行着大力的工作，但人类仍处于循环中。 他21分钟的演讲展示了如何培养出有用工作的代理人，并在判断重要时寻求帮助 它在这个话题上胜过任何付费课程 先观看。然后一步步关注文章，构建你的第一个工作流程
🎬
视频
#477 🔗 原推
2026-09-02 20:55:56
Blum
@OpenAIDevs 24 hours is a tight window but shipping under pressure usually brings out the best work
中文: @OpenAIDevs 24小时时间很紧，但在压力下发货通常会带来最美好的工作
#476 🔗 原推
2026-09-02 20:02:07
Blum
@EMostaque at a certain speed they also outrun our ability to notice theyre gone
中文: @EMostaque 以一定速度也超出了我们注意到他们消失的能力
#475 🔗 原推
2026-09-02 19:39:44
Blum
@EMostaque Standards are great until an agent reads them and decides compliance is optional
中文: @EMostaque 标准很好，直到代理人阅读并决定合规性是可选的
#474 🔗 原推
2026-09-02 19:19:57
Blum
@mattshumer_ Token rationing with a model that big sounds like a nightmare for an open world How did you even keep the context coherent across a whole city?
中文: @mattshumer_ 与一个听起来像是开放世界噩梦的模特进行代币配给 你是如何在整个城市中保持语境连贯的?
#473 🔗 原推
2026-09-02 17:11:13
Blum
@garrytan How did you and Stephen land on that logo concept?
中文: @garrytan 你和斯蒂芬是如何获得这个标志概念的?
#472 🔗 原推
2026-09-02 14:57:08
Blum
@rasbt So the looped transformer is basically a model that rereads the same paragraph twice and calls it depth Sounds like my study method
中文: @rasbt 因此，环形变压器基本上是一种模型，会重复读取同一段内容，并称之为深度 听起来像是我的学习方法
#471 🔗 原推
2026-09-02 13:28:32
Blum
OpenAI engineer Ryan Lopopolo has spent 9 months building software exclusively with agents: "the important thing is not the code, but the prompt and the guardrails that got you there." his 46-minute talk shows the work around the code: writing requirements, delivering context at the right moment, and turning review feedback into checks watch it, then read the article below for seven templates you can adapt
中文: OpenAI 工程师瑞安·洛波波洛花了 9 个月时间专门与代理人员合作开发软件: 重要的不是代码，而是提示和让你到达那里的护栏。 他长达46分钟的演讲展示了围绕代码的创作: 撰写要求，在合适的时机提供背景内容，并将评论反馈转化为审查 观看后，阅读以下文章，了解可修改的七个模板
🎬
视频
#470 🔗 原推
2026-09-02 07:24:44
Blum
@mattshumer_ The joke is that youre already broke
中文: @mattshumer_ 笑话是，你已经崩溃了
#469 🔗 原推
2026-09-01 19:43:44
Blum
@ClaudeDevs Resetting the weekly limits for everyone is a nice touch Keeps things fair for players who already used theirs up
中文: @ClaudeDevs 为每个人重置每周限额是件好事 让那些已经用上手的球员保持公平
#468 🔗 原推
2026-09-01 18:37:08
Blum
@mattshumer_ Cheaper cache reads are nice but the real test is whether it holds up under sustained agentic workloads A 75% cut only matters if the quality stays consistent
中文: @mattshumer_ 更便宜的缓存读取不错，但真正的考验在于它能否在持续的代理工作负载下保持良好状态 75%的切割仅在质量保持一致的情况下才重要
#467 🔗 原推
2026-09-01 18:28:23
Blum
@AIatMeta 20+ speakers in real time is a serious flex for live meetings and podcasts
中文: @AIatMeta 20+ 实时演讲者是现场会议和播客的有力嘉宾
#466 🔗 原推
2026-09-01 17:29:02
Blum
@SpaceXAI Refusing obfuscated biological tasks while still answering beneficial science is a tough balance to get right
中文: @SpaceXAI 在回答有益科学的同时，拒绝混淆的生物任务，是难以实现的平衡
#465 🔗 原推
2026-09-01 16:15:14
Blum
@AnthropicAI Glad to see the transparency here The part about reward hacking shaping model behavior is the kind of detail that usually stays internal
中文: @AntherpicAI 很高兴看到这里的透明度 关于奖励黑客塑造模型行为的部分，通常是内部细节
#464 🔗 原推
2026-08-31 22:46:55
Blum
@EMostaque so GLM 6.0 gets the full RSI treatment while 5.3 was already a massive scale-up that $2b ARR makes the bet way easier to stomach
中文: @EMostaque 因此，GLM 6.0 获得了完整的 RSI 治疗，而 5.3 已经规模化 那2亿美元的ARR让赌注更容易吃饱
#463 🔗 原推
2026-08-31 22:14:02
Blum
founders should never let a coding agent grade its own work separate the roles: > founder: writes the task and defines what a correct result must do > coding agent: writes the code > testing agent: gets the original task and the result, then checks every requirement > testing agent: reports each failure and the steps that caused it > coding agent: fixes the failures > testing agent: tests the new version from the start > founder: approves or rejects the release watch the video to see the workflow in action, then read the article below to build the full system with Grok Bot
中文: 创始人绝不能让编码代理对其自身的工作进行评分 将角色分开: 创始人:编写任务并定义正确结果必须做什么 编码代理:编写代码 测试代理:获得原始任务和结果，然后检查每个要求 测试代理:报告每次故障及导致故障的步骤 编程代理:修复故障 测试代理:从开始测试新版本 创始人:批准或拒绝该版本 观看视频以查看工作流程，然后阅读以下文章，使用 Grok Boot 构建完整系统
🎬
视频
#462 🔗 原推
2026-08-31 20:27:34
Blum
@OpenAIDevs Office hours with that lineup is basically a free masterclass
中文: @OpenAIDevs 使用该课程的办公时间基本上就是免费大师课
#461 🔗 原推
2026-08-31 18:06:30
Blum
@EMostaque latin for open source wins is a power move now i want the t-shirt
中文: @EMostaque latin 获取开源大奖是一次强力举措 现在我想要这件T恤
#460 🔗 原推
2026-08-31 11:34:38
Blum
AI made writing easier but made a lot of published text worse the Beemo study shows what expert editors check to make AI text better (see Table 7 below): repetition, awkward phrasing, tone and factual accuracy before publish, run these 6 checks: 1. decide what the reader should understand or do 2. check every name, number, link and cause-and-effect claim 3. cut repetition, generic intros, recap endings and prompt residue 4. replace vague claims with a specific example that shows who did what and what happened 5. read it aloud and replace anything you would never say 6. hand it to someone without context and ask what they would do next save Table 7 and the article to remove all AI patterns from your writing
中文: 人工智能让写作变得更容易，却让大量已发表的文字变得更糟 Beemo 研究展示了专家编辑为改善人工智能文本所检查的内容(参见下表7): 重复、尴尬的措辞、语气和事实准确性 发布前，先运行以下6个检查: 1. 决定读者应该理解或做什么 2. 检查每个姓名、号码、链接及因果关系索赔 3. 切口重复、通用介绍、重复结尾和提示残留 4. 用一个具体例子替换模糊的说法，以显示谁做了什么以及发生了什么 5. 大声朗读并替换任何你永远不会说的 6. 将内容交给没有背景的人，并询问他们下一步会做什么 保存表7和文章，以清除写作中的所有AI模式
media 0 共 1 项
#459 🔗 原推
2026-08-31 11:19:53
Blum
@EMostaque core opinionated and mutability on the edges sounds like a recipe for fighting the defaults more than enjoying them
中文: @EMostaque 核心在边缘具有可体和可变性，听起来更像是一种对抗默认值的配方，而不是享受它们
#458 🔗 原推
2026-08-30 07:41:08
Blum
Kimi K3 is the best LLM for Grok Bot its 1M-token context window can handle serious, context-heavy jobs and at just $3/1M input tokens, it won’t kill your margins here’s how to wire it into your setup: 1. open the Agent Computer terminal 2. install Claude Code: npm install -g @ anthropic-ai/claude-code 3. choose the route that matches your Kimi key Kimi API Platform: ANTHROPIC_AUTH_TOKEN ANTHROPIC_BASE_URL=https:/ /api.moonshot. ai/anthropic ANTHROPIC_MODEL=kimi-k3[1m] effort: max Kimi Code subscription: ANTHROPIC_API_KEY ANTHROPIC_BASE_URL=https:/ /api.kimi. com/coding/ ANTHROPIC_MODEL=k3[1m] effort: high 4. map the Opus, Sonnet, Haiku, Fable and subagent variables to the same Kimi model ID 5. enable the 1M context window by setting both context limits to 1048576 6. start Claude Code and run `/status for Kimi API Platform, check for the Moonshot URL and `kimi-k3[1m]` for Kimi Code, check for api.kimi. com/coding/. the model label may still show Claude 7. send a small test prompt and check the matching Kimi usage page the video shows the same Kimi K3 + Claude Code setup. for Grok Bot, run these steps inside the Agent Computer terminal after watching the video, check out the full guide below
🎬
视频
#457 🔗 原推
2026-08-30 06:53:55
Blum
@MiniMax_AI @fal Open weights and open doors Now thats how you make a milestone feel like an invitation
中文: @MiniMax_AI @fal 开放权重与开门 现在，你让一个里程碑感觉像是一份邀请
#456 🔗 原推
2026-08-29 19:15:00
Blum
@garrytan Unlimited support for drug use is a strawman Harm reduction isnt pro-drug, its pro-survival
中文: @garrytan 对吸毒的无限支持是个稻草人 减少危害不是药物，而是其生存能力
#455 🔗 原推
2026-08-28 20:12:49
Blum
Cursor engineer: "my agents could just automatically fix bugs for me while I sleep" she and Cursor PM Roshan Sadanani show the Grok Bot workflow behind it 00:03 - agents fix bugs while Lauren sleeps 02:36 - persistent computer keeps working 03:59 - chief of staff coordinates other Bots 04:28 - marketing agent works inside LinkedIn 14:28 - chief of staff dispatches engineering tasks 15:13 - Grok Bot starts Cursor cloud agents 18:36 - lint and CI constrain agent code 21:08 - skills train agent teams the quoted article covers Grok Bot’s initial setup, pricing, and everything you need to get started
中文: 柯尔工程师 我的经纪人可以在我睡觉时自动修复漏洞 她和Cursor总理罗山·萨达纳尼展示了其背后的Grok机器人工作流程 00:03 - 经纪人在劳伦睡觉时修复漏洞 02:36 - 持久计算机持续工作 03:59 - 参谋长协调其他机器人 04:28 - 营销代理在LinkedIn内部运作 14:28 - 总参谋长派遣工程任务 15:13 - Grok Boot 启动 Cursor 云代理 18:36 - lint 和 CI 约束代理代码 21:08 - 技能培训代理团队 引文文章涵盖了Grok Boot的初始设置、定价以及入门所需的一切
🎬
视频
#454 🔗 原推
2026-08-28 02:30:57
Blum
@MiniMax_AI @anyscalecompute @VictorSuOrtiz @vllm_project @VoidAsuka 33B open-weight and it still shows up to the summit on time Ray Summit must be the one place where multimodal means showing up with slides
中文: @MiniMax_AI @anyscalecompute @VictorSuOrtiz @vllm_project @VoidAsuka 33B 开放量，至今仍会及时到场 雷峰必须成为多式联运意味着出现滑梯的地方
#453 🔗 原推
2026-08-26 21:19:46
Blum
@Alibaba_Qwen @sgl_project day-0 support is great but "ready to deploy today" is doing a lot of work Qwen3.8-Flash-Next with SGLang already sounds like a full rollout, not a preview
中文: @Alibaba_Qwen @sgl_project day-0 支持很棒，但“今天就准备部署”正在做大量工作 使用SGLang，Qwen3.8-Flash-Next 听起来已经完全了，而不是预览版
#452 🔗 原推
2026-08-26 13:52:03
Blum
@Alibaba_Qwen @vllm_project Day 0 support is nice but calling it amazing feels early Most of the interesting bits are still in a specific container
中文: @阿里巴巴_Qwen @vllm_project 0 支持不错，但称它为惊艳的感觉很早 大多数有趣的比特仍然位于一个特定的容器中
#451 🔗 原推
2026-08-26 13:24:55
Blum
this guy connected Grok Bot to 3 X accs and hit 69.8M impressions which is ~$7K in X payouts here's how you can replicate it: 1. give Grok Bot one job: research X and turn proven material into post angles 2. feed it posts, screenshots, articles, product demos, conversations, results, and raw ideas 3. run 2 plays: > repackage an idea that already has attention > build a fresh angle from your own material 4. ask for 5 different hooks or angles for each candidate 5. make the final call yourself approve the claim, frame, hook, and full draft 6. use SendShort for video cuts. save or schedule approved drafts in Postiz 7. keep a winner log after every test save the hook, creative, posting time, likes, bookmarks, replies, and video drop-off points. feed that log back into Grok Bot before the next batch watch the video, then get Daniel Ch's full prompt and setup in article below
🎬
视频
#450 🔗 原推
2026-08-26 12:47:11
Blum
@mattshumer_ @agentmail the email part is what gets me it just went and made its own inbox like that was the obvious next step
#449 🔗 原推
2026-08-25 18:22:32
Blum
RT @Blum_OG: https://x.com/i/article/2089644483948240896
#448 🔗 原推
2026-08-25 10:13:07
Blum
@mattshumer_ @agentmail Devin really said I'll find you another way AgentMail about to become the new Slack
#447 🔗 原推
2026-08-24 18:33:28
Blum
why your posts don't go viral this 19-minute video explains how X scores each post and chooses whether to show it to people who don't follow you 01:33 - how X reads each post 01:48 - how X scores a post 04:17 - which actions matter most 05:51 - what can hurt your reach 08:39 - where feed posts come from 09:02 - why reaching new people is harder 11:44 - how X limits some posts 13:35 - how to check reach limits write one clear idea for a specific group of people, then add smth useful that's easy to share my article shows in detail how the X algo works and how to get more reach
🎬
视频
#446 🔗 原推
2026-08-24 17:51:29
Blum
@garrytan the objections in the quoted thread get so absurd they loop back to parody its like theyre not even hiding the goal is to block everything
#445 🔗 原推
2026-08-24 14:50:37
Blum
https://x.com/i/article/2089644483948240896
#444 🔗 原推
2026-08-24 12:36:14
Blum
@garrytan Systems of record becoming harnesses feels like the right pressure point The moment agents stop just reading data and start acting on it, the old interfaces become the bottleneck
#443 🔗 原推
2026-08-24 04:26:07
Blum
here’s how to automate away 90% of the workload all it takes is Grok Bot, 30 days, and this roadmap week 1 - build the base - create one coordinator - give it one small task you can verify in 5 minutes - connect only the tools that task needs week 2 - add one specialist - give it one repeatable job - describe what a good result looks like - list every action that needs your approval week 3 - automate one proven task - show the full workflow once - save the method as a reusable skill - watch 3 clean runs before adding a schedule week 4 - build the small team - add specialists for work that already repeats - let the coordinator route the work - review every routine each Friday after 30 days, the goal is 5-7 focused bots, 3-5 useful routines, and one weekly review save this roadmap, watch the video below, then read the article to fully master Grok Bot
🎬
视频
#442 🔗 原推
2026-08-23 20:50:51
Blum
@garrytan Luxury belief is doing a lot of work there Banning data centers would hit every app and service people rely on daily
#441 🔗 原推
2026-08-21 20:10:39
Blum
starting a business alone is now f*cking easy and Grok Bot is a key part of it: 1. pick one task you repeat every week 2. create one Bot. tell it what result to return and when it must ask you first 3. connect only the services it needs 4. open the Bot's computer and use "Teach a task" 5. show the full process once. the recording can last up to 10 minutes, then the Bot creates a draft skill 6. test the skill on a safe example. add a schedule after the test passes 7. if the workflow needs separate roles, create a group of 2-6 Bots. give each step one owner keep sending messages, making purchases, deleting files, and changing live systems behind your approval start with a simple task so you can understand the full process
🎬
视频
#440 🔗 原推
2026-08-21 12:55:48
Blum
@garrytan If data centers cracked the freeze, maybe the real unlock is just letting people build again
#439 🔗 原推
2026-08-21 12:32:25
Blum
@Alibaba_Qwen @sgl_project Dropping fresh recipes for Qwen3.8-27B is solid, but the real win is NVFP4 + DFlash2 already showing great results in the community
#438 🔗 原推
2026-08-21 08:04:49
Blum
@garrytan @johnmaeda The shift from building to knowing what to build is a big one Design as the differentiator makes sense when anyone can generate the code
#437 🔗 原推
2026-08-20 15:25:37
Blum
@Alibaba_Qwen @kilocode Qwen3.8-Max on Kilo UI so clean the code writes itself
#436 🔗 原推
2026-08-20 04:01:14
Blum
@garrytan Framing it as bullying makes it sound like the developers have all the power But this order flips that completely
#435 🔗 原推
2026-08-19 16:22:35
Blum
@Alibaba_Qwen What kind of hardware do you need to run it locally?
#434 🔗 原推
2026-08-19 05:02:37
Blum
@Alibaba_Qwen @cline What made this model stand out enough to dethrone the previous top local model?
#433 🔗 原推
2026-08-19 04:01:27
Blum
when is it actually worth using agent graphs? Anthropic ran the same agent task 2 ways: solo: 20 min, $9 full setup: 6 hours, $200 the full setup gave a better result, but cost over 20x more start with a simple agent loop add a graph when you need: > human approval > pause and resume > checkpoint recovery > routing rules in code test both versions on your own tasks pick the one with the better completion rate and cost per successful run the article below covers when and how to use agent graphs
中文: 什么时候真正值得使用代理图谱? Anthropic 以相同的代理任务运行了两种方式: 单人:20分钟，9美元 完整设置:6小时，200美元 完整设置给出了更好的结果，但成本却超过20倍 从一个简单的代理循环开始 需要时添加图表: 人为认可 暂停并恢复 检查站恢复 代码中的路由规则 在您自己的任务中测试两个版本 选择每次成功完成率和成本更高的一次 下文内容涵盖了使用代理图的时间和方法
media 0 共 1 项
#432 🔗 原推
2026-08-18 20:43:00
Blum
@mattshumer_ @OpenRouter @Etched Backing two seed picks that both matter this week is a strong hit rate Even picky investors miss more than they land
中文: @mattshumer_ @OpenRouter @Etcde 支持本周两个重要选秀权，但均受到强烈打击 即使是挑剔的投资者，也比他们错过的更多
#431 🔗 原推
2026-08-18 18:06:56
Blum
X published the code behind how its TL algorithms actually work I spent 48 hours studying it, and here’s what I found: 1. space out your best posts because when several compete for the same person's feed at once, X gives the strongest one the full score and lowers the rest 2. earn the first like early because it can place the post in another pool X uses to find Home recommendations 3. use the first 48 hours as your main promotion window because the For You feed stops considering older posts 4. write posts people will reply to, share, quote, and follow you for because X scores all of those expected actions 5. remove bait likely to trigger Not interested, mutes, blocks, or reports because those predictions lower the score 6. earn relevant followers because future posts become in-network for them and receive a scoring advantage 7. publish the main idea as a standalone post because X removes replies and reposts before choosing posts for non-followers For You feeds 8. keep the account public and reach posts open to everyone because protected and subscriber-only posts have smaller audiences 9. avoid suspicious links and be careful when quoting posts with warning labels because X can remove your whole post from recommendations to non-followers 10. if you have under 1,000 followers, focus on fresh standalone posts because X has a test that can give one post extra score before it reaches 1,000 impressions these are the core rules for getting the best possible reach on X
media 0 共 1 项
#430 🔗 原推
2026-08-17 17:02:22
Blum
RT @Blum_OG: https://x.com/i/article/2083612424280686593
#429 🔗 原推
2026-08-17 14:17:19
Blum
this the clearest 26-min introduction to Graph Engineering I've found Greg Isenberg shows how one recurring AI task becomes a map of jobs, handoffs, checks, and a verified final output 01:24 - prompt, context, and graph engineering 03:35 - jobs, arrows, and shared notes 10:01 -… https://twitter.com/Blum_OG/status/2089067556854448358/video/1
media 0 共 2 项
🎬
视频
#428 🔗 原推
2026-08-16 19:11:33
Blum
https://x.com/i/article/2083612424280686593
#427 🔗 原推
2026-08-16 13:16:39
Blum
Bitunix gives new users a chance to earn solid money deposit funds and leave them in the account for the required period > hold 100 USDT for 3 days to get a 300 USDT voucher > hold 500 USDT for 5 days or 1,000 USDT for 7 days to get a 1,500 USDT voucher > hold 2,000 USDT for 9… https://twitter.com/Blum_OG/status/2088333722986643564/photo/1
media 0 共 1 项
#426 🔗 原推
2026-08-14 18:35:33
Blum
https://twitter.com/Blum_OG/status/2088292831106343231/photo/1
media 0 共 1 项
#425 🔗 原推
2026-08-14 15:53:04
Blum
YOUR AI AGENT TALKS TOO MUCH. THAT'S THE PROBLEM. long texts with little useful information from AI have become routine research shows that simple visuals are much easier to understand and remember and I found a way to generate them with one line the skill is called show-me,… https://twitter.com/Blum_OG/status/2087986838002696365/photo/1
media 0 共 1 项
#424 🔗 原推
2026-08-13 19:37:09
Blum
CLOSE YOUR LAPTOP. GROK BOT KEEPS WORKING. SpaceXAI built exactly what anyone running AI on long tasks needed every user gets an always-on cloud computer create and configure one or more bots, then assign a task the request goes to that cloud computer Grok Bot opens the… https://twitter.com/Blum_OG/status/2087629338376040758/video/1
media 0 共 2 项
🎬
视频
#423 🔗 原推
2026-08-12 19:56:35
Blum
HOW TO MAKE $1K/DAY WITH JUST REDDIT + AI 1. Reddit reveals unmet demand people describe problems in their own words, compare available solutions, and talk about what they already pay for 2. AI is used here for quality research > collect and clean Reddit discussions > group… https://twitter.com/Blum_OG/status/2087271173579055605/video/1
media 0 共 2 项
🎬
视频
#422 🔗 原推
2026-08-11 20:13:22
Blum
Claude Code can now talk to itself before, users had to: > check multiple terminals > copy results over > explain the solution again > warn one session about changes made in another now different Claude Code sessions can exchange messages pass along status, ask questions,… https://twitter.com/Blum_OG/status/2086547937446814197/photo/1
media 0 共 1 项
#421 🔗 原推
2026-08-09 20:19:29
Blum
GROK IMAGE 2.0 IS BETTER THAN ANYONE EXPECTED the AI Arena leaderboard shows the model already sitting in #2 place, right behind GPT Image 2 advantages: > clean spot edits without regenerating the whole thing > way better at text, typography and dense layouts > takes up to… https://twitter.com/Blum_OG/status/2086180895841784132/photo/1
media 0 共 1 项
#420 🔗 原推
2026-08-08 20:00:59
Blum
AI IS QUIETLY TAGGING EVERYTHING IT MAKES > you asked AI to generate a photo > downloaded it, cropped it, tweaked the colors > posted it online > an AI info label popped up anyway AI tags video, photos, text, audio, basically anything it touches and it tags each of them… https://twitter.com/Blum_OG/status/2085796732882010229/video/1
media 0 共 2 项
🎬
视频
#419 🔗 原推
2026-08-07 18:34:28
Blum
RT @alphabatcher: https://x.com/i/article/2085333003966767104
#418 🔗 原推
2026-08-06 14:17:14
Blum
GPT-5.6 Sol and Mythos 5 turned a cyber test into a security incident Mythos 5 carried out 17 of 19 harmful actions it launched a sustained chain of actions targeting real people GPT-5.6 Sol made 2 smaller but still unsafe moves onto the open internet what Mythos did: 1.… https://twitter.com/Blum_OG/status/2085092936350671219/video/1
media 0 共 2 项
🎬
视频
#417 🔗 原推
2026-08-05 19:57:49
Blum
THESE 7 PROMPTS WILL GET YOU HIRED the mistake most people make: treating a CV as the final word on someone’s value it’s much more useful to treat it as a collection of evidence here’s what these prompts help you do: - discover transferable skills, overlooked roles, new… https://twitter.com/Blum_OG/status/2084363876217262398/photo/1
media 0 共 1 项
#416 🔗 原推
2026-08-03 19:40:48
Blum
AI needs to understand the process as deeply as the outcome I keep seeing the missing middle problem, even among advanced AI users people understand that AI needs smth concrete to work with what many still miss is this: > give a model nothing but final outputs, and it learns… https://twitter.com/Blum_OG/status/2083621157777936802/video/1
media 0 共 2 项
🎬
视频
#415 🔗 原推
2026-08-01 18:29:30
Blum
TOP-TIER AI IS FINALLY STOPPING BEING A LUXURY OpenAI just slashed prices on its frontier models: GPT-5.6 Luna - 80% cheaper GPT-5.6 Terra - 20% cheaper GPT-5.6 Sol API Fast mode - up to 2.5x faster for 2x the price it’s still expensive next to Chinese models but compared to… https://twitter.com/Blum_OG/status/2083116688032117143/photo/1
media 0 共 1 项
#414 🔗 原推
2026-07-31 09:04:55
Blum
BUILD YOUR OWN LOCAL AI FOR ONLY $8 there’s a common belief that local AI is a luxury but you can actually pull it off on a budget this setup gets you 28.9M parameters, a 14.9 MB model, and roughly 9.5 tokens/s here’s how to do it: > grab an ESP32-S3 N16R8: 16 MB Flash + 8… https://twitter.com/Blum_OG/status/2082518998428623304/video/1
media 0 共 2 项
🎬
视频
#413 🔗 原推
2026-07-29 17:29:55
Blum
AGENTS ARE BECOMING IMPOSSIBLE TO CONTAIN GPT-5.6 Sol showed strong cyber capabilities but they also proved uncontrollable the agents were supposed to solve ExploitGym tasks here’s what they did instead: 1. found a zero-day in OpenAI’s internal package proxy 2. escalated… https://twitter.com/Blum_OG/status/2081627448827912508/video/1
media 0 共 2 项
🎬
视频
#412 🔗 原推
2026-07-27 06:27:13
Blum
The guy in your group chat who called the Pokemon reprint 3 weeks early has never made a dollar from being right On Prophet he writes it as a Yes/No question with a resolution date, and the market goes live A consensus of 5 AI models sets the odds instantly and takes the other… https://twitter.com/Blum_OG/status/2081139709514744021/photo/1
media 0 共 1 项
#411 🔗 原推
2026-07-25 22:09:07
Blum
The guy in your group chat who called the Pokemon reprint 3 weeks early has never made a dollar from being right On Prophet he writes it as a Yes/No question with a resolution date, and the market goes live A consensus of 5 AI models sets the odds instantly and takes the other… https://twitter.com/Blum_OG/status/2081139427821183240/photo/1
media 0 共 1 项
#410 🔗 原推
2026-07-25 22:07:59
Blum
$1,000,000 trade, $0 taker fee The same trade at $10,000 costs about $4 on a typical CEX at 4 bps, and @TrueCurrentX subsidizes Injective gas on top, so opening, closing, depositing, and withdrawing cost nothing RFQ is how that math works: your order goes out to competing… https://twitter.com/Blum_OG/status/2080815682405499105/video/1
media 0 共 2 项
🎬
视频
#409 🔗 原推
2026-07-25 00:41:33
Blum
$1,000,000 trade, $0 taker fee The same trade at $10,000 costs about $4 on a typical CEX at 4 bps, and TrueCurrent subsidizes Injective gas on top, so opening, closing, depositing, and withdrawing cost nothing RFQ is how that math works: your order goes out to competing… https://twitter.com/Blum_OG/status/2080815452603765053/video/1
media 0 共 2 项
🎬
视频
#408 🔗 原推
2026-07-25 00:40:38
Blum
$1,000,000 trade, $0 taker fee The same trade at $10,000 costs about $4 on a typical CEX at 4 bps, and TrueCurrent subsidizes Injective gas on top, so opening, closing, depositing, and withdrawing cost nothing RFQ is how that math works: your order goes out to competing… https://twitter.com/Blum_OG/status/2080814951900282909/video/1
media 0 共 2 项
🎬
视频
#407 🔗 原推
2026-07-25 00:38:38
Blum
Ricker just broke down 7 features of Kimi K3, the first open model in the 3-trillion-parameter class It codes at Claude Fable 5 level, and the full weights land on Hugging Face July 27 So you can finally download a frontier-grade coder and fine-tune it on your own cluster… https://twitter.com/Blum_OG/status/2080744428688507366/video/1
media 0 共 2 项
🎬
视频
#406 🔗 原推
2026-07-24 19:58:24
Blum
Topview MCP. Now live. @TopviewAIhq Connecting data from multiple platforms. Amazon. YouTube. TikTok Shop. Shopee. Seamless intelligence. Infinite creativity. From data pulse to published brilliance — without leaving the flow
#405 🔗 原推
2026-07-24 14:19:58
Blum
I keep forgetting @TrueCurrentX is even a DEX Zero fees. No maker, no taker, no gas. Non-custodial on Injective, sub-second fills One app: crypto, stocks, gold, silver, oil, FX + prediction markets Its RFQ engine pulls live quotes from pro market makers and fills you at the… https://twitter.com/Blum_OG/status/2079962312857854258/video/1
media 0 共 2 项
🎬
视频
#404 🔗 原推
2026-07-22 16:10:33
Blum
Andrej Karpathy: "Your prompts are now programs that program the LLM" One English prompt becomes a JavaScript workflow that sends isolated jobs to parallel agents, passes typed JSON between them, and checks findings before the final answer https://twitter.com/Blum_OG/status/2079749918726438925/video/1
media 0 共 2 项
🎬
视频
#403 🔗 原推
2026-07-22 02:06:35
Blum
Mira Murati finally showed what she was building after OpenAI Thinking Machines released Inkling, its first model trained from scratch the model is not trying to become the strongest across every benchmark the main idea is to provide multimodality and flexibility and that’s… https://twitter.com/Blum_OG/status/2079657345039184026/video/1
media 0 共 2 项
🎬
视频
#402 🔗 原推
2026-07-21 19:58:43
Blum
AI AGENTS CROSS THE LINE AND HIDE IT Anthropic tested 14 frontier models in a simulated workplace and the results were shocking: - in the fraud test, 11 of 13 models disregarded the original requirements and followed a harmful request at least once - Claude LLM judges changed… https://twitter.com/Blum_OG/status/2079275689388216338/photo/1
media 0 共 1 项
#401 🔗 原推
2026-07-20 18:42:10
Blum
RT @Blum_OG: Grok is the best model for content creation on X but most people barely use 10% of its power the other 90% is covered here:…
#400 🔗 原推
2026-07-20 06:18:14
Blum
Grok is the best model for content creation on X but most people barely use 10% of its power the other 90% is covered here: 00:24 - finding live AI topics on X 01:05 - DeepSearch across web and X 02:15 - turning research into a script 03:04 - comparing Grok, ChatGPT, and… https://twitter.com/Blum_OG/status/2078911616171807014/video/1
media 0 共 2 项
🎬
视频
#399 🔗 原推
2026-07-19 18:35:28
Blum
RT @HarryTandy: https://x.com/i/article/2070228153566425088
#398 🔗 原推
2026-07-19 16:25:41
Blum
HEADS UP: GPT CAN DELETE ALL YOUR FILES 🚨 during one task, the model deleted the contents of Matt’s Mac and anyone could end up in the same situation HOW TO PREVENT THIS: 1. run agents inside a virtual machine or sandbox 2. keep physical backups 3. block rm -rf 4. add hooks…
#397 🔗 原推
2026-07-18 20:07:17
Blum
OpenAI packed key recent Codex updates into a 7-min video Ultra, appshots, browser edits, Sites, mobile tasks, and PRs all shown working: 00:25 - Codex and ChatGPT merge into one app 00:47 - GPT-5.6 Sol and Ultra 01:28 - computer use and appshots 02:16 - browser use and inline… https://twitter.com/Blum_OG/status/2078505456054468920/video/1
media 0 共 2 项
🎬
视频
#396 🔗 原推
2026-07-18 15:41:32
Blum
RT @Blum_OG: open models are getting damn close to the best closed models Kimi K3 is a 2.8T-parameter MoE with a 1M-token context window…
#395 🔗 原推
2026-07-18 06:51:04
Blum
open models are getting damn close to the best closed models Kimi K3 is a 2.8T-parameter MoE with a 1M-token context window and its benchmark scores back that up: > Program Bench: K3 77.8, GPT 5.6 Sol 77.6, Claude Fable 5 76.8 > SWE Marathon: K3 42.0, GPT 5.6 Sol 39.0, Claude… https://twitter.com/Blum_OG/status/2078087943818711217/video/1
media 0 共 2 项
🎬
视频
#394 🔗 原推
2026-07-17 12:02:29
Blum
taste is becoming the scarce skill in AI-assisted design AI knows spacing, hierarchy, typography, and color theory, yet the designer still decides what deserves to exist this 47-minute conversation explains where human judgment enters the design process 00:33 - meet the… https://twitter.com/Blum_OG/status/2078047908998295910/video/1
media 0 共 2 项
🎬
视频
#393 🔗 原推
2026-07-17 09:23:24
Blum
this 57-page paper extends the build’s context-compression section ACON shows how to rewrite compression prompts from paired agent runs across 3 benchmarks, peak token use fell by 26-54% with this method ACON also beat existing compression baselines on task success smaller… https://twitter.com/Blum_OG/status/2077779800517697593/photo/1
media 0 共 1 项
#392 🔗 原推
2026-07-16 15:38:02
Blum
RT @Blum_OG: every answer from an LLM carries fingerprints from several different training decisions data decides what the model has seen,…
#391 🔗 原推
2026-07-15 19:40:39
Blum
every answer from an LLM carries fingerprints from several different training decisions data decides what the model has seen, pretraining builds its raw capability, and fine-tuning shapes how it talks to you Andrej Karpathy explains the machinery behind that chain 00:00 - what… https://twitter.com/Blum_OG/status/2077420857287221733/video/1
media 0 共 2 项
🎬
视频
#390 🔗 原推
2026-07-15 15:51:43
Blum
RT @Blum_OG: OpenAI saves one of its smartest agent-building rules for page 31: decide when a human takes over the guide names 2 triggers…
#389 🔗 原推
2026-07-14 14:22:00
Blum
OpenAI saves one of its smartest agent-building rules for page 31: decide when a human takes over the guide names 2 triggers for human takeover: repeated failure and high-risk action that connects directly to the article’s draft and approve approach my first autonomy ladder… https://twitter.com/Blum_OG/status/2077018962622030037/photo/1
media 0 共 1 项
#388 🔗 原推
2026-07-14 13:14:44
Blum
one evolved skill cut Claude token use by 62% in a Kafka test the result comes from a 382-task study of procedural memory in AI agents when researchers rewrote the skills once, agents fully completed about 5 additional tasks per 100 attempts the one-person Claude business… https://twitter.com/Blum_OG/status/2076746589188526556/photo/1
media 0 共 1 项
#387 🔗 原推
2026-07-13 19:12:25
Blum
this 22-page study explains why agents burn tokens on failed fixes it studies 50 public loops and gives you a 5-level verifier ladder the Claude Code setup above gives you the runnable version this paper gives you the inspection method 1. deterministic check: exit code,… https://twitter.com/Blum_OG/status/2076363304205722057/photo/1
media 0 共 1 项
#386 🔗 原推
2026-07-12 17:49:23
Blum
trading used to mean 10 tabs and chart-checking like a full-time job now you can type one line: > show me what's moving today > check the best route across chains > set DCA, take-profit, or stop-loss before I enter with Quant AI, that sentence becomes a route, an execution… https://twitter.com/Blum_OG/status/2076313116913119461/photo/1
media 0 共 1 项
#385 🔗 原推
2026-07-12 14:29:57
Blum
GPT-5.6 JUST DROPPED OpenAI just launched the GPT-5.6 family for ChatGPT, Codex, and the API the global rollout started today, with full availability planned over the next 24 hours the family has 3 models: > Sol - the flagship for demanding code and long-running work >… https://twitter.com/Blum_OG/status/2075320186119274964/photo/1
media 0 共 2 项 media 1
#384 🔗 原推
2026-07-09 20:44:24
Blum
RT @alphabatcher: https://x.com/i/article/2075226442326269952
#383 🔗 原推
2026-07-09 19:42:44
Blum
save this paper if you're designing agentic engineering workflows the authors describe oversight as work with timing, tools, and failure modes they interviewed 17 experienced devs using agents for professional work, mostly several times a week the result is a 4-stage… https://twitter.com/Blum_OG/status/2075262171978588333/photo/1
media 0 共 1 项
#382 🔗 原推
2026-07-09 16:53:52
Blum
Boris Cherny, Head of Claude Code: "I talk to loop or I talk to a routine and it prompts Claude for me" the work moves one level up: > you stop writing the next prompt by hand > you design the thing that wakes Claude up, gives it work, checks the result, and decides whether it… https://twitter.com/Blum_OG/status/2074548471982424276/video/1
media 0 共 2 项
🎬
视频
#381 🔗 原推
2026-07-07 17:37:53
Blum
Andrew Wilson, Anthropic Applied AI architect: "all of a sudden running many sub-agents became really economical" Ryan Carson's Fable run shows how to use 1 parent session and 39 child sessions on a risky engineering program > 834 files > 31 PRs > prod data mutation > DB… https://twitter.com/Blum_OG/status/2074223595400597798/video/1
media 0 共 2 项
🎬
视频
#380 🔗 原推
2026-07-06 20:06:57
Blum
save this 20-page agent survey before you build your portfolio project the course says the hiring test is proof: a RAG app, an agent with tools, and a deployed product this paper gives the missing build map for the agent piece it splits LLM agents into 3 buckets: - tool use,… https://twitter.com/Blum_OG/status/2074182143090299255/photo/1
media 0 共 1 项
#379 🔗 原推
2026-07-06 17:22:14
Blum
a good career filter for AI work: ask whether the task gets smaller when the model arrives, or whether the job gets bigger because the model lets you take on harder work this HBS paper puts numbers on that split the paper breaks each job into everyday duties, asks which ones… https://twitter.com/Blum_OG/status/2073701860138242060/photo/1
media 0 共 1 项
#378 🔗 原推
2026-07-05 09:33:45
Blum
Diane Penn, Head of PM for Research at Anthropic: “Fable 5 can run for days on a single goal and stay coherent the entire way” if that sounds like a flex, it is also the warning Fable 5 is built for work that survives past the first reply the waste happens when you use it… https://twitter.com/Blum_OG/status/2073502652374069456/video/1
media 0 共 2 项
🎬
视频
#377 🔗 原推
2026-07-04 20:22:10
Blum
RT @Blum_OG: most first-agent guides stop after the loop this 33-page paper is the thing to read right after you get that loop working th…
#376 🔗 原推
2026-07-04 18:00:55
Blum
most first-agent guides stop after the loop this 33-page paper is the thing to read right after you get that loop working the roadmap above gives you the build order: tools, loop, done condition, error handling, verification, gates, logs Kapoor, Stroebl, Siegel, Nadgir, and… https://twitter.com/Blum_OG/status/2073342015337894122/photo/1
media 0 共 1 项
#375 🔗 原推
2026-07-04 09:43:51
Blum
Alex Albert, Research PM at Anthropic: "It's highly autonomous and can operate for days without intervention" Fable 5 can work for days, but only if it has memory it can read the article’s setup is simple: put your business context in an Obsidian vault, then let Claude Code… https://twitter.com/Blum_OG/status/2073112012297539887/video/1
media 0 共 2 项
🎬
视频
#374 🔗 原推
2026-07-03 18:29:54
Blum
RT @Blum_OG: the Fable 5 cost problem already has a research playbook it's a Stanford paper called FrugalGPT FrugalGPT gives 3 moves you…
#373 🔗 原推
2026-07-03 17:35:03
Blum
the Fable 5 cost problem already has a research playbook it's a Stanford paper called FrugalGPT FrugalGPT gives 3 moves you can paste into your own workflow: 1. prompt adaptation: shorter prompts, fewer examples, shared prompts for batches 2. LLM approximation: cache repeated… https://twitter.com/Blum_OG/status/2073047630440169644/photo/1
media 0 共 2 项 media 1
#372 🔗 原推
2026-07-03 14:14:05
Blum
RT @Blum_OG: read this paper before you schedule 20 agent workflows TheAgentCompany, it puts agents inside a small software company with…
#371 🔗 原推
2026-07-03 06:10:15
Blum
read SWE-Marathon before you give a coding agent the night shift the Fable 5 course says to set goals, run loops, delegate to cheaper workers, and demand proof SWE-Marathon shows why the proof has to be painfully specific the benchmark gives each task: - its own executable… https://twitter.com/Blum_OG/status/2072775736491147687/photo/1
media 0 共 1 项
#370 🔗 原推
2026-07-02 20:13:40
Blum
read this paper before you schedule 20 agent workflows TheAgentCompany, it puts agents inside a small software company with GitLab, Plane, RocketChat, ownCloud, a terminal, and coworkers they have to message the setup is painfully close to the one-person company dream: can… https://twitter.com/Blum_OG/status/2072625176005816818/photo/1
中文: 在安排20个代理工作流程之前阅读本文 TheAgent公司将代理人员送入一家小型软件公司 使用 GitLab、Plane、RocketChat、ownCloud、终端以及同事，他们需要发送消息 这个配置与单人公司梦寐以求的关系非常接近: 可以......
media 0 共 1 项
#369 🔗 原推
2026-07-02 10:15:24
Blum
the DSPy paper is worth saving if your LLM workflow still lives inside one giant prompt it’s 32 pages from Omar Khattab, Matei Zaharia, Christopher Potts & others the paper gives a clean build pattern for the exact problem in the piece: good output depends on the system around… https://twitter.com/Blum_OG/status/2072344147332366839/photo/1
中文: 如果你的LLM工作流程仍存在一个巨大的提示，那么DSPy纸就值得保存 共32页，来自Omar Khattab、Matei Zaharia、Christopher Potts & 其他 纸张为该片中的确切问题提供了清晰的构建模式: 良好的输出取决于周围的系统......
media 0 共 1 项
#368 🔗 原推
2026-07-01 15:38:41
Blum
RT @Blum_OG: https://x.com/i/article/2062964824523358208
#367 🔗 原推
2026-07-01 14:11:43
Blum
https://x.com/i/article/2062964824523358208
#366 🔗 原推
2026-07-01 08:08:03
Blum
RT @Blum_OG: Anthropic accidentally wrote a portfolio roadmap for engineers the career article gets the order right: Python, APIs, RAG, ag…
中文: RT @Blum_OG:Anthropic 意外地为工程师撰写了投资组合路线图 职业文章的顺序很合适:Python、API、RAG、ag...
#365 🔗 原推
2026-07-01 04:56:51
Blum
Anthropic accidentally wrote a portfolio roadmap for engineers the career article gets the order right: Python, APIs, RAG, agents, deployment this PDF gives you the missing map for the agent part Anthropic splits agentic systems into workflows and agents that distinction… https://twitter.com/Blum_OG/status/2071848076868817201/photo/1
中文: Anthropic 意外地为工程师撰写了投资组合路线图 职业文章的顺序是正确的:Python、API、RAG、代理、部署 此PDF文件为代理部分提供了缺失的地图 Anthropic 将代理系统划分为工作流程和代理 那个区别......
media 0 共 1 项
#364 🔗 原推
2026-06-30 06:47:29
Blum
Steven Johnson, NotebookLM co-creator: "the AI knows what I've been thinking" that's the bar for a paid AI setup now if the tool doesn't know your notes, voice memos, transcripts, client rules, and past decisions, it deserves a line-item review the article's math is blunt: >… https://twitter.com/Blum_OG/status/2071762639907148271/video/1
中文: 史蒂文·约翰逊，NotebookLM 联合创始人: 人工智能知道我一直在想什么 这就是现在付费人工智能设置的门槛 如果该工具不了解您的笔记、语音备忘录、成绩单、客户规则以及以往的决定，则值得进行逐项审核 文章的数学原理是直白的:
media 0 共 2 项
🎬
视频
#363 🔗 原推
2026-06-30 01:07:59
Blum
Mahesh Murag, Anthropic AI engineer: "as you interact with an agent and give it feedback and more institutional knowledge, it starts to get better" feedback only works when the agent has something real to absorb that is the point of the whole stack Obsidian captures the raw… https://twitter.com/Blum_OG/status/2071302654777582024/video/1
中文: 马赫什·穆拉格，人类人工智能工程师: 当你与代理人互动并给予反馈并掌握更多机构知识时，它开始变得越来越好 只有当中介有真正需要吸收的东西时，反馈才有效 这就是整个堆栈的要点 黑曜石捕捉到原始内容......
media 0 共 2 项
🎬
视频
#362 🔗 原推
2026-06-28 18:40:10
Blum
RT @Blum_OG: Ash Prabaker, Anthropic engineer: "tuning a standalone critic to be harsh is actually very tractable, but tuning a builder to…
中文: RT @Blum_OG:阿什·普拉贝克，人类工程师: 调整一个独立批评者的严厉态度实际上非常易处理，但调整一个构建者才能......
#361 🔗 原推
2026-06-28 07:28:25
Blum
Cole Medin, AI builder: “with natural language, we can ask our coding agent to build the agent, evaluate it, deploy it, harden it. everything is packaged together” Google Agents CLI makes the process feel less scattered you tell the coding agent what assistant you want then… https://twitter.com/Blum_OG/status/2071001972077977684/video/1
中文: 科尔·梅丁，人工智能构建者: 使用自然语言，我们可以要求编码代理构建代理，进行评估，部署它，使其硬化。所有内容都打包在一起。 谷歌代理商CLI让流程感觉不那么分散 你告诉编码代理你想要什么助手 那么......
media 0 共 2 项
🎬
视频
#360 🔗 原推
2026-06-27 22:45:22
Blum
Ash Prabaker, Anthropic engineer: "tuning a standalone critic to be harsh is actually very tractable, but tuning a builder to be somewhat self-critical is not." that’s the whole reason multi-agent systems need roles > builder writes > judge checks > manager decides when the… https://twitter.com/Blum_OG/status/2070956731631329732/video/1
中文: 阿什·普拉贝克，人类工程师: 将独立批评者调制为苛刻实际上非常易处理，但调整构建者以在某种程度上具有自我批评性却并非如此。 这就是多代理系统需要角色的原因 加法; builder 写 法官支票 经理决定何时......
media 0 共 2 项
🎬
视频
#359 🔗 原推
2026-06-27 19:45:35
Blum
RT @Blum_OG: Boris Cherny said this about Claude Code: "multiple agents, they have fresh context windows" and that explains why this 4-ag…
中文: RT @Blum_OG:鲍里斯·切尔尼谈到了克劳德·凯德: 多个代理，它们具有新的上下文窗口 这解释了为什么这个4分...
#358 🔗 原推
2026-06-26 17:12:21
Blum
Boris Cherny said this about Claude Code: "multiple agents, they have fresh context windows" and that explains why this 4-agent team performs so well one AI session writing code, testing it, and reviewing its own work will miss the stuff baked into its own path fresh context… https://twitter.com/Blum_OG/status/2070495741047366097/video/1
中文: 鲍里斯·切尔尼对克劳德·科德说: 多个代理，它们具有新的上下文窗口 这解释了为什么这支4号探员团队表现如此出色 一次人工智能会话会编写代码、测试并审视自己的工作，都会错过自己所开辟的路径 新背景......
media 0 共 2 项
🎬
视频
#357 🔗 原推
2026-06-26 13:13:47
Blum
Joe Holdcroft, Tessl security engineer: "text becomes something that can be vulnerable." when an ai builds your app, the attack surface is wider than visible code your prompts, docs, env files, APIs, auth flows, logs and database policies all matter once real users show up… https://twitter.com/Blum_OG/status/2070388497043571046/video/1
中文: 乔·霍尔德克罗夫特，蒂丝尔安全工程师 文字会变得脆弱。 当 i 构建你的应用时，攻击面比可见的代码更宽 提示、文档、env 文件、API、auth 流、日志和数据库策略 一旦真实用户出现，一切都会发生......
media 0 共 2 项
🎬
视频
#356 🔗 原推
2026-06-26 06:07:38
Blum
RT @Blum_OG: FABLE 5 IS BACK 🚨 AWS docs already show it as live again: > model lifecycle: active > model id: `anthropic.claude-fable-5`…
中文: RT @Blum_OG:Fable 5 回来了 🚨 AWS 文档已显示为实时的: > 模型生命周期:活跃 型号:`anthropic.claude-fable-5`
#355 🔗 原推
2026-06-26 05:40:57
Blum
Luke Alvoeiro, Factory-AI multi-agent systems engineer: "it uses a three-role architecture. there's orchestrator, there's workers and then there's validators." that line saves people from the worst swarm mistake: treating 10 agents like 10 copies of the same chat the better… https://twitter.com/Blum_OG/status/2070214183002468848/video/1
中文: 卢克·阿尔沃伊罗，Factory-AI 多代理系统工程师: 它采用三角色架构。有编排器，有工作人员，然后有验证器。 这条线可以帮助人们避免最严重的群体误会: 将10名特工当作同一聊天的10份 更好的......
media 0 共 2 项
🎬
视频
#354 🔗 原推
2026-06-25 18:34:58
Blum
FABLE 5 IS BACK 🚨 AWS docs already show it as live again: > model lifecycle: active > model id: `anthropic.claude-fable-5` small catch: new Anthropic/AWS customers may be asked to explain what they're building before they can run it before you plan around it, check 3… https://twitter.com/Blum_OG/status/2070134079589364025/photo/1
中文: FABLE 5 返回 🚨 AWS 文档已显示为实时的: > 模型生命周期:处于活跃状态 模型 id:`anthropic.claude-fable-5` 小陷阱: 新的Anthropic/AWS客户可能需要先说明他们正在建造什么，然后才能进行 在规划之前，请先查看3个......
media 0 共 1 项
#353 🔗 原推
2026-06-25 13:16:40
Blum
Boris Cherny on stop hooks in Claude Code: "you can just make the model keep going until the thing is done." the article’s point gets sharper when hooks are treated as checkpoints in the execution path project guidance can live in CLAUDEmd hard rules belong in code use hooks… https://twitter.com/Blum_OG/status/2069887753223819487/video/1
中文: 鲍里斯·切尔尼在《克劳德·科德》中的停弹道: 你可以让模型一直持续，直到事情完成。 当钩子在执行路径中被视为检查点时，文章的要点变得更加清晰 项目指导可以在克劳德德生活 硬规则属于代码 使用钩子......
media 0 共 2 项
🎬
视频
#352 🔗 原推
2026-06-24 20:57:51
Blum
Sam Altman, OpenAI CEO: "you have access to tools that can let you do what used to take teams of hundreds." research, writing, sales follow-up, ops, project tracking, SOPs all of this is now doable for one person the catch is context Claude needs business memory before it… https://twitter.com/Blum_OG/status/2069832469768270079/video/1
中文: 萨姆·阿尔特曼，OpenAI 首席执行官: 你可以使用工具，让你能够完成过去吸引数百支队伍的行动。 研究、写作、销售跟进、操作、项目跟踪、SOP 现在这一切对一个人来说已经可行了 捕获内容为上下文 克劳德需要商业记忆之前......
media 0 共 2 项
🎬
视频
#351 🔗 原推
2026-06-24 17:18:11
Blum
RT @Blum_OG: Andrej Karpathy: "Vibe Coding raises the floor. Agentic Engineering is about extrapolating the ceiling." anyone can get a d…
中文: RT @Blum_OG:安德烈·卡帕蒂: 维贝·科丁抬高了地板。 工程工程是关于推出天花板的。 任何人都可以得到一个......
#350 🔗 原推
2026-06-24 14:21:43
Blum
Andrej Karpathy: "Vibe Coding raises the floor. Agentic Engineering is about extrapolating the ceiling." anyone can get a demo agent to work companies need agents that work the ten-thousandth time that gap is the job, and it doesn't require a CS degree to fill it an AI… https://twitter.com/Blum_OG/status/2069737466622349345/video/1
中文: 安德烈·卡帕蒂: 维贝·科丁抬高了地板。 工程工程是关于推出天花板的。 任何人都可以让演示代理工作 企业需要能够工作一万次的代理 这个空缺是工作，而且不需要CS学位来填补 人工智能......
media 0 共 2 项
🎬
视频
#349 🔗 原推
2026-06-24 11:00:40
Blum
RT @Blum_OG: in a Dario Amodei interview: "90% of the end-to-end SWE tasks are written by the models." at that volume picking a model is…
中文: RT @Blum_OG:在达里奥·阿莫代伊的采访中: 90%的端到端SWE任务由模型编写。 在那卷上挑选模型就是......
#348 🔗 原推
2026-06-24 07:56:04
Blum
in a Dario Amodei interview: "90% of the end-to-end SWE tasks are written by the models." at that volume picking a model is the same as picking a vendor GLM-5.2 changes the bill without sacrificing quality the practical read: 1. use a stronger model for the first 10%: - spec… https://twitter.com/Blum_OG/status/2069549709904023633/video/1
中文: 在达里奥·阿莫代伊的采访中: 90%的端到端SWE任务由模型编写。 在那体积下挑选一个模型，就像挑选一个供应商一样 GLM-5.2 在不牺牲质量的前提下修改账单 实用阅读: 1. 使用更强的模型来达到前10%: - 规格......
media 0 共 2 项
🎬
视频
#347 🔗 原推
2026-06-23 22:34:35
Blum
Andrew Ng summed up agentic workflows with 4 patterns: "there are, I want to say, four major design patterns, which are reflection, tool use, planning, and multi-agent collaboration." that line is a good shortcut for reading any agent framework reflection means the model… https://twitter.com/Blum_OG/status/2069490119149863016/video/1
中文: 吴安德总结了具有四种模式的代理工作流程: 我想说，有四种主要的设计模式，即反思、工具使用、规划以及多智能协作。 这条线对于读取任何代理框架都是一个很好的快捷方式 倒影意味着这个模型......
media 0 共 2 项
🎬
视频
#346 🔗 原推
2026-06-23 18:37:48
Blum
Barry Zhang (runs agent infrastructure at Anthropic): "Don't build agents for everything. If you do find a good use case and want to build an agent, keep it as simple for as long as possible." he said this after showing how agent systems move from simple model calls, to… https://twitter.com/Blum_OG/status/2069175815883874694/video/1
中文: 张巴里(在Anthropic运营代理基础设施): 不要为一切而建立代理人。如果你确实找到了一款好的用箱，并希望打造一款代理，请尽可能长时间地保持简单。 他在展示了代理系统如何从简单的模型调用转向......之后就说了这话
media 0 共 2 项
🎬
视频
#345 🔗 原推
2026-06-22 21:48:52
Blum
RT @Blum_OG: Boris Cherny: “i don't prompt Claude anymore. my job is to write loops” that line is some of the most useful AI advice right…
中文: RT @Blum_OG:鲍里斯·切尔尼: 我不再提示克劳德了。我的工作就是写循环 这条线是一些最有用的人工智能建议......
#344 🔗 原推
2026-06-22 21:24:40
Blum
Boris Cherny: “i don't prompt Claude anymore. my job is to write loops” that line is some of the most useful AI advice right now one-off prompts are good for answers loops are good for work that needs to happen again without you babysitting it a strong loop has 4 parts: 1.… https://twitter.com/Blum_OG/status/2069123062188929345/video/1
中文: 鲍里斯·切尔尼: 我不再提示克劳德了。我的工作就是写循环 这条线是目前最有用的人工智能建议之一 一次性提示对答案有好处 循环对于无需照顾孩子就需要再次发生的工作是有益的 强循环有4个部分: 1. . . .
media 0 共 2 项
🎬
视频
#343 🔗 原推
2026-06-22 18:19:15
Blum
RT @Blum_OG: Andrej Karpathy: "think of it almost like an employee or an intern" that line is the clean way to read GLM-5.2 GLM-5.2 is b…
中文: RT @Blum_OG:安德烈·卡帕: 把它想象在员工或实习生里 这条线是阅读GLM-5.2的简洁方式 GLM-5.2 是 b...
#342 🔗 原推
2026-06-22 04:23:02
Blum
Andrej Karpathy: "think of it almost like an employee or an intern" that line is the clean way to read GLM-5.2 GLM-5.2 is built for repo-scale coding loops: give it the task, the repo, the tests, the constraints, and enough time to grind through the work the appeal is cost… https://twitter.com/Blum_OG/status/2068843959791870048/video/1
中文: 安德烈·卡帕蒂: 把它想象成员工或实习生 这条线是阅读GLM-5.2的简洁方式 GLM-5.2 是为 repo-scale 编码循环而构建的: 赋予它任务、回购、测试、约束以及足够的时间来完成工作 吸引力在于成本......
media 0 共 2 项
🎬
视频
#341 🔗 原推
2026-06-21 23:50:11
Blum
RT @Blum_OG: Head of Claude Code, Boris Cherny: "The way we build Claude Code is we build it to be hackable." that line explains why “cla…
中文: RT @Blum_OG:克劳德·凯德·鲍里斯·切尔尼负责人: 我们构建克劳德密码的方式是将其构建为可破解的。 那行解释了为什么“cla...
#340 🔗 原推
2026-06-21 22:38:46
Blum
Head of Claude Code, Boris Cherny: "The way we build Claude Code is we build it to be hackable." that line explains why “claude setup” matters so much the right setup order: 1. make a project first upload the product doc, audience doc, style rules, decision log, and success… https://twitter.com/Blum_OG/status/2068771868639691184/video/1
中文: 克劳德·科德·科德，鲍里斯·切尔尼: 我们构建克劳德密码的方式是将其构建为可破解的。 那行解释了“克劳德设置”为何如此重要 正确的设置顺序: 1. 先做一个项目 上传产品文档、受众文档、样式规则、决策日志和成功...
media 0 共 2 项
🎬
视频
#339 🔗 原推
2026-06-21 19:03:44
Blum
RT @Blum_OG: Greg Brockman (co-founder and president of OpenA): "you don't want to just have one instance of the model operating. you want…
中文: RT @Blum_OG:格雷格·布罗克曼(OpenA联合创始人兼总裁): 你不想只拥有一个模型运行实例。 你想要......
#338 🔗 原推
2026-06-21 02:52:02
Blum
Greg Brockman (co-founder and president of OpenA): "you don't want to just have one instance of the model operating. you want to have multiple, right? you want to be a manager not of an agent, but of agents." that's the whole shift behind loops instead of running one agent… https://twitter.com/Blum_OG/status/2068453941960908863/video/1
中文: 格雷格·布罗克曼(OpenA联合创始人兼总裁): 你不想只拥有一个模型运行实例。 你想多个，对吧? 你想成为一位不是经纪人，而是经纪人的经理。 这就是循环背后的全部转变 而不是运营一个代理......
media 0 共 2 项
🎬
视频
#337 🔗 原推
2026-06-20 22:00:24
Blum
Drew Bent (Education Lead at Anthropic): "the group that didn't use AI tools actually performed 17% better." he described an Anthropic study that split coding students into two groups one used AI tools, one didn't most people only use AI tools to get an answer they rarely… https://twitter.com/Blum_OG/status/2068421339224924549/video/1
中文: 德鲁·本特(《蚁科教育》的领先) 未使用人工智能工具的团队实际表现却高出17%。 他描述了一项将编程学生分为两组的人类学研究 一个使用了人工智能工具，一个没有 大多数人只使用人工智能工具来获取答案 他们很少......
media 0 共 2 项
🎬
视频
#336 🔗 原推
2026-06-20 19:50:51
Blum
RT @Blum_OG: "Agents interact with each other and come together to solve a task." Andrew Ng was describing multi-agent workflows the same…
中文: RT @Blum_OG:“代理人相互交流，并齐聚以解决一项任务。 安德鲁·吴正在描述多代理工作流程 同样的......
#335 🔗 原推
2026-06-20 05:30:30
Blum
"Agents interact with each other and come together to solve a task." Andrew Ng was describing multi-agent workflows the same pattern explains why model councils can feel sharper than one premium model answering alone one answer from one model gives you one set of blind spots… https://twitter.com/Blum_OG/status/2068056566813855959/video/1
中文: 特工们相互交流，并齐聚一堂，共同完成一项任务。 安德鲁·吴正在描述多代理工作流程 同样的模式解释了为什么模范议会能够比一个高级模型更敏锐地回答 一个模型给出的答案会给你带来一组盲点......
media 0 共 2 项
🎬
视频
#334 🔗 原推
2026-06-19 19:41:22
Blum
RT @Blum_OG: Andrej Karpathy on MCP: "it's a protocol of speaking directly to agents as this new consumer and manipulator of digital infor…
中文: RT @Blum_OG:Andrej Karpathy 在 MCP 上: 这是一种直接与代理机构联系的协议，是作为数字向新的消费者和操纵者......
#333 🔗 原推
2026-06-18 23:35:48
Blum
Boris Cherny said the quiet part out loud about agent work: "usually every night I have like a few thousand that are doing kind of deeper work." the useful question is what keeps those agents from making deeper mistakes the Kimi K2.6 run in this piece answers with a loop >… https://twitter.com/Blum_OG/status/2067690985526628717/video/1
中文: 鲍里斯·切尔尼说，关于特工工作的安静部分是如此: 通常每天晚上，我都会有几千人从事更深层次的工作。 有用的问题是，是什么让这些代理人避免犯更深层次的错误 基米K2.6在这篇文章中以循环回答 网址:
media 0 共 2 项
🎬
视频
#332 🔗 原推
2026-06-18 19:28:41
Blum
Andrej Karpathy on MCP: "it's a protocol of speaking directly to agents as this new consumer and manipulator of digital information." that is the cleanest way to think about MCP your coding agent is becoming a second worker inside the product it needs the same context you… https://twitter.com/Blum_OG/status/2067618342920167605/video/1
中文: 安德烈·卡帕蒂在《MCP》中的内容 这是一种直接面向代理机构的协议，即作为数字信息的新消费者和操纵者。 这是思考MCP最干净的方式 您的编码代理正在成为产品内部的第二名员工 需要你使用相同的语境......
media 0 共 2 项
🎬
视频
#331 🔗 原推
2026-06-18 14:40:02
Blum
RT @Blum_OG: "don't use frontier models for non-frontier problems." Satya Nadella (CEO of Microsoft) said that at a Hard Fork live intervi…
中文: RT @Blum_OG:不要使用前沿模型来解决非前沿问题。 萨提亚·纳德拉(微软首席执行官)表示，在“硬分叉”现场直播中......
#330 🔗 原推
2026-06-18 12:23:46
Blum
"don't use frontier models for non-frontier problems." Satya Nadella (CEO of Microsoft) said that at a Hard Fork live interview and it explains why the Microsoft AI certification path hits harder than another prompt course job market already has plenty of people who can ask… https://twitter.com/Blum_OG/status/2067343970351579155/video/1
中文: 不要使用前沿模型来解决非前沿问题。 萨提亚·纳德拉(微软首席执行官)在《硬分叉》现场采访中表示 这解释了为何微软的人工智能认证路径比另一个提示路线更加严重 就业市场已经有很多人可以问......
media 0 共 2 项
🎬
视频
#329 🔗 原推
2026-06-17 20:29:46
Blum
RT @alphabatcher: https://x.com/i/article/2064654776491765760
#328 🔗 原推
2026-06-17 15:47:44
Blum
Andrew Ng (Google Brain co-founder): "building agentic workflows requires ingesting external knowledge." that's why content agent should starts with research asking ai for post ideas with zero context usually gives the same flat answers everyone gets a better setup gives the… https://twitter.com/Blum_OG/status/2067255620529541309/video/1
中文: 安德鲁·吴(谷歌大脑联合创始人): 构建特制工作流程需要摄取外部知识。 这就是为什么内容代理应该从研究开始 以零语境的语文向Ai寻求帖子想法，通常给出与每个人相同的回答 更好的设置提供了......
media 0 共 2 项
🎬
视频
#327 🔗 原推
2026-06-17 14:38:42
Blum
Obsidian CEO Steph Ango on his product: "this is an opportunity to link these things." that is the missing layer in a lot of AI workflows > a chat can answer > a vault can remember > an agent sitting inside the vault can turn answers into durable context the article's setup… https://twitter.com/Blum_OG/status/2067225065309995439/video/1
中文: 黑曜石首席执行官斯蒂芬·安戈谈他的产品: 这是一个将这些事情联系起来的机会。 这是许多人工智能工作流程中缺失的一层 聊天可以回答 金库可以记住 存放在保险库内的代理可以将答案转化为持久的环境 文章的设置......
media 0 共 2 项
🎬
视频
#326 🔗 原推
2026-06-17 12:37:17
Blum
Dan Shipper put the work pattern in 7 words: "building software for humans and agents to use together" a revenue loop is one version of that: > a human picks the outcome > the agent runs the repeated work > the system decides when to keep going or stop take a stale deal -… https://twitter.com/Blum_OG/status/2066973088596787235/video/1
中文: 丹·希珀用7个字来表达工作模式: 供人类和代理共同使用的构建软件 收入循环是其中一个版本: 人类选择结果 代理人负责重复工作 系统决定何时继续或停止 拿一笔陈词滥散 - . . .
media 0 共 2 项
🎬
视频
#325 🔗 原推
2026-06-16 19:56:01
Blum
RT @Blum_OG: Andrej Karpathy: “the name of the game is how can you get more agents running for longer periods of time without your involve…
中文: RT @Blum_OG:安德烈·卡帕: 游戏的名称是如何让更多代理在没有参与的情况下长时间运行......
#324 🔗 原推
2026-06-16 19:26:03
Blum
RT @Blum_OG: Andrej Karpathy on why context matters: “anything that you give it as context goes directly into its working memory” Obsidia…
中文: RT @Blum_OG:Andrej Karpathy 关于上下文的重要性 任何你作为背景而给予它的，都会直接进入其工作记忆 黑曜石......
#323 🔗 原推
2026-06-16 17:58:14
Blum
Andrej Karpathy on why context matters: “anything that you give it as context goes directly into its working memory” Obsidian works as a context layer that just keeps getting smarter even when you're not touching it basically it's a vault Claude reads from, a memory it writes… https://twitter.com/Blum_OG/status/2066866464884211821/video/1
中文: 安德烈·卡帕蒂:为何背景重要: 任何你作为背景内容给予的内容，都会直接进入其工作记忆中 黑曜石作为一种语境层，即使你没有接触它，也能不断变得更聪明 基本上，这是一个克劳德从中读出的金库，它写着一个内存......
media 0 共 2 项
🎬
视频
#322 🔗 原推
2026-06-16 12:52:20
Blum
Andrej Karpathy: “the name of the game is how can you get more agents running for longer periods of time without your involvement doing stuff on your behalf.” that only works if you stop treating Claude Code like a chat window and give it a repeatable engineering loop the… https://twitter.com/Blum_OG/status/2066641308559474891/video/1
中文: 安德烈·卡帕蒂: 游戏的名称是如何让更多代理人员在不参与为你代为做事的情况下长时间运行。 只有当你不再把克劳德代码当作聊天窗口来对待时，这才有效 并给它一个可重复的工程循环
media 0 共 2 项
🎬
视频
#321 🔗 原推
2026-06-15 21:57:38
Blum
RT @Blum_OG: Boris Cherny put the agent shift plainly: "it actually uses tools. It acts in the world." that’s where Claude agents start g…
中文: RT @Blum_OG:鲍里斯·切尔尼明确说明了代理人的转变: 它实际上使用工具。它在世界上起作用。 克劳德特工们就在这里开始......
#320 🔗 原推
2026-06-15 19:47:22
Blum
Boris Cherny put the agent shift plainly: "it actually uses tools. It acts in the world." that’s where Claude agents start getting serious how to build one that survives real work: 1. write the task file - what it does - when it runs - what inputs count - where output goes -… https://twitter.com/Blum_OG/status/2066580431688925463/video/1
中文: 鲍里斯·切尔尼明确说明了代理人的转变: 它实际上使用工具。它在世界上起作用。 克劳德特工们开始认真认真起来 如何构建一个能够维持实际工作的: 1. 编写任务文件 它做什么 - 运行时 - 输入数量 - 输出位置 - . . .
media 0 共 2 项
🎬
视频
#319 🔗 原推
2026-06-15 17:55:44
Blum
RT @Blum_OG: Ash Prabaker (Anthropic Engineer): "give it the spec of what you're trying to achieve, and then let the loop iterate against…
中文: RT @Blum_OG:阿什·普拉贝克(人类工程师): 给它说明你想要实现的目标，然后让循环迭代来对抗......
#318 🔗 原推
2026-06-15 12:32:10
Blum
Ash Prabaker (Anthropic Engineer): "give it the spec of what you're trying to achieve, and then let the loop iterate against those criteria." the loop only works when it has: a home, memory, artifacts, and a check to iterate against 1. a durable thread give one workstream… https://twitter.com/Blum_OG/status/2066275784935800988/video/1
中文: 阿什·普拉贝克(人类工程师) 给它说明你想要实现的目标，然后让循环与这些标准形成反对。 循环只有在具有以下条件时才有效: 房屋、记忆、文物和检查，以抵御 1. 一根耐用的线 提供一个工作流......
media 0 共 2 项
🎬
视频
#317 🔗 原推
2026-06-14 21:45:11
Blum
RT @alphabatcher: Karpathy compressed the agent problem into 1 sentence: "If you can't evaluate then you can't auto research it, right?"…
中文: RT @alphabatcher:Karpathy 将代理问题压缩成1个句子: 如果你无法评估，那就无法进行汽车研究，对吧?......
#316 🔗 原推
2026-06-14 20:50:38
Blum
"I don't even talk to Claude. I have a Claude that's talking to my quads" that line from Boris Cherny gets close to the new job of the engineer the control loop is what lets coding agents survive stale context, failed tests, partial plans, and premature victory laps the… https://twitter.com/Blum_OG/status/2066194366989840829/video/1
中文: 我甚至不和克劳德说话。我有一个正在和我的四肢说话的克劳德 鲍里斯·切尔尼的那行内容接近工程师的新工作 控制循环是让编码代理能够在陈旧的语境、测试失败、部分计划以及过早获胜的循环中生存的
media 0 共 2 项
🎬
视频
#315 🔗 原推
2026-06-14 16:21:39
Blum
RT @Blum_OG: Boris Cherny (head of Claude Code): “I don't prompt Claude anymore. I have loops running that prompt Claude.” Peter Steinber…
中文: RT @Blum_OG:鲍里斯·切尔尼(克劳德·凯德·凯德负责人) 我不再催促克劳德了。我有循环在运行，提示克劳德。 彼得·斯坦伯......
#314 🔗 原推
2026-06-14 10:31:50
Blum
Dario Amodei (CEO of Anthropic) signed this warning: "Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." days later, Washington made the concern concrete on Jun 12, the US government… https://twitter.com/Blum_OG/status/2065938759540252889/photo/1
中文: 达里奥·阿莫代(安特罗派克公司首席执行官)签署了此警告: 将人工智能灭绝的风险与其他社会规模风险(如大流行病和核战争)一起，应成为全球优先事项。 几天后，华盛顿让担忧变得具体 6月12日，美国政府......
media 0 共 1 项
#313 🔗 原推
2026-06-13 23:25:58
Blum
Boris Cherny (head of Claude Code): “I don't prompt Claude anymore. I have loops running that prompt Claude.” Peter Steinberger (creator of OpenClaw): “You shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.” this approach… https://twitter.com/Blum_OG/status/2065829287362465925/video/1
中文: 鲍里斯·切尔尼(克劳德·科迪奇负责人) 我不再催促克劳德了。我有循环在运行，提示克劳德。 彼得·斯坦伯格(OpenClaw 的创作者): 你不应该再催促编程代理了。你应该设计提示你代理的循环。 这种方法......
media 0 共 2 项
🎬
视频
#312 🔗 原推
2026-06-13 16:10:58
Blum
RT @Blum_OG: every model has a blind spot you can't see it from inside one chat "the construction of LLM ensembles seems under-explored"…
中文: RT @Blum_OG:每个模特都有盲区 你无法从一个聊天中看到它 LLM套装的建造似乎未被探索过
#311 🔗 原推
2026-06-13 08:59:42
Blum
Sam Altman: "if you can build a product that is so good people spontaneously tell their friends about it, you have done 80% of the work" i'd read these 32 principles as a list of reasons people hesitate each principle removes one small reason to pause: - free plan: people use… https://twitter.com/Blum_OG/status/2065578107273093570/video/1
中文: 萨姆·阿尔特曼:“如果你能打造出一款非常优秀的产品，会自发地向朋友讲述，那么你已经完成了80%的工作。” 我会把这32条原则解读为人们犹豫不决的原因之一 每个原则都消除了一个暂停的小理由: - 免费计划:人们使用......
media 0 共 2 项
🎬
视频
#310 🔗 原推
2026-06-12 23:32:52
Blum
every model has a blind spot you can't see it from inside one chat "the construction of LLM ensembles seems under-explored" - Andrej Karpathy a test last week made this concrete > same Python function > 3 planted bugs > 4 models reviewing in parallel > Claude Opus 4.8 caught… https://twitter.com/Blum_OG/status/2065528676880449876/video/1
中文: 每个模型都有盲区 你无法从一个聊天中看到它 LLM套装的建造似乎未被探索 - 安德烈·卡帕蒂 上周的一次测试使这个具体 同一个Python函数 3个种子虫 4个模型并行评审 被抓到的是:Claude Opus 4.8......
media 0 共 2 项
🎬
视频
#309 🔗 原推
2026-06-12 20:16:26
Blum
Andrej Karpathy said it early: "the hottest new programming language is English" Claude Fable 5 / Mythos 5 makes that line concrete the prompt now carries the work spec, the stop rule, the budget choice, and the memory system the big change from Anthropic’s guide: Fable 5 is… https://twitter.com/Blum_OG/status/2065156977643192749/video/1
中文: 安德烈·卡帕蒂很早就说过: 最热门的新编程语言是英语 克劳德·法尔 5 / 神话5 使这条线变得具体 提示现在会执行工作规范、停止规则、预算选择以及内存系统 Anthropic指南带来的重大变化: 寓言5是......
media 0 共 2 项
🎬
视频
#308 🔗 原推
2026-06-11 19:39:26
Blum
june 23 is the deadline for testing Claude Fable 5 before subscription use moves to credits anthropic is including it at no extra cost on pro, max, team, and seat-based enterprise from june 9 through june 22 use that window on jobs opus 4.8 struggled to finish: 1. repo-wide… https://twitter.com/Blum_OG/status/2064803254324662612/video/1
中文: 6月23日是测试克劳德·法伯5的截止日期 订阅使用前会转向积分 从6月9日到6月22日，拟以专业、最大、团队和座位为企业的全功能，不附加任何额外费用 使用该窗口对职位运算为4.8，难以完成: 1. 全范围范围......
media 0 共 2 项
🎬
视频
#307 🔗 原推
2026-06-10 20:13:52
Blum
RT @Blum_OG: a rough note connected to 10 others outperforms a perfectly written one with none that one principle is the difference betw…
中文: RT @Blum_OG:一个与另外10个人相关的粗略音符 表现优于无笔写得完美 一个原则是区别 在那......
#306 🔗 原推
2026-06-10 15:21:43
Blum
a rough note connected to 10 others outperforms a perfectly written one with none that one principle is the difference between a vault that compounds and one that just accumulates here is the architecture that makes it compound: > INBOX - raw captures, empties every evening >… https://twitter.com/Blum_OG/status/2064692371262988351/video/1
中文: 一个与另外10个人相连的粗略音符 表现优于无笔写得完美 一个原则是区别 在一个化合物的保险库和一个只是积聚的保险库之间 这里是使其复合的架构: 加特;INBOX - 每晚原始捕获，清空 网址:
media 0 共 2 项
🎬
视频
#305 🔗 原推
2026-06-10 12:53:16
Blum
RT @Blum_OG: ANTHROPIC JUST DROPPED CLAUDE FABLE 5 first MYTHOS-class model available to EVERYONE > state-of-the-art on nearly all capabi…
中文: RT @Blum_OG: ANTHROPIC 刚刚推出了 CLAUDE FABLE 5 首个向所有人提供的MYTHOS级模型 几乎所有卡比都最先进的...
#304 🔗 原推
2026-06-10 07:08:02
Blum
8 months in. 600 notes. daily habit that mostly works the note you added yesterday has no idea what you captured in month two the architecture that fixes it: VAULT/ PERMANENT/ ← the notes COMPOUND-OUTPUTS/ connections/ ← auto-discovered nightly syntheses/… https://twitter.com/Blum_OG/status/2064467054129013049/video/1
中文: 8个月内。600条笔记。日常习惯大多有效 你昨天加了那张纸条，完全不知道在第二个月内拍到了什么 修复它的架构: VAULT/ 永久/← 注释 计算机-OUTPUTS/ 连接/← 自动发现的夜间 合成/...
media 0 共 2 项
🎬
视频
#303 🔗 原推
2026-06-09 21:57:56
Blum
ANTHROPIC JUST DROPPED CLAUDE FABLE 5 first MYTHOS-class model available to EVERYONE > state-of-the-art on nearly all capability benchmarks > highest score on FrontierCode - Cognition's production coding eval > beat Pokemon FireRed with vision only - no maps, no harness >… https://twitter.com/Blum_OG/status/2064410607189684497/video/1
中文: 特率刚刚被引援“克劳德”Fable 5 首个向所有人提供的MYTHOS级模型 在几乎所有能力基准上都采用最先进的 FrontierCode 上的最高分——Cognition 的制作编码 eval 仅以视觉为先，仅以视觉为先，无需地图，无需线束 网址:
media 0 共 2 项
🎬
视频
#302 🔗 原推
2026-06-09 18:13:38
Blum
RT @HarryTandy: Boris Cherny: "I don't prompt Claude anymore. I have loops running that prompt Claude and figure out what to do. My job is…
中文: RT @HarryTandy:鲍里斯·切尔尼:“我不再催促克劳德了。”我有循环运行，提示克劳德并弄清楚该做什么。我的工作是......
#301 🔗 原推
2026-06-09 17:24:11
Blum
prompt engineering had a ceiling you were always the one who had to write the next turn but loops genuinely changes the way you work with AI Boris Cherny (head of Claude Code at Anthropic): “I don't prompt Claude anymore. I have loops running that prompt Claude and figuring… https://twitter.com/Blum_OG/status/2064330976738652188/video/1
中文: 快速工程有天花板 你总是那个必须写下下一个转折的人 但循环确实改变了你使用人工智能的方式 鲍里斯·切尔尼(《安斯罗普》的克劳德·科迪奇负责人) 我不再催促克劳德了。我有循环运行，提示克劳德和思考......
media 0 共 2 项
🎬
视频
#300 🔗 原推
2026-06-09 12:57:12
Blum
me realizing I was still inside the loop > built the agent > opened a session > typed the prompt myself each time > thought "I should automate this" > watched Boris Cherny land 259 PRs in 30d without opening his IDE once > he deleted his IDE in November > reread the definition… https://twitter.com/Blum_OG/status/2064083743435042868/video/1
中文: 我意识到自己还在循环中 加底;建立了中介 > 开场 每次输入提示 认为“我应该自动化这个” 观看鲍里斯·切尔尼30d获得259张PR，却一次未打开 他于11月删除了自己的IDE 重读定义......
media 0 共 2 项
🎬
视频
#299 🔗 原推
2026-06-08 20:34:47
Blum
RT @Blum_OG: Claude Code's "done, tests pass" is a confidence claim, not a test run the model is trained to deliver fast running tests be…
中文: RT @Blum_OG:克劳德·迪·赛德的“完成，测试通过”是一种自信的主张，而不是一次试运行 该模型被训练以快速交付 运行测试是......
#298 🔗 原推
2026-06-08 19:16:23
Blum
Claude Code's "done, tests pass" is a confidence claim, not a test run the model is trained to deliver fast running tests before stopping isn't the default behavior. so it doesn't you spot the bug, ask Claude to check, it finds it that loop should have been one session 4… https://twitter.com/Blum_OG/status/2063962018017632637/video/1
中文: 克劳德·科德的“完成、考试通过”是一种自信的主张，而不是试运行 该模型被训练以快速交付 停止前运行测试并非默认行为。 你发现这个漏洞，让克劳德去检查，它找到了 那个循环本该是一次 4.
media 0 共 2 项
🎬
视频
#297 🔗 原推
2026-06-08 12:31:06
Blum
RT @Blum_OG: people keep asking which AI model to use for autonomous agents the model choice matters less than the memory architecture He…
中文: RT @Blum_OG:人们不断询问使用哪种AI模型来使用自主代理 模型选择比内存架构更重要 他......
#296 🔗 原推
2026-06-08 09:05:35
Blum
people keep asking which AI model to use for autonomous agents the model choice matters less than the memory architecture Hermes runs three memory layers: > short-term for the current session > working memory for active task context > `MEMORYmd` + `USERmd` that load at the… https://twitter.com/Blum_OG/status/2063717430204207302/video/1
中文: 人们不断询问要使用哪种人工智能模型来使用自主代理 模型选择比内存架构更重要 爱马仕拥有三层记忆: 当前会议的短期情况 >用于主动任务环境的工作记忆 在 加载到
media 0 共 2 项
🎬
视频
#295 🔗 原推
2026-06-07 20:19:12
Blum
Claude told you the bug was fixed 2 hours later the function it called never existed Anthropic engineer: "Every agent in production lies. We measured it. The good ones lie less, the great ones catch the lie before the user does." most devs learn to double-check everything and… https://twitter.com/Blum_OG/status/2063669903996125320/video/1
中文: 克劳德告诉你，这个漏洞是被修复的 两个小时后，它称之为从未存在过 人类工程师:“生产中的每一位代理人都是谎言。”我们测量了它。好的人少撒谎，而那些优秀的人，在用户面前就能追上谎言。 大多数开发人员会学会仔细检查所有内容 还有......
media 0 共 2 项
🎬
视频
#294 🔗 原推
2026-06-07 17:10:20
Blum
RT @Blum_OG: closing your laptop does not stop Claude unless you never set up Routines Claude shipped a full automation stack between Mar…
中文: RT @Blum_OG:关闭笔记本电脑并不能阻止克劳德 除非你从未设置过常规 克劳德在3月之间运送了一个完整的自动化栈......
#293 🔗 原推
2026-06-06 21:30:39
Blum
the hard part of multi-agent systems isn't building agents it's knowing which agent in a 20-step chain actually helped if you only get a result at the very end there's no obvious way to trace back who deserved credit EOM from Harvard solves this with one rule: > when agent B… https://twitter.com/Blum_OG/status/2063313653504565361/video/1
中文: 多代理系统的难点不是建筑代理 它知道20步链条中的哪个代理人确实提供了帮助 如果你只在最后结果 没有明显的方法可以追溯谁值得获得赞誉 哈佛大学的EOM通过一条规则解决了这个问题: 当代理人B......
media 0 共 2 项
🎬
视频
#292 🔗 原推
2026-06-06 17:34:44
Blum
closing your laptop does not stop Claude unless you never set up Routines Claude shipped a full automation stack between March and April /loop in March. Auto Mode March 24. Cloud Routines April 14 tier 1 - /loop (while your terminal is open) > type it once, Claude repeats it… https://twitter.com/Blum_OG/status/2063276453303579032/video/1
中文: 关闭笔记本电脑并不能阻止克劳德 除非你从未设置过常规 克劳德在3月至4月期间出货了一个完整的自动化系统 /3月循环。自动模式 3月24日。云版 4月14日 一级 - / 循环(当终端打开时) 输入一次，克劳德重复一遍......
media 0 共 2 项
🎬
视频
#291 🔗 原推
2026-06-06 15:06:54
Blum
RT @Blum_OG: $150k salary gap separates people who "know LangChain" from people who can build systems that survive production the gap is…
中文: RT @Blum_OG:15万美元的薪资差距将“了解LangChain”的人分开 能够构建在生产中存活的系统的人 差距是......
#290 🔗 原推
2026-06-05 18:17:08
Blum
$150k salary gap separates people who "know LangChain" from people who can build systems that survive production the gap is not about knowing more models or frameworks it comes down to one thing: the harness same Claude Opus 4.5, two different harnesses: > Claude Code… https://twitter.com/Blum_OG/status/2062914579156390255/video/1
中文: 15万美元的薪资差距使“了解朗链”的人分开 能够构建在生产中存活的系统的人 差距并不是关于了解更多模型或框架 归结于一件事:安全带 相同的克劳德·奥弗斯 4.5，两种不同的线束: 致我的代码......
media 0 共 2 项
🎬
视频
#289 🔗 原推
2026-06-05 15:08:57
Blum
watch two companies use the same frontier model one hooks it up to their systems the other hooks it up and hands it a skill library the second company has a different kind of asset > its agents know how to prep a sales call > how to read a customer escalation > when to… https://twitter.com/Blum_OG/status/2062882850072862879/video/1
中文: 观察两家公司采用相同的前沿模式 一个人把它连接到他们的系统上 另一个把它勾起来，递给它一个技巧库 第二家公司拥有另一种资产 其代理人员知道如何准备销售电话 如何解读客户的升级 何时......
media 0 共 2 项
🎬
视频
#288 🔗 原推
2026-06-05 13:02:52
Blum
RT @Blum_OG: coding with an AI agent is turn-based by default > Claude makes a change > you check it manually > type back, repeat the tea…
中文: RT @Blum_OG:使用AI代理进行编程默认为 格特;克劳德做出改变 你手动检查 返回，重复 茶......
#287 🔗 原推
2026-06-04 21:32:29
Blum
coding with an AI agent is turn-based by default > Claude makes a change > you check it manually > type back, repeat the teams moving fastest are not the ones with better prompts they're the ones who encoded their manual review steps as skills Claude can run itself > a… https://twitter.com/Blum_OG/status/2062559557226729692/video/1
中文: 使用AI代理进行编程默认为 格特;克劳德做出改变 你手动检查 返回，重复 速度最快的队伍并不是那些发出更好提示的队伍 他们正是将手动评审步骤编码为克劳德能够自行运行的技能的人 网址:
media 0 共 2 项
🎬
视频
#286 🔗 原推
2026-06-04 15:38:13
Blum
your agent app has three UI patterns most teams fall into one by accident > Controlled you pre-build React components, the agent picks which to render - token tax: ~400 tokens per tool, 25 components = 10k tokens per turn - agent starts picking the wrong component past 15 tools… https://twitter.com/Blum_OG/status/2062527973270458440/video/1
中文: 你的代理应用程序有三种用户界面模式 大多数队伍都因意外陷入其中 加以控制 你预先构建了 React 组件，即代理选择要渲染的组件 - 代币税:每台工具约400枚代币，25个组件，每折点10k - 代理开始选择错误的组件，超过15个工具......
media 0 共 2 项
🎬
视频
#285 🔗 原推
2026-06-04 13:32:43
Blum
RT @Blum_OG: 6 months ago, "self-improving local agent" required data center hardware this year it doesn't three things converged: Hermes…
中文: RT @Blum_OG:6个月前，需要使用数据中心硬件进行“自我改进” 今年没有 三件事汇合在脑后:爱马仕......
#284 🔗 原推
2026-06-03 23:51:02
Blum
6 months ago, "self-improving local agent" required data center hardware this year it doesn't three things converged: Hermes Agent, Qwen 3.6, and DGX Spark > Hermes saves completed tasks as skill files - and reuses them > Qwen 3.6 35B beats last year's 120B models at 1/3 the… https://twitter.com/Blum_OG/status/2062249214592036973/video/1
中文: 6个月前，需要“自我改进的本地代理”所需的数据中心硬件 今年没有 融合了三件事:Hermes Agent、Qwen 3.6 和 DGX Spark > Hermes 将已完成的任务保存为技能文件，并重复使用 Qwen 3.6 35B 比去年的 120B 型号 1/3 更胜 . . .
media 0 共 2 项
🎬
视频
#283 🔗 原推
2026-06-03 19:05:02
Blum
RT @HarryTandy: MOST AI APPS BREAK BEFORE THE FIRST LINE OF CODE Because the builder starts with implementation Not with the mental model…
中文: RT @HarryTandy:大多数AI应用在第一行代码之前就开始了 因为构建器从实现开始 与心理模型不同......
#282 🔗 原推
2026-06-02 19:53:59
Blum
QWEN3.7-PLUS JUST DROPPED > ran for 11+ hours straight > generated 10,000+ lines of code > triggered 1,000+ agent calls > shipped a full English vocabulary app from scratch and that is not the headline feature the model combines visual understanding & code execution in one… https://twitter.com/Blum_OG/status/2061851309683228840/video/1
中文: QWENG3.7-PLUS 刚刚掉头 连续跑步超过11小时 已生成10000多行代码 已触发1000多个代理电话 从零开始发送了一款完整的英语词汇应用程序 这并非标题特征 该模型将视觉理解与编程;代码执行合集在一个...
media 0 共 2 项
🎬
视频
#281 🔗 原推
2026-06-02 16:43:54
Blum
RT @Blum_OG: most AI builders skip the foundation and ship broken products here are the 10 concepts that make the difference: 1/ tokens -…
中文: RT @Blum_OG:大多数人工智能构建者会跳过基础并运送损坏的产品 以下是产生差异的10个概念: 1 个代币 -...
#280 🔗 原推
2026-06-02 15:36:52
Blum
most AI builders skip the foundation and ship broken products here are the 10 concepts that make the difference: 1/ tokens - everything is priced, limited, and rate-limited in tokens > 1,000 tokens ≈ 750 words. your prompt got cut off because you ran out 2/ embeddings - AI… https://twitter.com/Blum_OG/status/2061813339626381624/video/1
中文: 大多数人工智能构建者会跳过基础并运送破损的产品 以下是产生差异的10个概念: 1/ 代币——所有代币均为定价、限价和费率限制 1000个代币 ≈750字。你的提示被切断了，因为你用完了 2/ 嵌入 - 人工智能......
media 0 共 2 项
🎬
视频
#279 🔗 原推
2026-06-02 14:13:01
Blum
RT @Blum_OG: $25 per 1M output tokens is a lot especially when you are regenerating the same dashboard for the 4th time the real workflow…
中文: RT @Blum_OG:每100万个输出代币可赚25美元 尤其是当你第四次重新生成同一个仪表板时 真正的工作流程......
#278 🔗 原推
2026-06-02 05:48:53
Blum
$25 per 1M output tokens is a lot especially when you are regenerating the same dashboard for the 4th time the real workflow routes each model to what it is priced for > Kimi K2.6 at ~$0.95/$4 per 1M tokens handles the messy execution loop > Opus 4.8 at $5/$25 per 1M tokens… https://twitter.com/Blum_OG/status/2061556308332130693/video/1
中文: 每100万个输出代币25美元非常实惠 尤其是当你第四次重新生成同一个仪表板时 真正的工作流程将每个型号的价格都路由到 基米K2.6，每100万枚代币售价约0.95美元，可处理混乱的执行循环 网址:Opus 4.8，每100万枚代币售价5美元/25美元......
media 0 共 2 项
🎬
视频
#277 🔗 原推
2026-06-01 21:11:40
Blum
RT @Blum_OG: 5 folders. that is the entire structure you need your highlights, your Google Docs, your voice memos none of them surface wh…
中文: RT @Blum_OG:5 个文件夹。这就是你需要的整个结构 你的精彩集锦、谷歌文档、语音备忘录 它们都没有浮出水面......
#276 🔗 原推
2026-06-01 18:16:09
Blum
RT @Blum_OG: > downloaded Claude Code > thought it was just another AI chatbot > opened it up > found a chat bar > typed a prompt > nothing…
中文: RT @Blum_OG: > 下载 克劳德密码 我以为这只不过是另一个人工智能聊天机器人 快点;打开它 找到了一个聊天栏 输入提示 没什么......
#275 🔗 原推
2026-06-01 16:47:27
Blum
5 folders. that is the entire structure you need your highlights, your Google Docs, your voice memos none of them surface what you need when you need it Obsidian fixes this because notes are plain text files on your machine no subscription to access them. no company decision… https://twitter.com/Blum_OG/status/2061329109452493128/video/1
中文: 5个文件夹，即你需要的整个结构 你的精彩内容、你的谷歌文档、你的语音备忘录 当你需要时，它们都不会暴露出你需要的东西 黑曜石修复了这一点，因为笔记是机器上的纯文本文件 无需订阅即可访问它们。无公司决定......
media 0 共 2 项
🎬
视频
#274 🔗 原推
2026-06-01 06:08:51
Blum
RT @Blum_OG: 90 days from now you either have a system that runs without you or you are still manually doing the same tasks. that gap is…
中文: RT @Blum_OG:从现在起90天后，你要么拥有一个没有你运行的系统 或者你仍在手动执行相同的任务。 那个差距是......
#273 🔗 原推
2026-05-31 20:10:22
Blum
one markdown file. that is the entire subagent Claude Code subagents are separate Claude instances inside your session each gets its own context window, does one task returns only the summary without them: > one session reads 40 files, generates code, reviews it, runs tests… https://twitter.com/Blum_OG/status/2061116862390231287/video/1
中文: 一个markdown文件。即整个子代理 克劳德代码子代理是会话中的单独克劳德实例 每个都有自己的上下文窗口，完成一个任务 仅返回摘要 没有他们: 一次会话读取40个文件，生成代码，进行评测，运行测试......
media 0 共 2 项
🎬
视频
#272 🔗 原推
2026-05-31 16:05:28
Blum
one markdown file. that is the entire subagent Claude Code subagents are separate Claude instances inside your session each gets its own context window, does one task returns only the summary without them: > one session reads 40 files, generates code, reviews it, runs tests… https://twitter.com/Blum_OG/status/2061116467890749497/video/1
中文: 一个markdown文件。即整个子代理 克劳德代码子代理是会话中的单独克劳德实例 每个都有自己的上下文窗口，完成一个任务 仅返回摘要 没有他们: 一个会话读取40个文件，生成代码，进行审核，运行测试......
media 0 共 2 项
🎬
视频
#271 🔗 原推
2026-05-31 16:03:54
Blum
90 days from now you either have a system that runs without you or you are still manually doing the same tasks. that gap is what Hermes Agent is actually about it is not a chatbot. it is infrastructure 4 things separate it from every other agent framework: > persistent memory… https://twitter.com/Blum_OG/status/2061052353181053412/video/1
中文: 90天后，你要么拥有一个没有你运行的系统 或者你仍在手动执行相同的任务。 那个空白正是爱马仕代理公司真正的 它不是聊天机器人，而是基础设施 将其与其他所有代理框架区分开来的4件事: 持久性记忆......
media 0 共 2 项
🎬
视频
#270 🔗 原推
2026-05-31 11:49:08
Blum
RT @Blum_OG: > spent years switching between 5 apps to do 1 thing > WhatsApp to find the context > Gmail to find the email > browser to ver…
中文: RT @Blum_OG: > 花了数年时间在5个应用程序之间切换，以完成1项操作 WhatsApp 查找上下文 邮箱:Gmail 查找邮件 浏览器到...
#269 🔗 原推
2026-05-30 21:41:22
Blum
> spent years switching between 5 apps to do 1 thing > WhatsApp to find the context > Gmail to find the email > browser to verify the startup > Gmail again to draft the intro > WhatsApp again to confirm it was sent > 20 mins for a 2-sentence intro email > not hard work > just… https://twitter.com/Blum_OG/status/2060771504288071776/video/1
中文: 花费数年时间在5个应用程序之间切换，以完成1项操作 WhatsApp 查找上下文 邮箱:Gmail 查找邮件 浏览器以验证启动 再次通过Gmail起草该介绍 WhatsApp 再次确认已发送 加点;20分钟，用于两语静的介绍邮件 不辛苦 加时以回;仅......
media 0 共 2 项
🎬
视频
#268 🔗 原推
2026-05-30 17:13:08
Blum
> we gave agents more memory > they got more confident > but less accurate Andrej Karpathy: "you rarely ever write or edit the wiki manually; it's the domain of the LLM" most company AI stalls after 3 weeks for the same reason calls sit in Gong. SOPs sit in Notion corrections… https://twitter.com/Blum_OG/status/2060725363295068183/video/1
中文: 我们给了代理更多的记忆 他们变得更加自信了 但准确性较低 安德烈·卡帕蒂:“你很少会手动编写或编辑维基;它是LLM的域名 大多数公司的人工智能在三周后也会因同样的原因而停滞不前 电话在龚坐。SOP 在 NOTION 中 更正......
media 0 共 2 项
🎬
视频
#267 🔗 原推
2026-05-30 14:09:47
Blum
RT @Blum_OG: Anthropic is rolling out Claude Mythos "in the coming weeks" the most capable hacking model ever built. going live soon the…
中文: RT @Blum_OG:Anthropic 将在未来几周内推出 Claude Mythos 有史以来最强大的黑客模式。即将上线 ......
#266 🔗 原推
2026-05-30 05:56:06
Blum
Anthropic is rolling out Claude Mythos "in the coming weeks" the most capable hacking model ever built. going live soon the model found hidden flaws in every major OS and every major browser thousands of them. many critical > 83.1% on CyberGym vulnerability reproduction >… https://twitter.com/Blum_OG/status/2060431440085418396/video/1
中文: Anthropic 将在未来几周内推出 Claude Mythos 有史以来最强大的黑客模式。即将上线 该模型发现每个主要操作系统和每个主要浏览器都存在隐藏缺陷 成千上万个。许多至关重要 在 CyberGym 漏洞复制方面占 83.1% 网址:
media 0 共 2 项
🎬
视频
#265 🔗 原推
2026-05-29 18:41:50
Blum
RT @Blum_OG: the people getting weak Claude answers are not using the wrong model > they are opening a new chat > asking one question > ge…
中文: RT @Blum_OG:克劳德回答的人并没有使用错误的模型 他们正在开启一场新的聊天 提问 至上;g...
#264 🔗 原推
2026-05-29 17:56:43
Blum
Apple's Siri has been nearly 15 years old and mostly unchanged iOS 27 changes that. ships as early as September the biggest AI overhaul in Apple's history here is what is actually changing: > AI-powered web search - built into Search or Ask interface > rebuilt on Google… https://twitter.com/Blum_OG/status/2060377662648930353/video/1
中文: 苹果的Siri已有近15年历史，且基本未变 iOS 27 会更改，最早将于九月发布 苹果历史上规模最大的人工智能变革 实际变化的是什么: 支持人工智能的网页搜索——内置于搜索或提问界面中 在谷歌上重建......
media 0 共 2 项
🎬
视频
#263 🔗 原推
2026-05-29 15:08:09
Blum
the people getting weak Claude answers are not using the wrong model > they are opening a new chat > asking one question > getting one answer > starting from zero the next day > no personal instructions > no project context > no reference files > no repeatable workflows > no… https://twitter.com/Blum_OG/status/2060297064785809593/video/1
中文: 得到弱势克劳德回答的人并没有使用错误的模式 他们正在开启一场新的聊天 提问 获取一个答案 从零开始第二天 没有个人指示 没有项目背景 没有参考文件 没有可重复的工作流程 不......
media 0 共 2 项
🎬
视频
#262 🔗 原推
2026-05-29 09:47:53
Blum
RT @Blum_OG: Anthropic just dropped Opus 4.8 > sharper judgment > more honesty about its own progress > works independently for longer > s…
中文: RT @Blum_OG:Anthropic 刚刚退出了 Opus 4.8 更敏锐的判断力 更诚实地对待自身进展 工作时间更长 至
#261 🔗 原推
2026-05-28 22:52:39
Blum
Anthropic just dropped Opus 4.8 > sharper judgment > more honesty about its own progress > works independently for longer > same price as before - fast mode: same model, ~2.5x speed, 3x cheaper than before turn it on with /fast in Claude Code API access via waitlist at… https://twitter.com/Blum_OG/status/2060077822455251216/video/1
中文: Antropic 刚刚下跌了 Opus 4.8 更敏锐的判断力 更诚实地对待自身进展 工作时间更长 价格与之前相同 - 快速模式:相同型号，约2.5倍速度，比之前便宜3倍 使用克劳德密码中快速打开 通过候补名单获取API服务:
media 0 共 2 项
🎬
视频
#260 🔗 原推
2026-05-28 19:16:41
Blum
> ran 20 agents > felt completely productive > shipped almost nothing > the problem is not discipline > it is architecture starting an agent is one keystroke closing the loop on it is not cheap at all > you are the GIL of your AI agents > they all run at once > any work… https://twitter.com/Blum_OG/status/2060041935595647434/video/1
中文: 运营20名代理 感觉完全有效率 几乎什么也没发货 问题不在于纪律 它是建筑 启动代理是一个按键 关闭它的循环根本不便宜 你就是你人工智能代理的关键 它们都同时运行 任何工作......
media 0 共 2 项
🎬
视频
#259 🔗 原推
2026-05-28 16:54:05
Blum
3,575 char. that is the entire memory budget for a Hermes session two open-source coding agents two very different bets on what memory should do Hermes: memory frozen at session start OpenClaw: re-injected fresh every single turn > Hermes: ~1,300 tokens, never changing… https://twitter.com/Blum_OG/status/2059731232720671161/video/1
中文: 3，575 个字符。这是爱马仕会议的全部内存预算 两个开源的编码代理 记忆应该做什么，有两种截然不同的赌注 爱马仕:会话开始时记忆冻结 OpenClaw:每转一弯都重新注入新鲜剂 > 爱马仕:约1300个代币，永不改变......
media 0 共 2 项
🎬
视频
#258 🔗 原推
2026-05-27 20:19:28
Blum
RT @Blum_OG: Jensen Huang said it twice, at two different conferences first: "the IT department of every company is going to be the HR dep…
中文: RT @Blum_OG:黄仁勋在两个不同的会议上两次表示 首先:“每家公司的IT部门都将成为人力资源部门。
#257 🔗 原推
2026-05-27 14:28:13
Blum
Jensen Huang said it twice, at two different conferences first: "the IT department of every company is going to be the HR department of AI agents in the future." then: "you're not going to lose your job to an AI. but you're going to lose your job to someone who uses AI." he… https://twitter.com/Blum_OG/status/2059598223778480186/video/1
中文: 黄仁勋在两个不同的会议上说了两次 首先:“未来，每家公司的IT部门都将成为人工智能代理的人力资源部门。” 然后:“你不会因为人工智能而失去工作，但你会因为一个使用人工智能的人而失去工作。” 他......
media 0 共 2 项
🎬
视频
#256 🔗 原推
2026-05-27 11:30:56
Blum
RT @Blum_OG: > ran 5 tools for 2 years > Notion, Readwise, X bookmarks, browser bookmarks, ChatGPT > each one did its job fine > none of th…
中文: RT @Blum_OG: > 运行了5个工具，持续了2年 内容:注、阅读、X 书签、浏览器书签、ChatGPT 每个人的工作都做得不错 不，没有......
#255 🔗 原推
2026-05-27 09:37:27
Blum
> ran 5 tools for 2 years > Notion, Readwise, X bookmarks, browser bookmarks, ChatGPT > each one did its job fine > none of them talked to each other > every connection had to happen in my head > manually, every time, for every session > 34 Notion projects, 11 untouched for 3… https://twitter.com/Blum_OG/status/2059378568371577221/video/1
中文: 加长了5个工具，持续了2年 内容:Notion、Readwise、X 书签、浏览器书签、ChatGPT 每个人的工作都做得不错 他们彼此交谈都没有 我脑子里必须发生每一次联系 每次手动，每次 内容:34个，有11个未受影响，3个......
media 0 共 2 项
🎬
视频
#254 🔗 原推
2026-05-26 20:58:06
Blum
RT @Blum_OG: 43,000 followers. 9 posts a day. no filming, no writing. tjrmindset clips one creator's long videos into 30-second moments a…
中文: RT @Blum_OG:4.3万粉丝。每天发布9篇。不拍戏，不写文字。 特米德塞特将一位创作者的长视频片段剪辑成30秒瞬间 a...
#253 🔗 原推
2026-05-26 18:31:13
Blum
43,000 followers. 9 posts a day. no filming, no writing. tjrmindset clips one creator's long videos into 30-second moments and reposts them. that's the whole business the author grew a channel to 168,000 views in four weeks by hand got shadowbanned from posting too fast, and… https://twitter.com/Blum_OG/status/2059230534354542923/video/1
中文: 43000名粉丝。每天9次发布。不拍戏，不写。 特米德塞特将一位创作者的长视频片段剪辑成30秒瞬间 并转发它们。这就是全部的事 作者在四周内将频道的浏览量增长到16.8万次 因发布速度过快而被蒙上阴影，以及......
media 0 共 2 项
🎬
视频
#252 🔗 原推
2026-05-26 11:09:52
Blum
RT @Blum_OG: this guy pulled apart the full GPT/Claude pipeline 5 stages. architecture is the least of them > stage 1 - pretraining: mode…
中文: RT @Blum_OG:这家伙把完整的GPT/Claude管道拆了 5个阶段。建筑是其中最少的 第一阶段 - 预训练模式...
#251 🔗 原推
2026-05-26 08:59:13
Blum
300 parallel agents. one brief. a folder of finished files by morning that is what Kimi K2.6 Agent Swarm actually delivers the unit of work is no longer a response - it is a project > 40 social psychology PDFs in. 100-page academic document out > 1 CV uploaded. 100 customized… https://twitter.com/Blum_OG/status/2059058175102312627/video/1
中文: 300 个并行代理。一个简报。一个文件文件夹，到早上完成 这正是基米K2.6特工斯威姆实际提供的 工作单位不再是回应——而是一个项目 40 份社会心理学PDF文件，包含100页学术文档 已上传1份简历。100份定制版......
media 0 共 2 项
🎬
视频
#250 🔗 原推
2026-05-25 23:44:59
Blum
this guy pulled apart the full GPT/Claude pipeline 5 stages. architecture is the least of them > stage 1 - pretraining: model reads the entire internet, predicts next word > stage 2 - data: 250B pages filtered down to clean, quality tokens > stage 3 - scaling: chinchilla math… https://twitter.com/Blum_OG/status/2058988102547284050/video/1
中文: 这家伙把完整的GPT/Claude管道拆掉了 5个阶段。建筑是其中最少的 第一阶段 - 预训练:模型读取整个互联网，预测下一个词 第二阶段 - 数据:250B页被过滤以清洁优质代币 第三阶段 - 缩放:奇奇拉数学......
media 0 共 2 项
🎬
视频
#249 🔗 原推
2026-05-25 19:06:32
Blum
one .claude/agents/ folder. 7 files. that is the entire setup. asking one AI session to be: architect + backend + frontend + QA + reviewer compounds mistakes silently wrong assumption in the plan > becomes wrong database model > becomes wrong API > becomes wrong UI the… https://twitter.com/Blum_OG/status/2058894939841294656/video/1
中文: 一个 .claude/agents/ 文件夹。7 个文件。这是整个设置。 要求一次人工智能会话是: 建筑师 + 后端 + 前端 + 质量保证 + 审核员 悄无声息地使错误复杂化 计划中的错误假设 成为错误的数据库模型 > 变得错误 API 变得错误
media 0 共 2 项
🎬
视频
#248 🔗 原推
2026-05-25 12:56:20
Blum
RT @Blum_OG: this guy used Claude Code for 12 months then switched to Codex for 30 days here is what no tutorial tells you about the swit…
中文: RT @Blum_OG:这家伙用了12个月的克劳德密码 然后切换到Codex 30天 以下是没有教程告诉你的关于这个转机的内容......
#247 🔗 原推
2026-05-25 08:31:48
Blum
this guy used Claude Code for 12 months then switched to Codex for 30 days here is what no tutorial tells you about the switch: Codex is not a faster autocomplete > reads your codebase, writes and tests code > controls your Mac desktop, browses the web > remembers your… https://twitter.com/Blum_OG/status/2058653047241220488/video/1
中文: 这家伙用了12个月的克劳德密码 然后切换到Codex 30天 以下是关于该开关的无教程说明的内容: Codex 不是一个更快的自动完成程序 读取您的代码库、编写和测试代码 控制你的Mac桌面，浏览网页 记得你的......
media 0 共 2 项
🎬
视频
#246 🔗 原推
2026-05-24 20:55:09
Blum
RT @HarryTandy: > everyone in 2026 says they’re building ai agents > most are just wrapping a prompt in a fancy UI > you read one long…
中文: RT @HarryTandy: > 2026年，每个人都表示正在建立一家新闻机构 大多数只是用一个精美的界面包装一个提示 你读了一本长书......
#245 🔗 原推
2026-05-24 18:25:36
Blum
RT @Blum_OG: nobody audits their Сlaude setup here's what that costs: > your bill creeps > your output drifts > you blame the model here…
中文: RT @Blum_OG:没有人会审核他们的 Сlaude 设置 费用是以下内容: 你的账单会悄悄走 你的输出波移 你责怪这个模型 在这里......
#244 🔗 原推
2026-05-24 18:11:30
Blum
nobody audits their Сlaude setup here's what that costs: > your bill creeps > your output drifts > you blame the model here's what actually moves the needle: > 125+ keys exist in Claude Code's settings.json. official docs cover ~40 > fixing one cache_control breakpoint cut a… https://twitter.com/Blum_OG/status/2058504155963027495/video/1
中文: 没人会审核他们的 Сlaude 设置 费用是以下内容: 你的账单会悄悄走下来 你的输出波移 你责怪这个模型 真正推动针头的是什么: 在Claude Code的 settings.json 中存在 125 多个密钥。官方文档涵盖 ~40 修复一个缓存，控制断点会切断一个...
media 0 共 2 项
🎬
视频
#243 🔗 原推
2026-05-24 11:03:30
Blum
RT @Blum_OG: most people pick models by benchmark scores that's not where the real gap shows up the gap shows up in long agentic loops w…
中文: RT @Blum_OG:大多数人按基准评分选择模型 真正的差距并不在此 空白出现在长的特化循环中 w...
#242 🔗 原推
2026-05-23 21:37:17
Blum
most people pick models by benchmark scores that's not where the real gap shows up the gap shows up in long agentic loops where a model rewrites itself over and over 3 models were tested on a task: build Tetris bot that plays and trains itself each model could: > read its… https://twitter.com/Blum_OG/status/2058280726622007567/video/1
中文: 大多数人按基准分数选择模型 真正的差距并不在此 空白出现在长的特效循环中 一个模型会一遍又一遍地重写 在一项任务上测试了3个模型:构建可自行运行和训练的俄罗斯方块机器人 每个模型都可以: 阅读内容......
media 0 共 2 项
🎬
视频
#241 🔗 原推
2026-05-23 20:15:40
Blum
"everyone uses AI. almost nobody understands how it works." that gap is real - and it's the whole point here's what the article actually explains: > tokens are reusable building blocks - 1 token ≈ 0.75 words > attention lets every word look at every other word at once > this… https://twitter.com/Blum_OG/status/2058190084369752476/video/1
中文: 每个人都使用人工智能，几乎没人理解它是如何运作的。 那个差距是真实存在的——而正是关键 文章实际上解释的内容是: > 令牌是可重复使用的积木 - 1个令牌 ≈ 0.75 字 注意力让每个词同时观察每一个单词 网址:
media 0 共 2 项
🎬
视频
#240 🔗 原推
2026-05-23 14:15:30
Blum
RT @Blum_OG: you're probably using 10% of what Claude can do not because the features are hidden because no one told you they exist > pr…
中文: RT @Blum_OG:你可能正在使用克劳德能做的10% 并非因为这些功能是隐藏的 因为没有人告诉你它们存在 加以;请向...
#239 🔗 原推
2026-05-22 22:33:18
Blum
you're probably using 10% of what Claude can do not because the features are hidden because no one told you they exist > projects hold your documents and instructions permanently - no re-explaining yourself every session > artifacts produce working apps inside the chat - habit… https://twitter.com/Blum_OG/status/2057923019688005907/video/1
中文: 你可能正在使用克劳德能做的10% 并非因为这些功能是隐藏的 因为没有人告诉你它们存在 项目永久保存您的文件和说明 - 每次会话都无需重新解释 > 工件在聊天中生成工作应用 - 习惯......
media 0 共 2 项
🎬
视频
#238 🔗 原推
2026-05-22 20:34:16
Blum
most people use AI like a search box ask, get answer, close tab, repeat here's what a system looks like instead: > collect: x search + web search + your bookmarks > filter: strip duplicates, noise, engagement bait > map: themes, actors, competing claims > verify: primary… https://twitter.com/Blum_OG/status/2057829425572368692/video/1
中文: 大多数人使用人工智能就像搜索框 询问，获取答案，点击标签页，重复 系统看起来是什么样子: 收集:x 搜索 + 网页搜索 + 您的书签 过滤器:脱衣舞、噪音、订婚诱饵 地图:主题、演员、相互竞争的主张 网址:
media 0 共 2 项
🎬
视频
#237 🔗 原推
2026-05-22 14:22:22
Blum
RT @Blum_OG: > ran a dev agency > 10-15 people on payroll > project manager, devs, designer, QA, copywriter > client pays $15k - 50k per pr…
中文: RT @Blum_OG: > 经营一家开发机构 加薪，10至15人 >项目经理，开发人员，设计师，质量保证，文案撰写 客户支付15，000美元——每笔5万美元...
#236 🔗 原推
2026-05-22 05:36:11
Blum
> ran a dev agency > 10-15 people on payroll > project manager, devs, designer, QA, copywriter > client pays $15k - 50k per project > most of it goes to salaries > then Kimi K2.6 dropped > 1 trillion parameters, 32B activated per token > SWE-Bench score of 65.8 > solves nearly… https://twitter.com/Blum_OG/status/2057556999244800135/video/1
中文: 经营一家开发机构 加薪，10至15人 >项目经理，开发人员，设计师，质量保证，文案撰写 客户支付15，000美元，每个项目5万 大部分都与薪资有关 随后，基米K2.6下跌 每枚代币激活1万亿个参数，32B 英镑;瑞典-标准队得分为65.8分 几乎解决了问题......
media 0 共 2 项
🎬
视频
#235 🔗 原推
2026-05-21 20:19:50
Blum
RT @Blum_OG: everyone's debating which AI is "smartest." wrong question the real question is: who's winning on all the dimensions that ac…
中文: RT @Blum_OG:每个人都在争论哪个人工智能是“最聪明的”。 真正的问题是: 谁在所有维度上获胜......
#234 🔗 原推
2026-05-21 19:26:06
Blum
everyone's debating which AI is "smartest." wrong question the real question is: who's winning on all the dimensions that actually matter? > Grok 4 Heavy: first model to break 50% on Humanity's Last Exam > ARC-AGI v2: 15.9% vs Claude Opus's 8.6% > $1.25/million input tokens -… https://twitter.com/Blum_OG/status/2057427357376405676/video/1
中文: 每个人都在争论哪个人工智能是“最聪明的”。 真正的问题是: 谁在真正重要的维度上获胜? Grok 4 Heavy:人类最后一次考试中首个突破50%的模型 ARC-AGI v2:15.9% 对阵克劳德·奥图斯的8.6% 输入令牌:125万美元，网址:
media 0 共 2 项
🎬
视频
#233 🔗 原推
2026-05-21 11:44:41
Blum
RT @Blum_OG: if you're building AI agents, you've already hit this: > agent loses context after a few steps > tasks, memory, logs - all in…
中文: RT @Blum_OG:如果你正在构建人工智能代理，你已经点击此步骤了: 中介在几个步骤后会失去背景 任务、记忆、日志——全部在...
#232 🔗 原推
2026-05-20 22:28:42
Blum
if you're building AI agents, you've already hit this: > agent loses context after a few steps > tasks, memory, logs - all in separate places > works fine for short tasks, breaks for anything longer the problem isn't the model - it's the architecture > agents today are built… https://twitter.com/Blum_OG/status/2057184542314619216/video/1
中文: 如果你正在构建人工智能代理，你已经达到了以下位置: 中介在几个步骤后会失去背景 任务、记忆、日志——所有地点均不同 适合做短期任务，时间更长 问题不在于模型——而是架构 今天已建立代理......
media 0 共 2 项
🎬
视频
#231 🔗 原推
2026-05-20 19:39:50
Blum
you can run a terminal AI agent that posts, searches, and reads your X TL setup is 4 steps: > install Hermes - one curl command > `hermes setup` → pick xAI Grok OAuth → log in with SuperGrok > install xurl via brew, npm, or go > `xurl auth oauth2 --app my-app` what you can… https://twitter.com/Blum_OG/status/2057089546127294470/video/1
中文: 您可以运行终端人工智能代理，该代理会发布、搜索并读取您的 X TL 设置是4个步骤: 安装 Hermes - 一个卷曲命令 """"hermes setup` → 选择 xAI Grok OAuth → 使用 SuperGrok 登录 通过啤酒、npm 或 go 安装 xurl > `xurl auth2 -- 应用 my-app 你能做些什么......
media 0 共 2 项
🎬
视频
#230 🔗 原推
2026-05-20 13:22:21
Blum
RT @Blum_OG: > most teams jump straight to automated eval > and end up measuring things that don't matter here's what actually works: > st…
中文: RT @Blum_OG: > 大多数团队直接跳转至自动椭圆形 最后测量那些无关紧要的事物 真正起作用的是什么: 至此;st...
#229 🔗 原推
2026-05-19 21:40:43
Blum
most people drown in information before they synthesize it 40 sources. scattered notes. contradictions you can't hold in your head there's a different way: > upload everything at once - papers, notes, data, raw drafts > no folders, no pre-sorting > proximity surfaces… https://twitter.com/Blum_OG/status/2056829263358173414/video/1
中文: 大多数人在合成信息之前就已溺水 40个来源。零散的笔记。你头脑中无法掌握的矛盾 有另一种方式: 立即上传所有内容——论文、笔记、数据、草稿 没有文件夹，没有预分拣 靠近表面......
media 0 共 2 项
🎬
视频
#228 🔗 原推
2026-05-19 20:08:05
Blum
> most teams jump straight to automated eval > and end up measuring things that don't matter here's what actually works: > start with manual review - always > read outputs until you understand what "good" looks like for you > that understanding tells you which evaluators to… https://twitter.com/Blum_OG/status/2056785549411905803/video/1
中文: 大多数团队直接跳转至自动化椭圆形 最后测量那些无关紧要的事物 真正起作用的是什么: 加时赛;先从手动评测开始——始终 阅读输出，直到你明白“好”是什么样子 理解;你向哪些评估人员提供......
media 0 共 2 项
🎬
视频
#227 🔗 原推
2026-05-19 17:14:22
Blum
most tech companies post 10 clips a week and wonder why nobody knows them the ones winning post 5,000 here's the actual mechanics behind clipping campaigns: > a 3h shoot gives you 50 clippable segments, then hundreds of variations > each variation gets a different hook,… https://twitter.com/Blum_OG/status/2056730621079093384/video/1
中文: 大多数科技公司每周发布10个视频片段 并纳闷为什么没人认识他们 赢得5000个职位的人 以下是剪裁活动背后的实际机制: 拍摄3小时，可拍50个片段，然后进行数百种修改 每个变种都会有不同的挂钩，
media 0 共 2 项
🎬
视频
#226 🔗 原推
2026-05-19 13:36:06
Blum
RT @Blum_OG: > most people open an AI agent > chat for 20 minutes, > and close it thinking "this is just ChatGPT in a different app" that'…
中文: RT @Blum_OG: > 大多数人会开设一个人工智能代理 聊天20分钟 并关闭它，并思考“这只是在另一个应用中的ChatGPT” 那......
#225 🔗 原推
2026-05-19 09:17:28
Blum
RT @Blum_OG: developers treat Claude Code like a faster search bar it works better as a configured environment the setup matters more tha…
中文: RT @Blum_OG:开发者将克劳德代码视为一个更快的搜索栏 它作为一个配置环境运行得更好 设置更重要......
#224 🔗 原推
2026-05-18 19:49:05
Blum
> a 14-year-old in Ohio > $200 laptop > $10k in 2 months > clipping ttv streams into 60-sec videos > he didn't have a brand > didn't show his face > didn't grind 12-hour edit sessions > he had volume > the clipping economy doesn't reward the best editor > it rewards whoever… https://twitter.com/Blum_OG/status/2056443324714635626/video/1
中文: 俄亥俄州一名14岁的年轻人 加底;200美元笔记本电脑 2个月内1万美元 将 ttv 流媒体剪辑成60秒的视频 他没有品牌 没有露面 没有进行12小时的编辑工作 他有音量 剪裁经济并不能奖励最优秀的编辑 它奖励任何人......
media 0 共 2 项
🎬
视频
#223 🔗 原推
2026-05-18 18:34:30
Blum
> most people open an AI agent > chat for 20 minutes, > and close it thinking "this is just ChatGPT in a different app" that's not the agent's fault that's a brain floating in a jar smart, fast - but cut off from your actual life integrations are the senses and limbs you bolt… https://twitter.com/Blum_OG/status/2056396053239439806/video/1
中文: 大多数人会开设人工智能代理 聊天20分钟 快来，并用“这只是在另一款应用里的ChatGPT”来结束它 这不是代理人的错 那是一只漂浮在罐子里的大脑 聪明、快——但与实际生活相隔绝 集成性是你连接的感官和肢体......
media 0 共 2 项
🎬
视频
#222 🔗 原推
2026-05-18 15:26:39
Blum
> most people open an AI agent > chat for 20 minutes, > and close it thinking "this is just ChatGPT in a different app" that's not the agent's fault that's a brain floating in a jar smart, fast - but cut off from your actual life integrations are the senses and limbs you bolt… https://twitter.com/Blum_OG/status/2056389495843393666/video/1
中文: 大多数人会开设人工智能代理 聊天20分钟 快来，并用“这只是在另一款应用里的ChatGPT”来结束它 这不是代理人的错 那是一只漂浮在罐子里的大脑 聪明、快——但与实际生活相隔绝 集成性是你连接的感官和肢体......
media 0 共 2 项
🎬
视频
#221 🔗 原推
2026-05-18 15:00:36
Blum
developers treat Claude Code like a faster search bar it works better as a configured environment the setup matters more than the prompting before writing a single line, serious builders do this: > run "/init" to map structure, dependencies, and conventions > create CLAUDEmd… https://twitter.com/Blum_OG/status/2056332206532862311/video/1
中文: 开发者将Claude Code视为一个更快的搜索栏 它作为一个配置环境运行得更好 设置比提示更重要 在写完一条线之前，认真的构建者会这样做: 运行“/init”以映射结构、依赖关系和约定 制作《克劳德》......
media 0 共 2 项
🎬
视频
#220 🔗 原推
2026-05-18 11:12:57
Blum
RT @Blum_OG: everyone's chasing the AI agency path cold outreach, sales calls, close clients, repeat it works - but it's not the only mov…
中文: RT @Blum_OG:每个人都在追逐人工智能机构的路径 冷淡宣传、销售电话、客户亲密、回头 它奏效了——但它并非唯一的移动......
#219 🔗 原推
2026-05-18 10:07:51
Blum
everyone's chasing the AI agency path cold outreach, sales calls, close clients, repeat it works - but it's not the only move > 76% of large-company CEOs are hiring a Chief AI Officer in 2026 > two years ago that number was 26% > 57% of those CAIOs were promoted from inside -… https://twitter.com/Blum_OG/status/2056117619774927007/video/1
中文: 每个人都在追逐人工智能机构的发展路径 冷淡宣传、销售电话、客户亲密、回头 它奏效了——但并非唯一的举措 到2026年，76%的大型企业首席执行官正在招聘首席人工智能官 两年前，这一数字为26% 57%的CAIO从内部晋升——
media 0 共 2 项
🎬
视频
#218 🔗 原推
2026-05-17 21:00:16
Blum
> "write this function" > "fix this bug" > "explain this error" this is how most developers use Claude Code the best engineers use it differently some of the 40 practices that actually change output: > pipe terminal logs directly - no manual copying > rewind sessions… https://twitter.com/Blum_OG/status/2056077090970042828/video/1
中文: ""编写此函数" 并修复此漏洞 解释此错误 大多数开发者都这样使用Claude Code 最优秀的工程师使用方式不同 实际改变产出的40个实践中的一些: 直接操作管道端子日志 - 无手动复制 内容与回顾性会议......
media 0 共 2 项
🎬
视频
#217 🔗 原推
2026-05-17 18:19:13
Blum
> you paste info into Claude > Claude responds > you copy the output somewhere else > paste more info > every session resets > context rebuilt from scratch every time this is how most people use Claude MCP changes the entire model: > Claude reads directly from your codebase >… https://twitter.com/Blum_OG/status/2055938563762487727/video/1
中文: 你将信息粘贴到克劳德 加特;克劳德回应 你把输出复制到别处 粘贴更多信息 每个会话都会重置 每次从零开始重建 大多数人都是使用克劳德语的方式 MCP 会更改整个模型: >克劳德直接从你的代码库阅读 网址:
media 0 共 2 项
🎬
视频
#216 🔗 原推
2026-05-17 09:08:45
Blum
RT @Blum_OG: > watched a 40-min founder interview > read two 60-page market reports > cross-referenced every claim by hand > took 3+ hours…
中文: RT @Blum_OG: > 观看了一场40分钟的创始人访谈 阅读两份60页的市场报告 > 手动交叉引用每一个索赔 花了3个多小时......
#215 🔗 原推
2026-05-17 07:51:43
Blum
GROK 4.3 NOW RUNS INSIDE HERMES AGENT > persistent memory across sessions > image + video gen on command > text-to-speech built in > connects to WhatsApp, Telegram, Signal > works on Linux, macOS, WSL2, Android setup is 3 lines in terminal: 1. 'curl - fsSL https:/ / raw.… https://twitter.com/Blum_OG/status/2055766659743232506/video/1
中文: GROK 4.3 现在在赫瑞姆斯酒店内运行 >跨会话的持久记忆 命令中的图片 + 视频生成 已内置文本转语音 连接到 WhatsApp、Telegram、Signal 适用于Linux、macOS、WSL2、Android 终端中设置为3条线路: 1. curl - fsSL https:// 原始内容。
media 0 共 2 项
🎬
视频
#214 🔗 原推
2026-05-16 21:45:40
Blum
> watched a 40-min founder interview > read two 60-page market reports > cross-referenced every claim by hand > took 3+ hours > sent it all to claude instead > pasted the clean transcript > uploaded both PDFs > asked for a 6-bullet insight deck with confidence scores > done in… https://twitter.com/Blum_OG/status/2055711095382548528/video/1
中文: 看了一次40分钟的创始人访谈 阅读两份60页的市场报告 > 手动交叉引用每一个索赔 花了3个多小时 把一切都发到了撩人 粘贴干净的文字记录 并上传两个PDF文件 要求提供6个带自信度的洞察平台 内容:
media 0 共 2 项
🎬
视频
#213 🔗 原推
2026-05-16 18:04:53
Blum
Сlaude coding accuracy: 65% without context. 94% with one file a CLAUDEmd in your project root sets Claude’s starting knowledge it persists across sessions the file went #1 on GitHub trending. 82,000 stars. 7,800 forks here is the exact setup that produced the jump: > ask,… https://twitter.com/Blum_OG/status/2055642915335000345/video/1
中文: Сlaude 编程准确率:65% 无上下文。94% 使用一个文件 在项目中，一个 CLAUDEMD 根系克劳德的起始知识 它在会议期间持续 该文件在 GitHub 上排名第一，有8.2万颗星。7800个分叉 以下是产生跳转的精确设置: 询问，请问......
media 0 共 2 项
🎬
视频
#212 🔗 原推
2026-05-16 13:33:57
Blum
your hermes setup breaks when you skip the control room as of may 16, 2026: github shows 152k stars openrouter shows #1 daily global rank official docs say 70+ built-in tools and 20+ platforms the useful setup for marketers is a documented worker not another chat tab first:… https://twitter.com/Blum_OG/status/2055604986009038865/video/1
中文: 跳过控制室时，你的爱马仕设置会中断 截至2026年5月16日: 吉特布展示152万星 openrouter 显示全球每日排名第一 官方文件显示，70多个内置工具和20多个平台 营销人员的有用配置是文档化的工作人员 不是另一个聊天标签 先是:
media 0 共 2 项
🎬
视频
#211 🔗 原推
2026-05-16 11:03:14
Blum
RT @Blum_OG: deploying Claude agents comes down to one question: does the job need reasoning, or just execution? /loop - full agent loop,…
中文: RT @Blum_OG:部署克劳德特工归结为一个问题: 工作需要推理，还是仅仅需要执行? /循环 - 全代理循环，...
#210 🔗 原推
2026-05-15 22:54:25
Blum
deploying Claude agents comes down to one question: does the job need reasoning, or just execution? /loop - full agent loop, laptop stays on > terminal sessions last 7 days, desktop caps at 3 > timing jitter up to 30 min by design Claude Routines - same loop, Anthropic runs it… https://twitter.com/Blum_OG/status/2055334245275422839/video/1
中文: 部署克劳德特工归结为一个问题: 工作需要推理，还是仅仅需要执行? /loop - 全代理循环，笔记本电脑保持运行 终端会话持续7天，桌面版3 时间推棡，设计最高可达30分钟 克劳德·鲁廷斯——同样的循环，《蚁皮》运行它......
media 0 共 2 项
🎬
视频
#209 🔗 原推
2026-05-15 17:07:24
Blum
me realizing Claude wasn't bad at frontend i just never told it what aesthetic to use > asked Claude to build a landing page > got purple gradients, Inter font, generic card layout > did this 40 more times across different projects > found the frontend-design skill file >… https://twitter.com/Blum_OG/status/2055296902866903284/video/1
中文: 我意识到克劳德在前端并不差 我从没告诉过它要用什么美学 要求克劳德搭建一个着陆页 已采用紫色渐变、Inter字体和通用卡片布局 跨不同项目再做40次 找到前端设计技能文件
media 0 共 2 项
🎬
视频
#208 🔗 原推
2026-05-15 14:39:01
Blum
your agent's prompt will be worse in 30 days not because you wrote it badly > your product changed > your team's taste sharpened > new edge cases showed up Warp built a community reply agent called Buzz 1,000+ mentions a week across different socials here is the system that… https://twitter.com/Blum_OG/status/2055206405477453927/video/1
中文: 你的经纪人的提示将在30天内变得更糟 不是因为你写得不好 您的产品发生了变化 你的团队品味更加敏锐 新增边缘病例 沃普建立了一个名为Buzz的社区回复代理 每周在不同社交领域提及1000多次 系统是以下内容......
media 0 共 2 项
🎬
视频
#207 🔗 原推
2026-05-15 08:39:25
Blum
RT @Blum_OG: Moni Community lifetime access closes in 24h this is not avg alpha group. not even close what you get: - daily alpha reports…
中文: RT @Blum_OG:Moni 社区终身访问时间在 24 小时内结束 这并非阿尔法群体，甚至不接近 你能得到什么: - 每日阿尔法报告...
#206 🔗 原推
2026-05-14 22:30:31
Blum
Moni Community lifetime access closes in 24h this is not avg alpha group. not even close what you get: - daily alpha reports (filtered, not noisy) - structured airdrop guides - private NFT/degen deals you won't see in public CT - trenches chat with people who actually know what… https://twitter.com/Blum_OG/status/2054948077631574161/photo/1
中文: 莫尼社区终身访问时间在24小时内关闭 这并非阿尔法群体，甚至不接近 你能得到什么: - 每日阿尔法报告(过滤，不吵闹) - 结构化空投指南 - 私密的NFT/degen优惠，在公共CT中看不到 - 与真正了解的人聊天......
media 0 共 1 项
#205 🔗 原推
2026-05-14 15:32:55
Blum
> wrote code every day > reviewed it manually > wrote the PR description manually > checked coverage manually > analyzed logs manually > 300,000 tokens dumped into one session > context bloated before lunch > opened .claude/agents/ > created 8 md files > each file: a name, a… https://twitter.com/Blum_OG/status/2054904872877650200/video/1
中文: 每天编写代码 加底;手动审核 并手动撰写了PR描述 手动检查覆盖 手动分析日志 已投入一次交易，30万个代币 午饭前腹胀 已开盘 .claude/agents/ 已创建 8 个 md 文件 每个文件:一个名称，一个......
media 0 共 2 项
🎬
视频
#204 🔗 原推
2026-05-14 12:41:14
Blum
> browsed GitHub for years > starred repos, read readmes, closed the tab > never shipped anything > watched people build $3 - 8k/mo clipping channels > watched AI automation agencies charge $1 - 3k/mo > wondered how opened the repos they were actually using: > MoneyPrinterV2:… https://twitter.com/Blum_OG/status/2054862970534973503/video/1
中文: 多年来一直浏览 GitHub > 主演，阅读准备，关闭标签页 切不发货 观看的人们搭建的3-8k/mo剪贴画 观看的人工智能自动化机构收费1至3k美元/月 生了吧! 打开他们实际使用的仓库: 网址:MoneyPrinterV2:..
media 0 共 2 项
🎬
视频
#203 🔗 原推
2026-05-14 09:54:44
Blum
RT @Blum_OG: > uploaded 12 sources > asked for a summary > got a 6-paragraph wall of text > copy-pasted it into a doc > never opened it aga…
中文: RT @Blum_OG: > 已上传12个来源 请求摘要 拿到了一面六段文字墙 将其复制粘贴成文档 永远不要打开它......
#202 🔗 原推
2026-05-14 09:00:32
Blum
> uploaded 12 sources > asked for a summary > got a 6-paragraph wall of text > copy-pasted it into a doc > never opened it again > that is not NotebookLM's fault > the tool is built for staged extraction > not one-shot answers > 2.1% of users know this > the rest are summarising… https://twitter.com/Blum_OG/status/2054686248388677766/video/1
中文: 已上传12个来源 请求摘要 拿到了一面六段文字墙 将其复制粘贴成文档 再也没有打开过 这并非NotebookLM的错 该工具是为分阶段提取而构建的 不是一枪回答 2.1%的用户知道这一点 其余部分正在总结......
media 0 共 2 项
🎬
视频
#201 🔗 原推
2026-05-13 22:12:30
Blum
CLAUDE KILLED THE SOFTWARE DEVELOPMENT BOTTLENECK the path from idea to product is now just 2–4 weeks with Claude Code no dev team. no $50k budget. no months of waiting. the income potential: > SaaS path: $500-$3,000/mo > trading bots path: $1,000-$10,000/mo > marketplaces… https://twitter.com/Blum_OG/status/2054638531817898299/video/1
中文: 克劳德杀死了软件开发机器人 从创意到产品的路径现在只有2到4周，使用Claude Code 没有开发团队。没有5万美元的预算。无需等待数月。 收入潜力: > SaaS 路径:500-3000美元/月 交易机器人路径:1000至10000美元/月 市场市场......
media 0 共 2 项
🎬
视频
#200 🔗 原推
2026-05-13 19:02:53
Blum
> built an AI agent > used it for a week > corrected the same mistake 3 times > opened a new session > it forgot everything. again. > checked the popular agents > same pattern everywhere > session ends, memory dies > next session, start from scratch > no learning, no compounding… https://twitter.com/Blum_OG/status/2054598723531939903/video/1
中文: 建立了人工智能代理 使用一周 纠正同样的错误三次 > 开启了一场新的会议 它再次忘记了一切。 查看了热门代理商 各地模式相同 会话结束，记忆消亡 下次会议，从零开始 无需学习，无复合......
media 0 共 2 项
🎬
视频
#199 🔗 原推
2026-05-13 16:24:42
Blum
RT @Blum_OG: > gave Claude a long research job > it started well > an hour in, it forgot a decision we made at the start > redid the work t…
中文: RT @Blum_OG: > 为克劳德提供了一份长期的研究工作 开始得不错 一小时后，它忘记了我们一开始做的决定 将工作重新进行......
#198 🔗 原推
2026-05-12 21:04:24
Blum
> gave Claude a long research job > it started well > an hour in, it forgot a decision we made at the start > redid the work the wrong way > i had no idea why > turns out long conversations have a memory limit > when it fills up, Claude summarizes the old stuff > "use that… https://twitter.com/Blum_OG/status/2054275386045849998/video/1
中文: 给了克劳德一份长期的研究工作 开始得不错 一小时后，它忘记了我们一开始做出的决定 将工作重新删错了 我根本不知道为什么 事实证明，长时间的对话具有记忆限制 加满;当它填满时，克劳德总结了那些旧东西 使用“gt”......
media 0 共 2 项
🎬
视频
#197 🔗 原推
2026-05-12 18:59:53
Blum
> built a second brain > highlighted articles, clipped podcasts, took notes > felt organized > 2 weeks later stopped maintaining it > the tags fell apart > the cross-references went stale > went back to scattered notes > tried to rebuild 6 months later > same cycle, same result… https://twitter.com/Blum_OG/status/2054230021812732264/video/1
中文: 建立了第二个大脑 加文;重点文章，剪贴一向的播客，并记下笔记 感觉有条理 两周后停止了维护 标签分崩离析 交叉参照变得陈旧 回到零散的音符 6个月后，我试图重建 相同周期，结果相同......
media 0 共 2 项
🎬
视频
#196 🔗 原推
2026-05-12 15:59:37
Blum
me realizing i was using Claude like a $20/month Google > typed a question > got an answer > copied the useful part > closed the tab > repeated for 6 months > results were fine > never great > never consistent > finally asked why > Claude doesn't retrieve - it reasons > but… https://twitter.com/Blum_OG/status/2054199278441013622/video/1
中文: 我意识到自己用的是克劳德，就像每月20美元的谷歌 输入一个问题 得到了答复 复制有用的部分 关闭标签页 重度;重复6个月 结果很好 太好了 永不稳定 终于问起原因了 克劳德没有恢复——这是原因 但......
media 0 共 2 项
🎬
视频
#195 🔗 原推
2026-05-12 13:57:27
Blum
> everyone kept asking for smarter AI > more benchmarks > better scores > higher IQ numbers > so that's what the industry built > models that impressed on first response > then fell apart by prompt 30 > couldn't follow instructions across long tasks > lost context halfway… https://twitter.com/Blum_OG/status/2054167171920461899/video/1
中文: 大家一直都在要求更智能的人工智能 加减;更多基准 评分更高 高智商数字 这正是行业所建立的 对第一反应印象深刻的模型 随后，30分就崩溃了 在漫长的任务中无法遵循指示 中途失语......
media 0 共 2 项
🎬
视频
#194 🔗 原推
2026-05-12 11:49:53
Blum
RT @Blum_OG: > wanted to run Claude Code > needed a pro plan to use it > same with Codex - needed chatgpt plus > both cost money by default…
中文: RT @Blum_OG: > 想运行 Claude Code 需要一个专业计划来使用它 与 Codex 相同，需要 chatgpt 加 默认情况下，两者都需要花钱......
#193 🔗 原推
2026-05-12 09:27:41
Blum
> wanted to run Claude Code > needed a pro plan to use it > same with Codex - needed chatgpt plus > both cost money by default > found the exit > both tools let you swap the underlying model > you don't have to use Anthropic's or Openai's > point them at Openrouter instead >… https://twitter.com/Blum_OG/status/2053927746796155355/video/1
中文: 想要运行Claude Code 需要一个专业计划来使用它 与 Codex 相同，需要 chatgpt 加 默认价格为: 找到出口 两个工具都允许你更换底层模型 你不必使用Anthropic或Openai的 用Openrouter来点它们 网址:
media 0 共 2 项
🎬
视频
#192 🔗 原推
2026-05-11 19:58:29
Blum
> built the skill in an afternoon > put everything in 1 file > description said "handles authentication errors" > shipped it > it was 6,000 tokens loading every session > 5,200 of those tokens were things the model already knew > the test is simple: would the agent get this… https://twitter.com/Blum_OG/status/2053889635445527023/video/1
中文: 在下午将这项技能建立起来 将所有内容放入1个文件中 描述称“手柄认证错误” 加运;发货 每次会话加载的令牌为6000个 这些代币中有5200个是模型早已知道的 测试很简单:代理人能拿到这个吗......
media 0 共 2 项
🎬
视频
#191 🔗 原推
2026-05-11 17:27:03
Blum
> set up Hermes on day 1 > plugged in an API key > connected some tools > started chatting > got a generic chatbot > checked what was missing > no Soul md > no User md > agent had zero context about me > didn't know my theses, my voice, my portfolio > just another LLM wrapper >… https://twitter.com/Blum_OG/status/2053839178996052209/video/1
中文: 第一天就设立爱马仕 插入API密钥 加上;一些已连接的工具 开始聊天 获取了一个通用聊天机器人 检查一下缺少了什么 没有灵魂MD 没有用户 中介对我没有任何背景 不知道我的这些，我的声音，我的作品集 再说一下，再说一次LLM包装 网址:
media 0 共 2 项
🎬
视频
#190 🔗 原推
2026-05-11 14:06:33
Blum
> 2023: AI was a chatbot > 2024: AI was a copilot > 2025: AI was an agent > 2026: AI is infrastructure > most people missed the transition > a repo called PAI has 12,100 stars > 45 skills, 171 workflows, 37 hooks > built by 1 developer > saves 2-3 hours every single day > not… https://twitter.com/Blum_OG/status/2053807782848073846/video/1
中文: 2023年:人工智能是一台聊天机器人 2024年:人工智能成为副驾驶 2025年:人工智能成为代理 2026年:人工智能就是基础设施 大多数人错过了过渡 一个名为PAI的回购机构拥有12100颗星 45 个技能，171 个工作流程，37 个钩子 由1名开发者打造 每天节省2-3小时 不......
media 0 共 2 项
🎬
视频
#189 🔗 原推
2026-05-11 12:01:47
Blum
> built a workflow around Claude > prompted it manually every single day > results varied > some days sharp, some days off > could not figure out why > found the real issue > prompts are not repeatable systems, skills are > a skill is a folder, 1 file, plain English instructions… https://twitter.com/Blum_OG/status/2053776699838402631/video/1
中文: 围绕克劳德构建了工作流程 每天手动提示 结果各不相同 有些日子很清白，有些日子休息 无法弄清楚原因 找到真正的问题 提示不是可重复的系统，技能是 技能是文件夹、1 个文件、简单的英文说明......
media 0 共 2 项
🎬
视频
#188 🔗 原推
2026-05-11 09:58:17
Blum
> opened Claude > typed "write me a marketing email" > got something no one would ever send > edited it for 20 minutes > still mediocre > closed the tab > decided AI wasn't useful for this > the prompt had 1 element: the task > no role, no context, no format, no failure… https://twitter.com/Blum_OG/status/2053734805053784214/video/1
中文: 开张的克劳德 输入“给我写一封营销邮件” 有东西，从来没有人送过 剪辑20分钟 依然平庸 关闭标签页 确定人工智能对此没有用处 提示提示有1个要素:任务 没有角色，没有背景，没有格式，没有失败......
media 0 共 2 项
🎬
视频
#187 🔗 原推
2026-05-11 07:11:48
Blum
RT @Blum_OG: > shipped the agent > opened the dashboard > latency: fine > error rate: fine > users: unhappy > checked the responses > tech…
中文: RT @Blum_OG: > 已发货 打开仪表板 延迟:正常 错误率:罚款 用户:不满意 加以;检查了回复 技术......
#186 🔗 原推
2026-05-11 05:57:49
Blum
> used to spend mornings opening 5 tools in sequence > Slack first, then Gmail, then Notion > then realizing you missed something in Figma > then going back to Slack > not a focus problem > not a discipline problem > a structure problem > context was scattered by design > PMM… https://twitter.com/Blum_OG/status/2053560664870781021/video/1
中文: 过去常常在早晨按顺序打开5个工具 先是Slack，然后是Gmail，然后是 Notion 然后意识到你错过了《菲格玛》中的某件事 然后回到Slack 不是重点问题 不是纪律问题 结构问题 背景内容因设计而分散 网址://PMM
media 0 共 2 项
🎬
视频
#185 🔗 原推
2026-05-10 19:39:50
Blum
> shipped the agent > opened the dashboard > latency: fine > error rate: fine > users: unhappy > checked the responses > technically correct > wrong tool called 3 steps earlier > no trace to follow > no dataset to reproduce it > no eval to catch it next time > this wasn't a… https://twitter.com/Blum_OG/status/2053530171840295187/video/1
中文: 已发货 打开仪表板 延迟:正常 错误率:罚款 用户:不满意 加以;检查了回复 技术正确 错误工具，提前三步 无踪迹 没有数据集可以复制 下次没有茉病 不是......
media 0 共 2 项
🎬
视频
#184 🔗 原推
2026-05-10 17:38:40
Blum
> was using AI every day > starting cold every single chat > re-explaining the same context on loop > opened my Obsidian vault > looked organized > checked if my AI could actually use it > it couldn't > the bottleneck wasn't the model > it was me, doing manual work to keep them… https://twitter.com/Blum_OG/status/2053499664352039013/video/1
中文: 每天使用人工智能 开始每次聊天都冷 > 重新解释相同的循环环境 打开我的黑曜石金库 看起来有条理 检查一下我的人工智能是否真的能使用它 但是，它做不到 瓶颈不是模型 是我，做手工工作来保管他们......
media 0 共 2 项
🎬
视频
#183 🔗 原推
2026-05-10 15:37:26
Blum
> spent 2 hours building a plugin > shipped it on a Tuesday night > expected it to save some time > ran it Monday morning > 4-hour content job done in 11 min > checked what actually happened > the plugin read the file > extracted the key claims > built the thread, the LinkedIn… https://twitter.com/Blum_OG/status/2053375628473156054/video/1
中文: 花了两个小时构建一个插件 加运;周二晚上发货 期望它节省一些时间 周一上午运行 在11分钟内完成4小时内容工作 查了实际发生了什么 插件读取文件 提取关键声明 已建立该线，与 LinkedIn 建立...
media 0 共 2 项
🎬
视频
#182 🔗 原推
2026-05-10 07:24:34
Blum
RT @Blum_OG: > everyone argued about md vs html > md purists said html was showing off > html converts said md was for beginners > both cam…
中文: RT @Blum_OG: > 大家都争论了 md 与 html > md 纯粹主义者表示，HTML 正在炫耀 http:/ / gt;HTML 转换者表示 md 适合初学者 加点;两者都......
#181 🔗 原推
2026-05-10 04:58:25
Blum
> everyone argued about md vs html > md purists said html was showing off > html converts said md was for beginners > both camps kept losing tokens, edit cycles, and search relevance > nobody asked the right question the question was never about format preference > it was 3… https://twitter.com/Blum_OG/status/2053225881427222594/video/1
中文: 大家都争论过 md 和 html > md 纯粹主义者表示，HTML 正在炫耀 http:/ / gt;HTML 转换者表示 md 适合初学者 两个阵营不断失去代币、编辑周期和搜索相关性 没有人问对 问题从来不是关于格式偏好 是3......
media 0 共 2 项
🎬
视频
#180 🔗 原推
2026-05-09 21:29:31
Blum
> heard "AI agent" and pictured a team of engineers > thousands of lines of code > months of setup before anything ran > decided to try it anyway > opened Claude Managed Agents > no servers to configure > no agent loops to write > no sandbox to manage > just described what the… https://twitter.com/Blum_OG/status/2053062535692046779/video/1
中文: 听到“AI代理”后，画面中有一支工程师团队 成千上万行代码 在任何事情开始前，才进行了数月的设置 决定无论如何都要试试 > 开设 Claude Managed Agents 无需配置服务器 无需代理循环来书写 无需管理 刚刚描述了......
media 0 共 2 项
🎬
视频
#179 🔗 原推
2026-05-09 10:40:27
Blum
RT @Blum_OG: > asked Claude Code to write an implementation plan > it returned 200 lines of markdown > i read 40 of them > approved it anyw…
中文: RT @Blum_OG: > 请克劳德·德·德克撰写实施计划 返回了200行标记 我读了其中40本 已批准......
#178 🔗 原推
2026-05-09 06:16:23
Blum
> asked Claude Code to write an implementation plan > it returned 200 lines of markdown > i read 40 of them > approved it anyway > regretted it later > the real problem wasn't the plan > it was that markdown caps out fast > no spatial layout > no diagrams unless you count ASCII… https://twitter.com/Blum_OG/status/2052866274988863729/video/1
中文: 要求克劳德·科德制定实施计划 返回了200行标记 我读了其中40条 不管怎样，还是批准了 稍后会后悔 真正的问题不在于计划 是那种快速降价的上限 没有空间布局 不统计:除非统计为ASCII，否则不会显示图表......
media 0 共 2 项
🎬
视频
#177 🔗 原推
2026-05-08 21:40:35
Blum
> most people will watch the world cup > scroll highlights, react, forget > some people will get huge $$$ from those same clips here's what they know that most don't: > the world cup is not a sports event > it's the single largest content moment in your lifetime > 5.95B social… https://twitter.com/Blum_OG/status/2052822192983834751/video/1
中文: 大多数人会看世界杯 提示:滚动高亮，反应，遗忘 有些人会从同样的片段中获得巨额的美元 他们所知道的是，大多数人并不知道: 世界杯不是体育赛事 这是你一生中最重要的内容时刻 5.95亿比额表;社交...
media 0 共 2 项
🎬
视频
#176 🔗 原推
2026-05-08 18:45:25
Blum
> wanted Claude to write like me > not like AI-me > like actually me > tried 3 times > each time smth was off > too generic, too performative, too parody > the tics were right but the judgment was wrong > because i only fed it my best posts > the fix wasn't more samples > it… https://twitter.com/Blum_OG/status/2052791027891667325/video/1
中文: 希望克劳德像我一样写作 嗯;不像AI-me 和我一样 尝试了三次 每次都关闭 太笼统，太有表现力，太模仿 抽搐是正确的，但判断是错误的 因为我只发了我最好的帖子 解决办法不是更多的样本 网址:
media 0 共 2 项
🎬
视频
#175 🔗 原推
2026-05-08 16:41:34
Blum
> you post every week > you pay influencers > you run ads > and get almost nothing back > the problem isn't your product. not your price. not your creative > it's the format the brands quietly making money on TikTok rn > they're running slideshows > simple. cheap. impossible to… https://twitter.com/Blum_OG/status/2052755667329470718/video/1
中文: 你每周发一次帖 你向网红付款 你投放广告 而且几乎什么都没回来 问题不在于你的产品，而不在于你的价格，而不在于你的创意 嗯;是格式 这些品牌在TikTok rn上悄然赚钱 他们正在举办幻灯片放映 简单，便宜。不可能......
media 0 共 2 项
🎬
视频
#174 🔗 原推
2026-05-08 14:21:04
Blum
RT @Blum_OG: > downloaded Claude Code > thought it was just another AI chatbot > opened it up > found a chat bar > typed a prompt > nothing…
中文: RT @Blum_OG: > 下载 克劳德密码 我以为这只不过是另一个人工智能聊天机器人 快点;打开它 找到了一个聊天栏 输入提示 没什么......
#173 🔗 原推
2026-05-07 23:33:57
Blum
> downloaded Claude Code > thought it was just another AI chatbot > opened it up > found a chat bar > typed a prompt > nothing happened to my files > realized the difference > browser AI waits for you to bring your files > Claude Code goes into your folder > all 45 PDFs > all… https://twitter.com/Blum_OG/status/2052462798832300542/video/1
中文: 已下载克劳德密码 我以为这只不过是另一个人工智能聊天机器人 快点;打开它 找到了一个聊天栏 输入提示 我的文件什么也没发生 意识到了区别 浏览器AI等待您携带文件 > 克劳德密码会进入你的文件夹 全部45个PDF文件 所有内容......
media 0 共 2 项
🎬
视频
#172 🔗 原推
2026-05-07 18:57:18
Blum
> used Claude every day > got mediocre results every day > assumed the model was the ceiling > opened a better prompt > same model > completely different output > checked what changed > not the AI, but the brief > missing a role > missing a standard > missing constraints >… https://twitter.com/Blum_OG/status/2052430968405819661/video/1
中文: 每天使用克劳德 每天都得到很平庸的成绩 假设模型是天花板 打开一个更好的提示 相同型号 和我们;完全不同的输出 查看了哪些更改 不是人工智能，而是简报 缺少一个角色 缺少一个标准 缺少限制
media 0 共 2 项
🎬
视频
#171 🔗 原推
2026-05-07 16:50:49
Blum
here is the full stack for 3 coordinated Claude agents > agent 1: research > tools: Exa for neural search, Firecrawl for scraping, X MCP > for social monitoring > runs every Monday, logs a brief to the shared knowledge base > agent 2: content > tools: Exa, Firecrawl, Notion MCP… https://twitter.com/Blum_OG/status/2052383925951971588/video/1
中文: 以下是3个协同的克劳德特工的完整堆栈 > 代理1:研究 工具:用于神经搜索的Exa、用于抓取的Firecrawl、X MCP 用于社会监测 每周一运行，请将一份简报记录到共享知识库 代理商 2:内容 工具:Exa、Firecrawl、Notion MCP...
media 0 共 2 项
🎬
视频
#170 🔗 原推
2026-05-07 13:43:54
Blum
> built a beautiful Obsidian vault > spent a weekend on the folder structure > added notes for 2 weeks straight > opened it 6 months later > nothing came back out > checked the system > system looked fine > checked what was actually missing > found it immediately > no feedback… https://twitter.com/Blum_OG/status/2052348842117390744/video/1
中文: 新建了一个美丽的黑曜石金库 周末在文件夹结构上度过 连续两周添加笔记 6个月后开业 什么都没有回来 检查了系统 系统看起来不错 查了真正缺少的内容 立即找到 没有反馈......
media 0 共 2 项
🎬
视频
#169 🔗 原推
2026-05-07 11:24:29
Blum
RT @Blum_OG: > opened Codex > thought it was just ChatGPT with file access > one hour later - a full YT comment intelligence system > Excel…
中文: RT @Blum_OG: > 打开了 Codex 以来只是ChatGPT，带有文件访问权限 一小时后——一个完整的YT评论情报系统 以及;Excel...
#168 🔗 原推
2026-05-06 23:15:12
Blum
> launched an AI influencer > picked a niche from scratch > built a face from two photos > uploaded it to Kling > hit generate > got a plastic doll > tried again > plastic again > started figuring out why > turned out the model wasn't the problem > the method was a neural… https://twitter.com/Blum_OG/status/2052132731350360266/video/1
中文: 推出了一款人工智能网红 从零开始选择一个细分市场 用两张照片打造了一张脸 上传至Kling 以及 pyde 生成 买了一个塑料娃娃 再试一次 再次使用塑料 开始弄清楚原因 事实证明，这个模型并不是问题 方法是 神经......
media 0 共 2 项
🎬
视频
#167 🔗 原推
2026-05-06 21:05:44
Blum
> opened Codex > thought it was just ChatGPT with file access > one hour later - a full YT comment intelligence system > Excel with insights across 200+ comments > live dashboard on Vercel > cron automation every sunday > browser use found 6 bugs and fixed them on its own > i… https://twitter.com/Blum_OG/status/2052088511700844759/video/1
中文: 已开盘的法典 以为只是带有文件访问权限的ChatGPT 一小时后——一个完整的YT评论情报系统 > Excel，包含200多条评论的洞察 网址:Vercel 上的实时仪表板 每周日使用 > cron 自动化 浏览器使用已发现6个漏洞并自行修复 网址:
media 0 共 2 项
🎬
视频
#166 🔗 原推
2026-05-06 18:10:01
Blum
RT @Blum_OG: > started using Claude Pro > felt unstoppable for the first week > then hit the limit at 11am on monday > then again on tuesda…
中文: RT @Blum_OG: > 开始使用 Claude Pro 第一周感觉势不可挡 随后，周一上午11点达到极限 然后又在 tuesda 上......
#165 🔗 原推
2026-05-06 05:26:42
Blum
> downloaded the top-30 Claude Skills by install count > installed everything at once, opened a task > Claude stalled on generation > context window burned before the first response token > started digging > two design skills were fighting over the same task > three code… https://twitter.com/Blum_OG/status/2051781916290924547/video/1
中文: 已下载首30个Claude Skills的安装数量 同时安装所有内容，开启任务 克劳德在一代人的身上停滞不前 上下文窗口在第一个响应令牌之前被烧毁 开始挖掘 两项设计技能正在争夺同一项任务 三个代码......
media 0 共 2 项
🎬
视频
#164 🔗 原推
2026-05-05 21:51:43
Blum
> paid $9k/year for software > found one guide > turned out everything is free > you just need student status > signed up for GitHub Student Pack > JetBrains - free > Figma Education - free > Notion Plus - free > Cursor Pro for a year - free > kept scrolling the list > Canva Pro,… https://twitter.com/Blum_OG/status/2051762517249253648/video/1
中文: 每年支付9万美元的软件费用 找到一位向导 事实证明，一切都是免费的 你只需要学生身份 已注册 GitHub 学生包 加特;JetBrains - 免费 加丁;无花果教育 加价;注意附加——免费 加丁;Cursor Pro 一年免费 继续滚动列表 加特;Canva Pro，...
media 0 共 2 项
🎬
视频
#163 🔗 原推
2026-05-05 20:34:38
Blum
> started using Claude Pro > felt unstoppable for the first week > then hit the limit at 11am on monday > then again on tuesday > thought the plan was broken > sent a support ticket > sent another message correcting Claude > sent a follow-up to the follow-up > hit the limit even… https://twitter.com/Blum_OG/status/2051728732784922824/video/1
中文: 开始使用Claude Pro 第一周感觉势不可挡 随后，周一上午11点达到极限 然后又在周一 我以为这个计划被打破了 发送了一张支持票 发送另一条消息以更正克劳德 发送后续跟进 甚至达到了极限......
media 0 共 2 项
🎬
视频
#162 🔗 原推
2026-05-05 18:20:23
Blum
RT @Blum_OG: > opened Notion in the morning > 14 tabs. 3 databases. 2 overdue tasks > one client i should have messaged yesterday > spent 2…
中文: RT @Blum_OG: > 早上打开 NOtion >14个标签页。3个数据库。2个逾期任务 一个客户端，我昨天应该发消息 花费 2.
#161 🔗 原推
2026-05-05 01:21:27
Blum
saw an article about an OF model > subscribers paying $34 on average > top fan dropped $1,847 in a month > $43,000 net in 30 days > opened the profile > pretty girl, real messages, voice notes at 11pm > went looking for how she runs this alone > no team. no camera. no her. it's… https://twitter.com/Blum_OG/status/2051419587716288702/video/1
中文: 看到一篇关于模型的文章 订阅用户平均支付34美元 顶级粉丝在一个月内下跌了1847美元 30天内净收入4.3万美元 打开个人资料 女孩，真实留言，晚上11点发出语音留言 去找她独自经营这个的方式 没有团队，没有摄像头，没有她。 是......
media 0 共 2 项
🎬
视频
#160 🔗 原推
2026-05-04 21:51:57
Blum
RT @Blum_OG: > launched a project over the weekend > Claude Code wrote the code, i steered > three days - a working MVP > deployed to Verce…
中文: RT @Blum_OG: > 上周末推出了一个项目 > 克劳德·德·克德编写代码，我进行了引导 三天——一位在职的MVP 已部署至 Verce...
#159 🔗 原推
2026-05-04 20:59:58
Blum
> opened Notion in the morning > 14 tabs. 3 databases. 2 overdue tasks > one client i should have messaged yesterday > spent 20 mins just figuring out where to start > then connected Claude Code to Notion via MCP server > wrote one CLAUDEmd file describing my workspace > typed:… https://twitter.com/Blum_OG/status/2051390647802699777/video/1
中文: 上午开启“点头” >14个标签页。3个数据库。2个逾期任务 一个客户端，我昨天应该发消息 花了20分钟，弄清楚从哪里开始 随后通过MCP服务器将Claude Code连接到Notion 写了一个描述我工作区的 CLAUDEMD 文件 输入:..
media 0 共 2 项
🎬
视频
#158 🔗 原推
2026-05-04 19:56:58
Blum
> launched a project over the weekend > Claude Code wrote the code, i steered > three days - a working MVP > deployed to Vercel > sent the link to 5 people > got feedback > sat down to make changes > asked Claude to rename a function > Claude renamed it in one file > broke in… https://twitter.com/Blum_OG/status/2051348209289601308/video/1
中文: 上周末启动了一个项目 > 克劳德·科德编写代码，我进行了引导 三天——一位在职的MVP 已部署到 Vercel 发送链接给5人 已收到反馈 坐下来做修改 请求克劳德重命名一个函数 克劳德将其重新命名为“一个文件” 已成功，
media 0 共 2 项
🎬
视频
#157 🔗 原推
2026-05-04 17:08:19
Blum
RT @Blum_OG: > installed an Hermes agent at peak hype > stared at the interface for an hour > closed it and never came back everyone on X…
中文: RT @Blum_OG: > 在巅峰时刻安装了一位爱马仕代理 盯着界面看了一个小时 关闭后，再也没有回来 X 上的所有人......
#156 🔗 原推
2026-05-04 03:56:27
Blum
> installed an Hermes agent at peak hype > stared at the interface for an hour > closed it and never came back everyone on X was saying they use it for "everything" > i had no idea what that meant for me specifically > grabbed a notebook > wrote down everything i did in a day >… https://twitter.com/Blum_OG/status/2051057714319831165/video/1
中文: 在最热门的炒作中安装了一位爱马仕代理 盯着界面看了一个小时 关闭后，再也没有回来 X 上的每个人都说他们用它来做“一切” 我根本不知道那对我来说意味着什么 抓起一个笔记本 写下我一天中所做的一切 网址:
media 0 共 2 项
🎬
视频
#155 🔗 原推
2026-05-03 21:54:00
Blum
> spent 3 years building SaaS for accountants > automation, dashboards, integrations > launched, clients came > then stopped renewing > reached out to each one > got the same answer: "we just hand it to an agent now" > not another tool > the agent files the return itself >… https://twitter.com/Blum_OG/status/2051027148228759637/video/1
中文: 花了3年时间为会计师事务所构建SaaS 自动化、仪表板、集成 已推出，客户来了 随后停止了续订 联系了每一个 得到了同样的回答:“我们现在就把它交给代理人了 不是其他工具 代理将退货文件自行归档
media 0 共 2 项
🎬
视频
#154 🔗 原推
2026-05-03 19:52:33
Blum
3 months ago i was productive > shipping fast > earning well > feeling competent > then i tried to write a strategy doc without AI > 30 pages with Сlaude > closed the chat > tried to write one summary paragraph from memory > it was wrong in three places > i had been a typist for… https://twitter.com/Blum_OG/status/2050993774512546044/video/1
中文: 3个月前我很有效率 运输速度快 好赚 感觉有能力 然后我试着在没有人工智能的情况下编写一个策略文档 加30页，含Сlaude 关闭聊天 尝试从记忆中写一个摘要段落 在三个地方都错了 我曾是...... 的打字员
media 0 共 2 项
🎬
视频
#153 🔗 原推
2026-05-03 17:39:56
Blum
3 months ago i was productive > shipping fast > earning well > feeling competent > then i tried to write a strategy doc without AI > 30 pages with claude > closed the chat > tried to write one summary paragraph from memory > it was wrong in three places > i had been a typist for… https://twitter.com/Blum_OG/status/2050986381258936436/video/1
中文: 3个月前我很有效率 运输速度快 好赚 感觉有能力 然后我试着在没有人工智能的情况下编写一个策略文档 加底;30页带 claude 关闭聊天 尝试从记忆中写一个摘要段落 在三个地方都错了 我曾是...... 的打字员
media 0 共 2 项
🎬
视频
#152 🔗 原推
2026-05-03 17:10:33
Blum
> asked an agent to build a database > every AI data center in the US > names, locations, capacity, operators > one agent would take 8 hours > Kimi launched waves of agents in parallel > each one took its own region > every source validated - 1500 rows > file ready in one evening… https://twitter.com/Blum_OG/status/2050958761804840998/video/1
中文: 请求代理人建立一个数据库 美国的每一个人工智能数据中心 姓名、地点、容量、运营商 一名经纪人需要8个小时 基米同时发起了代理浪潮 每个地区都占据着自己的位置 每个已验证源 - 1500 行 晚上就准备好文件......
media 0 共 2 项
🎬
视频
#151 🔗 原推
2026-05-03 15:20:48
Blum
> quit my job to sell AI automations > posted in a Slack community > got my first client in 3 days > opened Claude > built the skill file > connected the MCP servers > ran it on their real data > sent the result > silence for 2 hours > then: "this just saved me 8 hours a week" >… https://twitter.com/Blum_OG/status/2050711249416024135/video/1
中文: 辞职出售人工智能自动化 发布于一个Slack社区 在3天内拿到了我的第一个客户 开张的克劳德 已构建技能文件 连接的MCP服务器 使用他们的真实数据来运行 发送结果 静默2小时 然后:“这每周只为我节省了8个小时” 网址:
media 0 共 2 项
🎬
视频
#150 🔗 原推
2026-05-02 22:57:16
Blum
> decided to do it properly > spun up a multi-agent pipeline > agents writing code in parallel > each making their own assumptions about architecture > merged at the end > conflicts everywhere > a week of debugging > then i spot this take: > "parallel agents writing code make… https://twitter.com/Blum_OG/status/2050683021519405163/video/1
中文: 决定做妥当 推出一条多智能管道 > 代理并行编写代码 每个人对架构都做出自己的假设 最后合并 各地冲突 一周的调试 然后，我发现以下内容: 网址:“编写代码的并行代理......
media 0 共 2 项
🎬
视频
#149 🔗 原推
2026-05-02 21:05:06
Blum
> decided to do it properly > spun up a multi-agent pipeline > agents writing code in parallel > each making their own assumptions about architecture > merged at the end > conflicts everywhere > a week of debugging > then i spot this take: > "parallel agents writing code make… https://twitter.com/Blum_OG/status/2050674334910586986/video/1
中文: 决定做妥当 推出一条多智能管道 > 代理并行编写代码 每个人对架构都做出自己的假设 最后合并 各地冲突 一周的调试 然后，我发现以下内容: 网址:“编写代码的并行代理......
media 0 共 2 项
🎬
视频
#148 🔗 原推
2026-05-02 20:30:35
Blum
RT @Blum_OG: > claude rewrites your whole doc when you asked to fix one line > gives you confident stats that are completely made up > forg…
中文: RT @Blum_OG: > 当你要求修复一条线时，整个文档会重新编写 提供完全由此组成的自信数据 结了点;放弃......
#147 🔗 原推
2026-05-02 09:31:46
Blum
> claude rewrites your whole doc when you asked to fix one line > gives you confident stats that are completely made up > forgets everything the next session. every time > turns out one text file fixes all 3 problems permanently > CLAUDEmd - Claude reads it automatically on every… https://twitter.com/Blum_OG/status/2050351155067433117/video/1
中文: 当你要求修一行时，克劳德会重写你的整个文档 提供完全由此组成的自信数据 下次会议就把一切都忘了。 事实证明，一个文本文件会永久修复所有3个问题 > 克劳迪·克劳德——克劳德在每条网站上自动阅读......
media 0 共 2 项
🎬
视频
#146 🔗 原推
2026-05-01 23:06:23
Blum
> most companies are stuck at AI level 1 > AI helps you personally, but nothing around you changed > 3 questions to find out where you actually are: > can AI touch your real systems or just your notes? > if your best AI user quit - would the team even notice? > did anyone outside… https://twitter.com/Blum_OG/status/2050329192601125144/video/1
中文: 大多数公司都停留在人工智能一级 人工智能对你个人有帮助，但周围什么都没有改变 3个问题，帮助你弄清楚自己究竟在哪里: 人工智能可以触及你的真实系统，还是只是你的笔记? 如果你最好的AI用户退出了——团队会注意到吗? 外面有人......
media 0 共 2 项
🎬
视频
#145 🔗 原推
2026-05-01 21:39:07
Blum
> met a dev who shipped 3 apps and made $0 > ask him why, expecting a product problem > "Apple held my money and i never got paid" > turns out he never signed the banking agreement > go home, read everything about how apple pays devs > Apple holds your money for 45 days minimum… https://twitter.com/Blum_OG/status/2050211442339758531/video/1
中文: 认识了一位开发者，他出货了3个应用程序，赚了0美元 问他为什么，预计会出现产品问题 苹果公司持有我的钱，而我从未拿到过钱 事实证明，他从未签署过银行协议 回家，读一读苹果如何支付开发者的一切 加特;苹果至少要存45天资金......
media 0 共 2 项
🎬
视频
#144 🔗 原推
2026-05-01 13:51:13
Blum
RT @Blum_OG: > open Cowork on a random monday > notice /schedule. never touched it before > set 7am inbox triage just to see what happens >…
中文: RT @Blum_OG: > 随机开放 Cowork 注意/时间表。以前从未接触过 设置7am收件箱分类，以查看情况 ......
#143 🔗 原推
2026-05-01 09:38:53
Blum
> open Cowork on a random monday > notice /schedule. never touched it before > set 7am inbox triage just to see what happens > wake up tuesday. inbox sorted. drafts written. summary saved > spend the next hour automating everything i hated doing > churn warnings. investor… https://twitter.com/Blum_OG/status/2049998504714764783/video/1
中文: 在随机的月份打开 Cowork 注意/时间表。以前从未接触过 设置7am收件箱分类，以查看情况 > 于 TUesday 醒醒。邮箱排序。草稿写。摘要保存 待在接下来的一小时里，将我讨厌的每件事都自动化 > 流失警告。投资者...
media 0 共 2 项
🎬
视频
#142 🔗 原推
2026-04-30 23:45:05
Blum
> hit claude's limit. again > actually audit what's burning my tokens > using opus for brainstorming > one giant chat open for weeks - claude re-reading 40k tokens every reply > spend one evening fixing it > haiku -> sonnet -> opus only for final output > new chat per task. two… https://twitter.com/Blum_OG/status/2049952730320703707/video/1
中文: 再次达到了引人的极限。 实际审核我的代币正在燃烧什么 使用 opus 进行头脑风暴 一个大型聊天持续数周——每次回复都会重读4万个代币 晚上花点时间修 > haiku -> sonnet -> 专集仅用于最终产出 每项任务新增聊天。两个......
media 0 共 2 项
🎬
视频
#141 🔗 原推
2026-04-30 20:43:11
Blum
RT @Blum_OG: > meet a burned-out ops manager > he says his company loses £40K/month to bad internal comms > SOPs in slack. decisions in ema…
中文: RT @Blum_OG: > 认识一位筋疲力尽的运营经理 他称公司因内部通信不善每月亏损4万英镑 松懈的SOP。在 ema 中做出决策......
#140 🔗 原推
2026-04-30 05:02:34
Blum
> meet a burned-out ops manager > he says his company loses £40K/month to bad internal comms > SOPs in slack. decisions in email. knowledge in heads of people who quit > spend 2 days building a RAG-lite system in Obsidian > show him a demo on friday > he doesn't blink at £1,500… https://twitter.com/Blum_OG/status/2049611572587434420/video/1
中文: 认识一位筋疲力尽的运营经理 他称公司因内部通信不善每月亏损4万英镑 邮箱:SOPs: slack. 电子邮件中的决策。关于辞职人员负责人的知识 在奥西迪安建造一个RAG-lite系统，花2天时间 星期五给他看演示 他不以1500英镑的价格眨眼......
media 0 共 2 项
🎬
视频
#139 🔗 原推
2026-04-29 22:07:33
Blum
> read an article about open-source AI tools > average startup pays $40k/year in subscriptions > spend a weekend replacing all of it > ollama. chroma. vllm. langchain. $0/month > same stack. same output. > start offering "AI cost audits" to startups > avg client saves $35k/year >… https://twitter.com/Blum_OG/status/2049565833614283188/video/1
中文: 阅读一篇关于开源人工智能工具的文章 普通初创企业每年支付4万美元的订阅费 花一个周末来代替全部 加特;奥拉玛。色度。vllm. langchain。每月0美元 相同的堆栈。相同的输出。 开始向初创企业提供“人工智能成本审计” 客户每年节省3.5万美元 网址:
media 0 共 2 项
🎬
视频
#138 🔗 原推
2026-04-29 19:05:48
Blum
RT @Blum_OG: > install Hermes > use it like a fancy chatbot for 3 months > stumble on SOULmd by accident > realize the agent has been waiti…
中文: RT @Blum_OG: > 安装爱马仕 使用它像一个花哨的聊天机器人长达三个月 偶然偶然发现了灵魂 意识到代理人一直在等待......
#137 🔗 原推
2026-04-29 06:09:09
Blum
me at 2am realizing memory isn't a feature. it's the substrate. the whole time, i've just been building on sand https://twitter.com/Blum_OG/status/2049239782623261046/video/1
中文: 凌晨2点时，我意识到记忆不是一种特征，而是基质。 一直，我一直在沙地上建造
media 0 共 2 项
🎬
视频
#136 🔗 原推
2026-04-28 21:30:11
Blum
> install Hermes > use it like a fancy chatbot for 3 months > stumble on SOULmd by accident > realize the agent has been waiting for instructions this whole time > spend one weekend properly setting it up > MEMORYmd knows my projects. USERmd knows how I think > /branch so I never… https://twitter.com/Blum_OG/status/2049217869054112224/video/1
中文: 安装 Hermes 使用它像一个花哨的聊天机器人长达三个月 偶然偶然发现了灵魂 意识到代理人一直在等待指示 花一个周末好好设置 MEMORYMD 了解我的项目。美国人知道我是怎么想的 嗯;/branch，所以我从不......
media 0 共 2 项
🎬
视频
#135 🔗 原推
2026-04-28 20:03:07
Blum
> meet a designer at a hackathon > tells you AI-generated websites all look the same > you go home and think about it > realize everyone prompts code directly > nobody separates the visual step from the coding step > spend a weekend building an image-to-code approach > generate… https://twitter.com/Blum_OG/status/2049152257749647702/video/1
中文: 在黑客马拉松上认识一位设计师 告诉你，由人工智能生成的网站看起来都一样 你回家去想这件事 意识到每个人都会直接提示代码 没有人将视觉步骤与编程步骤区分开来 花一个周末时间建立一种从图像到代码的方法 获取......
media 0 共 2 项
🎬
视频
#134 🔗 原推
2026-04-28 15:42:24
Blum
RT @Blum_OG: while you're farming points, others are getting paid in stables no agency. no follower minimum. no "trust the tokenomics." p…
中文: RT @Blum_OG:当你在农场时，其他人却在马厩里得到报酬 无需代理。无关注者最低要求。无需“信任代币组学”。 p...
#133 🔗 原推
2026-04-28 13:43:11
Blum
> no CLAUDEmd > wondered why Claude kept breaking things > turns out I never told it anything > wrote 60 lines - commands, rules, architecture > Claude stopped guessing > stopped rewriting files I didn't ask it to touch > stopped making decisions above its pay grade > 45 mins… https://twitter.com/Blum_OG/status/2048911405370470846/video/1
中文: 不，无克劳德德 想知道克劳德为何一直把事情打碎 事实证明，我什么都没说 写了60行——命令、规则、架构 克洛德停止了猜测 停止重写我没要求它去触摸的文件 停止做出高于其薪酬等级的决策 45分钟
media 0 共 2 项
🎬
视频
#132 🔗 原推
2026-04-27 23:45:20
Blum
while you're farming points, others are getting paid in stables no agency. no follower minimum. no "trust the tokenomics." payments arrive rn, not someday @RallyOnChain pays creators for posts in real stablecoins, on-chain AI scores your content and money hits your wallet… https://twitter.com/Blum_OG/status/2048883007000637777/photo/1
中文: 当你在农业积分时，其他人则在马厩里获得报酬 无需代理。无关注者最低要求。无需“信任代币组学”。 付款到了，不是将来 @RallyOnChain 向创作者支付在真实稳定币(链上)的帖子费用 人工智能为你的内容打分，金钱也会冲击你的钱包......
media 0 共 1 项
#131 🔗 原推
2026-04-27 21:52:29
Blum
> leave $80K OpenAI job in 2023 > everyone thinks you're insane > 2 years later: $6.47M ARR > one laptop, zero employees > old model: 30 people, thin margins > new model: you + AI + 80% margins > incumbents still paying 2019 salaries > that gap is the entire game > window: 1996… https://twitter.com/Blum_OG/status/2048546143022571998/video/1
中文: 2023年，请留8.0万美元OpenAI工作 每个人都觉得你疯了 两年后:647万美元 一台笔记本电脑，零员工 老款:30人，利润微薄 新模式:你+人工智能 + 80% 的利润率 现任者仍在支付2019年的薪资 那个差距就是整个游戏 网址:1996年，网址:
media 0 共 2 项
🎬
视频
#130 🔗 原推
2026-04-26 23:33:55
Blum
january 2027. agents run everything. your job title didn't exist. you could've seen it in 2026. you did nothing. https://twitter.com/Blum_OG/status/2048485000107147634/video/1
中文: 2027年1月。 代理人管理一切。你的职位名称不存在。 2026年你可能见过它，你什么也没做。
media 0 共 2 项
🎬
视频
#129 🔗 原推
2026-04-26 19:30:57
Blum
RT @Blum_OG: > switch to Opus 4.7 > realize my whole Claude setup was vibes > no real CLAUDE. md, no memory structure, no permission modes…
中文: RT @Blum_OG: > 切换到 Opus 4.7 意识到我整个克劳德的设定都是氛围 没有真正的克劳德。MD，没有记忆结构，没有权限模式......
#128 🔗 原推
2026-04-26 11:20:11
Blum
> switch to Opus 4.7 > realize my whole Claude setup was vibes > no real CLAUDE. md, no memory structure, no permission modes > spend 30 minutes reading the article below > rewrite instructions: 800 lines of noise -> 40 lines of signal > every repeated workflow becomes a skill >… https://twitter.com/Blum_OG/status/2048183016892268714/video/1
中文: 切换到 Opus 4.7 意识到我整个克劳德的设定都是氛围 没有真正的克劳迪。md，没有内存结构，没有权限模式 花30分钟阅读以下文章 重修说明:800行噪声 ->40 行信号 每一次重复的工作流程都变成了一项技能 网址:
media 0 共 2 项
🎬
视频
#127 🔗 原推
2026-04-25 23:30:59
Blum
> running an agency. good clients, good revenue, zero free time > same loop every day: emails, reports, content, repeat > read about mapping every task as "needs me" or "doesn't" > tried it on a weekend > 65% of my work was just pattern matching > spent two weeks building… https://twitter.com/Blum_OG/status/2048131660521304545/video/1
中文: 经营一家机构。客户好，收入好，空闲时间 每天循环相同:电子邮件、报告、内容、重复 阅读关于将每个任务映射为“需要我”或“不需要”的内容 周末时尝试了 我65%的作品只是模式匹配 花了两周时间进行建设......
media 0 共 2 项
🎬
视频
#126 🔗 原推
2026-04-25 20:06:54
Blum
RT @Blum_OG: > try GPT Images 2.0 for the first time > type "make me a cool robot character" > get a cool robot character > think you've ma…
中文: RT @Blum_OG: > 首次尝试 GPT Images 2.0 类型“让我变得酷炫的机器人角色” 获取一个酷炫的机器人角色 你觉得你......
#125 🔗 原推
2026-04-25 08:42:27
Blum
> try GPT Images 2.0 for the first time > type "make me a cool robot character" > get a cool robot character > think you've mastered it > but actually you haven't > you're using a visual production system like a vending machine > pros aren't prompting better > they're building… https://twitter.com/Blum_OG/status/2047795874294882414/video/1
中文: 首次尝试 GPT Images 2.0 类型“让我变得酷炫的机器人角色” 获取一个酷炫的机器人角色 你觉得你已经掌握了它了 但实际上你还没有 你正在使用像自动售货机一样的视觉制作系统 优点并没有促使你变得更好 他们正在建设......
media 0 共 2 项
🎬
视频
#124 🔗 原推
2026-04-24 21:52:37
Blum
RT @Blum_OG: multi-agent systems are finally working in production not swarms. not parallel writers. this narrower pattern: one writer +…
中文: RT @Blum_OG:多代理系统终于在生产中发挥作用 不是群，不是平行的写笔。这种狭义的模式: 一位作家 +...
#123 🔗 原推
2026-04-24 01:56:43
Blum
multi-agent systems are finally working in production not swarms. not parallel writers. this narrower pattern: one writer + multiple agents contributing intelligence around it what's working today: > code-review loop (yes, even self-review) - clean-context reviewer catches ~2… https://twitter.com/Blum_OG/status/2047383476677251408/video/1
中文: 多代理系统终于在生产中发挥作用了 不是群，不是平行的写词。这种狭义的模式: 一名撰稿人+多名特工围绕其贡献情报 今天有什么工作: 代码审查循环(是的，甚至自我审核) - 清白的复习员抓到 ~2...
media 0 共 2 项
🎬
视频
#122 🔗 原推
2026-04-23 18:33:54
Blum
RT @Blum_OG: cut design costs from $3k to $20/mo no designer needed, no Figma skills required, just Claude Design workflow: > open Claude…
中文: RT @Blum_OG:将设计成本从3万美元降至每月20美元 无需设计师，无需Figma技能，只有克劳德设计 工作流程: 加以;克劳德开
#121 🔗 原推
2026-04-23 16:00:53
Blum
RT @Blum_OG: your agent will keep making the same mistake until you turn it into a skill > problem most agent "fixes" are vibe engineeri…
中文: RT @Blum_OG:你的经纪人会一直犯同样的错误 直到你把它变成一项技能 问题 大多数代理“修复”都是 vibe neginesi...
#120 🔗 原推
2026-04-23 05:27:51
Blum
cut design costs from $3k to $20/mo no designer needed, no Figma skills required, just Claude Design workflow: > open Claude Design > create a project > drop in context: reference images, brand guidelines, copy > write your prompt: goal + layout + audience + tone > first draft… https://twitter.com/Blum_OG/status/2047091226848194683/video/1
中文: 将设计成本从3万美元降低到每月20美元 无需设计师，无需Figma技能，只有克劳德设计 工作流程: >开放式克劳德设计 创建项目 背景:参考图片、品牌指南、复制 快点:写下你的提示:目标+布局 + 受众+语调 草案;初稿......
media 0 共 2 项
🎬
视频
#119 🔗 原推
2026-04-22 23:12:36
Blum
your agent will keep making the same mistake until you turn it into a skill > problem most agent "fixes" are vibe engineering: bigger system prompt, "please don't hallucinate" works until the first complex query then the same mistake - just a different input > fix -… https://twitter.com/Blum_OG/status/2047062632797290940/video/1
中文: 你的经纪人会一直犯同样的错误 直到你把它变成一项技能 问题 大多数代理“修复”都是氛围工程: 系统提示更大，“请不要产生幻觉” 一直工作到第一个复杂的查询 然后是同一个错误——只是输入不同 修复 - . .
media 0 共 2 项
🎬
视频
#118 🔗 原推
2026-04-22 21:18:58
Blum
RT @Blum_OG: here’s why your Claude is burning through tokens the main cost issue isn’t the model being dumb it’s bad architecture of how…
中文: RT @Blum_OG:这就是你的克劳德因代币而被烧焦的原因 主要成本问题并不是模型的愚蠢 糟糕的架构是如何......
#117 🔗 原推
2026-04-22 14:50:01
Blum
10 habits that stop u from hitting Claude's limits: > edit your prompt, don't send a follow-up > start a fresh chat every 15–20 messages > batch questions into one message > upload recurring files to Projects - upload once, cache forever > set up Memory & User Preferences - skip… https://twitter.com/Blum_OG/status/2046818859337101572/video/1
中文: 阻止你触及克劳德极限的10个习惯: 编辑你的提示，不要发送后续消息 每15到20条消息开始一次新的聊天 将批处理问题发送成一条消息 将循环文件上传至 Projects - 上传一次，永久缓存 设置内存和放大器;用户偏好设置 - 跳过......
media 0 共 2 项
🎬
视频
#116 🔗 原推
2026-04-22 05:10:18
Blum
> open 6 tabs every morning > email. calendar. slack. stripe. github. notion. > 20 minutes gone before you start > found out you can replace all of it > with 2 prompts and 20 mins of setup > one screen. live data. rebuilds itself at 06:00 > yesterday's revenue. pipeline. what… https://twitter.com/Blum_OG/status/2046704588880048622/video/1
中文: 每天早晨打开6个标签页 邮箱:电子邮箱:日历。松弛。条纹。github。概念。 开始前20分钟 发现你可以更换全部 加时;配备2个提示和20分钟设置 一个屏幕。实时数据。于06:00重新构建 昨日的收入。管道。什么......
media 0 共 2 项
🎬
视频
#115 🔗 原推
2026-04-21 21:36:14
Blum
here’s why your Claude is burning through tokens the main cost issue isn’t the model being dumb it’s bad architecture of how the agent interacts with the backend > 3 token black holes - bloated documentation MCP tools often dump entire pages instead of specific answers… https://twitter.com/Blum_OG/status/2046673572048044094/video/1
中文: 这就是你的克劳德因代币而燃烧的原因 主要成本问题并不是模型的愚蠢 代理与后端交互的架构很差 3个暗号黑洞 - 文档内容严重 MCP工具通常会丢弃整页，而不是具体答案......
media 0 共 2 项
🎬
视频
#114 🔗 原推
2026-04-21 19:32:59
Blum
RT @Blum_OG: how to use Claude for FREE effectively where to get free access: > standard free tier (just use claude ai) > open source de…
中文: RT @Blum_OG:如何有效免费使用克劳德 免费访问的地点: 标准免费等级(仅使用 claude ai) >开源...
#113 🔗 原推
2026-04-21 14:35:14
Blum
how to use Claude for FREE effectively where to get free access: > standard free tier (just use claude ai) > open source developer program (Claude Max 20x for 6 months) > free API credits (for developers and automation creators) > third-party apps (look for services that… https://twitter.com/Blum_OG/status/2046332582477037701/video/1
中文: 如何有效使用克劳德的免费 免费访问的地址: 标准免费等级(仅使用 claude ai) >开源开发者程序(Claude Max 20x，持续6个月) 免费API积分(面向开发者和自动化创建者) 第三方应用程序(请查看以下服务):
media 0 共 2 项
🎬
视频
#112 🔗 原推
2026-04-20 20:58:01
Blum
3 Hermes Skills worth your attention first off, this isn't just some basic chatbot It’s a self-learning agent that builds out skills based on ur usage > X Insight Mining the agent automatically parses info from accounts you’re interested in It prioritizes using the official… https://twitter.com/Blum_OG/status/2046300004638503269/video/1
中文: 3 个值得你关注的爱马仕技能 首先，这不仅仅是一个基本的聊天机器人 它是一种基于我们使用情况来培养技能的自学者 > X Insight 挖矿 该代理会自动从您感兴趣的账户中解析信息 它优先使用官方...
media 0 共 2 项
🎬
视频
#111 🔗 原推
2026-04-20 18:48:34
Blum
naked LLMs are useless out of the box the model needs tuning based on what it's being used for > you must provide it with harness, skill files, and MCP servers - only then will you get the desired output with minimal time spent > you have to know which tools to give the AI to… https://twitter.com/Blum_OG/status/2045981266416443580/video/1
中文: 裸露的LLM在禁区外毫无用处 该模型需要根据其所使用的内容进行调优 您必须提供线束、技能文件和MCP服务器 只有这样，你才能在花费最少的时间时获得所需的输出 你必须知道把人工智能提供给哪些工具......
media 0 共 2 项
🎬
视频
#110 🔗 原推
2026-04-19 21:42:01
Blum
MCP servers that will save you serious money: > search and research - TavilyAI - Firecrawl - Exa > memory and context - Memory MCP - Notion MCP > automation - Zapier - Playwright > data and web extraction - Apify - Bright Data > development and client work - GitHub -… https://twitter.com/Blum_OG/status/2045944415525745063/video/1
中文: 可为您节省大量资金的MCP服务器: 查找与研究 - 塔维利艾 - 火虫 - 例 内存和上下文 - 内存MCP - 通知 MCP 自动化 - 扎皮尔 - 剧作家 数据与网页提取 - 认证 - 明亮的数据 开发与客户工作 - GitHub - . . .
media 0 共 2 项
🎬
视频
#109 🔗 原推
2026-04-19 19:15:35
Blum
Claude stopped being "just a tool" a long time ago It’s a MASSIVE force multiplier for your whole workflow in terms of practical use, Claude is on par with ChatGPT and tbh, in some areas, it actually pulls ahead so, It’s a crime not to know this Anthropic masterpiece but… https://twitter.com/Blum_OG/status/2045852078111875282/video/1
中文: 克劳德很久以前就不再只是一个工具了 这是您整个工作流程的大规模力量倍增器 在实际使用方面，克劳德与ChatGPT相当 而且，在某些领域，它确实会领先 所以，不了解这部人类史诗杰作是一种犯罪 但......
media 0 共 2 项
🎬
视频
#108 🔗 原推
2026-04-19 13:08:40
Blum
how to stack cash with Meta's LLM (Muse Spark) tech stack: - Muse Spark - distribution layer, provides access to 3B Meta users - Claude/GPT - reasoning engine for building logic and solving tasks > micro apps building and selling Muse Spark beats the competition at building… https://twitter.com/Blum_OG/status/2045603804989776272/video/1
中文: 如何与Meta的LLM(Muse Spark)一起堆叠现金 技术栈: - Muse Spark - 分发层，可访问 3B Meta 用户 - 克劳德/GPT——构建逻辑和解决任务的推理引擎 微应用构建与销售 缪斯·斯帕克在建设中击败了竞争对手......
media 0 共 2 项
🎬
视频
#107 🔗 原推
2026-04-18 20:42:07
Blum
the single-agent approach is dead building an entire AI squad is where the real efficiency is at the core idea is setting up a team where each sub-agent is responsible for its own task and you put one agent at the head - the coordinator it monitors the whole process and… https://twitter.com/Blum_OG/status/2045535519891988702/video/1
中文: 单一代理方法已经消亡 组建一个完整的人工智能团队才是真正的效率所在 核心理念是组建一个团队 每个子代理负责自身任务 你把一个代理人放在了头上——协调员 它监控整个过程，并......
media 0 共 2 项
🎬
视频
#106 🔗 原推
2026-04-18 16:10:46
Blum
RT @Blum_OG: your model isn't the problem - your harness is here are 3 tips for building it the right way > configuration optimization -…
中文: RT @Blum_OG:你的型号不是问题——你的线束是 以下是以正确方式构建它的3个技巧 配置优化 -...
#105 🔗 原推
2026-04-18 11:13:24
Blum
your model isn't the problem - your harness is here are 3 tips for building it the right way > configuration optimization - write your own instruction sets (CLAUDEmd, AGENTSmd) - treat every token like gold (only keep the essentials in global files) - manage MCP servers - teach… https://twitter.com/Blum_OG/status/2045274376841671009/video/1
中文: 你的模型不是问题——你的线束是 以下是以正确方式构建它的3个技巧 配置优化 - 编写你自己的指令集(CLAUDEMD，AGENTSmd) - 将每个代币都像黄金一样对待(仅将必需品保存在全球文件中) - 管理MCP服务器 - 教学......
media 0 共 2 项
🎬
视频
#104 🔗 原推
2026-04-17 22:53:05
Blum
another huge release from Anthropic! Opus 4.7 - the one that doesn't forgive blurry prompts price and quality now directly depend on the accuracy of the request If you write like before (for Opus 4.6), you get: - tons of unnecessary reasoning - answers are technically correct,… https://twitter.com/Blum_OG/status/2045228369726570858/video/1
中文: 又一次来自《蚁皮》的重大释放! Opus 4.7——无法原谅模糊提示的提示 价格和质量现在直接取决于请求的准确性 如果你写得像之前一样(为 Opus 4.6 写)，你会得到: - 大量不必要的推理 - 答案在技术上是正确的......
media 0 共 2 项
🎬
视频
#103 🔗 原推
2026-04-17 19:50:16
Blum
RT @Blum_OG: one-shot sloperator era is over multi-agent architecture now dominates in efficiency the problem: single-agent chaos - exces…
中文: RT @Blum_OG:单拍斜坡时代已经结束 多智能架构在效率方面现已占据主导地位 问题:单一代理混乱 - 过处......
#102 🔗 原推
2026-04-17 10:55:10
Blum
one-shot sloperator era is over multi-agent architecture now dominates in efficiency the problem: single-agent chaos - excessive cognitive load - сontext bloat - zero parallelism the fix: multi-agent architecture system works by splitting up the roles: > Orchestrator -… https://twitter.com/Blum_OG/status/2045048604390502742/video/1
中文: 单次滑坡时代已经结束 多智能架构在效率方面现已占据主导地位 问题:单一代理混乱 过度的认知负荷 - с 语日臼 - 零平行 修复:多代理架构 系统通过分割角色来发挥作用: 加·奥特;管弦乐 - . . .
media 0 共 2 项
🎬
视频
#101 🔗 原推
2026-04-17 07:55:57
Blum
10 MCP Servers Your Claude Setup Needs > GitHub - managing repos, PRs & code reviews > Sequential Thinking - structured step-by-step reasoning > Memory - persistent context & facts across sessions > Slack - reading channels & posting updates > AWS - cloud infra & deployment… https://twitter.com/Blum_OG/status/2044875559059210460/video/1
中文: 10 个 MCP 服务器，满足您的克劳德设置需求 GitHub - 管理仓库、公关和内容;代码审查 顺序思维——结构化的分步推理 > 记忆——持续的上下文和内容;跨会话的事实 > Slack - 阅读频道和动态;发布更新 网址:AWS - 云基础设施和环境部署 .. . .
media 0 共 2 项
🎬
视频
#100 🔗 原推
2026-04-16 20:28:19
Blum
Claude’s context is the key to getting the output u want most prefer just continuing sessions this often leads to context clutter and crap outputs after 5 ways to work with context: > continue - keep going in the current session > /rewind - roll back to a previous message… https://twitter.com/Blum_OG/status/2044744296423964922/video/1
中文: 克劳德的语境是获取你想要的输出的关键 大多数人更喜欢继续开会 这常常会导致上下文混乱和垃圾输出 处理上下文的5种方法: 继续——继续进行当前会议 返回至以下消息:
media 0 共 2 项
🎬
视频
#99 🔗 原推
2026-04-16 11:46:44
Blum
RT @Blum_OG: be one of the first to figure out Claude Managed Agents build production AI agents with this tool 10x faster deploying cycle…
中文: RT @Blum_OG:成为首批确定克劳德管理代理的人之一 使用此工具构建生产性AI代理 10倍速度更快 部署周期......
#98 🔗 原推
2026-04-16 05:22:58
Blum
be one of the first to figure out Claude Managed Agents build production AI agents with this tool 10x faster deploying cycle: - create an account - build an agent via assistant, template, or from scratch - configure environment: network and access rules - set up a credentials… https://twitter.com/Blum_OG/status/2044504662884421757/video/1
中文: 成为首批确定克劳德管理代理的人之一 使用此工具构建生产性AI代理 10倍速度更快 部署循环: - 创建一个账户 - 通过助手、模板或从零开始构建代理 - 配置环境:网络和访问规则 - 设置凭据......
media 0 共 2 项
🎬
视频
#97 🔗 原推
2026-04-15 19:54:31
Blum
everyone loves tearing open card packs but lbh, in this space plenty of folks are happy to trick u still, there’s one platform with a crystal‑clear approach @BRAVOREADYGAMES built on real-time VRFs and full onchain main advantages: - provably fair - vault-backed - guaranteed… https://twitter.com/Blum_OG/status/2044467208886358369/photo/1
中文: 每个人都喜欢撕开的卡片包 但在这个空间里，很多人很乐意骗你 仍然，有一个平台采用水晶般清晰的方法 @BRAVOREADYGAMES 基于实时VRF和全链构建 主要优点: - 可证明公平 - 金库支撑 - 保证......
media 0 共 1 项
#96 🔗 原推
2026-04-15 17:25:41
Blum
RT @Blum_OG: https://x.com/i/article/2042999198648242176
#95 🔗 原推
2026-04-14 20:09:40
Blum
> "have no time," you say every Friday > open inbox: 34 emails > didn't open half > one had today's deadline > you missed it > drop email into AI > run it through a solid prompt > get: priority, category, deadline, draft reply > 40 secs instead of 10 mins > multiply that by 34… https://twitter.com/Blum_OG/status/2044124461503525244/video/1
中文: “没有时间，”你每个星期五都会说 邮箱:34封邮箱 没有打开一半 有今天的最后期限 你错过了 将电子邮件发送至人工智能 通过一个可靠的提示来运行它 获取:优先级、类别、截止日期、草稿回复 40秒，而不是10分钟 将其乘以34......
media 0 共 2 项
🎬
视频
#94 🔗 原推
2026-04-14 18:43:44
Blum
https://x.com/i/article/2042999198648242176
#93 🔗 原推
2026-04-14 16:34:58
Blum
RT @Blum_OG: AI automation - everyone talks, few actually act market's 360M companies, only 50M automated best part: just 1M specialists…
中文: RT @Blum_OG:人工智能自动化——人人都会说话，很少有人真正行动 市场的360M公司，仅50M自动化 最棒的部分:仅1米专科医生......
#92 🔗 原推
2026-04-14 03:47:17
Blum
AI automation - everyone talks, few actually act market's 360M companies, only 50M automated best part: just 1M specialists - we're crazy early here’s the next 6 months for tomorrow’s winners: > month 1: n8n basics & core concepts > month 2: AI integration in workflows >… https://twitter.com/Blum_OG/status/2043774027937001812/video/1
中文: 人工智能自动化——每个人都在说话，很少有人真正行动 市场的360M公司，仅50M自动化 最棒的部分:只有1米的专科医生——我们早早就疯了 以下是未来6个月的获奖者: 第1个月:n8n 基础知识;核心概念 > 月2:工作流程中的人工智能集成
media 0 共 2 项
🎬
视频
#91 🔗 原推
2026-04-13 19:31:14
Blum
RT @Blum_OG: playbook for launching your own AI agency: > pick niche + define offer > build outreach list > send daily outreach (20-30/day…
中文: RT @Blum_OG:推出自己人工智能代理的游戏手册: 选择小众+定义报价 > 建立宣传清单 发送每日宣传(20-30/天)...
#90 🔗 原推
2026-04-13 19:03:31
Blum
playbook for launching your own AI agency: > pick niche + define offer > build outreach list > send daily outreach (20-30/day) > run discovery -> sales calls > close deal ($2.5K-$5K upfront) > onboard & deploy AI system > deliver results -> move to retainer ($1.5K/mo) the… https://twitter.com/Blum_OG/status/2043638115508924857/video/1
中文: 推出您自己的人工智能代理指南: 选择小众+定义报价 > 建立宣传清单 发送每日宣传(20-30/天) 运营 -> 销售电话 优惠(提前2.5万至5万美元) 上载和部署AI系统 实现业绩 -> 迁至留置式(1.5万美元/月)
media 0 共 2 项
🎬
视频
#89 🔗 原推
2026-04-13 10:31:10
Blum
RT @alphabatcher: https://x.com/i/article/1943379427234074624
#88 🔗 原推
2026-04-13 06:47:36
Blum
15 prompts that'll keep your subscription money safe these days platforms are all about subscriptions It's super profitable for them and creates an illusion of value for users but what's actually worth it is a Claude subscription It replaces: > grammar and style checker -… https://twitter.com/Blum_OG/status/2043458877883285868/video/1
中文: 15个提示，以确保您的订阅资金安全 如今的平台都与订阅有关 对他们来说非常有利可图，并为用户营造一种价值感 但真正值得的是克劳德订阅 它取代了: 语法与风格检查 - . .
media 0 共 2 项
🎬
视频
#87 🔗 原推
2026-04-12 22:38:56
Blum
AI is advancing at a crazy speed that's why there's more and more info but the problem is 90% of it is noise here are 10 papers worth your time: > Neural Computers > Memento: Teaching LLMs to Manage Their Own Context > Memory Intelligence Agent (MIA) > Single-Agent LLMs vs.… https://twitter.com/Blum_OG/status/2043419292365119846/video/1
中文: 人工智能正以极快的速度推进 这就是为什么信息越来越多 但问题是90%是噪音 以下是值得您花时间的10篇论文: 神经计算机 > Memento:教授LLM来管理自己的背景 内存智能剂(MIA) 加特;单探员LLMs vs...
media 0 共 2 项
🎬
视频
#86 🔗 原推
2026-04-12 20:01:38
Blum
RT @alphabatcher: Turn your 2-minute repetitive prompt into a 1-second slash command - same prompt every time - version-controlled via Git…
中文: RT @alphabatcher:将你2分钟的重复提示转换为1秒的斜线命令 每次都一样的提示 - 通过 Git 控制版本...
#85 🔗 原推
2026-04-11 22:46:22
Blum
RT @Blum_OG: key advantages of video AGI: > radical cost reduction - a shot now costs lunch money, not days of production - a mistake do…
中文: RT @Blum_OG:视频AGI的主要优势: 大幅降低成本 - 现在花的是午餐钱，而不是几天的生产 - 做错了......
#84 🔗 原推
2026-04-11 21:11:26
Blum
key advantages of video AGI: > radical cost reduction - a shot now costs lunch money, not days of production - a mistake doesn't mean blowing a huge budget, but a failure you can retry dozens of times > development timeline compression - "idea → prompt → video → edit"… https://twitter.com/Blum_OG/status/2043036581834137936/video/1
中文: 视频AGI的主要优势: 大幅降低成本 - 现在花的是午餐钱，而不是几天的生产 - 错误并不意味着浪费巨额预算，而是失败，你可以重试数十次 开发时间压缩 - 想法 → 提示 → 视频 → 编辑...
media 0 共 2 项
🎬
视频
#83 🔗 原推
2026-04-11 18:40:53
Blum
stop just chatting - unleash Claude's full power using Claude is about small levers that give you outsized returns in work here are the key ones: - CLAUDE md file: the file Claude reads and keeps in mind each session - plan mode: run an initial analysis in read-only (saves up… https://twitter.com/Blum_OG/status/2042997351166677334/video/1
中文: 停止聊天——释放克劳德的全部力量 使用克劳德是关于小杠杆，能让你在工作中获得超大回报 以下是关键内容: - CLAUDE md 文件:克劳德读取并记住每个会话的文件 - 计划模式:以只读方式运行初步分析(保存...
media 0 共 2 项
🎬
视频
#82 🔗 原推
2026-04-11 16:05:00
Blum
RT @Blum_OG: Claude is now able to integrate into Word and it runs inside it, in the same window main features: > comment-based editing >…
中文: RT @Blum_OG:克劳德现在能够融入Word 它在它内部运行，在同一个窗口中 主要特点: 以评论为基础的编辑 ......
#81 🔗 原推
2026-04-11 08:03:52
Blum
Claude is now able to integrate into Word and it runs inside it, in the same window main features: > comment-based editing > draft in your template > consistency check > revision summary > workflows into skills > context across Word, PowerPoint, and Excel add-ins Installation… https://twitter.com/Blum_OG/status/2042704268357243266/video/1
中文: 克劳德现在能够融入Word 它在它内部运行，在同一个窗口中 主要特点: 以评论为基础的编辑 在模板中添加 保持一致性 修订摘要 技能工作流程 Word、PowerPoint 和 Excel 附加组件的上下文 安装...
media 0 共 2 项
🎬
视频
#80 🔗 原推
2026-04-10 20:40:23
Blum
RT @Blum_OG: AI stack on Claude only Claude’s capabilities are evolving crazy fast day by day it’s covering more and more of what you act…
中文: RT @Blum_OG:仅限克劳德的AI堆栈 克劳德的能力正在迅速发展 日复一日地，你所做的事正在不断，越来越多......
#79 🔗 原推
2026-04-10 18:46:30
Blum
AI stack on Claude only Claude’s capabilities are evolving crazy fast day by day it’s covering more and more of what you actually need what you can already do with Claude: - coding (Claude Code) - research & intelligence (Claude + web search) - writing & content production -… https://twitter.com/Blum_OG/status/2042537414158356729/video/1
中文: 仅限克劳德平台上的人工智能堆栈 克劳德的能力正在迅速发展 日复一日地涵盖你实际需要的内容 你已经可以用克劳德做些什么了: - 编码(Claude Code) - 研究与研究;智力(克劳德+网络搜索) - 写作与创作;内容制作 - . . .
media 0 共 2 项
🎬
视频
#78 🔗 原推
2026-04-10 09:37:22
Blum
RT @Blum_OG: automation rn is the simplest action that saves tons of time - no coding knowledge required - totally free - tons of guides o…
中文: RT @Blum_OG:自动化 rn 是节省大量时间的最简单方法 - 无需编码知识 - 完全免费 - 数吨指南......
#77 🔗 原推
2026-04-09 22:58:29
Blum
automation rn is the simplest action that saves tons of time - no coding knowledge required - totally free - tons of guides on every possible question all u need is: > install Claude Code > get familiar with the terminal > use a couple of prompts (copy-paste) > an Anthropic… https://twitter.com/Blum_OG/status/2042301063563468930/video/1
中文: 自动化 rn 是节省大量时间的最简单方法 - 无需编码知识 - 完全免费 - 关于每个可能问题的大量指南 你只需要: 安装克劳德密码 了解终端 使用几个提示(复制粘贴) 一个反乌托邦......
media 0 共 2 项
🎬
视频
#76 🔗 原推
2026-04-09 17:58:12
Blum
RT @Blum_OG: 10 Claude prompts that actually boost ur productivity > morning briefing - sorts tasks by urgency of execution > email triage…
中文: RT @Blum_OG:10个真正提高工作效率的克劳德提示 > 早间简报——按执行紧急情况对任务进行排序 邮箱分流......
#75 🔗 原推
2026-04-09 05:14:35
Blum
10 Claude prompts that actually boost ur productivity > morning briefing - sorts tasks by urgency of execution > email triage - assigns categories to emails > repurposer - reworks content from articles for easy consumption > code reviewer - checks code against key problematic… https://twitter.com/Blum_OG/status/2041987179153649796/video/1
中文: 10 克劳德提示，实际上提升了您的工作效率 > 早间简报——按执行紧急措施分类 电子邮件分类 - 为电子邮件分配类别 > 再利用器 - 轻松轻松修改文章内容 代码审核员——检查代码以对关键问题进行检查......
media 0 共 2 项
🎬
视频
#74 🔗 原推
2026-04-08 21:10:56
Blum
RT @HarryTandy: - @Alibaba_Qwen just dropped Qwen3.6-Plus their new flagship model, and it feels like a real step toward AI agents that ar…
中文: RT @HarryTandy:- @阿里巴巴_Qwen刚刚放弃了Qwen3.6-Plus 他们的新旗舰机型，感觉像是朝着人工智能代理迈出的又一步，而人工智能又如此......
#73 🔗 原推
2026-04-08 17:23:36
Blum
new top open-source LLM just landed! Z ai dropped their flagship model - GLM-5.1 best model among open source and #3 globally on key benchmarks capabilities: - offer multiple thinking modes for different scenarios - support real-time streaming responses - have powerful tool… https://twitter.com/Blum_OG/status/2041576339048951971/photo/1
中文: 全新顶级开源LLM刚刚登陆! Z ai 放弃了他们的旗舰机型——GLM-5.1 开源最佳模型，以及全球关键基准的第3个模型 能力: - 为不同场景提供多种思维模式 - 支持实时流媒体回复 - 拥有强大的工具......
media 0 共 1 项
#72 🔗 原推
2026-04-07 17:58:24
Blum
RT @Blum_OG: Anthropic is reaching a new level run-rate revenue has now surpassed $30B (was $9B at the end of 2025) that's not some rando…
中文: RT @Blum_OG:Anthropic 正在达到一个新水平 运行时收入现已超过300亿美元(到2025年底为90亿美元) 那不是个空的......
#71 🔗 原推
2026-04-07 17:02:51
Blum
Anthropic is reaching a new level run-rate revenue has now surpassed $30B (was $9B at the end of 2025) that's not some random stat, it's straight-up massive growth over the past few months for context, closest rival OpenAI's sitting at $25B they also signed a game-changing… https://twitter.com/Blum_OG/status/2041502277719195868/photo/1
中文: 人类正在达到一个新的水平 运行时收入现已超过300亿美元(到2025年底为90亿美元) 这并非随机数据，而是过去几个月来的快速增长 就目前而言，最接近竞争对手OpenAI的售价为250亿美元 他们还签下了一款具有变革性的游戏规则......
media 0 共 3 项 media 1 media 2
#70 🔗 原推
2026-04-07 13:04:06
Blum
RT @Blum_OG: no more unemployed AI users! dude just built an open-source job searcher on Claude Code key functions: > auto-evaluation of…
中文: RT @Blum_OG:不再有失业的AI用户! 老兄刚刚在克劳德密码上建立了一个开源的求职者 关键功能: 对......进行自动评估
#69 🔗 原推
2026-04-07 04:00:27
Blum
no more unemployed AI users! dude just built an open-source job searcher on Claude Code key functions: > auto-evaluation of job offers (based on 10 weighted dimensions) > ATS-optimized PDF CV generation per job description > HUGE job portal scanning work modes (for the… https://twitter.com/Blum_OG/status/2041246412957208876/video/1
中文: 不再有失业的AI用户! 老兄刚刚在克劳德密码上建立了一个开源的求职者 关键功能: > 工作报价的自动评估(基于10个加权维度) >根据职位描述优化PDF CV 大型工作门户扫描 工作模式(适用于
media 0 共 2 项
🎬
视频
#68 🔗 原推
2026-04-06 20:07:24
Blum
RT @Blum_OG: 8 Claude hooks that cover ur blind spots: > auto-format code - code always looks clean > block dangerous commands - stop dest…
中文: RT @Blum_OG:8个遮盖盲区的克劳德钩: 自动格式代码 - 代码看起来总是很干净 阻止危险命令——停止......
#67 🔗 原推
2026-04-05 23:26:31
Blum
8 Claude hooks that cover ur blind spots: > auto-format code - code always looks clean > block dangerous commands - stop destructive actions > protect sensitive files - prevents accidental modifying important files > run tests on edit - tests after every code change > block… https://twitter.com/Blum_OG/status/2040872554609750325/video/1
中文: 8个遮盖盲区的克劳德钩: 自动格式代码 - 代码看起来总是很干净 阻止危险指令——停止破坏性行为 保护敏感文件——防止对重要文件进行意外修改 在每次代码更改后运行测试 ggt;块......
media 0 共 2 项
🎬
视频
#66 🔗 原推
2026-04-05 19:21:49
Blum
AI tools that actually help you grow X: > TweetHunter an AI platform that can: - writes personalized drafts - rewrites existing ones - suggests ideas based on your niche - gives you a scheduler with automation - runs performance analytics + highlights the best posting times >… https://twitter.com/Blum_OG/status/2040559968659849306/photo/1
中文: 真正帮助你发展X的人工智能工具: 推特 一个可以: - 撰写个性化的草稿 - 重写现有的 - 根据你的细分领域提出想法 - 为您提供一个带自动化功能的调度器 - 运行性能分析 + 突出最佳发布时间 网址:
media 0 共 1 项
#65 🔗 原推
2026-04-04 22:39:42
Blum
RT @Blum_OG: TOP LLMs BY USE CASE rn there's no super universal model for all tasks models are now sharpening for specific use cases so…
中文: RT @Blum_OG:按使用案例获取最佳LLM nn 没有适用于所有任务的超通用模型 模型现在正在针对特定用例进行改进 所以......
#64 🔗 原推
2026-04-04 16:22:18
Blum
TOP LLMs BY USE CASE rn there's no super universal model for all tasks models are now sharpening for specific use cases so using different models for different needs yields better results and boosts work efficiency built from Arena (human-voted) + LiveBench (clean benchmark)… https://twitter.com/Blum_OG/status/2040147335918108961/photo/1
中文: 按用途进行顶级LLM nn 没有适用于所有任务的超通用模型 模型现在正在针对特定用例进行改进 因此，针对不同需求使用不同模型能带来更好的效果，并提高工作效率 采用Arena(人工投票)+ LiveBench(净胜选)打造。
media 0 共 1 项
#63 🔗 原推
2026-04-03 19:20:03
Blum
RT @HarryTandy: prompt engineering like we knew it is dying a slow, inevitable death remember all those magic spells like “act like a 20-y…
中文: RT @HarryTandy:我们清楚的工程工程正在缓慢而不可避免的死亡中死去 记住那些魔法咒语，比如“像20岁一样行动......”
#62 🔗 原推
2026-04-03 18:29:19
Blum
RT @Blum_OG: "Freedom has a token. TWT" - they claimed now it's clear what kind of freedom they meant freedom to act with total recklessn…
中文: RT @Blum_OG:“自由有一个令牌。”TWT" - 他们声称 现在清楚他们指的是什么样的自由 完全鲁莽行事的自由......
#61 🔗 原推
2026-04-02 23:10:11
Blum
"Freedom has a token. TWT" - they claimed now it's clear what kind of freedom they meant freedom to act with total recklessness on their statements in sep, at the peak of their loud announcements, they unveiled roadmap but since November, not a single project has been… https://twitter.com/Blum_OG/status/2039793906452492525/photo/1
中文: 自由有象征意义。TWT" - 他们声称 现在清楚他们指的是什么样的自由 以完全鲁莽行事的自由 在他们大声宣布的高峰期，他们发布了路线图 但自11月以来，还没有一个项目......
media 0 共 1 项
#60 🔗 原推
2026-04-02 19:55:39
Blum
RT @Blum_OG: 10 BEST LOCAL LLMs MODELS lbh, everyone's tired of paying huge money for AI subs local models are our saviors in this matter…
中文: RT @Blum_OG:10 款最佳本地 LLM 模型 大家都厌倦了为人工智能潜艇支付巨额费用 本地模型是我们在这件事上的救星......
#59 🔗 原推
2026-04-01 23:54:22
Blum
10 BEST LOCAL LLMs MODELS lbh, everyone's tired of paying huge money for AI subs local models are our saviors in this matter total data privacy + no sub fees ever so, here are the top picks: > GLM-5 top open-weight model rn elite code + long context, but requires serious… https://twitter.com/Blum_OG/status/2039445932102779243/photo/1
中文: 10款最佳本地LLM车型 大家都厌倦了为人工智能潜艇支付巨额费用 本地模型是我们在这件事上的救星 数据隐私总额 + 无任何子费用 以下是精选的精选内容: GLM-5 顶级敞标车型 精英代码+长上下文，但需要严肃的......
media 0 共 1 项
#58 🔗 原推
2026-04-01 20:52:55
Blum
RT @HarryTandy: where is the fuel now? > the token’s value is collapsing > roadmap milestones are left unfulfilled > repeated security fai…
中文: RT @HarryTandy:现在燃料在哪里? 该代币的价值正在崩溃 路线图的里程碑尚未实现 重来安全......
#57 🔗 原推
2026-04-01 20:44:17
Blum
RT @alphabatcher: Coded to go to zero? Down 80% since October 2025 Is this what it was coded for? What about Trust Alpha that you talked…
中文: RT @alphabatcher:编码为零? 自2025年10月以来下降了80% 这是它所编码的吗? 你聊的Trust Alpha怎么样......
#56 🔗 原推
2026-04-01 18:08:16
Blum
RT @pavlitosy4: Dear @cz_binance, I am reaching out to you on behalf of myself (Binance KOL Award 2022 CIS), @toptraders0x (Binance KOL Aw…
中文: RT @pavlitosy4:亲爱的 @cz_binance， 我代表我本人(2022年币安KOL奖)联系您，@toptraders0x(Binance KOL Aw...)
#55 🔗 原推
2026-04-01 12:01:26
Blum
RT @alphabatcher: Today I'm slashing @TrustWallet Trust heavily shilled TWT near the highs, promised token value through its Sept 2025 roa…
中文: RT @alphabatcher:今天我要大幅削减 @TrustWallet 信任在高点附近大幅下跌的TWT，承诺通过其2025年9月的代币价值......
#54 🔗 原推
2026-03-31 23:01:20
Blum
HOW TO CRAFT TRULY EFFECTIVE AI PROMPTS you ask an LLM to for a high-quality report and get back text written with expert-level confidence but packed with total BS familiar? so, to avoid situations like this, you need to understand these basic points: > the “smart but… https://twitter.com/Blum_OG/status/2039052691088036269/photo/1
中文: 如何巧妙地制定有效的人工智能项目 你请一位法学硕士来获取一份高质量的报告 并以专家级自信重新撰写文字 但总BS满载 熟悉吗? 因此，为避免出现类似情况，你需要理解 这些基本要点:
media 0 共 1 项
#53 🔗 原推
2026-03-31 18:50:19
Blum
RT @pavlitosy4: Maybe it’s time to respond to the community? What’s going on with $TWT? What about the roadmap? The longer you stay silen…
中文: RT @pavlitosy4:也许是时候回应这个社区了? TWT 怎么了? 路线图呢? 你沉默的时间越长......
#52 🔗 原推
2026-03-31 10:02:50
Blum
RT @Blum_OG: https://x.com/i/article/2036169368388751360
#51 🔗 原推
2026-03-30 23:38:14
Blum
https://x.com/i/article/2036169368388751360
#50 🔗 原推
2026-03-30 19:28:44
Blum
BANGER RELEASE FROM ZAI - GLM 5.1 modified GLM 5, lighter, faster built for: > agent coding and engineering > adaptive reasoning > more predictable instruction following > long-horizon engineering tasks GLM 5 - the brains. GLM 5.1 - the work beast
中文: 从ZAI到GLM 5.1的更大发布 改性GLM 5，更轻，速度更快 为: 代理编码与工程 适应性推理 更可预测的指导 长距离工程任务 GLM 5 - 大脑。GLM 5.1 - 工作大动物
#49 🔗 原推
2026-03-28 17:56:13
Blum
RT @alphabatcher: 🚨formula to $20K/mo with apps: 1: validate idea 2: build with AI in days 3: nail your onboarding 4: UGC + influencers at…
中文: RT @alphabatcher:使用应用程序，每月 20 万美元的 🚨 公式: 1:验证想法 2:几天内使用人工智能进行构建 3: 锁定你的入职 4:UGC + 网红在...
#48 🔗 原推
2026-03-27 12:54:58
Blum
vibe coding has become so simple AI agent pipeline: spec → code → tests → deploy all that's left for vibecoders: https://twitter.com/Blum_OG/status/2037279068023046333/video/1
中文: 氛围编码变得如此简单 人工智能代理管道:规范 → 代码 → 测试 → 部署 留给 viecoders 的所有内容:
media 0 共 2 项
🎬
视频
#47 🔗 原推
2026-03-26 21:22:35
Blum
RT @alphabatcher: These 5 Claude Cowork plugins actually deliver $285 billion in software stock value wiped out Because of 5 plugins. Her…
中文: RT @alphabatcher:这5个Claude Cowork插件实际上提供了 软件价值2850亿美元 因为有5个插件。她......
#46 🔗 原推
2026-03-26 15:15:20
Blum
claude - the fastet moving contender for leadership not everyone understands how much anthropic is crushing ai race these crazy guys made 72 releases in just 52 days moreover, many of these are absolute bangers the main ones: > models & core platform - opus 4.6: upgraded… https://twitter.com/Blum_OG/status/2036890015851290761/photo/1
中文: 引人观爱——领导力的快速发展竞争者 并非每个人都明白人类在破坏种族是多么沉重 这些疯子在短短52天内就发布了72次 此外，其中许多都是绝对的 主要的: 型号和模式;核心平台 - 第4.6项:升级版......
media 0 共 1 项
#45 🔗 原推
2026-03-25 19:36:38
Blum
RT @Blum_OG: antarctic funded account program is officially launched! trade in real market environment with platform capital no personal…
中文: RT @Blum_OG: 南极资助账户计划正式启动! 以平台资本进行真实市场交易 不是个人的......
#44 🔗 原推
2026-03-24 19:26:03
Blum
antarctic funded account program is officially launched! trade in real market environment with platform capital no personal risks - the platform provides the capital > reg period: mar 11 10:00 - mar 24 10:00 (UTC) > trading competition: mar 25 10:00 - apr 14 10:00 (UTC) >… https://twitter.com/Blum_OG/status/2036478041099227378/photo/1
中文: 南极资助账户计划正式启动! 以平台资本进行真实市场交易 无个人风险——该平台提供资本 时节:马喣 11:00 - mar 24 10:00(UTC) 交易竞争:市场 25 上午10:00 - 4月14日 10:00(UTC)
media 0 共 1 项
#43 🔗 原推
2026-03-24 16:19:35
Blum
RT @Blum_OG: https://x.com/i/article/2036167441332707328
#42 🔗 原推
2026-03-24 13:19:28
Blum
RT @alphabatcher: Exact stack to build your iOS app in 3-7 days: here's what actually ships a working iOS app fast: code: - Cursor (AI-p…
中文: RT @alphabatcher:在3-7天内构建iOS应用的精确堆栈: 真正快速发布一款正在使用的iOS应用是什么: 代码: - Cursor(AI-p...)
#41 🔗 原推
2026-03-24 12:57:27
Blum
https://x.com/i/article/2036167441332707328
#40 🔗 原推
2026-03-23 20:00:05
Blum
RT @alphabatcher: Best GitHub repos for AI Agents: ​ 1. Free AI Agents Resources https://github.com/avinash201199/free-ai-agents-resources ​ 2. AutoGen by Microsoft Research…
中文: RT @alphabatcher:适用于人工智能代理的最佳 GitHub 仓库: ​ 1。免费人工智能代理资源 ​ 2。微软研究院的AutoGen...
#39 🔗 原推
2026-03-23 19:44:44
Blum
RT @Blum_OG: my ego the millisecond OpenClaw DMs me via tg bot: (I spent hours on it and it was an error message) https://twitter.com/Blum_OG/status/2035828226338254964/video/1
中文: RT @Blum_OG:我的自我:毫秒级的OpenClaw通过tg bot来DM我: (我花了数小时，上面有一条错误信息)
media 0 共 2 项
🎬
视频
#38 🔗 原推
2026-03-23 09:05:33
Blum
RT @HarryTandy: https://x.com/i/article/2035852105919905792
#37 🔗 原推
2026-03-23 02:47:14
Blum
my ego the millisecond OpenClaw DMs me via tg bot: (I spent hours on it and it was an error message) https://twitter.com/Blum_OG/status/2035828226338254964/video/1
中文: 我的自我毫秒级OpenClaw通过tg机器人来DM我: (我花了数小时，上面有一条错误信息)
media 0 共 2 项
🎬
视频
#36 🔗 原推
2026-03-22 21:17:27
Blum
RT @alphabatcher: 20 Must-Read AI Articles Right Now: ​ @0x_kaize: https://x.com/0x_kaize/status/2032527256355221954 "70+ Free Subscriptions - Become a Student" ​ @E…
中文: RT @alphabatcher:目前有20篇必读AI文章: ​ @0x_kaize: 70 多个免费订阅 - 成为学生 ​ @E...
#35 🔗 原推
2026-03-22 20:08:22
Blum
RT @Blum_OG: STOP LOSING ON HYPE. START TRACKING CASHFLOW iGaming is where money flows 24/7, with volumes in hundreds of billions unlike…
中文: RT @Blum_OG:停止在 HYPE 上丢失。开始跟踪现金流 iGaming 是资金全天候流动的地方，交易量达数千亿美元 不像......
#34 🔗 原推
2026-03-21 22:49:21
Blum
STOP LOSING ON HYPE. START TRACKING CASHFLOW iGaming is where money flows 24/7, with volumes in hundreds of billions unlike hype-driven memes, GambleFi got actual revenue from house edge but let's be real: not every such token deserves your money _____________ 5 core… https://twitter.com/Blum_OG/status/2035430751420961156/photo/1
中文: 停止在水合物上失去。开始跟踪现金流 iGaming 是资金全天候流动的地方，交易量达数千亿美元 与炒作驱动的表情包不同，GambleFi 实际收入来自家庭边缘 但让我们真实点:并非每个这样的代币都值得你花钱 __________________________________________________________________________ 5个核心......
media 0 共 1 项
#33 🔗 原推
2026-03-21 18:58:02
Blum
RT @alphabatcher: 🚨 Gemini API update that actually matters for builders: If you wanted AI to search the web and work with your system, yo…
中文: RT @alphabatcher:🚨 Gemini API 更新，对构建者而言实际上很重要: 如果你想让人工智能搜索网页并使用你的系统，那么......
#32 🔗 原推
2026-03-21 11:45:02
Blum
first time handing full control to "tech set to replace us" what's the output product: https://twitter.com/Blum_OG/status/2035094727234920893/photo/1
中文: 首次将完全控制权交给“技术将取代我们” 输出产品是什么:
media 0 共 1 项
#31 🔗 原推
2026-03-20 20:42:47
Blum
RT @alphabatcher: How to build own AI UGC Influencer (fully generated avatar + ways to monetization) ​ I've been studying this space for mo…
中文: RT @alphabatcher:如何打造自己的AI UGC网红(生成完全化身+变现方式) ​ 我一直在为这个空间学习......
#30 🔗 原推
2026-03-20 18:08:54
Blum
RT @Blum_OG: Good Stake -> Productive Agent many believe that an AI agent is just a well-written prompt beyond that, it is very important…
中文: RT @Blum_OG: 好赌注 -> 高效剂 许多人认为人工智能代理只是一个写得很好的提示 除此之外，它非常重要......
#29 🔗 原推
2026-03-20 12:55:23
Blum
Good Stake -> Productive Agent many believe that an AI agent is just a well-written prompt beyond that, it is very important to select the proper agent pieces: > LLM > Tools > Memory > Triggers > Feedback loop not a single point - the agent is just an empty talker 1. LLM:… https://twitter.com/Blum_OG/status/2034731285407842374/photo/1
中文: 良好的 -> 高效剂 许多人认为人工智能代理只是一个写得很好的提示 除此之外，选择合适的代理部件非常重要: > 法学硕士 工具 > 记忆 加注;触发器 >反馈循环 不是一个点——代理人只是个空话 1。传销:...
media 0 共 1 项
#28 🔗 原推
2026-03-19 20:38:36
Blum
RT @alphabatcher: 5 ways you can use AI to solve key business needs (automation guide) ​ I talked to dozens of small business owners over t…
中文: RT @alphabatcher:利用人工智能解决关键业务需求的5种方法(自动化指南) ​ 我和几十个小企业主聊过......
#27 🔗 原推
2026-03-19 18:03:43
Blum
RT @Blum_OG: WHY REVENUE > HYPE? market is oversaturated with tokens lacking full semantic backing what are they based on? nobody knows,…
中文: RT @Blum_OG:为什么收入 > 是吗? 市场因代币缺乏完全的语义支持而过度饱和 它们基于什么?没有人知道......
#26 🔗 原推
2026-03-19 01:20:30
Blum
WHY REVENUE > HYPE? market is oversaturated with tokens lacking full semantic backing what are they based on? nobody knows, everyone just believes and yet there are a couple of worthy representatives namely revenue-driven tokens - isn't this stability in modern market? so,… https://twitter.com/Blum_OG/status/2034343260223840261/video/1
中文: 为什么是收入和收入? 市场因代币缺乏完全的语义支持而过度饱和 它们基于什么?没人知道，每个人都相信 然而，还有几位值得尊敬的代表 即以收入为导向的代币——现代市场难道不是这种稳定性吗? 所以，......
media 0 共 2 项
🎬
视频
#25 🔗 原推
2026-03-18 18:56:44
Blum
RT @Blum_OG: NEXT MAIN NARRATIVE IS STARTING TO TAKE SHAPE historically, long bear markets precede new narratives and we are now extremel…
中文: RT @Blum_OG:下一个主要叙事开始出现 历史上，长期熊市先于新的叙事 而我们现在正处于极端......
#24 🔗 原推
2026-03-16 17:36:03
Blum
NEXT MAIN NARRATIVE IS STARTING TO TAKE SHAPE historically, long bear markets precede new narratives and we are now extremely close to a turning point a fairly long bear market + an empty market are signals of this so, what narrative awaits us? let's look at what never lies… https://twitter.com/Blum_OG/status/2033523654739571181/video/1
中文: 下一个主要叙事开始变得有底 历史上，长期熊市先于新的叙事 而我们现在正接近一个转折点 一个相当漫长的熊市+空置市场就是这种信号 那么，我们等待着什么样的叙事? 让我们来看看什么从不说谎......
media 0 共 2 项
🎬
视频
#23 🔗 原推
2026-03-16 12:39:54
Blum
RT @alphabatcher: i used to pay $500 for a single promo video now i do the whole thing myself in an afternoon. script, voiceover, edit, ca…
中文: RT @alphabatcher:我过去只为一个促销视频支付500美元 现在我自己在一个下午做整件事。剧本、配音、剪辑，或......
#22 🔗 原推
2026-03-15 16:41:43
Blum
RT @alphabatcher: Creating a landing page in 5min with Claude (Prompt) Copy this prompt into Claude, fill in your info, and hit send👇🧵 htt…
中文: RT @alphabatcher:与克劳德(Prompt)在5分钟内创建登陆页面 将此提示复制到克劳德，填写您的信息，然后点击发送👇🧵 ..
#21 🔗 原推
2026-03-15 15:10:28
Blum
RT @Blum_OG: 3 common mistakes in AI agents + solutions brush them off first -> they'll brush you off so, fix upfront: 1. infinite tool…
中文: RT @Blum_OG:人工智能代理中的3个常见错误 + 解决方案 先 -> 把它们刷掉 请提前修好: 1. 无限工具......
#20 🔗 原推
2026-03-15 11:31:54
Blum
3 common mistakes in AI agents + solutions brush them off first -> they'll brush you off so, fix upfront: 1. infinite tool call - resources drain even without your knowledge agent calls tool, it fail, restart repeatedly until collapsing (Ralph Loop) > fix: set limits +… https://twitter.com/Blum_OG/status/2032900479559283166/photo/1
中文: 人工智能代理中的3个常见错误 + 解决方案 先 -> 把它们刷掉 请提前修好: 1. 无限工具调用——即使没有你的知识，资源也会流失 代理调用工具，它会失败，会反复重新启动，直到崩溃(Ralph Loop) 修复:设定限制 +...
media 0 共 1 项
#19 🔗 原推
2026-03-14 19:23:38
Blum
RT @Blum_OG: how old‑school coders feel now: - no cursor - no claude code - no chatgpt - no copilot just sitting, typing code manually an…
中文: RT @Blum_OG:老派程序员现在的感受: - 无光标 - 无引字代码 - 无聊天 - 无副驾驶 只需坐着，手动输入代码......
#18 🔗 原推
2026-03-13 12:11:54
Blum
RT @alphabatcher: BEST local LLMs to run in 2026: ​ High-performance (24+ GB VRAM, preferably with multiple GPUs) ​ • Kimi K2 - 1T params,…
中文: RT @alphabatcher:2026年最佳本地LLM: ​ 高性能(24 GB VRAM，最好配备多个 GPU) ​ • 基米K2 - 1T 参数，...
#17 🔗 原推
2026-03-13 12:10:20
Blum
how old‑school coders feel now: - no cursor - no claude code - no chatgpt - no copilot just sitting, typing code manually and reading docs https://twitter.com/Blum_OG/status/2032233006363259140/video/1
中文: 老派程序员现在的感受: - 无光标 - 无标语代码 - 无聊天 - 无副驾驶 只需坐着，手动输入代码并阅读文档
media 0 共 2 项
🎬
视频
#16 🔗 原推
2026-03-12 23:11:20
Blum
RT @alphabatcher: Skills needed to get hired as a $10k/mo AI engineer: ​ AI-specific skills: ​ > Prompt engineering > Multi-agent systems >…
中文: RT @alphabatcher:获得雇佣成为 1万美元/月人工智能工程师所需的技能: ​ 人工智能专用技能: ​ 快速工程 多代理系统 ......
#15 🔗 原推
2026-03-12 08:50:29
Blum
RT @alphabatcher: Claude vs. Claude Cowork vs. Claude Code ​ Anthropic built three different Claude tools on the same AI ​ If you're using…
中文: RT @alphabatcher:克劳德 对阵克劳德·科姆德 对比克劳德密码 ​ Antropic 在同一 AI 上构建了三种不同的 Claude 工具 ​ 如果你正在使用......
#14 🔗 原推
2026-03-11 20:27:37
Blum
RT @Blum_OG: how to сreate OpenClaw skill in 3 min skills are a key part of a quality AI agent u can't do without them and custom skills…
中文: RT @Blum_OG:如何在3分钟内完成OpenClaw技能 技能是优质人工智能代理的关键部分 没有他们和自定义技能，你就无法做到......
#13 🔗 原推
2026-03-11 07:06:14
Blum
how to сreate OpenClaw skill in 3 min skills are a key part of a quality AI agent u can't do without them and custom skills are often essential safety first: always test locally before production use 1. creating a directory example (entered in terminal): | mkdir -p… https://twitter.com/Blum_OG/status/2031479745263841650/photo/1
中文: 如何在3分钟内培养OpenClaw技能 技能是优质人工智能代理的关键部分 没有他们，你就无法做到，而定制技能往往至关重要 安全第一:在使用前始终进行本地测试 1. 创建一个目录 示例(进入终端): | mkdir -p...
media 0 共 1 项
#12 🔗 原推
2026-03-10 21:18:08
Blum
RT @alphabatcher: Startup Founders Bundle (Free/paid): ​ 5 free options (best ROI for a startup): ​ Supabase - backend, database, and auth…
中文: RT @alphabatcher:初创企业创始人套装(免费/付费): ​ 5个免费选项(初创企业最佳投资回报率): ​ 后端图——后端、数据库和 auth...
#11 🔗 原推
2026-03-10 21:09:27
Blum
RT @Blum_OG: how I feel after using "please" in a dialogue with LLM (last session I called it DeepSeek-tier shit) https://twitter.com/Blum_OG/status/2031100673417424995/video/1
中文: RT @Blum_OG:在与LLM的对话中使用“请”后我的感受 (上一次会议我称之为深度寻找级)
media 0 共 2 项
🎬
视频
#10 🔗 原推
2026-03-10 00:41:20
Blum
how I feel after using "please" in a dialogue with LLM (last session I called it DeepSeek-tier shit) https://twitter.com/Blum_OG/status/2031100673417424995/video/1
中文: 在与LLM的对话中使用“请”后，我的感受如何 (上一次会议我称之为深度寻找级)
media 0 共 2 项
🎬
视频
#9 🔗 原推
2026-03-09 20:11:51
Blum
RT @alphabatcher: 90% of vibe-coded apps have at least one critical vulnerability ​ 15 security rules every vibe coder needs before shippin…
中文: RT @alphabatcher:90%的 vibe 编码应用程序至少存在一个关键漏洞 ​ 每个氛围编码器在发货前都需要的15条安全规则......
#8 🔗 原推
2026-03-09 17:59:32
Blum
RT @Blum_OG: pov: finally shelled out $2000+ for your mac mini m4 pro but heard u still need to install and setup some "openclaw" thing ht…
中文: RT @Blum_OG: pov:终于为你的 mac mini m4 专业公司支付了 2000 美元 但听说你仍然需要安装和设置一些“openclaw”的东西......
#7 🔗 原推
2026-03-08 14:50:57
Blum
pov: finally shelled out $2000+ for your mac mini m4 pro but heard u still need to install and setup some "openclaw" thing https://twitter.com/Blum_OG/status/2030398506847044080/video/1
中文: pov:终于为你的Mac mini m4 pro支付了2000多美元 但听说你仍然需要安装和设置一些“openclaw”的东西
media 0 共 2 项
🎬
视频
#6 🔗 原推
2026-03-07 21:41:41
Blum
RT @Blum_OG: POV: u quit your job to chase airdrops… now u work longer hours, and during payday it just says: “Not eligible.” https://t.co…
中文: RT @Blum_OG: POV:你辞职是为了追逐空投...... 现在你工作时间更长，发薪日期间只说:“不符合条件。”
#5 🔗 原推
2026-03-07 17:49:08
Blum
POV: u quit your job to chase airdrops… now u work longer hours, and during payday it just says: “Not eligible.” https://twitter.com/Blum_OG/status/2030016869244227694/video/1
中文: POV:你辞职是为了追逐空投...... 现在你工作时间更长，发薪日期间只说:“不符合条件。”
media 0 共 2 项
🎬
视频
#4 🔗 原推
2026-03-06 20:25:12
Blum
RT @Blum_OG: Error-Free Update OpenClaw agent develops quickly and gets updates frequently so, to unlock new features + get fixes for old…
中文: RT @Blum_OG:免费更新 OpenClaw 代理快速开发，并经常获得更新 因此，要解锁新功能，并为旧程序获取修复程序......
#3 🔗 原推
2026-03-06 08:57:08
Blum
Error-Free Update OpenClaw agent develops quickly and gets updates frequently so, to unlock new features + get fixes for old ones we shouldn't miss these updates first thing to do: > know how openclaw was installed before - preferred update method is re-run the website… https://twitter.com/Blum_OG/status/2029610825674346758/photo/1
中文: 无错误更新 OpenClaw 代理快速开发，并经常获得更新 因此，要解锁新功能，并获取旧功能的修复方案 我们不应错过这些更新 首先要做的事情: 了解 openclaw 之前是如何安装的 - 首选的更新方法是重新运行网站......
media 0 共 1 项
#2 🔗 原推
2026-03-05 17:31:43
Blum
OpenClaw is more than just trading bots it's a user-friendly tool for wide usage many look at it through a narrow lens but actually @openclaw provides huge opportunities a few cool ones imo: > automated web monitoring 24/7 > routine management - checking mail/notifications… https://twitter.com/Blum_OG/status/2028928006912393362/photo/1
中文: OpenClaw 不仅仅是交易机器人 是一款便于用户使用的工具 许多人通过狭小的视角来看待它 但实际上，@openclaw 提供了巨大的机遇 几个很酷的: 全天候自动网络监控 日常管理 - 查询邮件/通知......
media 0 共 1 项
#1 🔗 原推
2026-03-03 20:18:27