Keep on to blur preview images; turn off to show them clearly
RT @silsrc: 最近经常被请教职业规划和学习方法… 我草… 我是既没有职业规划，也没有学习方法，东一榔头西一棒槌，得着啥写啥… 上次跳槽的简历还是猎头帮我写的🤦，找工作是完全不会找的，只会默默等别人主动联系… emo了就自闭写代码，既不拉帮结派也不主动交流，一直在走弯路…
喜欢王小波，大概我们能成为朋友。 我的 2025 https://t.co/pAkSJnpKXA 我的 2024 https://t.co/HfDF6oduB7 我的 2023 https://t.co/QyV8PiZmOY ..............
We’re looking to hire a Software Engineer (Cloud Backend Engineer) to help work on Sakana AI’s advanced AI-driven discovery platform.
More information: https://t.co/P8ctt5KmaG
招聘之道：布莱恩·切斯基教你如何找到合适的人才 Airbnb 联合创始人布莱恩·切斯基曾说过：“招聘不是销售渠道，而是人才网络。” 十多年来，他不断完善自己的招聘理念，首先要了解成功是什么样子，然后再找到能够实现成功的人。 以下是他的观点，为了清晰起见，已进行简化。 从结果入手，而非简历 布莱恩借鉴了史蒂夫·乔布斯的智慧：“大多数人从简历入手——比如谷歌或亚马逊这样的公司。但你应该问自己：我欣赏哪些产品？然后是谁打造了这些产品？” 关键在于？关注影响力，而非资历。简历告诉你某人在哪里工作过，但只有结果才能揭示他们取得了*什么成就。例如，与其问“你是否在X公司工作过？”，不如问“你在那里解决了什么问题？你是如何知道自己成功了？” 面试：深入挖掘 布莱恩建议不要轻易相信候选人的第一个回答。“让他们解释一下他们是如何完成某件事的，然后追问两个问题——你需要的是第三个答案。” 为什么呢？人们通常可以复述（转述）一个故事，但只有真正理解的人才能回答更深层次的问题，例如：“最大的挑战是什么？”或者“下次你会怎么做？”如果有人在第二个追问后仍然支支吾吾，你就知道他们的经验可能不如他们声称的那么丰富。 推荐信：招聘中最关键的环节 布莱恩认为推荐信比面试更重要，尤其对于经验丰富的候选人而言：“高管们比你更擅长‘吹牛’。”他甚至建议花在核实推荐信上的时间与面试一样多。 以下是正确做法： 非正式沟通：首先说：“我需要一些真诚的反馈——以下内容与您无关。”人们在不担心遭到报复的情况下会更加坦诚。 询问“最佳人选”：“您合作过的最优秀的人是谁？”如果他们没有提到您正在招聘的候选人，那么他们的推荐就不够真实。 要求具体细节：如果推荐人说“他们太棒了！”，那就追问：“是什么让他们如此出色？是某个具体的项目、技能还是某个重要时刻？”如果没有具体细节，那么赞美就显得空洞无物。 询问潜在问题：“您建议我注意哪些方面？”或者“他们需要改进的方面是什么？”这会促使他们进行诚实的反思。 建立人脉网络：最后，问：“您能再提供两个名字让我核实吗？”这会将推荐人转化为不断增长的人才库。 CEO的角色：积极参与 布莱恩亲自面试了Airbnb的前400名员工，他坚信高层领导应该主导招聘。“你在面试和背景调查过程中投入的时间越多，招聘就越成功，”他说，“这项工作无法外包。” 简而言之，优秀的招聘意味着首先要对成功有清晰的愿景，深入了解候选人的工作细节，并相信推荐人能够揭示真相。正如布莱恩所知，合适的人才不仅仅是填补职位空缺——他们能够打造公司。
从投资领域转到创业：找工作、找面试题、改简历、模拟面试. 创业（冷启动）｜AI , AIGC ｜ 安全技术｜RAG | 时空智能 | 认知心理学｜智能体 ｜ 生命科学 ｜ 强化学习 I built open source software at https://t.co/b69DXZhcyR
Excalibur from Temu maximum range is about 50 km iirc in theory they could fire this from ships… They only have 130 mm H/PJ-45 gun on type 055 though
We're in a race. It's not USA vs China but humans and AGIs vs ape power centralization. @deepseek_ai stan #1, 2023–Deep Time «C’est la guerre.» ®1
Tech Twitter has been an entirely retarded mob on the Mamdani phenomena. Peter is unsurprisingly breaking from the herd with some serious analysis.
ceo @replit. civilizationist
I think a lot of people who might want to develop a new software project often get tripped up at the beginning in the process of starting it, since it seems so daunting to start out with a blank repo. So I thought I would quickly run through my latest workflow, which dramatically lowers the bar in terms of the effort and time needed to get started. The most important part by far is to have a good idea or what to make that would actually be useful to a lot of people if it really worked properly and did what it tries to do well. I can’t really help you with this part, but the common advice of scratching your own itch and solving your own (non-niche) pain points is a very good way to get started. I find myself thinking of project ideas constantly. Anyway, the next step is to write out the idea informally, as you might in a quick email to a close friend. You don’t try to make this a formal plan, just the quickest way to convey the basic idea and what it does and specify any parts of the tech stack or libraries that you know you want to use. The attached screenshots show an example of this for an idea I randomly had a couple days ago. It took me maybe 10 or 15 minutes to write up. It doesn’t need to be long, and it can reference other sources to keep it concise. This initial description then becomes the prompt for GPT-5 Pro. This usually takes at least 15 or 20 minutes to run (amusingly longer than it takes to write the prompt). You could try other models, but they will be a lot worse. I then often take the same prompt and give it to Grok4 Heavy or Opus4.1 and feed those ideas back into GPT-5 Pro and encourage Pro to take any good ideas it sees in the other proposals. If there’s actually something smart in those plans, GPT-5 Pro will recognize it and incorporate it. Then I’ll ask Pro to create a detailed, granular markdown plan document based on its first response, and I’ll save that as a file in a newly created project folder. Then I’ll often iterate on this a few times, starting a new Pro conversation in the web app and giving the entire markdown plan file and telling it to enhance the plan in various ways to make it more reliable, robust, performant, intuitive, user-friendly, and other good adjectives. And I’ll encourage Pro to do exhaustive web research on the latest documentation, blogs, tutorials, etc to find better libraries or ways to do things. I’ll then take its proposed revisions and paste that into codex and ask codex to integrate the revisions into the existing markdown plan document. After 2 or 3 rounds of this, things stabilize and you get a really good, fleshed out plan. This is the key to everything, because when things are still in the plan phase, it’s much easier to tweak and improve them because you don’t have any code yet. Measure twice, cut once, etc. Here’s a link to the resulting plan document that came from the initial prompt for this idea: https://t.co/mXHOZH9b2p At this point, I start adding an AGENTS dot md file; I start with an existing one I have and ask Pro (in the same session as the latest plan document was written) to customize it for this new project and tech stack while preserving anything generic. If there are some critically important libraries, I will also sometimes create a specialized best practice guide (say, if you’re making an MCP server, I’ll generate a best practices guide specialized for the fastmcp library, but where I also spell out how to structure the project, etc.) At this point, I then ask codex in a single session to start building out the project structure, creating folders and blank placeholder files, making the .gitignore file, etc. Here’s where my process diverges dramatically from the typical approach. I first use Steve Yegge’s beads project and tell codex to turn the plan document into a bunch of tasks and sub-tasks using beads. Then I use tmux to create a bunch of codex sessions, as many as 8 at once (I think more than that would also work well)…
Then I start up my mcp agent mail project and tell them to register with mail and introduce themselves to their fellow agents, to read the AGENTS dot md and the plan document in their entirety (referencing this file by name), and to review the beads tasks in an initial message. After that, it’s just a process of cycling through the same set of canned messages where I tell them to pick a bead task and start working on it, encourage them to continue on that a few times, then to review their work, then to check their agent mail and respond to any messages they’ve received that need a reply, and to communicate what they’re doing with their fellow agents. I recorded a 23 minute video yesterday showing how this part works for the sample project shown in this post which makes this all clear: https://t.co/KedUEzSRRG After recording that, I figured out how to automate even this last part so that I can automatically queue up a bunch of messages to all the agents to keep the busy for hours by using tmux to broadcast the commands. I’ll link to that post below. That’s the basic process. If you follow it, you can create shockingly complex and powerful software in a couple days with maybe 1 or 2 hours of actual “human time” required, but many, many agent-hours (like man-hours) of work. And each hour of autonomous agent work with gpt-5-codex when it’s not waiting around for human instructions constantly is probably more like 10 or 20 hours of human time. They think and type really fast!
