EY-X

in on tech product & strategy

[傅盛 @FuSheng_0306]

经过几天深度使用，我的工作生活已经全面 muse 化了🤣下面是 10 件它帮我干的事：👇

1.网上购物。香港这边的购物网站有多难用谁用谁知道。现在家具、日用品、食品全是它在 HKTVmall、京东上比价、装购物车，我看一眼点付款。去超市买了啥，拍张小票给它，它就知道哪些不用在网上重复买了。

2.盯股票账户。几个账户的持仓它实时同步，异动、浮亏、快到期的期权，半小时内推我。忙起来顾不上看盘，它替我盯着。

3.管邮箱。三个邮箱全归它看，银行每笔花销、结单它都盯着提醒，重要邮件分级，要我拍板的才推给我。

4.工作助理，接管飞书。审批、邮件、群消息它先筛一遍，早晚各一份简报，逐条给我建议，我只负责逐个拍板。以前这些能淹掉我大半天…

5.接管社媒。X 和 LinkedIn 的私信评论它按我定的口径全回，每个寻求合作的消息先通读分析作人员背景调查，再报我谁靠谱。香港 AI 项目的人就是这么攒起来的。

6.接管 mac 上的 Codex 和 Claude Code。我出想法，它整理成任务书，指挥那俩写系统，还负责验收盯进度。手机上就能调度，人在哪都行，香港还不用翻墙。

7.每天早上给我做一期 AI 早报播客。双人对话，5 分钟，过去 24 小时的行业动态，雷打不动 8 点发我。

8.教我粤语和英语。粤语整句翻+粤拼+录音；英语单词每天推，按记忆曲线考我，答错的反复追杀…

9.帮我写稿。诺贝尔奖科普口播、演讲稿都是它出稿我改。规矩一条：稿子里只能留能念出口的话。

10.管杂事。越野赛报名、租房找房源、水电自动转账、iphone duo 抢购，每天还去查门店有没有现货。

以上内容完全真实，因为这内容也是 muse 三万根据这几天的工作总结写的。能真干活的 agent 和聊天机器人，完全是两回事。用了几天，睡眠都颠倒了🤣

[Anton Ivanov @9Rasp]

@MaxForAI /goal hey deepseek. Make mod for infinite money for my bank account. Make no mistakes.

[BlackRay @nowayoutonline]

@Vincent_AINotes 幸灾乐祸的人忘了数学家是人类里最聪明的那一撮。 把数学这个能拴住他们的玩具抢了。先不考虑他们黑化，就是转身投入其他领域，也得搞一波摩擦性失业。

[数字生命卡兹克 @Khazix0918]

https://t.co/Cchay8aH3T

[LinearUncle @LinearUncle]

这周帮其他团队电话面试了几个应聘垂直业务 agent 研发的候选人，基本上我开头一个问题就能初筛掉 90% 的面试者：

/goal 命令是如何实现的？

真正做过 harness 的人才知道 goal 有多难搞：完成判断、验收标准、状态持久化、异常兜底、预算控制、上下文压缩不丢任务、缓存命中不被打乱。。。

[Pat Grady @gradypb]

The @BostonCollege Investment Committee (an LP and my beloved alma mater) asked for a few thoughts on what's happening in AI. I recorded a test run yesterday morning and then shared it with my partners, who encouraged me to share it more broadly... so here you go!

This is not a sales pitch, it's just a reflection on what we're seeing. And it wasn't intended to be shared, so please pardon the rough edges.

https://t.co/AuZqHMSSxD

[周先知 @xianzhizhou]

Anthropic CEO 达里奥没白去百度。

这套操作，百度十年前就玩过了。达里奥进化了。

但百度当年养莆田系做竞价排名，好歹明码标价。达里奥连价格都不让你看见。

美国调查博主Kevin Bass深扒了AI公司Anthropic的资金链。扒出一条完美闭环。

Facebook联合创始人Moskovitz早期投资Anthropic。Claude全球爆火，手里的股票从几亿涨到几十亿美元。然后宣布全部捐给自己的基金会。

自己的基金会。

基金会拿这笔钱干了两件事。

第一，资助一批号称独立的第三方评估机构，比如METR，专门给Claude最新模型做安全评测，出风险报告。

第二，养一批写AI末日论的内容团队，给主流媒体供稿，制造恐慌，呼吁监管。

钱是Anthropic股票变现的。评估对象是Anthropic的模型。结论永远是：需要更多评估。

Kevin Bass翻了National Philanthropic Trust的990税表。2023到2025财年，22258笔拨款流向METR关联网络。Founders Pledge拿了7880万美元。RAND拿了6160万。FAR AI拿了1440万。

这些机构的核心产出：告诉全世界AI很危险。

链条串起来。

恐慌越大，监管越严。监管越严，门槛越高。门槛越高，开源和小公司越难活。竞争对手越少，Anthropic护城河越深。估值越高，基金会越有钱。基金会越有钱，恐慌推文越多。

完美闭环。

此前有人扒过，Anthropic团队自导自演的AI恐慌推文拿了1.6亿阅读量。写文章的人，就是Anthropic自己出资培养的员工。目的只有一个：推动AI立法，让自己参与规则制定，把护城河写进法律。

写恐慌的人拿Anthropic的工资。做评估的人拿Anthropic股东的钱。呼吁监管的人拿Anthropic基金会的拨款。

三条线，一个钱包。

[EY-X @eyxuinc]

@skywt2003 可以记住该地点的名称或其他ID，到海外检测到IP change以后，再后台/弹出提示刷新一次坐标

[Rainier @mtrainier2020]

这能比吗？
Anthoropic，现在一个月多少收入，一堆人不顾IP限制国家限制，哭着喊着给它塞钱买它的服务。
客户是非常明确的，
营收增长是非常迅速的，
解决的问题是实实在在的。

而 人形机器人呢？
能干嘛？
能不能脱离遥控器，干点事情吗？
在工业领域里跟工业机器人比有任何优势吗？
有多少真实的客户？
有多少真实的杀手级应用场景？

即便是马斯克的transformer，装了逼去分拣包裹。
但是现在分拣包裹专业的分拣设备，效率比这种机器人高多了。 纯粹装逼。

家用这块，中国已经严禁装逼。
情趣机器人这条路，卒。

无论是资本市场还是政府投资基金，整个大的pool的容量是有限的。过量的资金进入这种注定会爆掉的项目，必然会挤占掉大量其他扎实有前景的项目。

苏联为啥失败？
苏联在特定领域中一直是力压美国的。
硬科技技术方面，美国很长一段时间内，在一些领域里，在苏联后面吃灰。

但是苏联硬科技的一大问题就是没有效率。
aka 资源配置效率极低。
存在大量的错配。

[EY-X @eyxuinc]

@zachlloydtweets I agree, just that they are not mutaually exclusive, automation and interaction

[EY-X @eyxuinc]

@yg100086 @BTCdayu 小米，破模型？OpenRouter是没有小米生态导流的吧

[Marcos Hernanz @MarcosHernanz]

After doing ~60B tokens, this is my full AGENTS.md

[EY-X @eyxuinc]

This is an extreme example of political retaliation and academic falsification, a waste of taxpayers' money in millions, and a material threat to all who are in a weaker position that should be condemned and actioned upon immediately.

[Nova @PSa_Cheng]

Alleged research misconduct by Sabine Kastner, and silencing whistleblowers by the psychology department at Princeton Univeristy.

Multiple former members of the Kastner Lab at Princeton University, including former postdocs and graduate students, have formally reported research misconduct by Professor Sabine Kastner, including alleged data falsification spanning over 20 years.

They describe a culture where people feared speaking up. Some remained anonymous because they feared retaliation.

My own experience: after I raised concerns, I was removed from Kastner lab. I was explicitly told by the psychology department leadership that this happened because I raised concerns anonymously: they told me “you questioned her integrity. PNI is a community. You attack one of us, you attack all of us.”
Afterward, I could not secure a new adviser and my PhD enrollment was terminated through the “no active adviser” rule.

Many affected trainees were international students. Losing an adviser means losing visa status, and international relocation, creating enormous pressure not to report concerns. This is exactly why Kastner lab recruits a disproportionately high number of international trainees. And why the lab website has most former members erased.

The research involved is supported by active NIH grants of over $20m. Involved publications include papers in Science, Nature Neuroscience, and Neuron.

The departmental leadership has been complicit, by helping remove students who raised concerns instead of protecting them.

Over the coming weeks, I will post documents, timelines, and evidence so people can examine the facts themselves.

#ResearchIntegrity #Whistleblower #Science #NIH #HigherEducation #GraduateSchool #PhD #Retaliation #OpenScience #ResearchEthics #AcademicFreedom @ivanoransky @NIH @HHSGov

[ali @waterloo_intern]

https://t.co/aivc7B4A7Z

[justin curl @curl_justin]

The legal profession is in a race between traditional firms adopting AI and AI-native firms scaling trust.

I can’t tell if a senior superstar attorney starting an AI-native firm is the best of both worlds or the worst.

Best case:

David starting an AI-native firm reduces the trust deficit because he has a book of business, and it’s easier to convince clients to follow the same lawyer to a new firm than to sign up with someone completely new.

And because he’s starting fresh, he doesn’t have the bureaucracy, legacy IT systems, and established ways of working that might inhibit his experimentation with / adoption of AI.

He would have trust from day one and could find someone good on AI to allow the firm to deliver quality legal services at unrivaled speed or cost.

Worst case:

David’s reputation doesn’t meaningfully close the trust gap. Clients might decide following their superstar lawyer from Wachtell to Kirkland & Ellis is perfectly fine, but following them from a massive firm to a new shop that’s experimenting with something new isn’t worth the risk.

And on tech adoption, becoming “AI native” is easier said than done. Although he’s starting fresh, he’s built a very successful career on a particular way of working. Breaking out of that will be tricky. A lot will depend on who he partners with, but his firm just might not be as good on AI as competitors like @crosbylegal.

He also won’t have as much data to draw on. I personally think analytics on top of legal services is the most valuable differentiator for an AI-native legal function: “This is the industry standard based on the past 23 deals we’ve done” is a much more compelling pitch in a billion-dollar deal than “I saved you $50K on associates time by using AI to draft and review clauses.”

At that point, you’re getting neither the experience / benefits of the best firms without the tech talent of the buzzy AI-native firms drawing from the SF startup scene.

[Austin @austinit]

想买一个macOS录屏软件，就是那种可以自动丝滑放大缩小聚焦的，大家有没有推荐的。

目前看 ScreenSage不错，但是不支持按周期订阅，长期划算，短期用不算便宜。

[Xudong Han @Xudong07452910]

我们之前设想的「强模型规划、弱模型执行」，可能本身就有一些问题。

现在常见的做法，是让前沿模型先读代码、写一份 plan.md，再交给便宜模型执行。听起来很合理，实际却可能更贵，因为执行模型拿到的只是一份几千 Token 的计划，后续要重新读取几十万 Token 的代码，前面的理解过程几乎又要在做一遍。

作者提出的 /prewalk 换了一种交接方式：

先让强模型探索项目、建立 TODO，并完成第一次有效修改。等方案真正接触过代码后，保留整个上下文，直接切换到便宜模型继续完成。

在 SWE-bench Pro 上，这种方式保留了 GPT-5.6 Sol 约 97% 的通过率，同时把成本降低了约 39%。

文章里有句话我很喜欢：交接时应该传递一段走过的轨迹，而不是一份描述旅程的明信片。

这对多 Agent 协作也很有启发。未来模型之间有价值的交接物，可能不只是计划文档，还包括它读过什么、排除了什么、当前做到哪里，以及已经验证过的第一步。

文章链接：https://t.co/YuTbScu0Me

[Tenobrus @tenobrus]

bro is getting +1 reward signal every time it retrieves a web result, frontier models gooning to diverse datasets of dictionary definitions

[EY-X @eyxuinc]

Chat compaction is a premature version of chat memory.

https://t.co/P0S9ZsMY2H

[Kangwook Lee @Kangwook_Lee]

https://t.co/6beJPutjiG

[tangjinzhou @tangjinzhou]

记录将 https://t.co/OW2sgE6Om3 做到百万用户得过程

我会将一些我的关键运营动作记录在这里

我不确定是否能成功，但我想记录他的成长过程

成王败寇

我不想成功后给大家分享过滤后的鸡汤，虽然美味，但无营养

更不想失败后没有留下任何痕迹，至少证明我努力过