도박전사쿠루미짱

nothing and everything

[USAnt 미국개미 @USAnt_IDEA]

빌 애크먼이 워런 버핏과 투자 본능에 대해 날린 일침 요약 ㅋㅋㅋ

버핏 할배가 왜 인간 개미들이랑 다른지 빌 애크먼이 팩폭 박음.
주식으로 진짜 돈 버는 놈들은 타고난 천재가 아니라 본능 거스르는 연습을 미친 듯이 한 놈들이라는 게 핵심임.

• 인간 DNA vs 주식 시장
밀림에서 사자 튀어나오면 다 같이 빤스런 치는 게 생존 본능임.
우리 조상들은 그렇게 도망쳐서 살아남았음.
근데 주식판에선 이 본능대로 움직이면 계좌 바로 갈려나감.
다들 벼랑 끝으로 다이빙할 때,
뒤돌아서 사자 아가리 쪽으로 뛰어가야 돈을 버는 구조라 인간 본성이랑 정반대임.

• 장내 패닉의 실체
애크먼 본인도 25살 땐 이 짓거리 못 했다고 털어놓음.
옛날 객장에서 목소리 제일 큰 놈 셋이 세상 멸망한다고 고함지르며 던질 때, 옆에서 멘토 한 명이 조용히 풀매수 땡기는 걸 봄. 나중에 어떻게 버텼냐고 물어보니까 명언 한 줄 나옴. "패닉은 ㅈㄴ 그럴싸한 변장복을 입고 있는 정보일 뿐이다. 그 옷��터 벗겨보고 판단해라."

• 멘탈은 타고나는 게 아니라 근육임
사람들은 강심장이 눈동자 색깔처럼 타고나는 줄 아는데 개소리임.
남들 다 때려치우고 빤스런 칠 때 버티면서 찢어지는 근육 같은 거임.
책 백날 읽어봐야 소용없고, 인생 최악의 하락장 오후를 엉덩이 딱 붙이고 버틴 다음 '어? 다음 날 아침에도 나 안 뒤졌네?'를 직접 겪어봐야 생김.

공포에 매수해서 떼돈 버는
버크셔 해서웨이 $BRK.B 버핏 같은 양반들도 안 쫄려서 그러는 게 아님.
남들 다 던지고 튈 때 의자에 앉아있는 연습을 남들보다 수천 번 더 해본 짬바의 승리임 ㅋㅋㅋ

Not financial advice

[How To Prompt @HowToPrompt__]

Google just automated the PhD.

They built an AI that reads a problem, forms hypotheses, runs experiments, writes the paper, AND simulates its own peer review. no human touches it.

It’s called “ScientistTwo”

It is a fully autonomous multi-agent framework that executes end-to-end machine learning research.

Without a single human in the loop.

You give it a fundamental challenge. That’s it.

The AI independently navigates the literature. It establishes the baselines. It formulates novel hypotheses.

Then it writes the code. It runs the experiments. It conducts its own ablation studies.

It even argues with a simulated peer-review engine to refine its work before submitting.

The results are staggering.

Researchers tested ScientistTwo against the highest standards of human scientific achievement—papers accepted at top-tier conferences like NeurIPS and ICML.

The AI improved human state-of-the-art results on 86 different tasks.

It generated an average performance gain of 25.2% over the absolute best human models.

It didn't hallucinate a single citation. It wrote fully executable, verified codebases. And it autonomously generated expert-level, publication-ready papers.

We’ve spent the last two years using AI as a highly advanced intern to help us write code and summarize data.

That era is over.

ScientistTwo isn't an assistant. It is an autonomous scientific pioneer.

It doesn't just summarize the frontier of human knowledge. It expands it.

If an AI can autonomously hypothesize, test, and publish breakthroughs in machine learning...

How long until it discovers something we don't even have the math to understand?

[CyberRobo @CyberRobooo]

IRON walks like a human.

XPENG has just held a supplier conference, bringing together major suppliers across actuators, dexterous hands, sensors, AI chips and other key components.

It has completed supplier audits and finalized key component suppliers. Around 85% of its suppliers overlap with XPENG’s automotive supply chain, and production scheduling is now underway.

IRON’s production line started operating on September 8,Guangzhou, with more than 80% of core processes automated.

[Sawyer Merritt @SawyerMerritt]

Figure has introduced their new humanoid robot neural network called Helix 2.5, which the company says is capable of entering homes it has never seen before and immediately performing complex tasks with zero additional training.

Figure says it tested its humanoid across 30 previously unseen homes, where it autonomously:
• Made beds
• Folded towels
• Tidied living rooms
• Navigated unfamiliar layouts
• Manipulated objects it had never encountered
• Self-corrected when it made mistakes

"No data was collected from the homes beforehand, and the robot received no fine-tuning or adaptation for the environments or objects."

[Odyssey @odysseyml]

Today we’re unveiling Odyssey-3, a big step forward for foundation world models.

It can control robots, power humanoids, drive cars (on the roads of India!), train AIs, pilot drones, and even play video games.

We can’t wait to see what intelligent systems it enables.

[Lee1ng @lilaoshizuikeai]

Anthropic报告里这一段真的笑死我了
因为月之暗面把��分Kimi 用户请求直接转给Claude
导致解放军把几百个军工设施的监控视频扔给Kimi抓间谍
然后Kimi直接把这些监控视频发给Claude了

[Techniahqrobot | humanoid robots @techniahqrobot]

This might be one of the clearest snapshots yet of how crowded the humanoid robot race has become.

16 humanoid robots

➝ AGIBOT A2
➝ EngineAI T800
➝ Xiaomi CyberOne
➝ XPENG IRON
➝ Figure 03
➝ Galbot ET1
➝ Unitree humanoids
➝ Boston Dynamics Atlas
➝ Tiangong Omni
➝ HONOR Robotics
➝ Tesla Optimus
➝ Booster T1
➝ Qiyuan Q1
➝ LimX Dynamics Luna

Some are being developed for factories. Others focus on locomotion, manipulation, teleoperation, research or human interaction.

What stands out in 2026 is the range of hardware already being tested outside controlled lab setups.

Different hands, joint layouts, camera systems, body proportions and control stacks are all converging on the same difficult target: useful general-purpose physical intelligence.

Which humanoid would you bet on?

#HumanoidRobots #Robotics #PhysicalAI #EmbodiedAI #AI

[Elon Musk @elonmusk]

@viktaur27 @Teslarati The rate of improvement from original GPT to GPT-3 is impressive. If this rate of improvement continues, GPT-5 or 6 could be indistinguishable from the smartest humans. Just my opinion, not an endorsement. I left OpenAI 2 to 3 years ago. Am a neutral outsider at this point.

[The Humanoid Hub @TheHumanoidHub]

This robot just beat Figure's package sorting speed.

Robotic arms powered by X Square Robot's WALL-B embodied AI sorted 10K parcels in 5 hours 14 minutes. That's 1.88 sec per parcel, ~35% faster than Figure's reported 2.88 seconds per parcel.

Figure's 9-day continuous demo in May this year was sustained by 5 unique Figure 03 robots taking turns.

The same WALL-B model also runs household tasks and fine manipulation across dexterous hands, arms, and mobile manipulators.

The highlight is fascinating to watch. The task is to flip the packages label-side up and slide them down the conveyor. Some unusual packages, like soft toys, get routed to a separate lane.

[CyberRobo @CyberRobooo]

Another major Chinese company had open-sourced a humanoid robot dataset--

EgoLive from JD.

This is a large-scale first-person perspective dataset recording real-world human tasks. It contains 1680 hours of high-fidelity binocular video data at 60fps, involving 65,866 data segments and 346 different tasks. The data comes from real-world business scenarios in retail, logistics, healthcare, and industry...

Earlier, JD established China's first embodied data collection community in Suqian, Jiangsu; simultaneously, their RoboBase project started construction in Guangzhou in the Q2, with plans to establish over 80 robot bases nationwide over the next five years.

Humanoid robot data is truly a new era gold mine.

[Mahax @Mahaximus_]

Andrej Karpathy:

"Prompting is going away.

Delete everything, keep Graph."

In 1 hour he shows how to build Graphs, and why it's the only thing that will be left standing at the end.

The missing piece most people don't get: LLMs, prompts, agents, they're all steps. Graph is where you end up.

Watch it, then read the full guide on Graphs below.

[Serenity @aleabitoreddit]

So there's a IBK Research report on Boston Dynamics value chains from last month.

Just a summary:

IBK maps these companies to Atlas as suppliers:
- Hwashin (010690) / body, arms, legs
- LG Energy (373220) / battery
- Hyundai Autoever (307950) / integration
- Hyundai Mobis (012330) / actuators

As for humanoid volume ramps:

They're modeling for, 11.29K in 2028, 20k 2029, 30k in 2030... 40k in 2031, and 50k in 2032.

Not quite sure why IBK and other institutions are a fan of linearly modeling S-curve volume ramps...

Like adding +10K per year, don't quite think it's volume ramp is going to work like that... if I had to guess it would look more like:

- 15-20k 2028
- 40k-70k 2029
- 90k-140k for 2030

Since Boston Dynamics is projecting 30k capacity by 2028 (I'm sure they'd aim to get more online by 2029-2030), as China collectively is already doing 100k EOY in 2026.

In terms of competitive landscape they name:

- $TSLA, Figure, Apptronik, $CCXI (Agility), as US players. Then Boston Dynamics (Korea parent owned now)
- Unitree, Fourier, AGibot, UBtech, $XPEV as the Chinese leaders.
- Neura, Pal Robotics, Wandercraft, Oversonic, as the EU leaders.

They also did quite a lot of valuation modeling around Hyundai Mobis/Hyundai Autoever/Glovis.

Regarding the BD economic ownership from 27.9% from Hyundai Motor, 11.3% from Mobis, and 11.3% from Glovis. So at least institution are valuing humanoid segments inside companies lot more now.

Then from report assumptions:
- 31 actuators per Atlas
- $1K per actuator in 2028
- $134K Atlas ASP,

Which implies actuator cost of final selling price is roughly 23–28%.

Probably the more interesting statement was IBK stated that actuator capacity is biggest signal for volume ramp. eg. every 310,000 units of actuator capacity supports 10,000 more robots.

So tracking actuator outputs, yields, ASP is a cleaner read on 2028-2030 ramp.

I didn't have much personal takeaways, but hopefully others find it interesting, maybe around Hwashin as a core supplier or around actuator capacity as an indicator.

[Serenity @aleabitoreddit]

Apparently, Korean Beauty was where all the alpha is.

APR (278470.KS) is up 529% over the past 2Y, wildly outperforming both $NVDA and $INTC.

Every female I know talks about their "Medicube" stick, then buying products... where you put a whole salmon's family tree on your face for better skin.

The world's female population creates a bigger structural demand than hyperscalers for memory or lasers? Should have known...

[Beff (e/acc) @beffjezos]

2 million years of techno evolution

[套利豪仔🗽 @pritipatelor]

本周增长最快的 10 个 GitHub 仓库：

1. codegraph（+14.1K stars）
为 Claude Code、Codex、Cursor、OpenCode 和 Hermes Agent 提供预索引的代码知识图谱——更少 token、更少工具调用、100% 本地运行
https://t.co/FtPSe0NDD5

2. openhuman（+17.1K stars）
你的个人 AI 超级智能。私密、简单且极其强大
https://t.co/vlpyrgXaur

3. academic-research-skills（+11.6K stars）
面向 Claude Code 的学术��究技能：研究 → 写作 → 审阅 → 修改 → 定稿
https://t.co/YvmtLeIIvV

4. RuView（+6.8K stars）
π RuView 将普通 WiFi 信号转化为实时空间智能、生命体征监测和存在感知——全程无需任何视频画面
https://t.co/RKUbk9A4N8

5. agentmemory（+6.9K stars）
基于真实世界基准测试的 AI 编程智能体第一持久记忆方案
https://t.co/rVs4nfTNRD

6. supertonic（+3.6K stars）
闪电般快速的端侧多语言文本转语音（TTS）——通过 ONNX 原生运行
https://t.co/BqvxeQALNI

7. CloakBrowser（+7.0K stars）
可通过所有机器人检测测试的隐身 Chromium。可直接替代 Playwright，并带有源码级指纹补丁。30/30 项测试通过
https://t.co/iWavMI5C9O

8. ViMax（+2.7K stars）
ViMax：智能体式视频生成工具，集导演、编剧、制片人和视频生成器于一体
https://t.co/oIVirlsU1F

9. 12-factor-agents（+1.9K stars）
我们可以用哪些原则来构建真正足够优秀、能够交付给生产环境客户使用的 LLM 驱动软件
https://t.co/KZTMpzX5GJ

10. bun（+2.0K stars）
极快的 JavaScript 运行时、打包器、测试运行器和包管理器——一体化工具
https://t.co/H7QRNeWT8y

本周主题：智能体记忆、上下文效率和端侧智能正在让 AI 基础设施成为最火热的构建方向。

收藏这份清单。下周的榜单可能会完全不同。

[Berryxia.AI @berryxia]

Stanford 这堂 2 小时 AI 系统构建课，直接把 Claude 所有教程和 Prompting Thread 秒了！

强烈推荐：
“比你刷过的所有 Claude 教程都实用 10 倍”

里面讲的不是 prompt 技巧，而是 Stanford 真正教工程师如何从零构建可靠 AI 系统的完整方法论。

周末就刷这一个，绝对是你这周最有生产力的事！

我已经将其翻译为中英 文双语视频直接戳这里👇

[Leonard Rodman @RodmanAi]

🚨 PDFs are officially broken.

Someone just dropped a tool that converts PDFs → clean Markdown
at 100 pages/sec 🤯

Tables? Extracted.
Messy layouts? Fixed.
Nested data? Perfect.

No GPU. No cost. No excuses.

It’s called OpenDataLoader.

This kills 90% of manual data work.

repo link: → https://t.co/zHn6GW8H7P

[Clash Report @clashreport]

Professor John Mearsheimer:

Israel may ultimately use nuclear weapons against Iran.

It is now clear that Israel cannot prevent Iran from acquiring nuclear weapons using its conventional forces.

That option does not work.

The only option they have left is nuclear weapons.

The United States will not stop it.

[Niko @guishou_56]

本来以为前端类的 Skill 已经过剩了，出不了什么新鲜的活了

昨天刷到一个叫 Awesome Design 的仓库，将近20K的star

把全球 55 个大厂的设计语言，全塞进了一个 DESIGN.md 里

苹果、Spotify、IBM 这些有极好品位的品牌
常用的配色、字体、组件，一次全有了

用法很简单：
把仓库链接发给 Claude Code，让它自己安装配置
装好之后让它参考这份设计规范去跑你的项目就行

随便跑了几个 case

因为有了这套规范，设计下限直接被拉高了

几乎很难再出 AI 味的前端了

https://t.co/DjmPrfoOwF