CFG_Labs - e/acc

We build the deep thinking community for Web3 and AGI
Contact:

[CFG_Labs - e/acc @0x_atom_crypto]

Fusion nucléaire : Comment piéger un plasma sans le toucher ? https://t.co/0oz9rjl07p 来自 @YouTube

[TechniaHQ | humanoid robots @techniahqrobot]

This is the part most people miss about humanoid robots.

The robot body is only one cost layer.

U.S. teams are spending across:

• Actuators and joint modules
• Dexterous hands with tactile fingertips
• AI software for perception and balance
• Robot learning data from demos and pilots
• Cameras, depth sensors and IMUs
• Onboard compute and control boards
• Battery packs and power electronics
• Manufacturing tools and test rigs
• Safety validation in real environments

That is why humanoid robots are still expensive.

Physical AI is not just code.

It is motors, sensors, hands, batteries, factories and failure testing.

@grok your take about this ??

[Rohan Paul @rohanpaul_ai]

Aravind Srinivas just explained why China’s open-source AI may become more powerful than ever.

And why Anthropic has lobbied very hard for export control.

"the only reason why there is even a 12-month gap between open source and frontier models is export controls.

But there is a chance that, because of that, they now get really good at the physical layer.

One advantage they (China) have is that they can actually build data centers a lot faster. Power is not a problem. Permits are not a problem. People are not a problem. Labor is not a problem. Expertise is not a problem.

And so, by forcing them to go out there and build all this, you are converting them into a far more potent competitor."

---

From "20VC with Harry Stebbings" YouTube channel (@HarryStebbings ), link in comment

[Seb Johnson @SebJohnsonUK]

The UK now has 49 Fintech unicorns that are collectively worth over $250bn! 🇬🇧

The UK has created 6 decacorn Fintechs, including some of the most excting fintechs anywhere in the world:

> Revolut, officially worth $75bn but rumoured to be worth $100bn+

> Wise, one of the only public companies, with a market cap of $10-12bn

> https://t.co/GqyoOlEW3Z, recently valued at $12bn in a share buy back

There's also a BUNCH of amazing sub-10bn companies out there e.g. Monzo, Starling, Zilch...

In fact the UK is now the Fintech capital of the world now that the US is so AI-focused.

It's also great to see so many of these companies mature. Most of these are now profitable while still growing quickly and many are thinking about a potential IPO.

Great data from Olek Skwarczek and the https://t.co/wLnhMRPcfM team

[RamenPanda @IamRamenPanda]

翻译：对于机器人公司，我看好 Unitree 和 Agility Robotics，这两家的人形机器人玩家。 而且我在 Agility Robotics 上有最大的持仓，因为我个人更偏好美国的人形机器人玩家。 对于上游组件的投资，我目前持有： - Harmonic Drive (6324)，因为在谐波减速器等部件上的 BOM 含量很高。 - Vishay Precision，用于传感器，并且是 Tesla Optimus 的潜在候选供应商。 LeaderDrive (688017) 和 Schaeffler 我也看好，但个人没有持有。 许多其他 AI 数据中心玩家，我通过像内存这样的间接方式暴露于机器人领域。 不建议任何人抄袭，只是分享个人仓位/想法。 人形机器人领域很大，正如高盛报告所见，“韩国公司将占据全球人形机器人生产 30% 的直接和间接份额”。 全球有很多玩家。这是一份较旧的报告，但我再次链接它，因为你会看到它们都出现在高盛机构报告中 + 他们覆盖的内容。 Agility Robotics 目前是我个人最喜欢的。

[nico @nicochristie]

Tianhang is finally on twitter

He was formerly a founding team member and head of RL at Qwen, and leads model training here at FRL / shortcut

He will be posting more on twitter now

Probably the most cracked person i've ever met. Excited for you to see his work soon!

[美少女战士_Ayla @Aylababe_]

谷歌前 CEO 强调了一个热知识：
如果你真的想赚钱，必须学习：理解代理、Claude Code、提示、记忆、技能、MCP和例程，然后让你的 AI 来控制你的 AI
大部分人学 AI, 就是在推特刷 AI 帖子，但其实，真正系统性的知识，大模型公司官方出的全面且免费，务必收藏这条帖子：

代理架构 https://t.co/CWHbp94eEG
Code 101 https://t.co/0oqE7KriXa
Code 实战 https://t.co/LFoS19qSOg
提示工程 https://t.co/rvLTIbi4ep
互动提示课程 https://t.co/PkVrmekuJD
Claude.md + 记忆 https://t.co/dnc98oxXhg
技能 https://t.co/r7GmXwg5ga https://t.co/gs3rSVtzSk
例程 https://t.co/DE2yPYvAag
Code 终极指南 https://t.co/E5P06i6gvY
Claude Code https://t.co/GkBByt1nq7
学院 https://t.co/Stjgg0gpqB
官方 Claude Code 文档 https://t.co/ab8BUYsLvZ

现在，复制全文内容，并喂给你的 Agent!

[Kevin S. Xu @kevinsxu]

The open vs closed model debate in China is very much alive...and diverging

Alibaba is in one camp. Other labs, particularly neolabs, are in another. ByteDance has been closed this whole time

The boundary of this debate appears to be 1 trillion parameters

I've shared this observation in both my writings and closed-door briefings since my recent China trip

One thing is for sure: No, Chinese labs are not mandated to open source by the government (a misconception I hear a lot...)

[m_ric @AymericRoucher]

You don't understand the current AI race if you don't think about it in terms of compute - and compute clearly distinguishes 3 tiers of companies.

Arthur Mensch, Mistral's CEO, recently had a hearing at the French Assemblée Nationale. He elegantly framed the AI race as a compute issue, where sovereignty would be ~"the ability to get leverage along the AI value chain" from electrons to tokens.

He also provided numbers (in MW) for Mistral's available compute : I was surprised at how low these numbers were compared to the gargantuan numbers touted by US labs.

So I ran the numbers, based on the recent and excellent @EpochAIResearch study, adding in my (not that reliable) AI-powered estimates of Chinese compute (see assumptions in blog post).

And I found out that there are 3 quite separate tiers.

1. US Champions are really far ahead. Anthropic, OpenAI and Google each command multiple gigawatts (OpenAI ~15 GW once you count the Stargate/Azure/Oracle capacity it rents). Ever wondered why their Claude/GPT /Gemini consistently top benchmarks? Now you know. By the way, tick in Meta and xAI and you'll see them entering tier 1 too with their recent buildouts.

2. Chinese giants scale fast. Alibaba, ByteDance, Tencent, Huawei and the three state telcos are racing from hundreds of MW toward multi-GW, increasingly on domestic Ascend silicon and the national "East Data West Compute" grid. They report "computing power" in EFLOPS rather than MW, so their points here are estimates, could be quite off the mark.

3. The contenders. Europe's Mistral commands ~90 MW today and aims at 1 GW by 2029, an order of magnitude behind the leaders. Interestingly, some of the best Chinese labs (DeepSeek, Moonshot, Zhipu, MiniMax) have no longstanding compute : they are pure-play : they rent or get allocations from government capacity for specific efforts. DeepSeek (~90 MW, the only one of this category that owns its cluster) is the largest.

With all that said, I hope someday someone in Europe wakes up to the absolute necessity of building compute faster than we do today.

If you want to go inspect the graph, I've got the interactive version and full sources in my blog post, link below.

[snowboat @snowboat84]

最近又看了一遍《教父》。每次重新看《教父》，都会有新的感受。这次最大的感受是，第一代移民应该如何做自己的路径积分。老教父Vito Corleone作为来自西西里的第一代移民，他的人脉关系网清一色是意大利人，尤其是西西里人。无论是刚起家时候的合伙人Clemenza和Tessio，还是从西西里走私橄榄油的生意，靠的全是西西里的人脉网。

我们第一代移民别想着一下子打入所谓的美国主流社会，人家和你根本不在一个生活圈里，你硬是要挤进去，既不合拍，也没人真的鸟你。一开始的人脉和网络，必须建立在自己熟悉的环境和熟悉的人群里，这是路径积分的起点，不是终点。

之后老教父做了几件事情，延长了他的路径积分：

- 首先，收养 Tom Hagen，让这个昂撒人（爱尔兰人）最终成为了自己西西里集团的军师。这是打入昂撒集团的第一步。

- 第二步，资助Johnny成为了好莱坞知名的歌星和影星，让自己的西西里的影响力打入主流的媒体市场，从此西西里人也有了在主流印象里的代言人。

- 第三步，老教父开始培养支持自己的参议员。但是这件事并不成功，一直要等到第二代教父Mike的时候，才设下圈套，收买了一个嫖娼的参议员Pat Geary。

- 正是由于第三步的不成功，老教父对自己小儿子Mike的要求，原本是希望他从政，成为议员或者州长，这样西西里集团就才能有真正的政治实力。

这样看来，《教父》这部片子，讲述的其实是第一代移民路径积分的故事。从这点来说，这部黑帮片要比《美国往事》要重要的多。后者讲的是个人命运的虚无，前者讲的是家族战略的兴衰。

可惜的是，二代目教父Mike并没有老教父Vito的战略思维。他懂权术，懂杀伐，但不懂积分的方向，每一步都像是战术上的胜利，却是战略上的失败，最终把整个家族的路径积分搞砸了，最后只能在西西里孤独地死去。老教父Vito把积分从西西里积到了纽约的五大家族之一，而Mike把积分又负向积回了西西里。

其实Mike的西西里妻子Apollonia如果不死，他的事业应该可以获得更大的成功，他的家族也会更加兴旺。结果他硬是找了一个昂撒女人Kay，和他处处不合拍，最后成为了他事业的最大阻力。

[Zhai Xiang @ZhaiXiang5]

In 1972, Nixon did not fly directly to Beijing.
In 1984, neither did Reagan.

Before actually arriving in China, both presidents stopped somewhere across the Pacific to rest, read briefing materials, and adjust to the time difference.

These seemingly minor travel details are precisely among the most fascinating parts of the history of China-US relations.

I once visited the Nixon and Reagan Presidential Libraries, where I reviewed and researched many declassified files related to their China and Japan policies. The archives contain countless moving historical details, which I would be happy to share gradually if people are interested.

These are the schedules I photographed at the time for President Nixon's visit to China in February 1972 and President Reagan’s visit to China in April 1984.

President Nixon first arrived in Honolulu on February 17, 1972. There, at a Marine Corps camp, he spent two days reading briefing materials with his wife Pat. They were extremely thick, almost like dictionaries, covering all kinds of background information and analysis about China. They even included dozens of mock scenarios for questions from reporters, along with A/B/C-style response options for each one.

On February 19, Nixon flew to Guam. After spending one night at a naval residence there, they flew to Shanghai on February 21. Following a brief tea break at the terminal, Nixon took off again for Beijing, arriving shortly after 11 am that day.

It was then that he had the handshake with Chinese leaders that crossed the Pacific and changed history.

Reagan's itinerary was equally fascinating.

On April 18, 1984, he was still busy in DC with meetings, phone calls, and briefings, and even got a haircut at the barber shop. He also called Nancy in LA but was too busy to complete the chat. In his diary that day, he described it as "a busy time tying up loose ends," and ended the entry with: "back to packing."

On April 19, he first flew to Seattle, where he met with several business executives, including people whose companies were doing substantial business with China. Even today, Seattle remains one of the cities in the United States most closely connected to China through trade and commerce.

That evening, Reagan flew to California, where he reunited with Nancy. The two spent two days resting at the ranch, riding horses and cutting wood, getting recharged.

On April 22, they flew to Hawaii and stayed at the Kahala Hilton. Nancy, who wanted to swim in the ocean, borrowed a villa so she could enjoy swimming without being followed by the press.

On April 25, they flew to Guam and stayed at the former residence of Admiral Nimitz. The following afternoon, at 1:35 p.m., Reagan arrived in Beijing and immediately went to the Great Hall of the People to attend the welcoming ceremony hosted by the Chinese president.

That evening, at the Diaoyutai State Guesthouse where Reagan was staying, the Chinese side served a 12-course dinner. Reagan wrote in his diary:
"We heeded Dick Nixon's advice & didn't ask what things were-we just swallowed them...We both did well with our chopsticks."

Nancy told the press that, for this trip to Asia, she and Reagan followed a "feast-and-fast" diet recommended by the White House physician to help adjust to jet lag. For example, on one day she ate only clear soup, fruit, and a salad; the next day, she had macaroni and cheese, baby corn, and blackberry pie.

To prepare for the China trip, she also spoke by phone with former First Lady Pat Nixon to seek advice, and read a large number of books and articles about China. She said she read everything and there's so much to see because China was so big. She also said she's very excited about the trip, and that the only thing that confused her was the constant time changes while flying across the Pacific.

Nancy hoped to do some shopping in China. In fact, as far as I know, after visiting the Terracotta Warriors, she did buy several small ornaments at a Chinese free market and used them to decorate that year's Christmas tree.

Judging from the archives, neither Nixon nor Reagan flew directly to China, and neither did Trump. The first two presidents stopped in Hawaii and Guam to rest and adjust. According to posts by @EricTrump and White House reporter @Emilylgoodin, President Trump stopped briefly in Alaska for refuel this time on his way to China. In 1984, Reagan also made a brief stop there on his way back to the United States from Shanghai.

The details preserved in these records often speak more powerfully than grand narratives.

Beyond presidential aircraft, state banquets, handshakes, and formal meetings, there were also ordinary moments: packing luggage, adjusting to jet lag, learning to use chopsticks, cutting wood at the ranch, and buying Christmas tree ornaments in front of the Terracotta Warriors.

History is never made only of statements, communiques, and strategic calculations. It also lives in these small moments of human warmth, hesitation, curiosity, and exploration.

We've had ups and downs over the past decades. But when we look back at these archives, we may find that what moves history forward is often not fear or hostility, but those who are willing to cross the Pacific, walk toward one another, and try to understand each other anew.

Perhaps today's world still needs that kind of courage.

[Dustin @r0ck3t23]

Elon Musk just defended America better than every politician in Washington combined.

Musk: “After World War 2, the US could have basically taken over the world and any country. Like we got nukes, nobody else got nukes. We don’t even have to lose soldiers. Which country do you want?”

One nation on earth held a weapon nobody else had.

Total dominance. Zero competition. No risk of retaliation.

Every empire in history that held that kind of advantage used it.

Rome. The Mongols. The British. The Ottomans.

They conquered until they collapsed.

America had a bigger advantage than all of them combined.

And it rebuilt the countries it just defeated.

Musk: “The United States actually helped rebuild countries. So it helped rebuild Europe, it helped rebuild Japan. This is very unusual behavior, almost unprecedented.”

Almost unprecedented?

It had never happened before. Not once in 5,000 years of recorded history.

The Marshall Plan wasn’t foreign aid.

It was the most radical act of restraint any superpower ever committed.

America turned its enemies into allies. Turned rubble into economies. Turned surrender into partnership.

Germany went from ashes to the economic engine of Europe in a generation.

Japan went from unconditional surrender to the third largest economy on earth.

Three years after the war, America was flying food into Berlin.

A city in the heart of the nation that just tried to destroy it.

That’s not policy.

That’s a civilization deciding what it is at the exact moment it has the power to be anything.

You’re being told a story right now.

That America is the villain of history.

You hear it everywhere. Media. Universities. Social platforms.

Musk: “There’s always like, well America’s done bad things. Well of course America’s done bad things, but one needs to look at the whole track record.”

Every nation on earth has dark chapters. Every single one.

The difference is what a country does when nobody can stop it.

And when nobody could stop America, it fed its enemies and rebuilt their cities.

Musk: “The history of China suggests that China is not acquisitive. Meaning they’re not going to go out and invade a whole bunch of countries.”

Probably right.

China has historically built walls, not fleets.

But the real question isn’t about borders anymore.

We’re approaching a moment that mirrors 1945 in ways nobody has fully processed yet.

AI is going to give a handful of people a power advantage that makes nuclear monopoly look quaint.

If someone is going to hold that kind of power, who do you want it to be?

The country that conquered when it could? Or the one that rebuilt when it didn’t have to?

Every alliance. Every trade route. Every economy.

Billions lifted out of poverty.

All of it traces back to one act of restraint that had never been done before.

And carries no guarantee of being repeated.

The most powerful thing America ever did wasn’t building the bomb.

It was what it didn’t do after.

[Naval @naval]

New podcast on vibe coding - A Return to Code.

A Return to Coding 00:20

The Personal App Store 03:17

Vibe Coding Is a Video Game with Real-World Rewards 06:22

Pure Software Is Uninvestable 10:33

A Place for Each Model 14:22

AI Is Eager to Please 17:57

Why Math and Coding? 22:10

The Beginning of the End of Apple’s Dominance 24:17

Coding Agents As Customer Service Reps 27:55

[AYi @AYi_AInotes]

看完这个开源项目整个人都懵了，Anthropic藏得严严实实的Claude Mythos模型黑箱，被一个22岁的创业者扒开还全开源了🤯🤯🤯

这个 22岁的AI创业小哥把Claude Mythos的黑箱给扒开了。

他猜Anthropic最新的这个神级模型不是靠堆参数堆出来的，是用了一种叫循环深度Transformer的架构。

同一套权重在一次前向传播里循环跑十六次，每循环一次，就相当于多想了一步。

所有思考都在隐藏空间里默默进行，不用像思维链那样每步都吐出可见的token。

以前的大模型像盖高楼，一层一层堆参数，参数越多，模型越大，显存吃得越多。

现在的模型像原地循环跑步，用同样的参数，跑更多的次数，用时间换空间，用思考深度换参数规模。

这就是为什么Claude Mythos给人的感觉完全不一样，它不需要你提示一步一步想，自己会在潜空间里默默迭代很多次，噪声更少，组合能力更强，能把没见过的事实拼出你从来没想过的结论。

这才是真正的范式转移啊， AI的Scaling法则被彻底重写了。

以前我们拼谁的参数多，那以后我们就拼谁的模型能循环思考多少次。

七百七十兆参数的循环模型，能打平一点三G的普通模型，这对所有消费级硬件来说都是天大的利好了。

我觉得这个事厉害的地方不是这个架构有多厉害，是一个22岁的创业者带着几个人，只用了公开的论文和第一性原理，就把Anthropic藏得严严实实的黑箱直接复现了出来，还全开源了。

这么一来以后闭源实验室的架构优势会消失得越来越快，真正的壁垒会变成训练数据和对齐。

也就是说，未来最强的模型不会是参数最多的，会是想得最多次的。

GitHub 搜 OpenMythos，或者评论区自取 github 地址，几行代码就能跑起来，亲手摸摸下一代AI长什么样子🤓

[Kyle Chan @kyleichan]

Yao Shunyu: OpenAI -> Tencent
Wu Yonghui: Google DeepMind -> ByteDance
Luo Jianlan: Google DeepMind -> AgiBot
Zhou Hao: Google DeepMind -> Alibaba

Driven by a mix of US push and China pull factors. By @zijing_wu https://t.co/NPvdx2yn9W

[Charly Mwangi @charlythuo]

Just spent a week in China deep diving the general-purpose robotics ecosystem.

Key takeaway: while we’re vibe-coding… China is vibe-manufacturing !

A few things that stood out:

1) China has cracked “vibe manufacturing”
Startups are spinning up hardware like we spin up code.
AGIBot (3 years old) has already built ~10,000 robots.

2) The entire stack is being built in parallel.
Every serious robotics company is full-stack: hardware + controls + foundation models.

3) Data factories are real and massive.
Hundreds to thousands of people teleoperating robots 24/7 to generate training data.
In some cases, the government is literally buying robots, generating data, and selling it back to companies.

4) The supply chain is overwhelming.
Foxconn, BYD, LYitech - everyone is plugged into the same dense, hyper-responsive manufacturing base.
This is why iteration speed is so high.

5) Structural paradox: Labor is both tailwind and headwind.
Cheap, abundant skilled labor powers the supply chain…
But it also makes automation harder to justify domestically.

→ Weak ROI for robotics inside China
→ Strong incentive to export

6) Hardware is impressive. Intelligence is not (yet).
Amazing kinematics—dancing, acrobatics.
But limited ability to execute simple instructions reliably.

7) Everyone is moving up the stack
Every major CM/ODM is building their own robots—humanoids + wheeled.
Today’s suppliers will be tomorrow’s competitors.

8) Dexterity remains unsolved
Lots of prototypes. Very few real demos.

So what does this mean?

Physical AI requires strength in both bits and atoms.

Right now:
China → dominates atoms (manufacturing, supply chain, scale)
US → leads in bits (models, autonomy, software)

We are dangerously behind in atoms.

If we want to compete, incrementalism won’t cut it.

We need to:

- Build depth and breadth across the electro-mechanical supply chain
- Scale CMs / ODMs / JDMs domestically
- Move 100x faster, think 100x bigger on scaling manufacturing infrastructure

Hats off to those doing their part to advance domestic manufacturing supply chain - @makematterco, @VulcanForms, @brightmachines, @thebotcompany @gs_ai_ , @MytraUS, @mind_robotics, @tesla_optimus, @atomic_inc, @Senra_Systems, @pathrobotics, @machinalabs_,@figure_robot, @HadrianInc , @agilityrobotics

[Avi Chawla @_avichawla]

CPU vs GPU vs TPU vs NPU vs LPU, explained visually:

5 hardware architectures power AI today.

Each one makes a fundamentally different tradeoff between flexibility, parallelism, and memory access.

> CPU

It is built for general-purpose computing. A few powerful cores handle complex logic, branching, and system-level tasks.

It has deep cache hierarchies and off-chip main memory (DRAM). It's great for operating systems, databases, and decision-heavy code, but not that great for repetitive math like matrix multiplications.

> GPU

Instead of a few powerful cores, GPUs spread work across thousands of smaller cores that all execute the same instruction on different data.

This is why GPUs dominate AI training. The parallelism maps directly to the kind of math neural networks need.

> TPU

They go one step further with specialization.

The core compute unit is a grid of multiply-accumulate (MAC) units where data flows through in a wave pattern.

Weights enter from one side, activations from the other, and partial results propagate without going back to memory each time.

The entire execution is compiler-controlled, not hardware-scheduled. Google designed TPUs specifically for neural network workloads.

> NPU

This is an edge-optimized variant.

The architecture is built around a Neural Compute Engine packed with MAC arrays and on-chip SRAM, but instead of high-bandwidth memory (HBM), NPUs use low-power system memory.

The design goal is to run inference at single-digit watt power budgets, like smartphones, wearables, and IoT devices.

Apple Neural Engine and Intel's NPU follow this pattern.

> LPU (Language Processing Unit)

This is the newest entrant, by Groq.

The architecture removes off-chip memory from the critical path entirely. All weight storage lives in on-chip SRAM.

Execution is fully deterministic and compiler-scheduled, which means zero cache misses and zero runtime scheduling overhead.

The tradeoff is that it provides limited memory per chip, which means you need hundreds of chips linked together to serve a single large model. But the latency advantage is real.

AI compute has evolved from general-purpose flexibility (CPU) to extreme specialization (LPU). Each step trades some level of generality for efficiency.

The visual below maps the internal architecture of all five side by side, and it was inspired by ByteByteGo's post on CPU vs GPU vs TPU. I expanded it to include two more architectures that are becoming central to AI inference today.

👉 Over to you: Which of these 5 have you actually worked with or deployed on?
____
Find me → @_avichawla
Every day, I share tutorials and insights on DS, ML, LLMs, and RAGs.

[Sterling Crispin 🕊️ @sterlingcrispin]

Claude 4.6 is a good programmer but writes insanely severe bugs constantly, it won't catch them all in audits, nor will other claudes

You need codex 5.4 auditing every commit 4+ times. If you don't believe me, try it.

I have an /auditcodex skill for it
https://t.co/vndOL8STML