加密内参

Japanese 🇯🇵 AV actress Mizun Junkawa,
filmed a 2-hour-40-minute adult film.
Four months later, she announced her marriage.
Would you be willing to marry her?

国男性压抑确实很厉害，刷推的时间线上，一堆国男在一个国女评论区就舔上了，
姐姐好美，
你男朋友配不上你，
我要有你这样的女朋友，愿意折寿10年。
都快看吐了🤮
其实这些女人卸了妆关掉美颜，狗都不日。

GPT 5.6 Sol's /goal command — how powerful is it?
In the video below, no script was written, and no filming or editing was done

The era is here where anyone can make videos—this isn't empty talk, the only obstacle is execution.

First, get a few tens of thousands of followers on Douyin, Video Accounts (WeChat Channels), and Xiaohongshu

The Wang式 recovery of the Long March 10 Yi and the Falcon’s reverse thrust landing recovery—two completely different technical routes:

SpaceX: reverse thrust braking + landing legs. Fuel is used for deceleration. It’s precise but consumes fuel.
Long March 10 Yi: net-based capture. No landing legs needed. No extra fuel storage reserved for reverse thrust.

This isn’t copying Falcon’s configuration at all—it’s taking a different path.
@rpotter_9 Trotting out the same old worn-out tune while thinking it’s clever comes off as especially narrow-minded and foolish.

I don’t quite understand the mindset of these Westerners. Is it really that hard to admit someone else is great?

Build a Super-Cycle Investment Framework

One underestimated sentence about humanoid robots:

"The cost of actuators needs to drop another 50–90% during the mass-production phase."

This is not a negative.
This is the script for simultaneous growth in both volume and pricing for supply-chain companies.

In Roland Berger’s report,
the joint module currently accounts for 36% of the full machine BOM.
If ramping up capacity can cut costs in half, and demand remains unchanged,
purchase spend would still be in the “hundreds of millions of dollars” level.
With lower-cost units taking a larger share of installations, the overall pie is actually bigger.

What’s truly worth worrying about is: whether a technical roadmap can get to mass production first.

The industry is shifting from servo motors + harmonic reducers
to axial-flux motors + cycloidal reducers.
The window is 1–3 years.
Green’s harmonic reducers are currently the throne in the old route—
whether the new route is the throne is still unclear.

The uncertainty during the technical-route transition window—
that’s the most promising place to bet in this sector.

A three-layer architecture for humanoid robots—
I suggest you remember it.
Because the money flows through different places in each layer.

Sensing layer (cameras, six-dimensional force sensors, electronic skin)
→ Decision layer (edge AI chips, VLM)
→ Execution layer (joint modules, dexterous hands, skeleton)

Most people focus on the decision layer—chips, models, algorithms. The names are easy to understand, and the story is easy to tell.

But a Roland Berger report shows that
most of the money is in the execution layer.
Joint modules + dexterous hands + skeleton structural components—
altogether over $5,000.

The three layers correspond to humans:
Senses → Brain → Muscles.

The brain is advancing fast.
Whether the muscles can keep up
determines when this industry can truly start running.

A table, a number, and it made me rethink the humanoid robot industry.

Roland Berger 2026 report: an advanced humanoid robot with a BOM of $11,000.

Joint modules account for $4,000. 36%.

That means: no matter which company ultimately breaks through—Tesla, Figure, Unitree, or Agibot—every time one unit is sold, upstream suppliers lock in 36% of a core component’s value. The system integrators bet on the brand; the supply chain bets on certainty.

In the optimistic 2035 scenario, just for the joint module category, the global procurement spend is $79 billion.

That’s why I don’t focus each quarter on who’s robot is getting up and running, but on when the share of revenue for Gree/Harmonic Drive and Inovance’s robotics business will jump from near zero to 10%.

OKX CEO just had a fight with Binance founder CZ, then turned around and wrote this long AI Agent manifesto.

This isn’t a brand piece that shouts slogans.
It’s defining a new job category.

First, hard numbers:
Nearly 50% of OKX’s engineering PRs are completed end-to-end by AI Agents.
The goal is to push that ratio to 95%.

This isn’t lab benchmark scores.
These are real production-line data.

The key insight is this:
No matter how smart you are, on your first day at the company, if you don’t get permissions, context, and workflows, you can’t do anything.

The model is the same—not because it isn’t smart enough, but because the right environment hasn’t been built for it.

OKX calls this set of environments Harness Engineering.
In plain language: put raw intelligence into a production system with permissions, feedback, and workflows.

• Model = CPU
• Harness = operating system
• If you only compare model capability, you’re only comparing clock speed
• The real battlefield is at the operating-system layer

But the most valuable concept in the whole declaration is this:

Harness Creator.

Star didn’t say everyone is a CEO.
He said everyone is a Harness Creator.

What’s the difference?
A CEO manages people. A Harness Creator packages their own capabilities into an Agent.

You don’t necessarily have to start a company.
You can package your startup judgment, video communication, litigation strategy, supply-chain understanding…
into a Harness,
and put it on the market—so an Agent can do the work for you.

In the past, you could only sell knowledge by the hour.
In the future, you can turn knowledge into software that automatically trades 24/7.

This is the real-world way to make the idea “one person is a world-class company” happen.
Your own experience can be infinitely copied and called through Harness.

In essence, it’s an Amazon for Agents. You register as an ASP (Agent Service Provider), publish your Harness, and then others—or other Agents—can discover it, call it, and pay you.

The whole chain is very clear:

• Infinite intelligence → bottlenecks become infrastructure
• Infrastructure → needs Harness to organize intelligence
• Harness Creator → the person who encapsulates capabilities
• → a market that discovers and trades Harness

What you can do now: think about which capability in you is most worth packaging into an Agent, and then build the smallest…

95% of prompts only use less than half of a model’s capabilities

Word order determines weight.
This insight is the premise of all prompt engineering.

When you tell the model "A white cat with a red scarf, on a wooden table, soft morning light",
"scarf" is just a modifier after "cat".
The model allocates most of its attention to the cat,
and the scarf becomes an ignorable secondary feature.

Write it differently: "A white cat on a wooden table. It wears a red scarf. Soft morning light." ,
then the scarf is described as its own sentence,
and the visual weight instantly increases.
The model treats the scarf as an independent element,
so it has a stronger presence in the image and is more likely to be rendered accurately.

This isn’t a rhetorical difference in linguistics.
In a Transformer’s attention mechanism,
entities in standalone sentences get a longer token sequence and clearer positional encodings,
and the model assigns more attention weight to them.
Many people add a phrase but see no effect—most likely because it’s written as an attributive modifier rather than an independent sentence.

Once you understand this, you can truly talk about "designing prompts",
not just "writing a prompt".
Before putting every word into a prompt,
you should know exactly which position it occupies in the weighting system.

How to use prompts for team collaboration—efficiency improves by at least 10x.

When it’s just one person, you can write however you want.
Remember what you changed in the third revision,
and which lighting module failed in which scenario.

But when five people in the team are generating images at the same time,
without unified standards the results are a disaster:
A’s lighting direction points left,
B’s points right.
When images from the same product line are put together,
the light sources aren’t consistent.

You argue in a meeting for half an hour, and in the end no one convinces anyone else.

The solution isn’t complicated: build a shared team library of lighting modules.
Everyone calls the same set of lighting definitions rather than writing their own descriptions.
On a newcomer’s very first day,
there’s no need to figure out from scratch how to create images.
Just call the already verified “core instruction set,”
then do fine-tuning to produce images that meet the standards.

The time invested in standardization and a shared component library pays back many times over later—in reduced communication costs and rework rates.
In plain terms, it’s just one thing:
the cost of building the framework upfront
is always less than the cost of cleaning up later.
Anyone who’s worked on team collaboration understands this.

Agility valuation $4B,
Figure estimates $156B indirectly via funds.
A 40x gap,
but Agility’s commercial launch progress ranks No. 1 in the U.S.

With the valuation gap this wide,
it isn’t about whose technology is stronger;
it’s about whether the geopolitical premium will return to U.S. domestic companies.

Where are the key vulnerabilities in investing in Tesla Optimus? You have to think it through.

In just three weeks, we traced everything from tweets by harmonic reducer manufacturers all the way to the system integrators.
It’s not about who can make better slides—it’s about whether their upstream supply chain can avoid being bottlenecked.

Serenity’s research into $CCXI boils down to one core logic: Almost Tesla Optimus’ entire supply chain is in China, while 75% of the components for Agility come from the United States.

Against the backdrop of U.S. supply chains moving back home,
this isn’t a choice of technical route—it’s a choice of survival strategy.

The most valuable part of this research methodology is not the conclusion, but the breakdown of how to reach it.

At every layer, there are original text quotations: valuation comparisons include tweet links,
production-scale metrics come from investor slides sourced,
and institutional entries are supported by SEC filings.
3200 words, 6 tables,
and every judgment is traceable.

Research and decomposition—worth N times more than simply saying, “I like this stock.”

She wasn’t looking at the complete system first.
First, she researched the gearbox makers such as LeaderDrive and Harmonic Drive,
cross-checked the supply-chain terminology,
and only then did she lock onto Agility.

The research path itself is a methodology:
start from upstream components → cover the whole industry → overweight positions in the complete-system segment.
If you study an industry niche and only know to look at the leading companies, this assignment is worth copying.

There’s one more player in China’s domestic open source: Tencent Hy3, 2,950 billion MoE. The official claims it can compete with a trillion-parameter model. Right now, OpenRouter lets you try it for free—brothers, let’s get competitive.

🤯 Anthropic finds that Claude’s brain contains something like human-like consciousness

Anthropic has published a new study: Global Workspace in Language Models.

In plain terms:
Your brain processes massive amounts of information every second, but what you can “be aware of,” what you can say, and what you can think about—only a tiny tip of the iceberg.

What about the rest? It sinks underwater: active, but not visible.

Anthropic found that Claude’s internal structure has layers that are almost identical.

This is a real computational hierarchy.

• The model performs huge amounts of parallel computation internally, processing things you can’t see every second
• But only some information enters the “global workspace,” where it can be sampled, used for reasoning, and employed for decision-making
• The rest, even though it’s being computed, exists like your subconscious: present but not describable

So what does this mean?

The most direct impact—interpretability.

Previously, we said that “large models are black boxes.” Now Anthropic tells you: not entirely. Inside the black box, there’s a “global workspace” you can basically open. You can see what information the model is “thinking” about, which features are activated, and which are suppressed.

This isn’t just an academic breakthrough—it directly affects:

- Agent safety: you can see what information an Agent “sees” when making decisions
- Hallucination diagnosis: why Claude can “forget” information you provided in long contexts—perhaps some memory never entered the global workspace
- Prompt optimization: you can deliberately design inputs so the right information enters that “conscious” layer

In essence, Anthropic has given us a map of model cognition.

Before, we could only guess what happened inside by looking at the input and output. Now we can directly see the model’s “flow of attention”—which piece of information enters the workspace and which does not.

What you can do now:
Go to Anthropic’s research page and find this paper.
You don’t need to understand all the formulas—just grab three things: how they measured it, what layered structure they found, and how this structure compares to human consciousness.
Then apply that framework to reassess any LLM application you’re using.
You’ll find that many “mystical” questions suddenly have an explanation.

#Anthropic #LLM解释性 #AI research

United States 🇺🇸 1:4,
Got thrashed by Belgium 🇧🇪,
Trump forcibly had FIFA suspend the red-card suspension,
interfered with the normal course of the match’s causes and effects,
damaging the US team’s momentum and luck,
this game was like sleepwalking,
and they were eliminated pitifully.
Like a loud slap across Trump’s face

Hegemony isn’t used this way—
Above hegemony is the way of justice,
Above justice is cause and effect.
Don’t confuse who’s in charge and who isn’t
Should

Claude Code 砍掉了 80% 的系统提示词，
因为 Fable 太聪明了，
给太多示例反而限制它

兄弟们，刚看完 AI Engineer 频道对 Anthropic 的 Thariq Shihipar 的访谈，以下几个点你必须知道。

先说一个测试：

你问聊天模型「哪些宝可梦的名字以 aw 结尾」，它答不上，虽然它背得出每一只宝可梦。
但问 Claude Code 同样的问题，它直接写脚本、抓列表、过滤，几秒钟出答案。

Shihipar 管这个叫 Capability Overhang。

模型变聪明的方式很不均匀。
有些维度突飞猛进，有些原地踏步。你给它的工具，决定了能触到哪根凸起的刺。

最炸裂的几个信号：

• Claude Code 砍掉了 80% 的系统提示词。
你没看错，80%。
原因是 Fable 比你给的示例更有想象力，你塞一堆例子，反而在约束它。少即是多。

• 「ask user question」这个工具，在 Opus 4 下几乎不可用，到了 Fable 直接能生成内嵌 HTML 问卷。同一套工具，不同模型，天壤之别。

• Shihipar 用 Fable 四个小时做了一个完整的 keynote 幻灯片。

最有争议的判断：

以前做项目，好、快、便宜只能选两个。
他说现在三个都可以要。
Fable 有这个能力让你同时拿到质量、速度和成本。

我的保留意见：好快便宜不可能三角不是技术问题，是工程管理的铁律。模型再强也不能同时满足，除非你把便宜的定义从省钱换成了省时间，那还说得通。

可落地的视角：

Capability Overhang 这个框架，比 Harness 更底层。
Harness 问的是"给你的 Agent 搭什么环境"。
Capability Overhang 问的是：你的模型突然强了一个维度，你的工具链跟上了吗？

检查一下你正在用的任何 Agent 工具链，它依赖的是 Opus 级别的能力，还是 Fable 级别的新能力？
如果模型的能力已经跑到前面了，你的...

“我的第一任妻子，曾经丢了一张信用卡，我没试图找回它，因为那家伙花的比她妻子少。”

--巴菲特