加密内参

I have a plan: Unitree flexing to the world, selling to overseas markets—by within two years, its market value will break one trillion.

Jimei is fantasizing again—does the West Coast really lack bitches that much?

Complete video URL:

Where is Silicon Valley in the AI era?
Huawei Ascend’s chief scientist answers: Not in California—it might be in Wudaokou.

At the end of a five-hour interview, this part made me sit up straight.

His reasoning was especially specific, not just a matter of sentiment:

His own child is studying in the third grade at Wudaokou, and he often goes there to take part in exchanges.

Just find any restaurant or café to sit down—at the table next to you, two people are arguing fiercely:

Why did that loss run off course yesterday?
How should reinforcement learning be done better?

His logic for what makes an innovation center is only one rule: wherever the atmosphere and the people are, that’s the center.

So why not Silicon Valley?

He didn’t say Silicon Valley is no good. He gave a more counterintuitive reason:

It’s not that Silicon Valley no longer has the talent-gravity—it’s that these companies themselves have a desire to monopolize, which limits mobility.

Big-company employees wouldn’t discuss the latest reinforcement learning techniques with Stanford undergrads—that’s a confidential competitive advantage.

And China is currently furthest from monopoly, with a talent supply that’s crushing-level. He said that 70% of the world’s top algorithm talent is Chinese, and a Berkeley professor is still teaching the distributive law of multiplication—so the gap is huge.

He also predicted something that’s happening right now:

In these three to five years, we’re in a period where the stars are shining brightly, where a hundred flowers bloom, and where exchanges of ideas are extremely active.

The window is only these few years.

After the whole five-hour interview, I still remember these:

• He said Moore’s Law should be measured in “how many atoms.” Economic milestones like 16/7nm have already stalled, and only energy efficiency is still profiting
• Electricity costs for China’s data centers are one quarter of those in the US; the real gap between China and the US chips lies in energy consumption
• The benchmark for whether you’ve survived a desperate situation: people start criticizing you
• A closing line: when the reporter asked what the most crucial judgment is right now, he only said three words: Betting on China

At 14, he entered Tsinghua’s talent program. In 1996 he went to the US with $1,000. In 2008 his wife pulled him back to China. In 2016 he joined Huawei HiSilicon—he said himself that at the time he “completely didn’t believe China had this capability.”

A decade later, Ascend runs on DeepSeek.

That entire process—an individual’s understanding being slapped by reality and rebuilt again—was all in those five hours.

The biggest magic in Sun Yuchen’s little essay isn’t the $50 million.
It’s the bunch of people in the comments who “manage his mindset.”
An eyebrow pencil set costing 79 yuan wants to drive the streamer off the platform,
while a bride price of 350 million yuan “isn’t anything.”
His own account balance is four digits, yet his valuation of someone else’s marriage promise is in the hundreds of millions.
Has Xiaohongshu’s sense of deserving inflated to this extent?

The core of this “bookkeeping” set from the Sweet Potato Girl is the “proportion method”:
The money you spend on me, as a proportion of your total assets—that’s what counts as sincerity.
Is that right?
By this logic, if Musk buys you a 30,000-yuan bag, it’s zero cost—so that means he doesn’t love you.

The proportion method only works when it’s used to calculate someone else’s wallet unilaterally.
When it’s your turn to split the bill, you can’t squeeze a cent out.
Double standards aren’t scary.
What’s scary is being so self-righteous about the double standards—and even turning them into a whole methodology.

Most ironic of all: the ones truly trapped by this “sense of deserving” are precisely the girls who believe they deserve it.
Quote a price based on fantasy, and the market responds with auctions that get no bids.
Value things based on daydreams, and you miss every real match.
This is the biggest situation in China right now: demanding prices are wildly misjudging reality.

The platform sells a “sense of deserving” as traffic.
And the influencers sell it as lessons.
In the end, the one who pays is the fool who still believes.

That high-upvoted comment on Xiaohongshu cursing “Sun Ge” for being stingy got 100,000 likes.
And everyone who liked it is fantasizing that they’ve got a diamond-studded—well, you know what.

If the 30 million betrothal gift isn’t refunded, what should this charge be considered as?
1. Breakup fee
2. Money for paying for prostitutes
3. Compensation for lost youth
4. Other

Which fee is more reasonable to choose?

Oh my God, that’s terrifying. The power of the landslide is 10 times greater than a magnitude 10 earthquake. In an instant, everyone was gone—absolutely horrific!

50 million for an egg? Sun Ge is just bad, not stupid.

With just a few lines of code, you can give your website the vibe—
transform it into a million-dollar web project.

I’ve done vibe coding on many websites.
Can you tell it instantly—the rich “AI flavor”?
It looks too similar to other generic sites.

Today I found a hidden gem:
→ Over 160 free components and website templates, including 3D, icons, and dynamic designs.
→ You can copy prompts or source code and hand it to an AI agent.
→ Afterwards, you can have it tailor themes, lighting, animations, or layouts to your project.

If you want to build a high-quality website, save this now!

It is said Byte has started training a model with 100 trillion parameters.
Its parameter count is catching up to Mythos.

Training costs have risen from "hundreds of millions" to "tens of billions,"
and China’s compute-power arms race has escalated.
#字节 #Doubao

2026 Beijing Robot Sports Games,
hosted the world’s first official human-robot table tennis match.

In matches involving low-pressure balls,
robots can basically handle it.
For balls with slightly higher pressure,
the robot technology still needs improvement.

This was done entirely autonomously by the robot,
with absolutely no human intervention.

It’s expected that within 1 year,
robot tennis sparring practice will become mature.
Within 3 years—can robots defeat the strongest human players?

Soon, robots defeat humans on a tennis court

Galbot robots
at the Beijing World Humanoid Robot Sports Conference
practice tennis—
completely autonomous.

You can now subscribe to Gemini models for free
If you’re a student,
you can subscribe for one year for free.

This is definitely not AI-generated

There are political commissars in the night shift too

In the end-to-end era, driving is evolving from a specialized system into a sub-skill of a general-purpose model.

With the same world model, driving, carrying boxes, and folding clothes are different outputs.

When that day truly comes, whether you put humans or objects into the vehicle won’t matter—it’s the model that matters, not the carrier.

Right now, for Optimus to learn to drive, it’s likely that storytelling comes before engineering.

Tesla’s robotaxi has been painting a picture for years without reaching mass production.

Learning to drive for humanoids is mostly for the capital markets—it hasn’t reached a road-safety milestone yet.

This timeline comparison has too much information.

Falcon 9: first flight in 2010, first successful recovery in 2015.

That’s a gap of 5 and a half years.

Jiefang 3: first flight in December 2025, successful recovery in August 2026.

That’s a gap of more than 8 months.

On the same route, SpaceX took five and a half years, while Blue Arrow took eight months.

Yesterday morning, Jiefang 3 Y2 took off at 07:35, and at 07:41 the first stage landed steadily on the Minqin range.

In 6 minutes: orbit insertion + landing recovery—everything went perfectly.

China’s first rocket that can return itself after reaching orbit.

Why is it “the first”? Because this closed-loop works:

• The second stage really sends the Honghu 03 satellite into orbit
• The first stage follows the planned trajectory and returns to the land landing site
• Both things happen simultaneously in the same flight

Engines and the rocket body are the most expensive parts of a rocket—reusability is basically picking up money.

But don’t rush to hype it yet.

Falcon 9 has achieved success for 676+ flights, while Jiefang 3 has only flown twice.

Catching up in 8 months is matching the first time—not achieving the 676th flight.

Next, an even more important thing: for the same first stage, when will it fly again.

Only when it’s reused does the cost accounting really hold up.

Walking on the grass, I happened to meet two little chicks 🐥 eating grass

RT @PeterDiamandis: Memory, not compute, is the rate limiter of the Agentic Era.

Why Are More and More People Using Pi?
While nearly all code agents are desperately piling on features, Pi chooses the opposite path: doing subtraction.

How thorough is this “subtraction”? Just look at the data: Claude Code’s system prompt reaches as high as 14,000 Tokens, while Pi uses only 200 Tokens. That means Claude Code has to cram 14,000 Tokens into the context every round of conversation, whereas Pi needs just 200 Tokens.

Behind these numbers is Pi’s radically different design philosophy.

"For agents, intentionally not doing certain things is often more important than doing things."

As agents become more and more complex, they eventually turn into an opaque black box that nobody can fully understand or predict. No one can clearly explain why it makes a given decision, and no one can guarantee it won’t go off track.

Pi’s solution is to converge the system into a minimal core—so simple that anyone can understand it, master it, and control it. Then, it grows capabilities through scalable mechanisms. The result is surprising: the simpler the system, the more controllable it is.

This philosophy reminds me of Unix design: each program does one thing well, and infinite possibilities emerge through combination. Pi isn’t building an all-encompassing super agent; it’s creating a clear, transparent, and predictable intelligence foundation so users can truly be in control—not just spectators guessing what happens inside a black box.

In this era where functionality reigns supreme, Pi proves through a subtraction mindset: sometimes, less is more. Simplicity doesn’t mean weakness—on the contrary, real strength comes from the parts you can fully control.