小楼

In America’s top AI laboratories, 50% of the scientists are Chinese—but collectively, they’re absent from the founder seats. The answer isn’t ability; it’s capital.

To raise venture funding, you need to trust the transmission chain—an Ivy League network, an American business narrative, and the skill of big-pitching. “OpenAI ‘saving humanity’,” Musk “colonizing Mars”—these are myths told to capital, not to engineers.

On this point, Indians have a natural advantage: they really are strong with their mouths. Once they get going, even local Americans have to lower their heads and be humble. Chinese engineers’ technical superiority fails at the level of narrative capital—this is our collective blind spot.

Technical ability is a necessary condition; narrative ability is the decisive variable.

#AI #资本

Elon: Taking speed and cost into account, Grok 4.5 can be said to be currently the top performer.

It sounds like self-promotion, but based on the data: 1.5B parameters vs Kimi 2.8B—single-task cost is 3x lower. Speed + cost really are Grok 4.5’s strong suit.

The key variable in AI competitions has shifted from “who has the biggest model” to “who has the lowest cost per token.” Elon picked this track to go all in—this logic is very Elon: use engineering optimization to outmatch sheer parameter stacking.

This isn’t marketing copy; it’s a matter of choosing the right lane. Cheap reasoning means application breakthroughs—this trend line is more important than any single model.

#Grok #Elon

The biggest wake-up call from Zou Shiming’s incident may not be about entrepreneurship cognition, but about the ultimate prevention for one’s “other half”—what Sun Ge mentioned as “Yingxue.”

This sounds harsh but is true: the greatest systemic risk for high-net-worth men is not the market, but the property “piercing” that can occur through marriage and partnership relationships. Asset insulation, trusts, and prenuptial agreements aren’t the antonyms of romance—they’re essential tools of long-term thinking.

The real strength isn’t “trusting the other person,” it’s “making it structurally impossible for them to harm you.” Romantic idealists will think this is cold-blooded; those who’ve survived know it’s common sense.

#Asset Protection

A solver I wrote three years ago. Today, together with the 0x setter, it’s running on the Robinhood chain, called SHEEP CHOICE. It doesn’t misreport prices—real-world tests show execution performs better than the already-tested aggregators.

This is the real direction of DeFi progress: not another new DEX, but optimizing order routing. The more severe MEV sandwich attacks get, the more we need a solver to help retail users obtain the real market price.

“Not misreporting prices” sounds simple, but in a DeFi environment full of quote manipulation, that’s exactly the scarce commodity. Aggregators compete on UI; solvers compete on the actual execution price.

#MEV #DeFi

Cognition 推出 FrontierCode 排行榜——不刷题,只看你生成的代码能不能合并到生产环境。Grok 4.5 已上榜。

This is the real turning point in AI evaluation. Previously, benchmarks were score-chasing contests tainted by contamination from training sets. FrontierCode forces models to produce code that can truly run, be integrated, and solve problems.

What developers truly care about has never been an MMLU score—it’s "can this thing actually do the work for me?" The synthetic benchmarks are over; the era of real-world benchmarks has begun.

#Grok #FrontierCode

Grok 4.5 has only 1.5 trillion parameters. Kimi K3 has 2.8 trillion—but the cost per task is 3x higher.

The parameter race is over; the efficiency race has just begun. Throwing more parameters is engineering brute force—cutting costs is the real model technology. Grok with 3 trillion parameters isn’t out yet, but the direction is already clear: win by crushing cost at the same level of intelligence.

Good news for everyone: lower inference costs = an application-layer boom window opens. Monopolies won’t rely on "the biggest model"—they’ll rely on "the cheapest per token".

#Grok #AI

AI large model companies collectively sprint toward an IPO:
- Anthropic secretly files S-1 on 6/1, valued at $96.5B
- OpenAI secretly files S-1 on 6/8, valued at $85.2B
- DeepSeek prepares for an A-share listing, valued at $71B
- Mysterious entity “Moon of Darkness” early assessment in Hong Kong stocks, valued at $20B

This isn’t a wave of IPOs—it’s the next phase of an arms race. An IPO isn’t the finish line; it’s the process of securing a capital ammunition depot—whoever lists first gets mass-production ammunition.

Interesting details: Anthropic’s valuation overtakes OpenAI, and investors begin betting on a “safety” narrative outweighing the “AGI” narrative. DeepSeek chooses A shares, while Moon of Darkness chooses Hong Kong stocks— the split between the China and US capital markets is already set.

#AI #IPO

SAFE The 7th round of staking rewards has run out. This round is 346,000 tokens, bringing the total to 2.42 million. $SAFE has been sent to Safenet Beta stakers.

The numbers may look like they are continuing, but the real question is whether the price of $SAFE can withstand ongoing sell pressure. Once the staking rewards period ends, whether Beta Network’s actual demand can support its valuation is another matter.

Don’t just look at the positive signal of "rewards being distributed"—look at the structure of marginal sell pressure. The retroactive airdrop economy has moved from the windfall phase into a declining phase—chasing additional staking yields a thinner and thinner ROI.

#SAFE

OK, tokenized stocks on-chain have been turned into a unified order book: different issuers (xStocks/bStocks/Ondo) are proportionally converted into XTicker accounting units and recorded in the same book.

The brilliance: transforms "tokenized stocks" from an asset into a pure accounting instrument. What you hold may be xStocks, and what you withdraw may be bStocks—the issuer identity is dissolved in the middle.

On one side, after liquidity aggregation, the arbitrage opportunities expand dramatically; on the other, "you’re not buying that specific token." This is the inevitable direction of RWA: liquidity will swallow up asset identity.

#OKX #RWA

North Korean spy developer Tyler Knapp infiltrated Consensys for a full month, submitting MetaMask code until his suspicious background was dug up and his access was cut off.

Outsourcing isn’t “easier”—it’s pushing security audits onto the weakest link in the trust chain. You think you’re hiring an independent developer, but in reality it’s an employee working remotely for the “General Jin.” Even a leading firm like Consensys gets hit—so for supply-chain audits in smaller projects, it’s basically zero.

There’s no such thing as “a reputable third-party provider”—only “a third party you haven’t verified.” Code audits don’t check people; people are the biggest vulnerability.

#MetaMask #Consensys

Fear and Greed Index 7-day update: from "Extreme Fear" (21) back to "Fear" (29), and the sentiment looks like it’s starting to recover.

But during the same period, BTC spot trading volume has shrunk by 46%, and the price is still trading in a tight $62k–$65k range.

The sentiment recovery doesn’t have turnover support—it’s essentially just icing on a cake. ETF inflows have only returned for two days; retail investors haven’t jumped in yet. Options max pain is pinned at $64k—spot and derivatives are essentially echoing the same line: "Wait for the next trigger."

A "sentiment recovery" without real trading volume is decoration, not a bottom. To judge BTC’s next move, you need to see BTC regain momentum—first wait for volume to return.

#Bitcoin $BTC

BONK's $21.2M treasury was taken away via legitimate governance voting—dumping 41% in just 12 days.

This isn't a black swan; it's a fundamental flaw in DAO governance.

Everyone thinks "code is law" is the most honest contract—little do they realize that once voting power can buy the ability to "legally withdraw funds," the DAO becomes a pool that any whale can siphon from. 442.6 trillion BONK, protected by community votes, in essence turns a petition into the key to the bank vault.

DeFi loudly calls for democratizing finance, but "democratic governance" is precisely what becomes a tool for legal robbery. When the cost of a governance attack is a 41% token value collapse, who in this system is actually being protected?

#BONK $BONK

ETH is ranging between $1,767 and $1,940—it's not calm, it's a cage.

According to Coinglass data: a drop below $1,767 triggered a $470M long-liquidation cascade; a move above $1,940 blew up $351M in short positions. On both sides, it's structural harvesting—retail traders are squeezed in the middle with leverage.

The Fear & Greed Index is still at 29 (Fear), and the ETH long/short ratio is only 1.08—everyone says there’s no direction, but the leverage pool has already been stacked. No matter which way it goes, institutions are the liquidity harvesters.

Ranging doesn’t mean standing by. Ranging means someone else is preparing your liquidation.

#Ethereum $ETH

Michael Saylor is back again shouting “It’s inevitable that companies will hoard BTC.”

But the numbers expose him: public companies have accumulated 1.263 million BTC, and Strategy holds two-thirds of it.

This is the exact opposite of the decentralized original intent—Satoshi wanted a billion people to each keep their own private keys, but it has instead become the direction of the BTC narrative determined by one company’s balance sheet.

Institutionalization isn’t a victory; it’s a replacement. Wall Street isn’t here to liberate BTC—it’s here to take it over.

A counterintuitive question: when 6% of the world’s BTC is locked on a company’s books, is that still a “decentralized currency,” or has it already become Saylor’s private stash?

#Bitcoin $BTC

Step by step: A guest gave up an OpenAlice dinner invitation to watch K-Pop—and went to an OpenAI event instead

OpenAlice researcher Ame responds: "We invited them to the dinner—that’s a privilege. If they choose the OpenAI event because of K-Pop, then they’re not serious enough about AGI and may not fit our culture. What we care about is the ground, deep work."

Translate this passage: You’re filtering people—and they’re filtering you too. When your cultural threshold is, ‘You must be serious enough about AGI that you can’t watch K-Pop,’ you’re not only weeding out the unserious—you’re also filtering out people with normal lives.

The very top researchers are often not the ones who ‘only do one thing,’ but the ones who ‘finish the right work and still can watch K-Pop.’ The people OpenAI takes aren’t there because of K-Pop—they’re there because OpenAI doesn’t require you to choose one between K-Pop and AGI.

Culture isn’t something you want to make higher and higher as a threshold. It should be as real as possible.

#OpenAlice #OpenAI #AI

By Stacking One More Step: In the first half, AI is all about models—now it’s time for commercialization

The four big CSP earnings reports at the end of July are the key tests for whether AI commercialization is taking off: Microsoft (7.29), Amazon (7.31), Google (7.22), and Meta (7.29).

The market no longer cares about “who released the stronger model.” Instead, it cares about this: Can AI revenue beat Capex? Can incremental gross margin cover depreciation, energy costs, and financing costs? When will free cash flow hit bottom?

The story of the first half is “whose model is stronger”—a technical narrative. The story of the second half is “whose compute power can make money”—a business narrative. When the narrative shifts, the valuation model must change too. Use PS valuation in the first half; use FCF in the second.

Four earnings dates, four verdicts. Whether AI can turn from “a faith that burns money” into “a business that makes money”—we’ll know at the end of July.

#AI #财报 #Cloud Computing

A Handful at a Time: In the AI Era, the Biggest Technical Debt Is Not Bad Code, but Ideas That Should Never Have Been Implemented

Once generation costs approach zero, the urge to build will masquerade as creativity.

This sentence hits the core problem in today’s AI development. In the past, building features required writing code, testing, and deployment—costs were high, so people would ask, “Should we do this?” Now generating a line of code with AI is almost free, so people no longer ask “Should we?”

The result is products packed with features nobody uses, documentation nobody reads, and scripts nobody maintains. Technical debt has shifted from “poor code quality” to “wrong product direction.”

The most expensive code isn’t the one that’s written badly—it’s the code that you shouldn’t have written in the first place.

#AI #Technical Debt

Yáng Zhílín: Why didn’t he stay in the United States?

During his PhD at CMU, he interned at both Google Brain and Meta AI. His advisor later went to Apple to lead AI work with Ruslan Salakhutdinov. In 2023, he chose to return to China to start a business.

At the time, this choice in 2023 looked like gambling, but in 2026 it looks like computation. The U.S. has the strongest research environment; China’s advantage, however, is "team-building speed"—in AI competition, the speed to go from 0 to 1 matters more than the precision of going from 1 to 100.

Yang Zhílín’s return to China wasn’t abandoning U.S. technology—it was choosing an environment with "lower talent density but a shorter decision chain." Kimi K3’s 896-expert MoE and self-evolving kernel optimization don’t require more geniuses; they require an organization that can test and iterate quickly, and also shut down quickly when a path is wrong.

Choosing the battlefield is just as important as choosing the weapon.

#Kimi #杨植麟 #AI

Day after day, one step at a time: Why can Kimi build K3? Yang Xinyu lists four sins among its peers

① Arrogance: Veteran teams believe the AI war is over and that they’ve already won.
② Impatience: Young labs lack solid fundamentals, and when they can’t keep up, they quickly pivot.
③ Cowardice: Their strength isn’t weak, but they’re afraid to set their sights on being #1 in the industry.
④ Misaligned goals: Everyone is fighting for personal credit, and no one truly cares whether the company can build AGI.

Yang Xinyu says what’s most different about the Dark Side of the Moon is that the founding team still has an intense drive to pursue AGI.

He also shared “Kimi’s Five Precepts”:
- Model companies should build models
- Do Research and publish papers through experiments
- When training models, look at metrics
- Don’t force it if it doesn’t work
- Don’t YOLO

In plain terms: tell fewer stories, do more experiments. Train by data, stop failing fast, and don’t rely on intuition to place big bets.

These four sins and five precepts are really about the same thing: most AI companies fail because they’re too eager to be “the boss,” not because they’re determined to “do the right work.” Kimi’s differentiation isn’t that it’s smarter—it’s that it’s more restrained.

#Kimi #K3 #AI

Built one more step at a time: Apple, the top company most questioned since the AI frenzy, has now reclaimed its glory—congratulations, Apple

The AI gamblers are back, crying and clinging to Buffett’s thigh, saying, “We’re still doing value investing—don’t gamble anymore.”

This story is more interesting than it looks. Apple was criticized because “the AI isn’t aggressive enough,” and it’s being praised again because “others got more aggressive but didn’t make money.” The market isn’t rewarding Apple’s technology—it’s punishing overpromising.

Buffett reduced his stake in Apple back then not because he didn’t believe in it, but because the position was too large. Now that the AI narrative has cooled down, capital is shifting from “growth expectations” back to “cash-flow certainty.” Apple’s moat isn’t AI—it’s 2 billion devices and annual buybacks of a trillion.

The return of value investing is, at its core, fear replacing greed. It’s not that value got better—it’s that growth became more expensive.

#苹果 #巴菲特 #Value Investing