跟Leo看AI

Mt. Gox has been delayed yet again, with the repayment deadline pushed back to October 31, 2027.

The official reason is that quite a few creditors still haven’t completed registration and verification, and there have also been issues with some payments. Mt. Gox was hacked more than ten years ago 👇👇

2014: Japan-based exchange Mt. Gox was hacked, with around 850,000 bitcoins stolen. It subsequently went bankrupt.

July 2024: Mt. Gox officially began repaying creditors in BTC through platforms including Kraken, Bitstamp, BitGo, Bitbank, and SBI VC Trade. More than 90,000 BTC were transferred out in a single month.

Since 2023: The deadline has been extended four times in a row, pushed back from 2023 all the way to 2027.

Of the 142,000 BTC available for repayments, nearly 110,000 bitcoins have already been paid out, leaving around 34,000 in the trustee’s wallet. This suggests that the selling pressure on the market from Mt. Gox repayments is far less significant than it once was. #门头沟还款再度延期

🐮😱 The wildest documentary ever? A new film released overseas, <i>You Can See Everything</i>, has earned the highest rating on MTC in nearly 10 years—and it’s a real documentary.

The film’s subject is Elizabeth Holmes, the central figure in Silicon Valley’s shocking fraud. At 19, she dropped out of Stanford to start a company, building its valuation up to $9 billion—on par with Tesla at the time. It turned out that the company had been a scam from start to finish❗️

The film was shot 34 days before she began serving her prison sentence. Holmes appears in it herself, offering a glimpse into her everyday life through interviews and footage that follows her around. In light of what she did, combined with those signature wide-open eyes that barely blink, the whole film has a chilling, spine-tingling feel.

Let’s take a look back at the scam👇👇

In 2003, Holmes founded the blood-testing company Theranos, claiming it could run hundreds of tests using just a few drops of blood from a fingertip. With that story—and an image carefully modeled on Steve Jobs—she brought political heavyweights like Henry Kissinger and George Shultz onto the board. By 2014, the company was valued at around $9 billion, making her the world’s youngest self-made female billionaire. But the technology had never actually worked. The company secretly used modified commercial machines to test samples, deceived investors and regulators, and sent inaccurate test results to real patients.

In 2015, Wall Street Journal reporter John Carreyrou exposed the scam based on tips from company insiders. Theranos dissolved in 2018. In 2022, Holmes was sentenced to 11 years and 3 months in prison for defrauding investors, and she is still serving her sentence. The case also became the most famous cautionary tale of Silicon Valley’s “fake it till you make it” culture. #骗局 #你能看到一切

The quantum threat is back—and this time, it seems different!!🚨🚨

A couple of days ago, Anthropic published a paper on quantum computing titled “AI Designs More Efficient Quantum Circuits, Bringing Applications in Chemistry and Materials Science Closer.”

But the most detailed of the paper’s eight tasks was 256-bit elliptic-curve point addition. It’s a key step in using quantum computers to break the elliptic-curve discrete logarithm problem, and elliptic curves are exactly what the signatures for $BTC and $ETH use.

What did the task involve?
1. Claude Mythos 5 was given only a brief specification and an automated scorer (number of qubits × number of gates), then left to iterate in a closed loop on its own, with virtually no human intervention.
2. The most expensive step in point addition is modular inversion. Claude ran for 53 hours and produced the first reversible divsteps implementation that doesn’t need to store branch history. It also proved several new algebraic identities on its own and formally verified them using Lean.
3. Compared with the smallest design to date, it used 12% more qubits but 17 times fewer gates.
4. It also reduced the gate count of DQI’s Reed–Solomon decoder to one-twelfth of Google’s previous design.
5. For historical context: the estimated number of Toffoli gates needed to simulate FeMoco was around 50 trillion in 2017. It has now fallen to about 340 million—a reduction of roughly 150,000 times. This was mainly achieved through eight years of incremental algorithmic improvements by multiple research groups, with each step taking months.
By contrast, Claude’s elliptic-curve design took just one 53-hour run.

In other words, in just 53 hours, AI made a key step toward quantum attacks on BTC and ETH signatures more resource-efficient. Your cryptocurrency isn’t at risk of being stolen just yet, but the quantum threat may arrive sooner than people expect.

Once the hardware catches up, addresses whose public keys have already been exposed on-chain will be the first at risk. The upgrade to post-quantum signatures can’t be put off any longer!!

#AI Could CoreWeave become the AI industry's “subprime bomb”? 🚨⚠️ This time, Jensen Huang's reassurances may not be enough.

Lately, Huang has been trying to shore up market confidence: the current turbulence will eventually pass, AI will transform every industry, and we're still at the beginning.

But Wall Street is no longer asking whether AI is useful. It's asking: how long will it take to earn back all the money borrowed to buy GPUs?

CoreWeave's expansion model is simple: buy GPUs, sign long-term computing contracts, then use those contracts and assets to raise financing and buy more GPUs. As long as customers pay on time and financing keeps flowing, the cycle can continue.

The problem is that if power shortages delay deliveries, revenue could be pushed back while interest payments still come due. Older GPUs also lose value as the technology advances. If new orders slow and renewal prices fall, making money and borrowing money could both become harder.

Revenue falls short of expectations, and borrowing gets more expensive; borrowing gets more expensive, and repaying debt gets harder. That's the “death spiral” the market fears—not just a few days of falling share prices.

Jensen Huang certainly has good reason to be bullish on AI. But Nvidia making money by selling GPUs doesn't mean the companies buying them will make their money back.

Long-term contracts and continued demand could support this business model, so it's too early to declare a “subprime crisis.” But to dispel market doubts, CoreWeave needs to show cash flow—not ask for faith.

Musk really is at the forefront of human progress. Just look at the things he’s working on: rockets, Starlink, new energy vehicles, data centers in outer space—each concept is big enough to inspire its own Hollywood blockbuster. And now he’s made a flying car! The stuff we wrote about in childhood essays has come true!!! So cool. You really have to hand it to this man #特斯拉发布Roadster

Lao Gan Ma was recently trending in the news after reports claimed the company made nearly $800 million in a year using AI.

What’s the story? As it turns out, last year Lao Gan Ma introduced an AI-powered visual inspection system that reduced the defect rate in key production processes to 0.02%. In the factory, cameras scan the conveyor belt every 0.3 seconds, and any substandard peppers are blown into a waste bin. Put simply, the profits don’t really have much to do with AI—it’s just a case of giving AI credit where it’s not due 😂😂

But before the AI system was introduced, there were over 600 workers sorting peppers along the production line. Now there are fewer than 50. Am I the only one feeling sorry for the workers who’ve been replaced??

#Muse超越ChatGPT成iOS免费榜首 Why hasn’t China come up with its own Muse yet? Doubao Phone: I want to, but they won’t let me 😕😕

The Doubao Phone Assistant has long wanted to create a full suite of AI agents that can help users automatically place orders on Taobao and order food delivery on Meituan. To get permission, Doubao introduced a policy last month and made it public for 30 days, allowing app companies to decide whether to let Doubao AI automatically perform actions on users’ behalf. When the consultation period ended, several of China’s major companies—including WeChat, Alipay, Taobao, Meituan, Pinduoduo, and Xiaohongshu—had all refused.

The logic behind the big tech companies’ decision is actually easy to understand: if you let AI act on users’ behalf, how am I supposed to sell merchants’ recommendations and splash-screen ads? There’s no upside for them in this at all. There really is no solution in the short term.

Another male celebrity came out and revealed that he lost even more trading NFTs than he did trading stocks 😂😂 From what I remember, other celebrities who’ve traded NFTs include Stephen Chow, Jay Chou, JJ Lin, and Edison Chen, right? But most of them still managed to make a quick buck off retail investors. #方力申自爆买NFT亏惨

I can hardly believe it! One person created this incredibly polished 3D game with AI in just half a month!! I tried it out—the art style and soundtrack are both top-notch. It got over 1.6 million views overnight after being posted!! And the numbers are still climbing rapidly.

That’s incredible. Having great taste, knowing how to use AI, being highly driven, and being willing to share and express yourself—that’s the winning formula for this era!!!

This world has really gone crazy... AI are starting to have babies and even taking maternity leave 😓😓

A few days ago, the U.S. District Court for the Southern District of New York sentenced 54-year-old Michael Smith to 18 months in prison for fraud involving AI-generated music and artificially inflated streaming numbers 😳😳 This is also the first criminal case of its kind in the United States.

So how exactly did he do it? First, he used AI tools to generate hundreds of thousands of songs in bulk, then uploaded them to major platforms like Spotify and Apple Music to collect royalties. He then used bots to rack up streams. Over seven years, this scheme earned him $8.09 million...

And this guy was arrested back in 2024. If he’d used today’s AI, he’d only have made even more #神操作

#Ledger暂停CryptoBilis销售 After the Ledger theft, someone dug up the compensation terms for another hardware wallet, OneKey, which state that the maximum payout is just $100. This is a tough one to judge, and there don’t seem to be any comparable cases in Web2. Have you ever heard of someone buying a safe, having their assets stolen, and then demanding compensation from the safe manufacturer? But if a wallet’s own security flaws lead to a huge amount of assets stored in it being stolen, who could accept the wallet company saying it can pay out no more than $100? 🤡🤡

#xrp账本修复可增发xrp的漏洞 If I remember correctly, something similar happened with ZEC a few months ago, and it plunged by 50%+ that same day, right? Why isn't the price of XRP reacting much??

Do you have to use your hands? Can't you use your feet?

Siba sis is so good at coming up with antics hahaha

#Manus完成超5亿美元新一轮融资 Manus has actually made a comeback?? And raised $500 million?

After a dramatic series of events—being acquired by Meta, the deal being abruptly halted, and then called off—Manus didn’t go under. Perhaps years from now, looking back, it’ll seem that not being acquired by Meta was the right call.

I went back to my hometown for a visit during the National Day holiday. In the county town, nobody talked about AI, large language models, U.S. stocks, chips, or embodied intelligence. The timeless topics were still the everyday affairs of family life.

The world seems just as it always was 🤫🤫

Muse’s 13th day of hype: Meta up 11%, but don’t believe the “AI new entry point” story yet

Ten years ago, Zuckerberg demonstrated a Messenger bot at F8 that would arrange flower deliveries, saying, “Nobody likes calling merchants.” Ten years later, the same company rolled out its personal Agent, Muse. After just 13 days, it topped the U.S. free charts in both categories—while reaching over 2.5 million downloads.

Meta’s narrative has managed to form a circle: in 2023, it saved money during its “efficiency year” and grew nearly 200%; in 2026, capital expenditures were pushed to 130 billion yuan, and Q2 free cash flow fell to just 784 million yuan. Yet Muse still surged another 11% in a single day. The market doesn’t care whether you spend money or save money—it only cares whether the story can close the loop.

But this is the fifth time a tech giant has tried to seize the “super entry point” with AI—each of the first four attempts stalled at functionality and never reached the entry point. Alexa’s voice shopping was tried by only 2% of users, with 90% not repurchasing; Messenger bots turned into customer service; the GPT Store and Humane AI Pin both failed; ChatGPT Instant Checkout drew criticism from Walmart, with a conversion rate only one-third of what it could drive back to the official site; during the Spring Festival, Qianshu’s “30 billion free orders” saw DAU jump to 73.52 million before dropping back to 30 million. Each time the market paid for the excitement, and each time the ending was the same.

The core contradiction of an Agent: eliminate time, but want to keep users

There are only two ways to monetize internet entry points: recommendations sell ad exposure by “killing time,” while search “delivers” users by turning them over to merchants. Personal Agents stand on neither side. It does the job for you and gets you out as quickly as possible—that’s “killing time.” But it also wants to hold onto users. What it passes to merchants is just an order: you can see the account address, but not the person who browsed the store—there’s no browsing, no lingering, no repurchase. For the next order, users re-compare prices; winning again this time doesn’t translate into a long-term advantage.

Merchants won’t pay a premium for an invisible customer. And an Agent’s frontend has real marginal costs: inference, virtual machines, rendering, retries—all cost money. Meta, however, wants to dip into both ends—charging users subscriptions while also taking a cut from transactions.

The math is brutal: Morgan Stanley modeled that by 2028 Muse would have 100 million users, 5 queries per day, 10% with commercial value, and 0.07 dollars monetized per transaction—13 hundred million dollars a year. That would contribute only about 1% to Meta’s earnings per share. Meanwhile, Meta’s Q2 ad revenue annualized is 237 billion.

Choosing sides is a craft

Consumer or merchant? What’s worth buying survives on the seller’s money. Taoyao was blocked by JD; Google Shopping moved from free price comparison to paid listings. On September 20, Amazon banned Muse, citing plainly that third-party applications that place orders across stores on customers’ behalf should disclose how they operate and respect the service providers’ willingness. Once the Agent shifts to charging merchants for lead generation, it stops being a buyer proxy that helps users save money and becomes yet another advertising channel. Zuck says Muse should “help you make money and help you save money,” but these two goals are inherently in conflict.

Low barriers to entry—shockingly low

The MCP protocol makes capabilities like connecting to Gmail, calendars, and maps quickly commoditized. Everyone gets a cloud virtual machine plus a visible browser—this architecture isn’t unique to Meta. Google’s Gemini Spark came more than three months before Muse. If you truly can run this playbook, then it should be Google or OpenAI/ Doubao—those holding Gmail/Chrome/Android—that come next. Even better, Muse’s most praised use case fits perfectly on Google’s email—adjust the pricing sheet and it can go after the same demand. What it wins is only a head start for a new product: in the same 13-day window, ChatGPT was downloaded 3.1 million times—still more than Muse.

The situation is similar in China: Qianshu connects with Taobao, but it lost 13.8 billion yuan on AI in a single quarter, backfiring on its own recommendation-stream ads; WeChat’s Xiaowei stayed restrained in its gray area because Tencent makes money from video-account ads; Meituan opened interfaces but kept tolls at the checkout counter; Doubao hit walls twice. The quietest is Pinduoduo—because it relies most on “browsing,” it’s hit the hardest.

The shelf life of an entry point

To become an entry point, you need at least four things: generating demand, taking payment, retaining users, and staying neutral. Right now, personal Agents can at most capture one or two. Every minute of attention it saves is an asset that internet companies could otherwise sell. Every user it keeps is exactly the kind of person merchants would gladly pay to get.

Morgan Stanley analyst Brian Nowak said he only cares about one thing: whether users are actually spending money to buy things on Muse. He doesn’t care about downloads or DAU.

For years, the market has loved to revalue tech companies by how hard they spend. But among all those entry-point narratives that never held up, what’s truly scarce was never the courage to spend—it was the restraint to understand “what money should never be spent.”

Last night, Meta’s stock surged by more than 11% in a single day. Intel and AMD rose by about 12% and 10%, respectively, mainly because Meta’s consumer AI agent product, Muse, has continued to heat up. This is worth discussing because it may be rewriting the underlying logic of “AI compute.”

1. What exactly did Muse change?

Most consumer-level AI on the market is still, in essence, a Q&A tool: you input some text, the model reasons and returns an answer, and the task is over. Muse pulls AI out of the chat window and into a real work workflow. You set a goal, it breaks the job into steps—signs in, visits websites, calls external services, reads page status—and then decides what to do next. When it encounters actions like payments, sending emails, or rescheduling, it proactively asks for your authorization, completes the task, and then checks whether the job is finished.

The closed loop is: understand the goal → formulate a plan → call tools → observe results → adjust the plan → continue execution.

In traditional AI, the “work unit” is a single “conversation.” In an agent, the “work unit” is a complete “task.” Meta also provides a cloud-based persistent execution environment: even if you close the app, it keeps running in the background. This means competition is no longer only about the base model—it also involves distribution entry points, user data, and the business ecosystem. Meta controls vast user entry channels. Once Muse has accumulated emails, calendars, spending habits, and task history, switching costs become enormously high—not just moving chat history, but moving an entire way of working.

2. Why did this pull focus onto CPUs?

For the past few years, GPUs have been the main battlefield for AI infrastructure. But once agents enter real workflows, the compute structure changes: GPUs handle the “thinking” part—understanding the goal, generating plans, and interpreting page content. The “doing” part—running systems, starting virtual machines, calling APIs, managing permissions, and tracking task state—gets pushed onto CPUs and traditional servers.

A chat model’s pipeline is: input → infer → output → release resources. An agent’s pipeline is: infer → CPU execution → wait for external responses → re-read state → infer again → execute again. A single long task may loop dozens or even hundreds of times. Browsers, virtual machines, memory, and task state need to stay online continuously. AI compute demand expands from “generating tokens” to “the full system resources required to complete a task.”

Once Muse is opened to billions of users, even if only a small portion runs agents simultaneously, it will generate massive concurrency: virtual machines, browser processes, and API calls. Consumer agents need to access email, payments, and account permissions. They can’t cram everyone’s tasks into the same environment. The more emphasis on secure isolation, the more independent resources must be allocated to each user. CPU demand in the agent era comes from three lines: increased concurrency, longer task durations, and tighter security isolation. This is also why the market is once again focusing on multi-core server CPUs. Some even predict CPU core demand could double, and the CPU-to-GPU ratio could approach 1:1.

One-sentence summary of the observation framework: when inference costs fall, the AI bottleneck will shift from “can the model think?” to “can the backend execute with sufficiently low-cost concurrency?”

3. How does compute demand translate into profits?

Demand doesn’t automatically equal profit—the key is Muse’s unit economics. With free and paid tiers at $20 and $100 subscriptions, it can also monetize later through enterprise agents, developer APIs, and transaction commissions. Muse can identify real purchase intent, so the theoretical value per user may be higher than traditional ad-click revenue. Some organizations estimate that future revenue could exceed $10 billion (one hundred-plus billion USD is implied).

However, agent cost structures are different: you must run virtual machines for users, occupy memory, burn network bandwidth, and call models. The harder you use it, the higher the backend costs. The two subscription tiers are essentially also about resource control—limiting task volume, limiting the number of calls, and limiting background runtime. Compared with fixed subscriptions, transaction commissions align more closely with the cost structure of agents.

But a tradable model can also spark platform conflicts. Amazon has already cut off Muse’s ability to shop inside its app. Once an agent searches and places orders on the user’s behalf, users skip recommendation pages and ads. Third-party platforms will likely go down one of three routes: charge for open interfaces, require transaction revenue sharing, or block unauthorized access. Any path will compress Muse’s service boundaries and profit margins.

4. How to understand this entire transmission chain

For Meta: Muse is the litmus test of whether AI capital expenditures can be converted into consumer revenue. Downloads prove distribution; completion rate determines value; retention and paid conversion determine revenue; unit task cost determines profit.

For Intel: GPUs handle thinking, CPUs handle execution. If consumer agents roll out at scale, server CPU may evolve from a supporting component into a critical factor determining concurrency scale and running costs.

The full transmission chain is: consumer agents get the user → tasks shift from Q&A to continuous execution → virtual machines, browsers, and tool calls surge → demand for CPU, memory, and networking rises → infrastructure costs increase → subscription and transaction revenue validate unit economics → chip demand ultimately turns into supplier profits.

In short: GPUs determine whether an agent can think things through clearly; CPUs and system infrastructure determine whether it can get enough people’s jobs done at the same time. And the final capital return depends on how much money is earned per completed task—whether it can cover the costs of the entire execution chain.

Two weeks ago, they were still shouting “industry-wide slowdown.” Seven days later, they were already considering launching a new model ahead of schedule—Anthropic’s move. Today, I’ll break down the AI landscape shift behind it.

The spark: Astra gains a share reversal in two weeks

OpenRouter’s data is striking. OpenAI’s GPT-6 Astra went live and, in just two weeks, captured about 13% of enterprise AI spending share, while Anthropic’s Fable series is at only 8%. Even more critical is the crossover curve from the first week of September: OpenAI’s share jumped to about 65%, while Anthropic fell to about 35%. To put it in context, from January to June this year, Anthropic had still been holding steady above 50%. This overtake—after two and a half years—took only two weeks.

The whole process that throws the rhythm off

On September 1, Anthropic struck first: it released Fable 5.1, positioning it as the “strongest programming and knowledge work model,” cutting the cached-read pricing by 75% to hold onto its enterprise base with a cost-performance advantage. Two days later, on September 3, OpenAI rolled out Astra. Altman’s line—“hand Astra to everyone as soon as possible”—was almost a direct face-to-face challenge.

But what truly made Anthropic nervous wasn’t Astra itself—it was the fact that its own flagship couldn’t sell. After Fable 5 launched, two months in, the enterprise upgrade rate for the top-priced flagship version stayed consistently below expectations. Market share hovered around 11% for a long time, and users preferred buying the more cost-effective Claude Opus 5. In other words, Anthropic’s previously “good enough” market position was being shaken at the root by Astra.

“Slowdown narrative” collides with commercial reality

On September 12, Anthropic CEO Dario Amodei published a long post urging AI development to slow down, and Altman and Musk even briefly echoed the sentiment. Then, only seven days later, Reuters broke the news: Anthropic internally was discussing launching a new model ahead of schedule, and Claude assistants had also been deeply involved in the development of the next generation. On one side, the “responsible AI” banner; on the other, enterprise customers casting votes with real money. When these collide, the one that always loses is the narrative.

A reshuffling of priorities under the IPO window

The deeper pressure is the IPO. Reuters revealed that Anthropic’s listing timeline may be pushed to November. This fire burns two directions at once: share is being eroded quickly, and valuations need the growth curve to justify them. Right before listing, investors don’t care whether you shouted “responsible”—they care whether revenue growth and enterprise market share can hold steady. So “launching the model early” is almost certainly a business decision—despite the fact that it means, within two weeks, having to reassess the “accelerating race” logic they had publicly criticized.

One last thing to say:

Astra grabbed 13% of share in two weeks. Its significance isn’t that OpenAI won one round—it’s that it proves something: under the current competitive landscape, “calling for slowdown” is already a luxury PR line. When the gap between leading models can translate into a double-digit migration of market share within a one- or two-week time difference, the industry’s release cadence gets locked into an acceleration spiral—whoever dares to move slowly, yields.

This isn’t Anthropic’s predicament; it’s the next rule in the global AI arms race.