Alpha哥AI投资日记

New iPhone Duo is suitable for every NASDAQ YOLO investor and crypto all-in gamblers

#iPhoneDuo

I think Zuckerberg’s idea is quite right: personal AI assistants for ordinary people should be free, and only heavy users should pay.

As tokens get cheaper and cheaper, the cost for regular users to chat, ask questions, and write a bit of content will keep dropping. Charging a fixed $20 per month is completely unnecessary.

I think it’s more reasonable to charge for high-intensity use cases—like writing lots of code, running complex tasks, or having an agent work all day—because those consume a lot of resources.

And when it comes to commercialization, I’m more bullish on Meta. It already has so many merchant resources and a mature advertising system. In the future, AI can help users find products, book hotels, and choose services, and there’s an opportunity to earn money from facilitated transactions. Merchants are willing to pay for customers and orders, so ordinary users can continue to use it for free.

As AI capabilities gradually approach parity, competition will increasingly test who can turn user needs into a business. Meta has built up in this area for so many years—I think it’s entirely possible for them to do better serving C-end users than Anthropic and OpenAI.

Reference: https://x.com/i/web/status/2097705128606462078

$META just launched its personal AI assistant Muse, offering up to 100 million tokens free every week—this quota is pretty aggressive

Following Zuckerberg’s playbook, ordinary people really shouldn’t need to pay for AI. Spending $20 a month to subscribe to ChatGPT or Claude’s basic tier shouldn’t be the default way to use AI; everyday needs of regular users should be met for free

AI companies can charge heavy users, or they can make money by charging fees and taking a cut through helping users shop, book services, and run businesses. Give it away for free first to win users and build usage habits, then build a business around that entry point. Once free AI is already good enough, OpenAI and Anthropic will have to prove their value again every month with that $20

Today $META is up 7%, it seems the market is still buying Zuckerberg’s promise

A magical world, a meme posted by Biden’s son $LAPTOP market cap (of course it’s fake) once exceeded Solana: Es9vMFrzaCERmJfrF4H2FYD4KCoNkY11McCe8BenwNYB market cap https://twitter.com/nicksolrik/status/2097658042838077772

Cow incoming! 🫡 https://twitter.com/aleabitoreddit/status/2097601244898775465

Garrett Jin currently has about 39,760 ZEC:native short positions, with an average price of $576, and strong liquidation prices around $2,292 under Cross Margin (cross margin / full-position margin mode). He also has BTC long positions worth $100 million.

If ZEC really had a main force trying to mess with him, the most ruthless play would actually be not to directly pump ZEC, but to wait for BTC to pull back a bit, then hard-stabilize ZEC and push it upward against the trend.

For example, when BTC dips, $ZEC instead keeps running it up all the way to $1,700–1,800. Then it would make him feel terrible: his ZEC shorts would be floating in losses growing to around $45–$49 million, while his BTC longs would also be losing money—both sides would drain margin together.

At that point, even if it hasn’t reached the displayed liquidation price yet, he might still have to proactively cut his ZEC shorts. And a big short near a $50M position being forced to cover is, by itself, fuel for the next leg up.

If I were the market maker for ZEC, seeing such a big fish hanging on-chain, I might genuinely be tempted to give it a try 😂

Source: https://x.com/i/web/status/2096866422559252609

For the Democratic Party, Jon Ossoff may be more suitable than AOC for 2028

AOC’s strengths are that she has strong appeal among young people, the left-wing base, and on social media. But the problem is that her label is too far left. Issues like Medicare for All, immigration, and policing are very easy for Republicans to target and attack heavily in a general election

Ossoff’s path is much smarter. In essence, he is still a center-left Democrat, but on issues like healthcare, energy, public safety, and immigration he is more moderate than AOC. He is somewhat like Obama back then, a young technocrat-style candidate. His biggest asset is his real-world electoral success in a swing state like Georgia, which is much more convincing than winning 70% in a deep-blue New York district

Recently, Ossoff’s momentum has also been very strong. Polymarket has already risen to 12%. If he can win re-election to the Senate again in Georgia this year, I feel he may quickly become the top choice for 2028

Pure analysis, does not represent my political stance

Source: https://x.com/i/web/status/2096791552953897316

zcash:native Weekend crushes the shorts

AI short sellers should be careful about one thing: once any sufficiently large interest group takes shape, it does not break apart easily. Especially now, AI is no longer just the business of a few startups. OpenAI and Anthropic have raised tens of billions of dollars. Nvidia, cloud providers, VCs, storage, and optical modules are all tied into this AI industrial chain. Every company in it has huge interests, so the whole system will naturally try to extend this cycle as much as possible, rather than simply watching it collapse because of a single factor, such as short-term revenue failing to materialize quickly enough.

Now token sales are becoming more and more competitive. As model capabilities converge, prices will only get cheaper. If ARR cannot be sustained, OpenAI and Anthropic could very well switch to another pricing model. If the game can no longer be played, then the rules of the game can be changed. Not just selling tokens, but charging by task, by outcome, or by how much human labor has been replaced. A job that originally needed a white-collar worker for one hour and that a company would be willing to pay $50 for, an AI system might only spend $1 in inference cost to complete, yet still charge you $5 or $10. In this way, even the way the market does the math changes, shifting from “how much can tokens still be sold for, and is ARR growth slowing?” to “how much of global white-collar wages can AI ultimately take?”

Changing the way the game is played can buy time for the entire industry. When the old business model starts to be questioned, a new charging method, new KPIs, and a new TAM calculation can be introduced. The market then has to reassess adoption, success rates, and how much companies are willing to pay. These things themselves then take another half year or a year to be disproven.

And for such a large interest group, half a year or a year is already very valuable. Because during that time, model capabilities can keep improving, inference costs can keep falling, and computer use, AI video, and all kinds of non-coding scenarios may suddenly emerge. As long as one or two real major breakthroughs appear in the meantime, the whole cycle can continue moving forward, and the party can keep going.

Once a huge industrial interest community has formed, it will constantly look for new tools, new business models, and new valuation methods to extend the cycle. China’s real estate market once kept going with shantytown redevelopment, credit expansion, and land finance.

If AI short sellers realize that everyone has changed the rules, they may need to cover their shorts first in order to survive.

Before the non-farm payrolls report, I saw Binance push me this notification, and I felt something was off, so I quickly reduced some positions 😂

“BTC rose in the first two non-farm reports, come and position early” — when exchanges start reminding retail traders to pre-trade the logic of “weak NFP = no rate hike = BTC up,” it basically means this expectation has probably already been fully priced in

Tonight we got a direct blow from the opposite direction: August non-farm payrolls came in at +162k, versus expectations of +56k, nearly 3 times the forecast; the unemployment rate was still 4.1%, and wage growth was also slightly above expectations. The previous two months of employment were revised up by a combined 55k, and July was revised straight from -23k to +21k

Judging from this, U.S. employment is still very strong. I guess we’ll keep oscillating for now, and then wait for next Friday’s CPI to decide whether there will be a rate hike

Recently, $HOOD has been rising very strongly, but I feel the market may not have fully priced in the value of Robinhood Chain yet. Using two of the simplest methods to value Robinhood Chain separately, the fair HOOD price implied by both approaches ends up around $140–160, with the midpoint right around $150.

First, use HYPE directly as a benchmark. HYPE currently has a market cap of about $19B and annualized revenue of roughly $710M, which means the market is valuing it at about 27x revenue. Robinhood Chain generated $11.2M in revenue over the past 7 days, which annualizes to about $580M; at the same 27x multiple, that would be worth about $16B. HOOD’s current market cap is about $112B, so adding back that $16B implies a share price of roughly $142.

Second, an even simpler method is to value it based on existing public-chain valuations. SOL currently has a market cap of about $61B. If Robinhood Chain can ultimately reach only half of SOL’s value, that would be about $30B. Adding that $30B to HOOD implies roughly $34 of value per share, giving a fair share price of $159.

One method gives $142 based on revenue, and the other gives $159 based on half of SOL, with the midpoint landing right around $150. HOOD at $125 likely already reflects part of the Robinhood Chain expectation. If Robinhood Chain can keep gaining traction for a while and also pick up some sentiment premium, then $HOOD $150 may actually be a fairly normal case.

zcash:native 1000 is here, the privacy edition starts from $BTC 🛫

I started buying zcash:native around last year—bought around 60, sold some at 600, then bought back 400 again.

The long march of privacy coins has just begun https://twitter.com/alphaguytrading/status/2090997661688352873

GPT-6 Astra has been released. The main improvement is that AI operating a computer is really starting to get close to “being able to get things done”

In OpenAI’s tests, for the same type of computer operation tasks, GPT-5.6 Sol took about 75 minutes, while Astra reduced that to around 40 minutes, and the completion rate was even higher; with the new Codex, the completion speed on another set of web tasks was nearly doubled directly. In the future, having AI open a browser by itself, fill out forms, work with Excel, operate CRM systems, look up information, and modify websites will start to shift from demos into truly usable productivity. (They’ve basically built in Manus’s core capabilities 😅)

Citation: https://x.com/i/web/status/2095595741528125780

The DeepSeek, GLM, and Kimi behind the transit station didn’t go down 😅 https://twitter.com/_FORAB/status/2095531562720973182

This so-called “Delta neutral” strategy actually has serious hidden risks. It’s a market maker trap—while the mantis stalks the cicada, the oriole comes after.

The losses from shorting and holding a position in a funding-rate futures contract are settled by the exchange at extreme prices in real time. On the other side, the profit from spot is only an accounting gain. For a small coin like AKE, the spot order book depth is terrible. If you have a few million dollars worth of spot, it’s simply impossible to sell it all near 0.04. By the time Binance’s short positions get liquidated and the price then falls from 0.045 back to 0.013, the floating profit on the spot side is gone too. If you also layer in Cross Margin, a sudden AKE pump could blow up the entire account margin, and the other dozens of originally independent arbitrage positions would be liquidated in a chain reaction.

Quote: https://x.com/i/web/status/2095371870258516372

Google Gemini’s world No.1 ranking only lasted 30 hours.

Google just released Gemini 3.8 Flash last night, reaching about 74% on DeepSWE and briefly taking the world No.1 spot for Coding Agents. DeepSWE means giving the model 100 fairly complex, real software development tasks; about 74 of them it can finally complete correctly and pass the tests independently. Not long after, Meta released Muse Spark 1.3 as well—its score reached 75.4%, directly pushing Google out of the top place.

The coding abilities of these cheap models from Google and Meta are increasingly getting close to the most expensive tier like Opus, but at a fraction of the price. After running a task on DeepSWE, Gemini 3.8 Flash averages only $2.36, while Opus 5 with similar capabilities costs $11.84. The programming ability of Google and Meta’s two models is now at roughly the same level as Claude Opus 5, but the average cost to finish a task is only about one-fifth of Opus.

Big models are really going all-in right now—coding intelligence is rapidly becoming commoditized.

Reference: https://x.com/i/web/status/2095232032896946311

Many technological revolutions first found their fastest adoption in industries like “porn, gambling, and drugs”—for example, in the early days the majority of internet traffic came from porn sites; darknet transactions often used BTC; and online gambling once grew faster than e-commerce.

When the internet first emerged, in 1999 adult websites accounted for only 2%–3% of commercial sites, yet they generated 10%–20% of searches on search engines. In 1998, just the top 10 adult sites reached 15% of all Web users at the time. In 1996, adult entertainment already made up 10% of total B2C e-commerce revenue.

Where porn was truly important to the early internet was that it was among the first internet applications to have massive demand—users were willing to pay—and at the same time it consumed bandwidth extremely heavily. Today we take for granted paid subscriptions, credit card payments, affiliate commission systems, and the transmission of high-definition video; adult sites were among the earliest to use these at large scale.

Now AI may be retracing the same path. When AI-generated adult short dramas can already be produced in bulk and people have started making money from them, that suggests the cost, speed, and controllability of AI video may have begun to cross the threshold of truly consumer-grade applications.

Everyone has been looking for new AI application scenarios besides coding. There’s a chance that the scenario might be right here...😃

With the development of AI technologies, combined with AR technologies like Apple Vision Pro, the world of the future may really become the Metaverse.

A world model can directly generate environments, a video model can generate scenes in real time, and an LLM can give every NPC its own memories, personality, and action logic. Add Blender/game-engine agents on top, and in the future it may even be that you can say something like: “Generate 1980s Tokyo Ginza, raining, with 500 people who can freely move.” It could come out in minutes—or even seconds.

You walk down the streets of Tokyo, and your glasses already know where you are, who you’re looking at, and what stores are next to you. Real-world buildings can be re-rendered. Virtual people appear along the way. Your friends’ avatars can sit next to you.

AI world model × AR × AI Agent—stacked together. If AR glasses become something ordinary people wear for 6–8 hours a day, it could be a revolution on the level of the iPhone.

Reference: https://x.com/i/web/status/2094884634719133925

Anthropic has released a new model again

The most outrageous part is the research work. Terminal-Bench-Science: Fable 5 is only 24.7%, while Fable 5.1 directly reaches 52.6%. Opus 5 is 29.0%, and GPT-5.6 Sol is 22.4%.

Coding has also improved significantly. Terminal-Bench 4.0 goes from 42.0% to 55.8%, and AutomationBench increases from 17.1% to 31.4%. Looking at just these numbers,

the model’s capabilities are rising, but the real-world usage cost is actually going down. The cache read price is 75% lower than Fable 5. Anthropic estimates that for a typical workload, the actual cost drops by about 25%.

Fable 5.1 is usable now—everyone can try it and see whether it’s that much stronger.

Source: https://x.com/i/web/status/2094848581425377479

Fireworks AI CEO Lin Qiao recently shared a video about "post-training" that I found quite good. I think there’s an important trend here: in the future, it may not be a few super large models that eat up all AI applications; instead, we may see millions of vertical models.

This change has already started. Cursor began using its own user data for post-training long ago, and updates its model every few months. Doximity trains its own clinical AI in the healthcare domain. Legal, finance, recruiting, sales, customer service—these fields are also starting to do similar things.

Today, the UI, features, and even workflows of an AI product can be copied by others in a matter of weeks. What can’t really be copied is the data accumulated after a company has served tens or hundreds of thousands of users, as well as that company’s judgment and understanding of what the “right answer” is for the industry. Post-training is about writing those judgments and intuitions directly into the model.

And there’s a very practical benefit too: Lin Qiao mentioned that some models that have undergone post-training can achieve performance close to—or even surpass—the strongest general-purpose models, while reducing inference costs by 5–10x.

In the future, once many vertical AI companies reach a certain scale, they won’t be satisfied with just continuously calling the APIs of OpenAI and Anthropic. They’ll use the strongest model to find product–market fit first; then, once they have the data and users, they’ll start training their own models.

Original video: Post-Training Is How You Keep Your Taste :https://youtu.be/yAvJ7b_FxUA?si=8IZ_nJiu_avnyBTM