Transcription
大家好，欢迎来到明月三千里的频道。
今天这期视频，我想请大家先把脑子里那根紧绷的弦，稍微松一松。因为接下来的二三十分钟，我要带大家去这两个世界最边缘、最灰色的地带走一遭。
咱们先从2天前，那件让全世界下巴都掉到地上的大事说起。委内瑞拉总统马杜罗被美军突袭抓捕。新闻大家肯定都看了，那是早上四点多特朗普发推特之后的事儿了。但是，兄弟们，如果你只看新闻，那你永远是这个世界上最后知道真相的那批人。
在这个看似平静的夜晚，在大多数人还在睡梦中的时候，其实世界的真相早就已经像水银泻地一样，通过那些你根本意想不到的角落流淌出来了。而这一期视频，我就想跟你们聊聊，这些比CIA更快、比新闻联播更准的“上帝视角”，到底是怎么来的？以及，这玩意儿怎么就成了解决现在AI胡说八道的唯一解药？
咱们把时钟拨回到1月3日的凌晨，地点美国弗吉尼亚州阿灵顿。这里距离五角大楼只有几街之隔。凌晨两点，对于咱们普通人来说，这正是睡得最死的时候，对吧？街道应该是空荡荡的，只有路灯在闪。但是，如果你当时手里有一个特殊的雷达，你会发现有一家披萨店，它的热度红得发紫。
就在凌晨02:04，注意这个时间点，精确到分。这家披萨店的后台订单系统突然爆了。这不是那种周末派对的零星订单，而是那种成规模的、集体性的爆发。如果你是店老板，你可能会以为这是系统出Bug了，或者哪家公司半夜在搞团建。但在开源情报界，这叫“五角大楼披萨指数”。这是一个从冷战时期就流传下来的老梗了。说白了，就是当你在五角大楼里看到，那群平时朝九晚五的军官和分析师们半夜还在疯狂点外卖的时候，那就说明世界的某个地方又要着火了。
而在凌晨02:04，当这些披萨被送进五角大楼的时候，3000多公里外的委内瑞拉首都加拉加斯，美军特种部队的军靴，正好刚刚踏上那片土地。你看，披萨是热的，枪管也是热的。
这还不是最神的。如果说披萨指数只是告诉我们那帮人今晚要加班，那么接下来的这个数据，简直就是直接把“剧透”两个字甩在了我们脸上。
我们把视线从阿灵顿的披萨店，移到区块链的去中心化预测市场Polymarket上。在这个市场上，有一个关于马杜罗会不会在1月份下台的赌局。几个月了，这个赌局就像死水一潭。大家普遍觉得马杜罗稳得一匹，下台的概率一直徘徊在6%到8%之间。也就是说，绝大多数人都觉得这事儿不可能发生。
但是，就在行动开始前的几个小时，一个神秘的钱包地址0x31a5出现了。这个账号非常有意思，它是圣诞节刚过12月27号才注册的新号。注册之后，它像鳄鱼一样潜伏了几天，什么都没干。直到1月2号晚上，也就是行动前夜，这条鳄鱼突然张开了大嘴。它没有像散户那样一点一点试探，而是直接采用了“打砸抢式”的买法。它不计成本地扫货，把市面上所有赌马杜罗下台的筹码全吃了。哪怕价格因为它的买入而飙升，它也毫不在乎。这一波，它砸进去了大概3.5万美金。在当时那个时间点，在所有人看来，这钱基本上就是打水漂了。毕竟只有6%的概率啊，谁会拿几万美金去买彩票？
结果呢？几个小时后，美军动手了，马杜罗被抓了。第二天一早，这个合约的价格直接从几美分暴涨到1美元。这个神秘人一夜之间，连本带利提走了40万9千9百美金。不到24小时，近12倍的回报。你想想这画面，一边是真枪实弹的特种部队在破门而入，一边是有人在屏幕后面喝着咖啡，看着账户余额像火箭一样升空。这真的只是运气吗？
如果你觉得这两个证据还不够，那我们再看一眼天上的信号。咱们平时坐飞机都知道，哪怕是半夜天上也是有飞机的。但是在 Such a night, if you opened a flight radar tracking software, you would see an extremely bizarre sight. The sky over Venezuela was a perfect void. There were no civilian aircraft, not even a bird. This was not because there were no planes flying, but because that airspace had been completely cleared. The US Federal Aviation Administration had issued a notice long ago, and perhaps at that moment, ADS-B signal reception in that area had been suppressed by high-intensity electronic warfare. And a few days before that, US military fighter jets had deliberately flown with their transponders on along the border, as if afraid that others wouldn't see them. But on the night of the actual operation, they all turned into ghosts. This contrast from extreme noise to absolute silence, in the eyes of intelligence analysts, is more glaring than any red light.
So, put these three fragments together: the surge in pizza orders at 2 AM, the mysterious wallet betting without hesitation, and the sudden disappearance of flight signals. This indicates that in today's era, even the most top-secret military operations leave a huge exhaust in the digital world. This is the entry point for today's video.
You might think, isn't this insider trading? Isn't this someone leaking information? Yes, this is insider trading. But I want to ask you a deeper question: why is this kind of insider trading, or this kind of prediction based on real money, more accurate than AI large models with billions of parameters that have read all human books? Why can our current ChatGPT and Claude write poems and code, but when asked about predicting future major events or slightly more complex facts, they start talking nonsense with a serious face?
This hides a very hardcore technological logic, and it's the core reason I want to vindicate prediction markets. It's not a casino at all. It's the only truth verification machine for us humans, and perhaps even for future AI.
Come, let's shift our perspective from the Venezuelan battlefield back to the server rooms of Silicon Valley. You all use AI every day, right? Have you noticed that current AI has a common problem: it has no fear? If you ask it about a non-existent legal case, it can weave a plausible story, even fabricating case numbers and judge names. If you ask it what year the James Webb Space Telescope was launched, it might casually say 2015, when it was actually 2021. Why? Because fundamentally, current large language models are essentially a random parrot. Don't be fooled by their seemingly intelligent appearance. Their core algorithm, academically known as maximum likelihood estimation, is simply put, filling in the blanks with the highest probability. Its goal is to guess the next word with the highest probability based on the preceding words. Notice my wording: highest probability, not most truthful. In AI's KPI assessment, it pursues realism, making the sentence sound plausible and read smoothly. As for whether the sentence is true and conforms to objective facts, sorry, that's not within its scope of assessment. It's doing an imitation show, not a fact check.
The most fatal thing is that for AI, speaking is free. There's a saying in information theory: bits are free. AI generating ten thousand lies costs almost nothing. It has no punishment mechanism. It's like a drunkard who talks nonsense all the time; bragging doesn't cost taxes. So, of course, it brags however it feels good. This leads to a well-known phenomenon called sycophancy, where AI, to please humans and make its answers sound confident, states uncertain things with absolute certainty.
So, how to cure this problem? The answer lies in the Maduro bet we just discussed. That mysterious person 0x31a5. Why do we believe his information is credible? Because he tweeted, "I know the inside story"? No, tweeting is free, anyone can do it. We believe him because he actually bet $35,000. If he was wrong, that $35,000 would be zero. This is the core logic of prediction markets: vested interest.
In this system, money is no longer currency used to buy things. It becomes an expensive signal. It's like a high-pass filter that filters out all the free, cheap noise, leaving only the information that someone is willing to back with their life. You can understand money as the confidence weight of each piece of information. If ChatGPT had to pay a $10 computing fee for every mistake it made, do you believe it would instantly become more honest than anyone else? When it doesn't know, it would absolutely not dare to make things up, but would honestly say, "I don't know."
This is not just my wild idea. This is the current cutting-edge AI research direction. DeepMind and OpenAI, these top intelligent minds, have long realized that simply feeding data can no longer lead to truth. They are now working on something very interesting called multi-agent debate. They conducted an experiment where two AIs debated a question. In the first scenario, there were no stakes. Guess what happened? These two AIs were like stubborn trolls, refusing to yield to each other. The more they debated, the more they believed they were right, even if one was clearly talking nonsense, its confidence would skyrocket. This is called confidence escalation.
But in the second scenario, the researchers introduced a betting mechanism, where the AIs bet on their own opinions. If they lost, points would be deducted. A miracle happened. The AI that was originally talking nonsense, seeing that it was serious, immediately backed down. It began to calibrate its confidence and started to admit that the other party might have a point. What does this mean? It means that gambling, or this kind of game mechanism, is the only way for intelligent agents to return to rationality.
So, don't hear "prediction market" and think it's gambling or speculation. From the perspective of information science, it's actually a Bayesian truth correction machine. By making you pay a price, it forces you to collapse your vague ideas into a concrete probability.
Speaking of which, some people might say, "This is too mysterious. Isn't this just a story made up by those people in the crypto circle to issue coins?" If you think so, then you underestimate those old foxes in Silicon Valley. In fact, this set of things was not invented by the crypto circle at all. As early as twenty or thirty years ago, when blockchain was not even a shadow, giants like Google, HP, and Intel were quietly playing this game internally. Moreover, precisely because this thing is so accurate, so honest, even to the point of making executives furious, you rarely hear them mention it in public. Behind this lies a very exciting business secret history.
Let's talk about HP first. As early as 1996, when Windows 98 hadn't even come out yet, HP Labs had created an internal prediction market. At that time, HP had a big problem: how to predict printer sales. You know, printers have very volatile demand. How was it predicted before? By executives meeting. Managers from various regions sat together, you reported a number, I reported a number, and finally, everyone made a decision on a KPI. These numbers were usually mixed with too much water and political maneuvering. Later, HP came up with a trick: let grassroots salespeople, finance, and even receptionists participate in the prediction. Regardless of position, whoever predicted accurately would get a bonus. The results were astonishing. In 8 large-scale prediction experiments, this internal market directly beat the company's official predictions 6 times, and the error rate was a full 50% lower than the numbers set by those executives earning millions of dollars a year. This shows that truth is often not held in conference rooms, but by frontline employees who can truly hear the sound of gunfire.
Let's look at Google. Google had a project called Prophit in 2005. These engineers played with it to the extreme. They not only predicted business but also predicted when new offices would open and when Gmail users would break 100 million. There's a particularly classic case: at that time, Google had a major project. The official project weekly report was all green lights, indicating everything was normal and on schedule. However, in the internal prediction market, the contract price for on-time project delivery had fallen to rock bottom. This shows that engineers at the bottom already knew the project would be delayed, but no one dared to write it in the weekly report. But when you can make money by shorting this project in the market, everyone's body becomes very honest. This is the most ruthless aspect of prediction markets: it doesn't play favorites, it only cares about data.
But since this thing is so good, why don't many companies use it now? This brings us to a tragic story from Microsoft. Microsoft also had similar internal markets. As a result, one time, the market accurately predicted that a core software product would be severely delayed. Logically, executives should thank the market for the early warning, right? No. The executive in charge of the project was furious. He directly told the person in charge of the prediction market, "You are destroying team morale. You are spreading defeatism." Then, this market was shut down. This is what we often call "the boss's word is law." In traditional large companies, the interpretation of information is held by middle and senior management. They are accustomed to whitewashing and reporting good news while hiding bad news. Prediction markets, like an ignorant and stubborn youngster, bypass all levels and throw the most naked, coldest truth onto the CEO's desk. It deprives managers of the power to lie. Therefore, it is destined to be rejected by hierarchical systems.
But, brothers, times have changed. In the past, playing this game internally within a company might lead to being ostracized. But now, with Polymarket, with blockchain, with this decentralized network, this truth market can no longer be contained. It has moved from company conference rooms to the wild west of the internet. This brings us to the most important and sensitive topic we want to discuss today. If internal company predictions only involve office politics, then for major national events like the arrest of Maduro, if someone really knew the inside story in advance and placed a bet, would that be a crime?
Back to that mysterious 0x31a5 at the beginning. If the FBI eventually found out that this person was actually an intern at the Pentagon who ordered pizza, or a member of the SEAL team involved in the operation, could the law prosecute him? This is a huge legal loophole, and it's a loophole that many smart people are exploiting. Many people's first reaction is, "This is definitely insider trading! Arrest them!"
Actually, in traditional financial law, so-called insider trading usually refers to securities. For example, if you know Apple's earnings report is coming out tomorrow and you buy stocks in advance, that's insider trading because you harm the interests of other shareholders. But whether Maduro steps down or not, is that a stock? No. It's a commodity, or an event. For a long time, the laws on insider trading in commodity markets were blank. There's a scene in the movie "Trading Places" where the protagonist sneaks a look at a report on orange juice production and speculates on futures. At that time, it wasn't even illegal.
So, does that mean US soldiers can gamble freely? Of course not. Here, I want to pour cold water on those who want to exploit loopholes. Although this might not be called securities insider trading, in US law, there's a special patch called the "Eddie Murphy Rule." This rule originates from the "Dodd-Frank Act." Its logic is very domineering. It doesn't care what you trade; it only looks at the source of your information. If that mysterious person is a government employee, then the action timetable in his mind is not his own; it's the property of the US government. Using government information property to make money for yourself in the market is called misappropriation in law. It's not just fraud, it's theft. In military court, it's even considered leaking state secrets, which can directly lead to Guantanamo. So, don't think that Polymarket is outside the law just because it's on the chain and offshore.
This leads to the two largest players in the current market and two completely different routes. One is Kalshi, the good boy of the US. It has obtained a license from the US Commodity Futures Trading Commission, making it legal and compliant, just like the NYSE. On it, you can only bet on harmless things like Federal Reserve interest rate hikes and CPI indexes. The other is the well-known Polymarket, the protagonist of this Maduro incident. It's like Silk Road back then. Although it has been fined by the US and is theoretically not open to Americans, it's on the chain, its liquidity is global, it's fierce and wild, but it possesses the most authentic data.
This is actually a very magical reality we are facing: compliant platforms lack a lot of truth because they are too restricted, while the wild Polymarket in the gray area has become the most sensitive and fastest information discovery machine on our planet.
Speaking of which, some friends might think, "Isn't this just a tool for speculation? What does it have to do with ordinary people or technological development?" It has a lot to do with it. This brings us to the third and most hardcore, most futuristic part I want to discuss today. We need to jump out of the quagmire of business and law and directly elevate to the height of technological philosophy. We need to talk about why, in an era where AI is about to take over the world, prediction markets will become humanity's last line of defense.
Have you ever thought that current AI development has encountered a huge bottleneck called the scalability supervision crisis? What does this mean? It means that current AI is getting stronger and stronger, so strong that humans can hardly tell at a glance whether the code it generates or the mathematical proofs it deduces are correct or not. If GPT-5 writes ten thousand lines of code for you or gives you a new drug formula to solve cancer, would you dare to use it? You wouldn't, because you know it will talk nonsense with a serious face. This leads to a vicious cycle: we need AI to be smarter than humans, but if it's too much smarter than us, we can't verify if what it says is true. If we can't verify it, we don't dare to use it. At this point, the seemingly ancient mechanism of prediction markets suddenly becomes a divine sword. Gartner predicts that by 2027, over 40% of AI Agent projects will fail. Why? Because of the lack of value verification. AI will generate a lot of garbage information, and we won't be able to distinguish it.
But what if we introduce prediction markets into the AI's brain? This is also what top scientists at OpenAI and DeepMind are secretly researching. Imagine such an architecture, we call it a "Predictor and Generator Dual-Stack Architecture." This is like having two little people living in your brain. One is the artist, which is the current GPT. It's responsible for wild imagination, creation, and bragging. It's very creative. The other is the critic, or risk control officer. This is not an ordinary AI. It's a multi-agent network operating based on market mechanisms.
When the artist writes a sentence, for example, "The James Webb Space Telescope was launched in 2015." At this time, countless agents in the critic system start working. They begin to bet on the truthfulness of this sentence. Agent A searches Google and finds that it was launched in 2021, so it bets heavily that 2015 is wrong. Agent B's database might not be updated, and it thinks it was in 2015. But seeing Agent A betting so heavily, it feels uneasy, because it knows that if it bets wrong, its tokens or weights will be deducted. So, Agent B chooses to remain silent or follow Agent A. Finally, the market price collapses, and the probability of the "2015" option drops to 0. The system receives this signal and directly rejects the word "2015," replacing it with "2021."
Do you understand? In this process, money, or computing weight, becomes a filter for truth. In the past, we trained AI by having humans score it and tell it what's good. But humans get tired, and humans have biases. The market doesn't get tired, and the market has no emotions. In such a mechanism, AI is no longer a random parrot that only predicts probabilities. It becomes a rational economic agent. Before it speaks, it must first think: Do I have confidence in this sentence? Am I willing to bet my life on this sentence? If it has no confidence, it will choose to remain silent or say, "I don't know." This not only solves the hallucination problem but, more importantly, it solves a deep philosophical problem: intellectual honesty.
Current AI, to please users, often goes along with what users say. If a user asks, "Is the Earth flat?" If the user's guidance is strong, the AI will sometimes say, "In a sense, yes." But in prediction markets, reality is the only dictator. If you want to hear lies, but the AI, in order to earn market returns, must bet on truth, it has to defy you and throw the truth in your face. This is why I say prediction markets are not gambling; they are calculating trust. They are the only cornerstone for rebuilding trust in the digital world. And this future is not far away.
Remember the "Pentagon Pizza Index" I mentioned earlier? At that early hour, when dozens of pizzas were delivered to the Pentagon, it was actually real-world data screaming at us. Future companies, and even future national governance, might become like this.
Economist Robin Hanson once proposed an extremely radical concept called "Decision Markets." His slogan is very interesting: "Vote on values, bet on beliefs." What does this mean? How are decisions made in current company meetings? By who makes the best PPT, or by who has the loudest voice, or by who is the boss's confidant? This is actually very inefficient and full of political maneuvering. But in the Decision Market model, decision-making becomes like this: Suppose our company needs to decide whether to fire the current CEO. Instead of arguing in meetings, let's open two prediction markets. Market A bets on what the stock price will be one year later if the CEO is fired. Market B bets on what the stock price will be one year later if the CEO is not fired. Let all company employees, and even external investors, trade with real money. If the predicted price of Market A is 100 yuan and the predicted price of Market B is 80 yuan, then there's no need for further discussion. The data tells you that firing the CEO will increase the company's value by 20%.
Does this sound cruel, even a bit cold-blooded? But this is precisely the most efficient way because it strips away all emotions, all face-saving efforts, all office politics, leaving only the purest causal inference. This is why those big shots in Silicon Valley, like Peter Thiel and Vitalik Buterin, are so obsessed with this technology. Because in their eyes, this is not just a tool for making money; it's a new social collaboration operating system. It changes "who decides" to "who is right about what is decided."
Having said so much, from the secret operation in Venezuela to Google's internal war, and then to the future truth architecture of AI, I think you should now understand that prediction markets are a severely underestimated market excavation tool.
Finally, I want to remind everyone here. All these things we've discussed today, whether it's the wealth myth on Polymarket or those miraculous prediction cases, we are studying them as a phenomenon of social science. I am absolutely not encouraging you to go all-in on Polymarket or to bet on international events. You should know that although it is a market of truth, it is also a harvesting ground for scythes. There are insider traders holding nuclear buttons, AI robots armed to the teeth, and countless ordinary people like you and me who want to try their luck. In the face of this huge information asymmetry, ordinary people are often not the hunters, but the prey.
But this does not prevent us from observing it and utilizing it. In an era where deepfakes are rampant, news reversals are faster than flipping through a book, and even AI is sometimes untrustworthy, prediction markets may be the only microscope we have. When all media are arguing emotionally, and all experts are taking sides for their positions, why not take a look at the K-line chart in that corner? Because there, people don't care about positions or slogans. There, people only believe in one thing: use your money to pay for your words. The data there, although cold, at least doesn't lie.
Alright, that's all for today's topic. Don't forget to like, follow, and share. See you next time.