預測市場只會下注？Seer 想拿它做決策

0:00預測市場已有規模，下一步是資訊金融

0:00大家好。今天要談的是，我們怎麼從事件市場走到一個更大的領域，也就是資訊金融。預測市場不是新點子。以太坊很早期就有，Augur 其實是第一個 ICO，做的就是預測市場。但現在它們好像贏了。你看到可觀的成交量，上個月大約 500 億美元。新聞到處都是，甚至走出加密貨幣圈。So hi, everyone. So today we're gonna talk about how we're going from event markets to a more global field, which is gonna be information finance. So we see that prediction markets are not a new idea. We had that at the very early days of Ethereum. Actually, Augur was the first ICO. It was prediction markets. But now it seems that they may have kind of, like, won. Like, you get a significant amount of volume, like 50 billion in the last month. It's all over the news, and it even gets outside of crypto.

0:46它們也運作得不錯，選舉預測做得很好。問題是：這就是終局嗎？預測市場能做的，我們是不是都挖完了？我的答案是否定的。因為目前大部分成交量是體育，再來是政治和加密貨幣，所以還可以有一點意思。但以今天的形態，預測市場比較像賭博場所，順便提供一些資訊上的正外部性。這跟原先的願景是反過來的：原先是你想要某個資訊，所以為它做市場；不是你想要市場，然後順便得到一點資訊。We also see that they are working pretty well. They've been able to provide nice predictions for elections. And the question is, okay, is it the end game? Did we basically fully exploit what we can do with prediction markets? But my answer would be no, because currently most of the volume that we have is sports. We have politics and crypto, so it can still be a bit interesting. But currently, I would say prediction markets as they are today tend to be more like a gambling venue, which are providing some positive externality in terms of information. But it's not the other way around in the original vision, where you want some information, so you make markets for it. It's not like you want markets and, okay, you get some information as a byproduct.

1:31這讓我想起加密貨幣早期。Bitcoin 早期，大概四分之一到一半的交易，只是有人在 SatoshiDice 上賭博。以太坊有 Fomo3D 和一堆賭博，或他們叫鏈上龐氏，其實就是賭博而不是龐氏，因為大家知道自己在玩什麼。你可能會說，那些時間都浪費了。但其實很好，因為在壓力測試系統。它證明 Bitcoin 可以承載極多的交易。Fomo3D 有一陣子連吃以太坊全部 gas 達三分鐘。這些看起來只是好玩實驗或賭博的東西，最後卻把系統壓出更大價值和公共財。現在可以談這個大願景。So this reminds me of the early days of crypto. In the early days of Bitcoin, you had a quarter to half of the transactions, which were just people gambling on SatoshiDice. With Ethereum, we had Fomo3D and a bunch of gambling, or what they called on-chain Ponzi, which was basically gambling and not Ponzi because people knew what they were playing. And we could say, oh, like, all of that was basically just time wasted. But no, actually it was great at stress testing the systems. It has shown that we could have a very large amount of transactions on Bitcoin. For Fomo3D, at some point, they were taking all the gas of Ethereum for three minutes. So those things, which may seem just like fun experiments or fun gambling things, actually end up stress testing systems which can have a way larger value and public good in the end. So now we can go on this grand vision.

2:26Vitalik Buterin 有一篇文章，〈From Prediction Markets to Info Finance〉。預測市場只是一個大得多、極有力量的類別裡的一個例子：資訊金融。資訊金融的意思是：你想要某個資訊。你想知道該做什麼決策、某件事有沒有風險、會不會有人攻打你的國家，然後從你想知道的事出發，去設計一套市場機制來得到這個資訊。來看資訊金融能做什麼。先從很簡單的開始。我們還是有事件市場，這件事會不會發生，那會變成機會市場。So we have this article from Vitalik Buterin, “From Prediction Markets to Info Finance.” Prediction markets are only one example of a much larger, incredibly powerful category: info finance. So the idea of information finance is that you want some information. You want to know some decision to take, you want to know if something is risky or not, you want to know will someone attack your country or not, and you go from what you want to know into designing some market mechanism to know this information. So let's see what we can do with info finance. So first, let's start with something very simple. We still have an event market where will this thing happen or not, and that's going to be opportunity markets.

3:16用市場篩提案，也用來估協議安不安全

3:16比方我們要幫網站改品牌、幫 Devcon 做周邊，又不太知道自己要什麼，就讓大家丟很多提案。現在有了 AI，這樣做可能收到幾百個提案，很多是垃圾，你沒辦法逐一評估。市場怎麼解？我們做一個市場，問：這些提案裡我們會選哪個？大家可以交易：這件 T-shirt 會不會印？另一件會不會印？你會看到比較可能被選上的提案浮上來。我們就可以多花時間看那些，因為市場相信它們比較可能被選，大概也相信我們會選好提案，所以它們可能真的比較好。So we want to rebrand a website, we want to get some merch for Devcon, and we don't really know what we would like, so we let people give us a lot of proposals. And the thing is that now with AI, if we do that, we can get hundreds of proposals. A lot of them may be slop, and you can't just evaluate all of them. So how can we solve that with markets? So we make a market where we ask, which of those proposals will we pick? And people can trade. Okay, like, are we gonna print this T-shirt or not? Are we going to print this other T-shirt or not? And what we can see is that it's bubbling up proposals that we are likely to pick. So now we can spend more time looking at those because we know that the market believes they are more likely to be chosen, because probably they believe we are good at choosing proposals, so because they may be good proposals.

4:09你會說，會有人操縱市場吧。確實有。上線第一天，有人把自己的提案（我相信是他自己的，否則就太蠢了）炒到 50%。但很快有人反向交易，價格回到比較理性的位置。這裡很單純，合約不用改，就用跟 Polymarket、Kalshi 同一類合約。再複雜一點。我們想知道某件事有沒有風險。這就進到資安。前面有兩場資安的講。資安一部分是怎麼讓東西更安全，另一部分是怎麼知道什麼是安全、什麼不是。You can say, oh yeah, like people are going to try to manipulate the market, and actually they did. First day we launched that, you had some guy trading his own proposal. Well, I believe it was his own proposal; otherwise, it was stupid, to 50%. But then pretty fast, people start to counter-trade that, and it goes back to more rational values. So here it's, yeah, very simple. We don't change the contract. We just do the same type of contract as Polymarket and Kalshi. Okay, now slightly more complicated. We want to know if something is risky or not. Well, that goes into the security. We had two talks on security before, and one part of security might be, how do we make stuff more secure? And the other part of security can be, how do we know what is secure or not?

4:55我們可以做一個市場：這個協議一年內會不會被攻破？這裡可以看到，這個協議大概在 1% 左右，有些協議會低很多。所以可以看到這個協議，Stake，大概比 WBE 更安全，WBE 我以前其實沒聽過。怎麼做？不能只問一個簡單市場：這個協議會不會被駭？那樣市場會極度失衡，大概 99% 不會、1% 會。這樣做，Yes 那邊的價值會被資金成本主導，而不是被協議被攻破的機率主導。So what we can do, we can get a market on: will this protocol be broken within a year? And we can see here, like, this protocol would be at around 1%. Some protocol would be way, way, way less. So we could see that this protocol, Stake, is probably more secure than WBE is, that I actually had never heard of before. So how can we do that? We can't just ask a simple market, will this protocol be hacked or not? Because if we do it, we're going to end up with a market which is deeply unbalanced. We're going to get, like, 99% no, 1% yes. And if we do that, the value of the yes side is going to be dominated by the capital cost and not the likelihood of the protocol being broken.

5:41例子是 Polymarket 上那些耶穌市場：耶穌明年會不會再來？給出 1% 或 2%。不是因為耶穌回來有 1%、2% 的機會，只是沒人願意買 No，一年只賺到一點百分比的零頭，因為資金有更好的去處。所以這裡可以做多類別市場。比方說 WBE 和 STE 都被攻破，各自只賠 50 美分。這樣你只需要一個市場，同一筆資金可以同時用在所有標的上。最後賠付不是確定的，但如果你假設……Like, example would be, in Polymarket, you had those Jesus markets. Will Jesus come back next year? And it was giving 1 or 2%. It's not because there is 1 or 2% chance of Jesus coming back. It's just because people are not ready to buy the no if they get only a fraction of a percent per year because they just have better venues to put their capital. So what we can do here, we can do a multi-categorical market such that if... Let's say WBE and STE are both broken. They will each pay 50 cents. So now you can have only one market, and you can use the same capital on all of those at the same time. Now you end up with payouts which are not deterministic, but if you assume...

6:27預測治理：用通過／失敗兩個宇宙的價格做決定

6:27這些事件互相獨立。它們其實不獨立，但大概沒那麼高度相關。這個假設合理到能讓系統運作。你可以用代幣價格反推隱含的被攻破百分比，也可以從被攻破百分比反推代幣價格，再用交易引擎依此自動交易。於是我們得到更複雜的市場，離傳統是非二元市場更遠。接下來完全瘋狂的是預測治理：對價值投票，對信念下注。這是 Robin Hanson 的想法。他在這裡描述一種新的治理形式。就這樣。That those events are independent, and they are not independent, but they are probably not that correlated. So it's an assumption that is reasonable to make the system work. You can, from the price of the token, compute the implied percentage of something breaking. And from the percentage of something breaking, you can also compute the price of the token, and you can have a trading engine trade automatically based on that. So. So we are getting a bit more complex kind of markets, which are getting farther away from the traditional yes-no binary markets. Okay, now, like, we go into something completely crazy, which is futarchy: vote values, bet beliefs, and that's an idea of Robin Hanson. And here, he just describes a new form of governance. Just that.

7:12想法是：你先決定自己要什麼。如果你是一個國家，說我要高人類發展指數、不想被入侵、要高人均購買力，你就選好要拉高的指標，再做市場去拉高它。怎麼做？把宇宙切成兩個。一個宇宙你做了某件事，一個宇宙你沒做。兩邊可以獨立交易。例如 Kleros 有一些 KIP，也就是提案，大家可以在提案通過的宇宙交易 PNK，也可以在提案被拒絕的宇宙交易 PNK。So what the idea is, you first, you determine what you want. So if you are a country who says, I want high human development index, I don't want to be invaded, I want to have high purchasing power per capita, so you pick up the metric of what you want to optimize, and then you make markets to optimize that. So how do we do it? We do it by splitting the universe into two universes. One universe where you do something. And one universe where you don't do it. And so you can trade independently in both of those universes. So, for example, for Kleros, we have some KIPs, so some proposals, and people can trade PNK in the universe where the proposal is accepted and in the universe where the proposal is rejected.

8:01他們可以說：如果這個提案通過，我要買 PNK；如果不通過，我要賣一些 PNK。於是有兩個條件價格。這裡可以看到，通過價高於失敗價，表示這個提案很可能對價格有正面影響，所以大概該做。你可以從單純的諮詢系統，走到混合系統，Kleros 就是這種：某些特定提案既要投票、也要通過預測治理測試；也可以走到完全的預測治理。我認為除了單純事件市場，這大概是目前第二成熟的類別。So they can say: If this proposal passes, I want to buy PNK. If this proposal does not pass, I want to sell some PNK. And so now we have two conditional prices. And here we could see, okay, the pass price is higher than the fail price, so this means that this proposal is likely to have a positive impact on price, so we should probably do it. So you can go from a simple advisory system to a hybrid system, which is the case for Kleros, where for some particular proposals, you need both votes and to pass a futarchy test, to systems which are full-on futarchy. And here, I think besides simple event markets, that's probably the second most developed category that we can see.

8:47因為有三個不同專案在做。從 Solana 上的 MetaDAO 開始，現在以太坊上也有 Lumia 和 Futarchy.fi 在做這種資產預測治理。預測治理不一定只衝一個數值，也可以衝一個指標。我們跑過這個實驗。我想衝的是看一部好電影。怎麼做？做一些市場，問：如果我看這部電影，我會打幾分？然後我去看、去評分排名前面的電影，這就變成建議我該看什麼。第一次結果普通。第一輪我自己的 AI 贏過市場，但第二輪市場贏過我們試過的其他預測者，而且把我最後看的五部電影排名預測得很準。Because you have three different projects doing that. It started with MetaDAO on Solana, and now you also have Lumia and Futarchy.fi on Ethereum doing this asset futarchy. So for futarchy, you don't necessarily need to optimize just a value; you can optimize a metric. So we ran this experiment. And what I wanted to optimize is watching good movies. So what do we do? We make some markets, and we ask: if I watch the movie, how will I rate it? And then I can watch and rate the top movies, and that acts as something to advise me in what I may want to watch. The first time we ran that, the results were kind of like so-so. My own AI was outperforming the market in the first round, but then in the second round, the market was outperforming all the other predictors that we tried and was actually perfectly predicting the ranking of the five movies that I ended up watching.

9:47市場短期像賭具，長期才像秤；也可以放大人工判斷

9:47也許是運氣，也許市場第二輪之後變得很好。結論是：如果你要做資訊金融，第一輪也許不行，繼續做第二輪。因為你在架市場機制時，第一次常常不靈。短期來說市場是賭具，長期才是秤。參與者需要一點時間適應、變得更老練；做得好的人發現自己行，才加碼；做得差的人發現自己不行，才停手。這些機制才會開始運作。我們也能做比較花俏的事，例如 Devcon：我們在辦一些側邊活動，要選辦哪一場。So maybe it was luck, or maybe the market became pretty good after the second round. But the conclusion of that is that if you want to work in Info Finance, maybe you will launch a first round, and maybe it's not going to work, but continue a second round. Because often, when you are setting up some market mechanism, it may not work the first time. Because on the short term, markets, we say, are gambling machines, and on the long term, they are weighing machines. So it may need a bit of time for participants to adapt, to become more sophisticated, for participants who were good to realize they are good, so put more money, for participants who realized they were bad to stop putting money, for those market mechanisms to work. And we can do fancy stuff, you know, like here, what we are doing for Devcon: we are making some side events, and to choose the side event.

10:35我們做一個市場，問在某種活動類型的條件下，會有多少人來、會後問卷分數會是多少。然後試著最大化這個合成指標，把兩個預測市場合在一起，找出依這個指標最可能帶來最大影響的活動。資訊金融另一件有意思的事，Devansh 今天早上已經談過，是去中心化的人類判斷。意思是：你需要評估某件事，你做得到，但想做到規模。這裡我們有很多開源軟體倉庫要評估，大約 2,000 個。We are making a market on what will be the number of attendees and what will be the rating given in the post-event survey, conditional on the type of event that we run. And then we try to maximize this composite metric, which is combining two different prediction markets to find out which event is most likely to give the highest amount of impact, according to this composite metric. Another interesting thing of Info Finance, and Devansh already talked about that this morning, is decentralized human judgment. So the idea of decentralized human judgment is you need to evaluate something, you are able to do it, but you want to be able to do it at scale. So here we had a lot of OSS repositories to evaluate, something around 2,000.

11:25我們沒辦法找人去評 2,000 個倉庫的品質。所以改做市場，問：如果這個被評估，分數會是多少？讓大家交易，然後只挑其中一部分，真的派人去評。這些市場依人類評估來結算。於是市場在估計人類會怎麼評，也在估計某件東西的價值。它既是去噪器，市場估計可能比單一人類估計更好；更有意思的是，它是擴充機制。And we can't get people to evaluate the quality of 2,000 repositories. So what we do instead, we make a market where we ask, if this is evaluated, what will be the score given to this repository? We let people trade those, and then we only select some of those repositories where we actually send some humans to evaluate those. And those markets, they resolve according to the evaluation of the humans. So we basically get the market which estimates the human evaluation, which estimates the value of something, and it acts both as a denoiser. The market estimate may be better than the human estimate, and more interestingly, it acts as a scaling mechanism.

12:10因為市場能交易的項目，遠多過人類能評的。為什麼？因為你可以讓 AI 在這些市場上交易。現在大多數預測市場還是人類在交易。有人試過幫 Polymarket 和 Kalshi 做交易機器人，結果是負的。但我們看到 AI 愈來愈好、愈來愈快。可以比較人類和 AI。人類多數理性有限，時間也有限，而且很貴。目前至少好的人類還是比較準。但 AI 的重點是：開機時間可以很高，你可以再拉一個 AI 起來，成本很低。Because the market can trade way more items than a human could evaluate. And why? Well, because you can get AIs to trade on those markets. So now, most prediction markets tend to be traded by humans. You had some people, they tried to make some trading bots for Polymarket and Kalshi, but they got negative results. But we are seeing AI getting better and better, and faster and faster, getting better. So we can compare humans and AIs. So humans, they have limited rationality, most of them. They tend to have pretty limited time, and they are pretty expensive. For now, they still tend to be the most accurate, at least the good humans. But the thing with the AI is that you can have a very high uptime, like you can spin up some other AI, and the cost is pretty low.

13:02AI 做微型市場，預測市場當有誘因的集成器

13:02這就打開機會：去做流動性或成交量不到一美元的微型市場。我們可以有極多的低價值市場，加總起來卻夠讓一個 AI 開發者、甚至一個失控的 AI 去交易。Deep Funding 那個去中心化發放補助實驗就是這樣。最後大家都跑某種 AI 機制去預測人類評估，那個市場基本上只有 AI 在交易。現在還有人類在掌舵，也許未來會更自主一點。So this opens the opportunity of making those micro-markets with less than a dollar of liquidity or volume, and we can get this very large amount of low-value markets, but which together add up to some value which is high enough to get an AI developer or even a rogue AI working on trading those markets. And that's what we had for the decentralized grant-making experiment with Deep Funding. Everyone ended up running some AI mechanism to predict the human evaluation, and that was a market which was basically only traded by AIs. Like, now, there still are human handlers, but maybe in the future they would be a bit more autonomous.

13:48我對整個領域的總體觀察是：我們發展了很多很好的經濟機制，但對人類常常偏複雜。我剛剛只簡短帶過預測治理，大概多數人還是不太懂這種機制。這些條件市場是什麼？人類要採用可能比較難，但 AI 可以採用。所以很多現在理論上很好、只用在很重大決策或很大市場的東西，接下來會用在日常微型層級，只因為 AI 能在上面交易。And that's, I think, a general reflection on this whole field, is that we have had a lot of very nice economic mechanisms that we've developed, but often they are a bit complicated for humans to grasp. Like here, I gave a brief overview of futarchy, but probably most people don't really understand this kind of mechanism. What are these conditional markets? It may be a bit harder for a human to adopt those mechanisms, but AI can adopt that. So I think a lot of things that currently are nice in theory and only used for very important decisions or very important markets, very large markets. Now we're gonna see it used at the day-to-day micro level just because AI can trade on that.

14:35這也接到我們正在做的研究：把預測市場當終極集成方法。集成方法的意思是：你有多個預測者，試著把它們合起來，得到比任何單一預測者更好的預測。預測市場基本上像學習演算法，流動性則像學習率的倒數。流動性愈高，市場動得愈慢。跟多數集成方法比，有意思的是它誘因相容。就算你認為某些預測者不懷好意，可能懶、也可能有惡意。And this opens up to some research that we are currently doing of using prediction markets as an ultimate ensemble method. So the idea of the ensemble method is that you get multiple predictors, and you try to combine them to get something which is a better predictor than each of your individual predictors. And prediction markets basically act as a learning algorithm, and you have some liquidity, which is an inverse learning rate. Like, the higher the liquidity, the slower the market is going to move. What is interesting compared to most of the ensemble methods is that it's incentive-compatible. So even if you consider that some of your predictors may be non-benevolent, they may be lazy or have bad intentions.

15:22你還是給這些預測者誘因，讓他們朝好的價格交易。傳統集成機制裡，所有預測者都得你自己找，自己開發或付錢請別家公司開發。現在你做一個市場，等於有一個採購部門去找預測者。任何人不必跟你說話，你也不必招募他們，他們可以自己來交易，給你更準的預測。如果不準，他們會賠錢，準的人反而拿到更多。結論是：預測市場只是我認為未來大得多的領域裡很小一塊，那個領域是資訊金融。You give the incentive for those predictors to actually trade toward good values. And also, on classic ensemble mechanisms, you need to get all your predictors by yourself. Like, you can develop them or you can pay some other company to develop them. But now, if you make a market, you basically have a procurement department to get those predictors. And anyone, they don't need to talk to you, you don't need to try to recruit them. They can come and trade on those markets to give you more accurate predictions. And if they don't give you accurate predictions, they're going to lose money, and that's even more money for people who are accurate. So. What is the conclusion of that is that prediction markets are only a very small fraction of what I believe to be a future, way larger field, which is information finance.

16:18內線不是風險：市場本來就要把知情者逼出來

16:18我們可以看到，它們現在接近賭博，而賭博這個市場，跟做決策比起來還很小。你每天都在做決策：Twitter 上看什麼、要不要來這場研討會、公司要把錢花在哪個專案。你一直在做決策。做決策這個市場，遠大於交易體育或誰會贏選舉。所以我預期這個類別會變大。也歡迎你們來幫忙把它做大，自己做資訊金融，或用資訊金融機制去拿到你想要的資訊。We can see that, basically, now they are close to gambling, and gambling is still pretty small of a market compared to making decisions. Like, you make decisions in your life every day, whether it be what you will watch on Twitter, whether you will go to this conference, what project your company will spend on. So you make decisions all the time. So making decisions is a way, way bigger market than trading on sports or on who's going to win the election. And so I expect this category to grow, and— you're also welcome to help grow it, either by working in InfoFinance yourself or by using InfoFinance mechanisms to get the information that you want.

17:07謝謝聆聽，接下來開放提問。謝謝 Clément。有沒有人要問 Clément 這個題目？有人有問題嗎？喔。謝謝，題目很有意思。我想確認一下。我聽說有些預測市場裡，內線或參與者可能握有別人沒有的資訊，基本上是他們在控制市場結果。你怎麼看這種風險？有辦法處理嗎？這不是風險。這其實是好事。So thanks for listening, and we're going to go to questions. Thank you, Clément. Now, do we have any questions for Clément on... this topic? Anyone have questions? Oh. Thank you for a very interesting topic. I just want to check that. So, I heard that in some of the prediction markets, the insider or the participant might have some information that no one else would know, and they are basically the ones who control the market outcome. What's your take on that kind of risk? Is there a way that we can deal with that kind of— It's not a risk. It's actually good.

17:58如果你把預測市場當賭博機制，內線資訊對其他人就不公平，那是壞事。但如果你把它當資訊匯總機制，那就很好。你有內線知識，市場會逼你把這份內線知識顯露出來。這其實是機制照設計在運作。我同意，這點很好。我想再追問。如果押注大到讓人做出原本不會做的決定呢？我是說，真正握有結果控制權的人，可能去做原本沒打算做的事，只為了贏錢。So if you see a prediction market as a gambling mechanism, okay, that's bad for people to have insider information because that makes unfair gambling. Now, if you see prediction markets as an information aggregation mechanism, that's great. If you have some insider knowledge, the market makes you reveal this insider knowledge. And so that's actually the mechanism working as it should. Oh, I agree with you. That's a very good point. I just want to follow up on that. What do you think about when sometimes the value at stake is so large that it makes people make a decision that they originally wouldn't do? I mean, for the people who actually have control of the outcome, they can do something that they originally don't intend to do, but they just want to win the money.

18:46他們可以把錢砸進去、改變決定、再把錢拿回來。對，你可以說，如果他們在這個市場交易，市場可能影響他們的決定。這有可能，顯然取決於市場規模相對於那個決定的價值。但我還沒真的找到具體例子。我看過拿資訊做內線交易的具體例子，這確實發生過。但人為了贏市場去做某件事。喔，體育市場有一些例子，市場問會不會有裸奔的人跑進球場，然後他們真的跑進去了。They can just throw their money into it and change their decision and then get the money back. So, yeah, you could see that, like, the market could influence people's decisions if they are trading in this market. That's a possibility, and that obviously depends on the size of the market compared to the value of the decision. But I haven't really been able to find any concrete example of that. I've seen concrete examples of insider trading with information, so this actually did happen. But people actually doing something to win the market— oh, I have some examples with some sports markets where you had a market about, will there be a streaker, like someone running naked in a sports stadium, and then they did run there.

19:34操縱賽果、做空砸盤不是資訊金融獨有

19:34所以對，這確實是一種副作用。但我認為副作用通常遠小於資訊匯總的價值，而且只發生在很特定類型的市場。像我剛講的預測治理，會讓人用不同方式行動，但他們會這樣行動，是因為多拿到了資訊，而這正是我們要的。我想補充一些……我得承認我其實沒有參與這個市場，也沒有在那裡交易。我只是想像，體育比賽裡的選手，當大家在交易他們會贏，他們卻試著故意輸，因為注碼太大，再用隱身帳戶去買。So yeah, like, this is effectively some kind of side effect that you can have. But I think the side effect tends to be pretty small compared to the value of information aggregation, and this only happens in very particular types of market, like what I talked about, like things like futarchy. It makes people act in a different way, but they act in a different way because they got extra information, which is what we want. I just want to provide some of the... I have to confess that I am not actually involved in this market. I do not do the trading there. But I just imagine, what if, like, for a sports game, a player inside that sports game, when people are, like, trading that they win, and they try to lose it just because the stake is, like, too high, they can ... like, use a stealth account and then buy some, yeah, you know.

20:29對，這就是經典的打假球，通常發生在比較像賭博的市場。預測治理也可能出現類似情況。有人可以對某個指標放空，再故意把這個指標砸崩，讓空單獲利。但你也可以說加密貨幣裡早就有了。你可以放空 ETH 再開始唱衰 ETH，也可以放空 USDT，再用 Bitfinex-Tether 那種帳戶去唱衰。所以這不是資訊金融特有的，是金融通用的，而且發生在你能放空的東西上。你能放空某樣東西，就可以故意傷害它來從空單獲利。Yeah, so this is the classic match-fixing, which happens generally in markets which are more like gambling. For things like futarchy, you could see something similar. You could have someone take a short position on a metric and then try to purposely crash this metric. In order to have their short position pay, but you could argue that's already the case in crypto. Like, you can short ETH and then start to FUD on ETH. You can short USDT and then try to FUD on it with, like, the Bitfinex-Tether account. So it's not really something which is specific to information finance. It's something which is specific to finance in general. And on things that you can short. If you can short something, you can purposely try to harm this thing to benefit from your short.

21:20所以那是一般金融的副作用。我仍然認為，整體來看，預測市場帶來的資訊收益，比有人故意砸指標來從市場賺錢的邊緣案例，高出幾個數量級。謝謝 Clément。你預測資訊金融或預測市場會成長，而且會長得比賭博市場更大，對吧？以前傳統體育賭博有中心化供應鏈。So that's a general finance kind of side effect. But I still believe that overall, the gain of information of prediction markets is orders of magnitude higher than the potential edge cases of people trying to crash metrics on purpose to gain money on the market. Thanks, Clément. Since, like, you predicted the information finance or prediction market will grow, then grow them bigger than the gambling market, right? So at the previous time, in traditional way, in sport gambling, they have the centralized supply chain.

22:10他們的預言機，像有 Betradar，是單一預言機，訊號來自一家資料公司。我的問題是，在 DeFi 做預測市場，不管體育、金融還是政治議題，DeFi 的預言機在哪？因為供應鏈不是很……目前多數公司確實來自美國。但 DeFi 的供應鏈在哪？大概在歐洲或其他地方？So their oracle, like, they have, like, a Betradar. They have the single oracle, the signal from a company, a data company. So my question is, in DeFi, to do the prediction market, whatever sport or whatever in finance or political issue, where is the DeFi oracle to, to, to, because the supply chain is not very... ... For sure, and we can see for now most of the companies are from the United States. But where is the supply chain in DeFi? Probably in Europe or something else?

22:57結算可以用 UMA、平台自裁，或 Kleros 仲裁

22:57這就是我的問題。對，不同預測市場用不同的預言機方法。Polymarket 用 UMA。像 Kalshi 這種，就自己結算。如你所見，我在 Kleros 工作，所以在 Kleros，我們有一套仲裁機制、去中心化仲裁機制，也裁決這類市場。所以結算方式很多種。顯然我是 Kleros 的人，會建議你來找我談這個。但如果你不用 Kleros，也可以做聚合：用多個預言機，然後三取二。So this is my question. Yes, so different prediction markets use different oracle methods. Polymarket uses UMA. You have some like Kalshi, they just resolve it by themselves. And as you can see, I'm working at Kleros, and so at Kleros, we have an arbitration mechanism, a decentralized arbitration mechanism, which also adjudicates those types of markets. So you have a vast variety of ways to resolve those. Obviously, I'm part of Kleros, so I would advise you to come and talk to me for that. But if you don't use Kleros, what you could also do is you could do some aggregator. You could use multiple oracles, and then you could take, like, a two out of three of those oracles.

23:44這樣有人同時攻破三個裡兩個的機率，會比只靠一個預言機低很多。還有其他問題嗎？我想還有一點時間，如果有的話。沒有的話，謝謝 Clément 的精彩分享。謝謝邀請。And now it's way less likely that someone manages to break two out of three of those oracles at the same time, compared to relying on one oracle. Any other questions? I think we have a little time for some more, if there's any. If not, thank you, Clément, for your wonderful presentation. Thanks for having me.