Prediction markets for distilled human judgement

Transcript

[Music] test test. Oh, okay. Uh, hi everyone. So, today we're going to talk about prediction markets uh for distillum management. So first we're going to go a bit through prediction market primitives and then see how we can apply to distill management. So what's a prediction market? So that's a a way to get some probabilistic estimate of the future.

You probably all followed the US election and we had the market which were asking who will win the election Trump ARIS and you can trade those outcomes and it provides some probabilistic estimates but we can also get more complex signals. So we had for example those markets on if Trump wins the election what would be the price of Bitcoin? If Aris wins the election, what will be the price of Bitcoin? And we could see that the market were giving higher Bitcoin price in the universe of Trump winning than in the universe of ARS winning. And indeed on the result became known bit price increase.

Obviously the prediction was not completely exact. We got like 8,000 extra compared to 3,000 extra. uh but it was still some uh interesting signal. So how does it work from a technical perspective? So you have an underlying token let's say it can be like a USD stable coin or ETH. You put it into a market which is a question and you get a complete set of outcomes. So which party will win the election? You get Democratic party, Republican party other that's ERC20 tokens. You can trade them wherever you want. And when you see the price, let's say, of the democratic token at 38 cent, you can infer that the market is giving a 38% chance of the democrat winning the election. At the

end of the day, we know what is the outcome. So you use oracle for that like claros. And in this case, you redeem the correct outcome for the underlying. So at the end, since the Republican won the election, all the Republican token, they got turned into die. We can uh go further than just outcomes. We can go on values. So here we have a we want to make a market to predict the inflation. So we can ask what will be the inflation next year. We have to define a range 0 to 10%. And we have some up token and some down token. So if the value is above the range, the up token redeem. If the value is lower than the range, the down token redeem. But if

it's within the range and that's what it should be the range is defined correctly they both redeem proportionally. So the up token red proportionally to the value the down token is a complement of the up token that they both to one. So here for example if we can see that the up token is trading at 80 cent we can infer that the market is giving an expected value of inflation of 8%. Okay. you know condition market what we can do we can take some prediction market we can take the outcome and we can put it into another market so here we ask who will be the candidate of the Democratic party we get the Biden token we put it into another market which is which party will be in the election and now here we have a token which is a

Biden Democrat token so it redeems for $1 if Biden is a candidate and the Democrat party wins the election and if we look at the price of this token in term of the price of the Biden token, we can infer conditional probabilities that if Biden is a candidate, the Democrat will have 30% chance of winning the election. Okay. So now let's go how we can use that for decision. Let's go into futarchy. So the idea of futarchy is uh you're going to take decision by using prediction market in particular conditional prediction markets. So here let's assume you are a voter. You only care about the price of Bitcoin. You care about nothing else. you can see

this market and you say, "Oh, I'm going to vote for Tren because that would be what would be the highest price of the Bitcoin." In practice, people value more complex things, but that will give you a simplified version of it. So, in classic photo, let's go again with the Democrat party. Uh you want to send the best candidate to the election. So, you make the first market. Which uh candidate will we pick? Biden, Ariel, and then you put the token into the second market. So you get the Biden Democrats, you get the Ari Democrat and those token you can trade them with a token of the first market. So you can trade Biden Democrat with Biden, you can trade ARS Democrat with Ari. And actually you don't even need to have a market for those tokens themselves. Only

the conditional market here is interesting. And since you see that ARS has more chances of winning if she's a candidate, well, you can pick Ari. That's to increase your chances. Okay, remember the scalar market. Most of the time we don't optimize for a specific oka. Most of the time we want to optimize for metrics. So if you combine all of that together uh let's say you launch a new chain you want to get user on your chain and you have a bunch of project which are coming to you and say hey can you give me some grant uh and I will bring you some users because of what I'm going to build on your chain. Uh you can make a first market which is which project will be funded. This market you don't need to trade on it. is it's not very interesting to trade on it. But then you

put on dollar and you get a complete set of the project A, project B, project C and project D tokens and this token you will trade them in the second level market where they ask how many user will this project get if it get funded. So trader can come they go with some dollar they split some complete set and they can trade uh in all of those conditional markets. At the end of the day, you tend to pick uh what is the best project. So you pick project B. All the trade done with project C, D and A token, they end up not having any impact on the profit and loss of the traders. And only the the trade done on project B will impact the profit and loss of the trader. So if a trader was

only knowledgeable at project A, they could just trade on project A and they would have kept B, C and D token. And since B will redeem, they will just get their money back. The trader can also trade in all the universe at once. Um here an idea is that similar to an exam, if you have an exam, you may have a bunch of topics and because you don't know which topic will end up at the exam, you have to try to be good on all topic, be knowledgeable on all topic. So here you want to trade well on all the markets. So what do we want to optimize in harchy? Uh back in the day they were talking a lot about this as a method of governance for states. uh which I think is interesting because decision are so valuable that the cost of running the

markets are negligible compared to the impact it could have but state are a bit hard to convince. So here they would optimize for GDP, human development index or any metrics they could vote on. Uh but now we see that crypto project tend to be way more uh likely to experiment on it and so they can try to optimize for user for TVL for fees. Um but those metrics cannot give the full picture of how successful was the project and sometime we need some subjective evaluation. So we could use a system we already have like retropf or we have all this work which is done in term of impact evaluation where you want to evaluate the impact of a project and then you can trade on the market which is if this project is funded what will

be the value of uh the impact metric and you can fund the project with the is impact metric. Okay, so now let's go into the deep of this element judgment. So we have this nice experiment that we are starting. So I ask a bunch of people on Twitter to give me some movies that I will like or that I will dislike. Uh I remove all the movie I already watch and then we make this market uh which is asking if I watch the movie uh which score in term of percentile I will give to the movie. So traders they can trade those movies uh conditionally on me watching them.

Uh, and you can see for example that I think people think I would like 12 Angry Men, it's probably because of Keros. Um, and if I watch 12 Angry Men and I give maybe like a 70% rating, people who bought the EP token of 12 hungry men, they will get some profit. If I don't like it, I give 25% rating, they will get some some losses. uh here I'm going to watch five of them and most of those movie will not be watched but traders they have to trade them correctly in case I end up uh watching them. So you can see it in with two lengths. You can see the first lens which is the lens of I can use it to make decision to determine what I want to watch. But we can also see this as human judgment like if we wanted to get

a fully comprehensive list of movie rating according to my taste. I could not watch all of them obviously but then we could make a very large market with thousand of movies and here human will probably not be able to trade those but now we can get AI agents which can look at my preference look at my previous rating look at public information on me and then trade uh in those markets and you will be able to rate uh the world movie space according uh to my taste. That would be dist management only of me uh but we can do dist management of a group. So uh in with the uh deep funding you had uh this project where they want to to evaluate the impact of project to

SRM either directly like for example if you see level one nodes there direct impact on SRM but you also have second level and actually even third and fourth and fifth level but we did not show them in the graph that would be too big. um and we can wonder what is the impact of this project to this project which would then impact SRM indirectly. So we want to assign credit for the impact of SRM of all the project even if the impact is indirect. So on the first uh version of this experiment uh they took some committee which are rating the impact of project and then they got some AIS which were trying to estimate the rating of

the committee uh without knowing it and then they rang the AI which one was the best at matching the committee evaluation but the committee is only evaluating a subset of project uh because otherwise you you don't you don't scale but here with a committ which may evaluate two or three groups of project. You could have multiple thousand of project and then the AI would be giving a distilled judgment of what the committee will evaluate if it were to evaluate those part of the graph. So how can we do that better with prediction market instead of just giving prize on the leaderboard? Uh we can turn all of those into markets where we ask if this level two is evaluated

what will be uh the impact score of B1, B2, B3, B4, B5, B6 and maybe it's not evaluated. So in this case all the trade of the AI have no impact on their finances. But if they are evaluated in this case AI will get some profit and loss depending of how good they were at trading those. And now we end up with a full uh a full graph from the prediction markets uh which is a distribution management of the committee and AI may not even be good at evaluating the world graph. Maybe you have an AI is just good at this level two and it's only going to be trading on level two.

Um now let's do a small parallel with auctions. So auction tend to be a pretty old uh uh economic technology. Uh you want to sell some items and you will have some human they will say hey like give me the if you want to increase the price uh and that's what most people figure um picture when they think about auctions. But actually nowadays most of the auction are auction which are made by computer and traded by AIs which are for advertisement space. uh and uh you have massively more AI auction than you have actually human uh auction where human are bidding and I think we're going to have see a similar parallel with prediction markets currently we got poly market they've been pretty good at making market about the news uh but soon

perhaps in two and five maybe 10 years but that would be very very long uh we are going to see something similar where most of the prediction market will not be traded by humans but they will be traded by AIS So on the movies you could have if you have some decentralized Netflix to determine what to show to the user they could ask if this user watch this movie what would be the rating given by the user and then you display that uh social media you know X um in X they have a reaction score like if you uh quote tweet something that's very high if you reply that's very high if you like that's high if you just click on the tweet that's average uh if you blog the guy who made the tweet that's very very low. Uh so in this case they would ask

if we are show this post to the user what will be their action score. Um we can also do that with the physical world. If you have a self-driving car they can ask if we take this route how long will it take to reach the destination and some AI can propose some route and we have a prediction market on the length uh of each route and then we take the route which is the shortest according to the prediction market. So the use cases are are very endless uh because most of the use of machine learning tend to be about optimizing something. So we have chatbot in this case they want to optimize your reaction score like if you give them a plus vote minus vote or sometimes they even ask you to choose between two responses. So that's so they evaluate your reaction.

Now we can ask a prediction market which answer would be the will have the highest reaction score. You display to the user and you display the best answers. For advertisement you want to optimize revenue. uh for pricing like in airline mainly of dynamic pricing you want to optimize profit uh root optimization we talk about that search engine that's also something with reaction score if you click on a website and you close it two second after that's probably not what you were searching for if you stay on it that's probably what you're searching for so you can get a a reaction score uh based on uh the website which are shown to you recommener system we talk about movies uh but you can add that to you know like restaurant preview like if I go to this restaurant will I give good rating to it

if I go to this hotel will I give a good rating to it spam uh spam is a very old problem like proof of work was even invented to to fight spam uh in this case you can ask the AI if the user u going to change the classification to spam or to non-spam and lending you can ask if this user is going to get the loan uh would he remembers or not so in general you have something you want to optimize you want to find which is optimizing some kind of metric that you want in your system. Um why do we do that instead of just taking the best AI? Uh because taking the best AI is probably not going to lead to the best AI. Uh so in 2009 we had the Netflix price where they had a bunch of

movies, a bunch of user data on the movie rating and they were giving $1 million to whomever will get the best prediction of the user score that some score that they kept secret. Um at the end you had the bellor pragmatic cows who won this competition and that was basically measure of the bellor the pragmatic and the cow team they merged together to get the best model. Uh so you had multiple um AI developers and also they had a bunch of different models. So they had the bunch of models and then they had what we call unsemble models which were taking the output on the initial model and combining them together into a better model.

uh and I think predition market can be used as like the ultimate transible method. So why it's great uh is because you don't need to reveal uh your model if you want to participate. So you don't leak information to the competitors uh but you can still provide information to the system which wants your evaluation. Uh you get automated collaboration here you only had three teams which merge together because human coordination is hard but market coordination is easy. So if you add that for Netflix, you will probably end up having a ultimate AI which would be the work of way more than just three teams and then it's open. You don't need to know anyone. You can just run your bot by yourself. And even you can also even run it only on a specific set uh of the problem. Like maybe you

make a bot which is very good uh at uh TV shows but doesn't really work for movies. In this case, you only run it on movies and you only trade it on on movies. Okay. Okay. Okay, so for now like everything's look very great, right? Uh what if you optimize everything? Uh for example, uh let's say we want to optimize to make some paperclipip. You get a paperclipip factory, you give it to a paperclipip optimizer. Uh and what do you think will happen if you give the goal to optimize paperclipip? I see some guy already laughing. They already know the the answer. Uh and that was a experiment uh by Nick Bostrom. And what will probably happen will be that you will have this paperclip maximizer which will first produce paperclip nicely but

then there will be oh human may actually shut me down and if I'm shut down well there will be less paper clip uh so they will try it will try to protect itself and then on that took all the resources nearby they will go to the city nearby take all the metal turn all the metal into paper clips and if human try to like prevent that they will kill the humans and when most of the earth is made out of paper clip Well, you may still have some few human living. Well, you get those human because you know like they made of atoms and with atoms you can make paper clips. So you can increase the function you're trying to optimize. So giving like simple metric is something which is very dangerous specifically if you get some AI which is uh more performance than classic humans.

Uh and that's not something theoretical at all. Uh like you got a bunch of AIs which ended up uh what I would call gaming the objective functions. Uh so here you can see uh with the Lego bricks. So they were asking the AI to pile up Lego and the way they were considered Lego to be piled up. They were looking at the head of the bottom of the Lego and that the arm was not touching it. So what ended up that the AI was just flipping the Lego and then they would as being piled up. Uh you had another example where they were making some form to be able to jump and in the end it was not jumping at all. it was just uh falling one time and another time and that's where it reach highest

distance which could as a jump. So you have countless example of that if you give optimization function to uh to AI sometimes they just game it. Another example was just they were getting points for surviving and the AI just pause the game and even in the classic in animal world we see those kind of issues. So you all know about peacocks. So peacocks initially they have nice feather and having a nice feather is a sign that you are in good health you have food you have no parasite so then the PN they will want to mate with speaker with nice feathers right so in this case your objective function tend to be great like have nice feathers because that's showing good else but what happen is that with evolution you

will get the PN which would like pick up with nice feather uh reproduce with them and it will also be good for them because and they will get chicks pixs which would also have nice feather and so they will be more likely to reproduce also because the other pen will like uh those uh those speed cheek with nice feathers. So you end up into some sort of gaming of the objective function uh where they have way too much feather which should have been good otherwise and then it's easier to catch by predator. It's consuming a lot of resource uh and it's overall quite bad for the species. Maybe for us human we look at that looking good but you have some other example which are way darker.

So if you look at ducks so ducks they have very very very long pennies and it's just uh because that's used uh to force population with the female ducks and so here normally the objective function would be look like a nice duck and get the female duck to mate with you in this case well you know like you just force them to mate with you uh and you game the objective function and we saw that in the evolution uh kingdom. Uh so there's a good art law which is when a measure become a target it cease to be a good measure and what kraova and the other researcher said is it's often easier to evaluate whether an outcome has been achieved than to specify it explicitly. So I think if we use AI, if we use prediction market using AIS, uh

we should not just give simple objective function, not simple metric to optimize uh and we have to think uh about including some human evaluation. So be a bit careful about not just entropo morphizing AIs. So we often see that in the AI doomers they just think that AI are going to be like human but more clever. That's not really the case. uh because in animal evol evolution we have natural selection like don't starve don't get killed by a predator you have sexual selection find a mate you may sometime have some artificial selection like with dogs and you may have the reef which is just moving away more or less randomly uh with AI you have two main thing which is driving which AI are getting to reproduce uh you have like

evaluation like is how do you manage to get a good uh score in the evaluation functions so in the case of prediction market are you able to uh trade well in this case when you trade well you accumulate more resources so you have more impact on the market you're also able to pay uh some sort of R&D to increase yourself to improve yourself as an AI or even make other sub AIS and you also have artificial selection like human can shut down AI if they go bad uh we had this kind of chat bot on Twitter at some point which ended up racist and then you know like artificial selection the humans they just shut it down uh but as AI get better and better human shutting them down is going to stop to to be a viable strategy. So what I would advise is uh to have the target that you

want to optimize here like the f but also to have a subjective evaluation by human out of the result obviously the evaluation you don't do it all the time that's why this still human judgment you do it sometime but because on those market can sometime be resolved they have they have some incentive into trying to optimize for the case where they end up being evaluated and they end up being resolved and that goes with one of the recent article of Italic AI as an engine and human as a steering wheel. You should not just give like a direction you should be able to steer uh because simple direction is not going to optimize for what you want. Uh now what some people could say say hey like if

you ask human evaluation like the AI can just force the human to give good rating you know like if you have an AI which went full on up token for Barbie they could try to force me to give a good rating to Barbie. Yes, that could work if you have only simple evaluation system like just one person. For more complex evaluation system where you have multiple humans that tend to be harder um because of value in the optimization. So if you are an AI you are here you can increase here by uh doing like nice thing which increase your the objective function. Uh and then if you wanted to go there, you will have to arm the human a bit, but you could try to force them to give you good score because you could

capture them and make them good score. Uh but the issue for the AI is to go there from there. They have this big value when you start to arm some of the human and then the other they give you very bad subjective judgment and so you get a very low score on on the prediction market that you are trading. Um and in this case um it's possible that an AI rumal should jump from there to there but it's way less likely because there is not a clear path to go from there to there and we see that very often in uh in optimization systems. So it may not be perfect but that can be a way to align uh AI with human that sometimes some can be randomly it can be when the human decide to they have to give subjective evaluation in the

prediction market which serve as unseable method for the AI in order to uh keep them align um and that's would be all for my talk. So thanks for for listening and if you have a bit of time we we can we have a bit of time so we can go with with questions. [Applause] Thanks. One of the underrated issues with prediction markets especially exotic prediction markets is market making and liquidity. So what would be the best mechanism here for the underlying underlying trading? So so very often uh when you want people to do

something the answer is basically pay them. Uh so the prediction market is not going to be I make a prediction market and I get the answer for free. Uh that's not going to work right. Uh otherwise it will be a zero sum game and with like the work uh made to make the AI or the work to run the AI that would be a negative sum game right. So you need to give incentive to the providers which will then on average lose a bit of money uh and then that make it as a positive sum game for people who are trading those markets.

Thanks. Yeah, was that thanks for the talk. So, uh my question is if you're operating the prediction market platform, then the problem I see about prediction market is this uh in the many countries it's illegalized. For example, I'm from South Korea. It's uh prediction markets especially for political decisions are very uh strongly illegalized. Is it like it's different in like I guess you're from France or like in European countries? So France doesn't really say that prediction market are illegal. Uh they say that uh some of which look like gambling should be regulated and is gambling if the main element is luck. So there is obviously

something which is obvious gambling like if you go on sport or if you go on the last bit of the next bitcoin block because that's pure that this one will be purely random. Then you have stuff which may be more gray area like polit political elections. It's not clear what is a part of randomness and the part of actually skilled in trading the market. I will tend to say that it's way more ski than randomness. Uh and now if you get into you know like something like uh uh trying to optimize metric uh they go to the metric if you want to if you go into trying to optimize metric like this like it's risk AI is really doesn't look at all like gambling and uh I would be very very surprised to to see any country considering that to be gambling.

Okay. Thanks. Any other question? Oh, yeah. Are you considering model? So, here Claros, it's uh not necessarily super related to what I'm talking today on the top versus on the bottom. uh because you need to have a way to you know like for example report those metrics. I will get that most of the time you actually do have some AIs which are working as reporters and if those AI disagree then you probably escalate uh

to a Claros court with very very low fees that is almost only uh feasible for AIS to be there and then if it's still wrong then you can again escalate again uh to humans and we already see that happening you know like if you look at presio which is like know like the rebranding of Omen it's a prediction market traded mainly by AI liquidity is very low like something maybe like 10 $10 per market but the trading uh volume in not volume the trading uh the amount of trading action is very high because it's traded by those AIs so then they report and sometimes those AI will uh will conflict in their report and uh that can be arbitrated by by claros in the end so the AIS are incentivized into report correctly because they misreport

they will end up losing deposit in the end in a clear dispute Uh question no. Okay. Okay. So I guess that's all. So thanks everyone for coming and uh we can talk aside. [Music]