Agents in the Shell: TEE-Attested AI as On-Chain Oracles

Transcript

[music] Okay. So, first like um prediction markets like prediction markets have gone mainstream. So, but they only cover big events like US elections, BTC prices and major sports and most like longtail events in people's lives are not captured there. So, let's make a hypothetical like thought experiment on what can be possible if like more longtail things are captured.

So, what a longtail prediction market can enable. So let's think of this case like you want to fly for a job interview tomorrow and in the flight if the flight is more than one hour late you miss the interview and you get a x% chance of getting that job which would give you like a y dollar raise. So one thing you would want to do like here you know that you have like expected loss that you can calculate based on like missing a flight and you would want like want to hedge it against it. So like basically give you insurance.

So this would be all nice but like this is pretty hard to do and I'll list a couple of like reasons why this is hard. So why longtail markets are hard like it makes such bet like making such bet requires like a lot of cognitive load. So like you have to think about like what are my like chances of like missing this flight? what are the historical like flight misses like percentage what are like my expected like chance of getting this job in the first place and there's like a bunch of things that you have to considerate that's way too much cognitive load for a human to do but one nice thing that I think about agents is that once you have AI agents like your personal AI agent that AI

agent can handle all this cognitive load which will be very nice another problem is that it requires like settlement of results of the longtail events so like a lot of these oracles they only handle these huge events but like how about the long tail and here I think AI and TE will be a good like solution so what AI and TE can enable is that it can enable like credibly committed intelligence what I mean by this is that in blockchains it's very nice you get if it it's credible so like if it says something like you know that everyone agrees that it says something but one problem is that it's not intelligent it can hold logic but it can't make like decisions based on looking at the world and like make fuzzy decisions. One thing

about humans is that humans are intelligent but we cannot make credible commitments. Meaning that like we can like do calculations in our neuros system but we can't like make proofs that we did like a specific calculation in our head. So the very nice thing about AI and TE and blockchain is that it enables both. So it enables a credibly committed intelligence and by being intelligent like it can also enable like more broader forms of market like you know fuzzy things like like please create a TR that handles like all these kind of specifications and the agent can make a call on whether that was met.

So here is like one way that we can implement it and I I implement it as a PC. That's a very simple thing like a user here is a user who would want to like create a market or like a create a prompt and they will post it to like a blockchain to a contract and some like offchain relay will like read that and then it will like decide to commit to it and then it will commit to it onchain saying that when the time comes I will provide the answer to this prompt and then it'll wait time passes the time comes and then the prompt will be sent to a TE and inside the TE like you'll have uh code that will then call into like some external AI server.

So it'll send out the prompt. It'll get back the result. So the nice thing here is that this prompt and this result it cannot be tampered with because it's running inside a TE. So and then the result will go back all the back way back and then you can verify that it was run inside a TE and this is the prompt and this was the result in onchain. So this is like a quick screen screenshot of the PC. Like you can see that there's a prompt and there's a result. So like taking a step back like what I think is that this is like one example of a broader system like pattern that is emerging or I like really hope that will emerge and it's like AI making decentralized long-term coordination practical.

So like we always talk about this openness and decentralized in Ethereum, right? But the thing that we often like miss is that being open and being decentralized has a lot of cognitive load. It would be super cool if you don't have to use Amazon and you can like just find people like online and interact with them and pay pay them on blockchain. Well, that' be super hard to do because like you have to like assess, hey, is this guy trustworthy? You have to first like discover in them in the first place. you'll have to like say that like what kind of like like contract should I have with this guy and a bunch of stuff. This like way too cognitively loading for a human to do and I think like a lot of like the early Ethereum vision like this was the bottleneck. We had this cool technology

that can enable like this like decentralized credible platform where you can create credible commitments and enable coordination but the bottleneck was human like cognition. So what I think is super cool is that AAI could like be this missing piece in this decentralized world where like it could like lower the the cognitive load that is required to have open and decentralized world and like enable like decentralized coordination practical even in the colonail. Thank you.

>> Thank you so much Lynn and almost perfectly on time. We have time for questions. I will throw this. >> Thank you. What about the problem? I could bet $10 million on the flight being late and then call in a bomb threat. So, who would take that bet against me? >> Yeah. So, so I I think like one thing that I did not cover here, which is like a hu that's a very good question, is like how do you create markets in this like thin longtail like uh like events. And there's like a bunch of like interesting research here. Like one way you like one naive way that you can

think of is you could just like pull put your all your money in like buckets. Then whoever like in the winning bucket can like take all the money from all the other buckets. But this has like obvious problems where like incentivize people to bet late and like because like the more you wait the more information you gain and like then like the better bet you can make. And there's like more interesting solutions like what's called like logarithmic LMSR market scoring rule like where like basically the property that you would want in like a AM style prediction market is that you would want people to like bet what they really believe in instead of like trying to game the market like if you think like 70% something will happen and 30% that it won't you want that 73 like

allocation instead of like just going to 100 in that seven part and like using logarithm that you can actually like enforce this kind of properties. There's a bunch of like interesting research there, but yes, it's I didn't >> time for more question. >> Does someone want this box? I promise to be good better at throwing. No. All right, then we will [music]