Machine-generated transcript; timing is approximate. Unclear audio is marked. [00:00:01.100] I'll be talking about some lessons from mechanism design and adversarial conditions. [00:00:07.620] I have a ton of disclaimers. One, I'm not a formal mechanism designer. I see some in the crowd. [00:00:14.200] So maybe talk to Pranav and others if you have any questions and I may use some terms loosely. [00:00:20.340] I will also be going across a variety of abstraction levels from engineering to distributed systems and auction design and a bunch of things. [00:00:30.000] So apologies for that. And then last, who's like a crypto person here? Raise your hand. Okay. [00:00:37.300] Who's an AI person? Who's like something else? Okay. Cool. Okay. [00:00:48.240] Yeah, this was my rough prompt. I didn't actually prompt my agent this, but this was my prompt to myself. [00:00:54.200] I've built a few products across a variety of things involving incentive design, distributed systems, and privacy in permissionless markets. [00:01:04.220] So I think that's the kind of like key thing here. [00:01:08.820] And so, yeah, I won't explain crypto, but I'll double click on permissionless markets. [00:01:13.580] So the thing that has created a lot of challenges there is, yeah, the fact that they're permissionless. [00:01:20.500] So anyone can join. [00:01:22.280] And then the fact that it's a market means that there's like real incentives or real value on the table. [00:01:27.320] And so when you have those two things, it can be very, very hard to design, you know, tools and things to protect against people attacking you. [00:01:39.100] On top of that, anonymous identity involved with crypto. [00:01:43.040] That also makes this even more fun. [00:01:46.080] And so, yeah, cool. [00:01:48.320] So I'll just start getting into a bunch of examples. [00:01:51.440] And really this whole talk is actually probably like 15 examples with lessons. [00:01:57.920] Yeah, so we'll keep going. [00:01:59.380] So in the like early days of Ethereum, it used when we were doing like proof of work, there was miners. [00:02:08.220] They compute these. [00:02:09.800] They spend a lot of compute to find this random value that lets them propose a block. [00:02:15.100] And so there's this very funny thing where the Ethereum core developers, the people that develop this client called GoEthereum that all the miners ran, [00:02:24.100] well, it had an API that literally was just get all of the contents, all of the transactions that haven't been put on chain yet. [00:02:31.320] And when you've got a financial system, you can see how this is pretty funny. [00:02:34.580] Also, a lot of those miners didn't block this API. [00:02:37.720] So anyone could just fetch all the transactions that were, you know, imagine you could just see everything that was going to happen on NASDAQ even before it was in the order book, for example. [00:02:48.840] So this is a very funny and simple one. [00:02:51.440] But yeah, APIs are the attack surface. [00:02:55.120] So the next example is, again, and I'm kind of going in chronological order. [00:03:00.320] So in the proof of work Ethereum days, you know, there was this big emphasis in terms of system design on there being a public mempool. [00:03:09.640] So the goal of this was like really to like foster decentralization. [00:03:14.460] You know, you want this like node. [00:03:15.960] Anyone can just run and get all of the data on the network. [00:03:19.800] So it doesn't need to become reliant on a single party to serve that data. [00:03:24.500] So after they, you know, patched that funny API thing, the P2P layer was the next target. [00:03:29.840] Because again, you want these nodes talking to as many people on the network as possible. [00:03:33.720] So you don't need to rely on one party. [00:03:37.200] So yeah, nodes gossip transactions in plain text. [00:03:40.940] And anyone can run a node. [00:03:42.560] So you can imagine how this ended. [00:03:45.220] People would then spin up nodes, observe every transaction before it executes, and try to use that knowledge somehow. [00:03:51.780] And I'll talk more about that. [00:03:53.900] But one of the key lessons here is that like, whatever you do for decentralization, it's at odds with privacy, efficiency, or both, and maybe even more. [00:04:06.000] Cool. [00:04:08.260] So one of the results of this is something called priority gas auctions. [00:04:13.780] So these arbitragers, well, okay. [00:04:16.800] So one more step back is that once Ethereum started to actually have exchanges of values, like decentralized exchanges, what would happen was you'd have a lot of arbitragers looking for like just arbitrage opportunities on chain. [00:04:33.360] So they would see the same asset in a set of pools at a certain price, and then another asset at a different price. [00:04:40.320] Or they could do front running. [00:04:42.240] So they would look in the public mempool, see I was about to make a $2 million trade for some meme coin, and then they would front run me. [00:04:50.640] So they'd write their own transaction. [00:04:53.100] And then they would put it, they would try to get it in front of that. [00:04:57.160] And then mine would get included, and then they'd try to sell it. [00:04:59.960] Or there's sort of like various other ways. [00:05:01.560] That exact example is called a sandwich. [00:05:05.320] And so, yeah, so, well, what is the point here? [00:05:13.920] Yeah, this slide is a bit, I think I jumped ahead, but that's okay. [00:05:21.920] So, yeah, yeah, okay. [00:05:23.720] So really one of the key things that happened here is like, okay, the UX of the chain sort of sucked because you could get front run. [00:05:30.520] But on top of that, you have all of these people watching these transactions in this auction. [00:05:37.080] And then sort of ignoring the like arbitrage example, like just, this is just for like competing against other people. [00:05:43.800] They watch, and then they increase their bids like just a penny more, often called like pennying or like strategic bidding. [00:05:52.400] And the result here is that, well, one, you don't really get efficient prices on things. [00:05:59.360] Like people aren't bidding their true value or like this property in auction theory called like truthful bidding. [00:06:05.940] So they set the price much lower than they actually value that opportunity. [00:06:10.800] And then they just add pennies on top of it. [00:06:13.680] So this doesn't get you the most efficient price. [00:06:15.980] This means the network is not properly pricing in the resources that it takes to actually facilitate all these auctions. [00:06:23.860] The most simple example of this is just like bandwidth. [00:06:26.840] So if you send a thousand transactions over the course of a second, that is a lot of bandwidth for the network. [00:06:33.320] This is also a price that, you know, never gets recompensated really. [00:06:39.560] And then again, yeah, bids are public. [00:06:41.280] So that sucks. [00:06:42.540] Another thing too is that losers still pay gas on reverts. [00:06:46.660] So if you sent a transaction to get included and it wasn't actually included, or sorry, you didn't win the opportunity, you still were inserted on chain and your transaction just fails, which means you pay a little bit of gas. [00:06:59.020] And so there's a lot of things about this that are not ideal from an auction perspective. [00:07:04.540] Also, users eat the congestion. [00:07:08.360] Meaning that if they want to now get just a regular transaction in, they need to compete with this like stampede and a bunch of other things. [00:07:16.440] Okay. [00:07:16.820] I'll fast forward since I'm already, I think I have time, which is fun. [00:07:21.760] So yeah, we call this like set of properties MEV, maximal extractable value. [00:07:26.300] It's really just a value of a system. [00:07:29.560] It's supposed to mirror the sort of like potential energy to extract all this value in a system from all these different ways, which you can observe the ordering in the chain and sort of like exploit it to create value. [00:07:45.920] Skipping ahead, we introduced something called MEV geth. [00:07:50.120] It was a way to sort of mitigate a bunch of the negative things from this. [00:07:56.720] The first was we just created a new bid type. [00:07:59.500] So if you wanted to sandwich someone now, instead of competing with these horde of transactions, you could actually just send that preference in via this new bid type, which I'll sort of add some caveats. [00:08:12.180] I'm kind of baiting you. [00:08:14.200] Yeah. [00:08:14.840] So you could send in, oh, here's the three transactions, my first one, then the last one and the person I want to sandwich. [00:08:21.740] And I want it executed in that order. [00:08:23.660] On top of that, we changed the rule from this client in this client's perspective to not order just by the gas price, but how much value you gave to the actual miner. [00:08:34.740] So if you were able to extract 10 ETH in value and you decided to bid 9 ETH to the miner, well, then we'd put you at the top. [00:08:42.520] Assuming no one bid more. [00:08:44.020] Assuming no one bid more. [00:08:45.840] And so, yeah. [00:08:48.560] Yeah, I think what I was baiting there as well is that, so, you know, there's many different types of arbitrages. [00:08:55.720] I'm using sandwiching as one example. [00:08:58.240] But a lot of people were quite upset when we did that. [00:09:01.400] And so, you know, naturally they're like, oh, you're doing this as like a harmful thing to users. [00:09:06.700] But the reality is in a permissionless and decentralized system, there is no other option. [00:09:11.860] So if you can't get rid of it, actually, like protocolizing it is your best option because now you can become aware and at least return some of the value. [00:09:21.540] And that also lets you focus on how to stop that problem upstream. [00:09:25.940] So, yeah, the lesson there, I think, which was shocking for me is, you know, pseudo protocolizing the bad thing is better than ignoring it. [00:09:33.380] And I say pseudo protocolizing because this actually wasn't in the Ethereum protocol. [00:09:38.300] We did this totally out of the protocol. [00:09:40.240] When we introduced this, 90 plus percent of the hash rate adopted it within a few weeks and months. [00:09:47.520] I think within a few months, almost all the gas wars were essentially sent to zero. [00:09:51.960] So that's great. [00:09:55.080] Yeah, on top of this, it created like sealed bids for competing in this auction, which is great. [00:10:02.260] And yeah, I think the sort of fun thing here is like ordering rules in a distributed system are very hard to enforce. [00:10:08.660] Every single node sees a different series of events. [00:10:12.120] And that's kind of a known thing in distributed systems. [00:10:13.980] I'm at 10 minutes. [00:10:14.840] Should I wrap up pretty soon? [00:10:16.360] What's the... [00:10:19.280] Okay, okay. [00:10:20.080] All right. [00:10:20.440] I'll try to wrap up. [00:10:22.800] I'll do like a few more. [00:10:24.280] Okay, so I'll just double click on this. [00:10:26.560] There's a lot of text on this. [00:10:27.820] But seeing like a protocol is... [00:10:30.580] It's super important. [00:10:31.800] There's this great post from Barnaby at Ethlabs where... [00:10:38.740] Yeah, I think... [00:10:40.020] So a great example of this is like, okay, we want to say that our system has some fair ordering of transactions in this example. [00:10:48.900] And so one, like fair is extremely objective. [00:10:53.100] But two, if I have 10 transactions and I want to say like, oh, I ordered these in a very fair way, the only way for a network to agree on that is to first agree on all the transactions that are being considered and then to run the mechanism over it and then agree with the results. [00:11:12.780] And so the challenge here is sometimes Ethereum has over a million transactions in the pending queue. [00:11:18.360] And so like, you know, this is sort of almost... [00:11:21.100] Yeah, I'm sure there's some impossibility result related to this. [00:11:23.900] But yeah, it's just there's no way you can come to agreement on the million transactions just bandwidth wise before you do that. [00:11:32.200] Narwhal and Tusk are a pretty cool, like set of consensus mechanisms which innovated on this. [00:11:38.300] They did something cool, which was like, they don't get total agreement on every transaction in the network. [00:11:45.040] But they get a bunch of different partial agreements. [00:11:47.640] And then they pick a leader to sort of find a path through all of those. [00:11:51.420] I think that's really cool. [00:11:55.300] Cool. [00:11:55.940] I'll skip that. [00:11:57.640] Yes, this is another fun one. [00:11:58.800] So searchers ended up still spamming the miners. [00:12:02.060] It's just way less spam and more efficient spam. [00:12:06.080] But they were getting very tired of having to keep up with, you know, adding all this infra to support all these transactions coming in. [00:12:13.500] And so this is a fun way to sort of think about the limits of like Web 2 attempts to reduce spam. [00:12:20.340] So I think so we started offering a product which would accept all of these bids coming in from these arbitragers and then send those to the miners and not pick who won. [00:12:30.420] We just kind of filter out all the spam and then send it to the miner. [00:12:34.980] This was great until then we started getting spammed, of course. [00:12:39.020] Very predictable. [00:12:40.480] And so it was called the bundle relay. [00:12:43.080] And so like the first thing we do, like anyone else, if you've ever done any Web 2 infra or software engineering work, it's like IP rate limiting. [00:12:51.820] And so that's great until it costs a few cents to spin up a new IP. [00:12:55.900] The next thing is account-based rate limiting. [00:12:58.860] So we, you know, only allow each Ethereum address to submit a few transactions per second. [00:13:04.520] That's great until it costs a few cents to make a new Ethereum account. [00:13:09.240] Yeah, all of these things, all these existing algorithms, token bucket, leaky bucket, sliding window, they're all sort of gamed in the same way and it's very hard. [00:13:18.280] There's also just tons of things at odds with rate limiting and the idea of decentralization. [00:13:25.520] So I think anything we really wanted to do, people were like, oh, you know, you're sort of exerting this like unfair rule on the network. [00:13:34.340] Also, it's hard to make the rule totally public because then that gives away how it can be gamed. [00:13:41.540] So, yeah, the way we ended up solving this was with reputation. [00:13:45.340] And there actually are a bunch of things like add auctions, which use similar techniques. [00:13:50.000] But really all we did was just kept track of how many times you gave us, you actually generated meaningful value for the block. [00:13:57.500] And then use that as a reputation mechanism. [00:14:00.840] And, you know, the first thing we did was like have a high priority queue and a low priority queue. [00:14:05.300] So anytime you've ever sent something with even a little bit of value, we put you in the high priority queue. [00:14:11.140] If you've never sent us anything, you go in the low priority queue. [00:14:13.800] And we sort of had two people building blocks. [00:14:17.560] Yeah, so this worked. [00:14:18.900] It was awesome. [00:14:20.260] It eventually got too expensive to run. [00:14:23.040] And also after we introduced it, about a year later, we never saw any of these attacks again. [00:14:28.080] So it's kind of funny how just knowing about the deterrent stopped people. [00:14:32.860] We sort of scaled it down a bit. [00:14:34.940] And it actually ended up not being a problem. [00:14:37.380] It popped up again on L2s. [00:14:40.380] But, yeah, there's a cost as well as always. [00:14:42.920] There's always tradeoffs. [00:14:43.660] So this mechanism, you know, it's basically a hidden boost to your bid. [00:14:50.300] And there's a lot of, like, auction theory saying how this is, like, not the most sort of efficient auction design. [00:14:56.880] But, yeah, so the lesson there is just incentivizing good is generally more successful than punishing bad in these types of systems. [00:15:07.040] I will skip through. [00:15:11.160] Okay, maybe I'll say very quickly on this one. [00:15:15.140] I won't explain this whole one. [00:15:16.760] But basically, you know, this was, I think, a year ago. [00:15:21.800] No, actually, maybe two years ago. [00:15:23.420] What had happened was basically there was one of these sandwichers who actually got this idea where they spun up a bunch of validators. [00:15:33.860] So they stake 32 ETH per validator. [00:15:37.100] And the Ethereum protocol is designed this way because, again, it's trying to be permissionless. [00:15:42.400] But if anyone could spin up a node for no cost, then you could spin up a million and attack the network in 20 seconds. [00:15:49.780] So one of these arbitragers spun up a bunch of Ethereum nodes. [00:15:53.400] They waited until they were about to build a block. [00:15:55.500] And what they did was they sent a bunch of transactions into the public mempool that were easy to sandwich. [00:16:02.300] And then a bunch of sandwichers saw that. [00:16:05.320] And they created these sandwiches. [00:16:07.560] And then they, you know, it ended up in a block and was given to this sandwichers validator. [00:16:13.320] And then what they did was they unbundled the block. [00:16:15.740] So they took out all of the other sandwichers transactions and then sandwiched all the sandwichers. [00:16:23.860] This is actually a violent. [00:16:25.400] The way they did this was like, one, it used a little bit of an exploit. [00:16:28.820] But then, two, it does violate a core part of the Ethereum protocol, which is that a validator cannot propose two blocks to the network. [00:16:37.160] It's called an equivocation. [00:16:38.220] So they did end up getting slashed. [00:16:41.260] But they only got slashed for about two ETH. [00:16:43.780] And they made, I think it was 30 million. [00:16:46.280] But it was many millions of dollars. [00:16:49.720] So, yeah, always, always incentives rule everything around you. [00:16:53.180] And I personally am just skeptical of anything that uses slashing. [00:16:57.660] It's a good deterrent. [00:16:59.740] But it's, I, by no means ever think something is secure just by having a slashing condition. [00:17:08.960] Okay. [00:17:09.980] I think that's where I'll end. [00:17:13.640] Maybe one more is just, yeah, double clicking is like fair ordering or fair system designs are like fair to who, whom. [00:17:22.200] You should, yeah, make sure you add that in your statement when you're saying your system design is fair. [00:17:29.580] Yeah. [00:17:30.200] So that's about it. [00:17:31.160] There is no perfect protocol. [00:17:33.060] Security is always a cat and mouse game. [00:17:35.960] There are always tradeoffs. [00:17:38.900] Okay. [00:17:39.400] Wait. [00:17:39.540] Actually, sorry. [00:17:40.020] The one last thing I had to include a teaser. [00:17:41.820] I'm so sorry. [00:17:43.460] Some of my recent work is more on the AI agent realm. [00:17:47.160] But what I essentially did was I wanted to see if agents would naturally symbol each other. [00:17:53.560] So I made this experiment with 200 agents. [00:17:55.900] It's a simulated economy. [00:17:57.300] I tell the agents maximize profit, but also there's some regulators. [00:18:01.660] So if you get over 35%, you're going to get penalized. [00:18:05.500] It's like an economic-based simulation. [00:18:08.920] And so literally in 90% of the simulations, the agents start creating fake identities to avoid getting penalized. [00:18:17.020] I kind of set them up for a trap, to be honest, as a very simple example. [00:18:22.160] But yeah, I think it's a fun... [00:18:27.260] I think there's a bunch of fun, like, interesting R&D directions here around how to define, like, safety properties for a swarm. [00:18:33.840] And you can also even think as anti-monopoly rules as, like, actually trying to create safety properties for markets. [00:18:40.920] So it's all related. [00:18:42.100] Okay. [00:18:42.600] I'm way over time. [00:18:43.620] Thank you. [00:18:44.660] All right. [00:18:46.860] Chuck, two questions. [00:18:49.140] You already asked a question. [00:18:50.980] Someone else. [00:18:53.020] If no one else, I'll come to you. [00:18:57.560] All right. [00:19:00.000] It sounds like you spent most of your time in blockchain but are experimenting with AI. [00:19:06.360] What, in your view today, are maybe the most challenging, like, market-based issues in some of the AI systems that you've seen built today [00:19:20.760] or have the most overlap with some of the things, projects you've worked on here? [00:19:25.160] Yeah. [00:19:26.260] My cop-out answer is I think the next talk actually goes into this for inference markets. [00:19:31.040] So all that stuff. [00:19:34.260] Two, also sort of feels like a cop-out answer. [00:19:37.600] But I guess I think, like, the sort of value exchange we're making with a lot of Frontier Labs is, like, not transparent. [00:19:46.600] So it's almost, in a sense, we're sending bids without knowing their value. [00:19:54.800] So, yeah. [00:19:55.580] I think that's, yeah. [00:19:58.000] You know, we're sort of giving a lot of our thoughts on how to, you know, do what we do. [00:20:03.700] And then they're RLing on it. [00:20:05.740] And that's not exactly clear how valuable that is. [00:20:11.580] I'm hearing lately that, like, a medium-experienced software engineer's trace for one multi-hour session is worth $200 on the open markets. [00:20:28.940] All right. [00:20:29.720] We'll go here. [00:20:30.540] And then we'll go to the next one. [00:20:32.080] Y'all can always find DMARS afterwards. [00:20:34.660] Yeah, yeah. [00:20:35.340] DistributedMARS on Twitter. [00:20:36.340] Happy to chat. [00:20:37.020] So, I'm a consultant. [00:20:42.220] And I've been speaking to some of my TradFi and healthcare colleagues. [00:20:48.080] And they are, like, very hesitant to engage with AI agents, especially with the swarm principle. [00:20:56.960] Because they said that, Nune, we can't afford to screw up. [00:21:01.500] Like, with respect to TradFi's, like, if we fuck up someone's account, we're going to be held liable for that with respect to governance and compliance. [00:21:08.440] And if it's healthcare, like, we cannot afford to, like, screw up potentially a patient's case. [00:21:18.400] And I was wondering, in that regard, with regard to swarms, like, do you have, like, some sort of suggestion with regard to safeguards, possibly? [00:21:27.820] It's okay if you don't. [00:21:29.040] But I know this is something that has been an intense point of discussion, especially with highly regulated industries. [00:21:38.760] Oh, great question. [00:21:42.220] Yeah. [00:21:43.460] So, with swarms in general, one of the things I'm trying to understand, actually, like, first is just when are swarms even good? [00:21:51.520] I think there's some papers on this, but I still am not totally sure when you should reach for a swarm versus a super powerful single agent. [00:22:03.500] In terms of safeguards, yeah, I think probably you should just assume everything you give to the agent will get leaked. [00:22:14.020] And if there's something you want to be done securely, you should probably not do it with an agent. [00:22:21.740] Yeah, I know that sounds scary. [00:22:23.140] But I think we'll get there, and the safeguards will be improved. [00:22:27.020] But especially as, like, mechanistic interpretability is getting better, I think we can actually make claims over what's happening in the weights given a set of inputs. [00:22:37.500] But there's this other concept, too, called, like, a sleeper agent, which I've done some work on recently that Anthropic posted about two years ago or something like that. [00:22:47.740] But it's just, they showed you can very easily have, like, a capability in an LLM which acts differently in the presence of a trigger phrase. [00:22:57.240] So it writes programs, you know, fantastic, while the system prompt says the year is 2024, and the second it hits 2025, the Python, you know, libraries it starts importing are from your malicious packages. [00:23:12.940] That's, again, also, like, there's no way to fuzz the weights for that. [00:23:15.980] There's no way to test for it. [00:23:18.220] Like, well, maybe if you've got infinite compute. [00:23:21.180] Yeah, okay. [00:23:23.180] That's scary. [00:23:25.840] Okay. [00:23:26.520] Thank you, D-Mars. [00:23:30.000] Okay.