[REDACTED] 2024 | The Future of DeFi in the AI Age

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:00

   right great great cool so we have half the chopping block I'll let everyone introduce themselves we have Dean from agoric systems uh Tom from dragonfly L from caliber uh and ton from Gauntlet and robot so why don't we start with you tun like in 30 seconds uh what do you do at Gauntlet what's Gauntlet about and then we'll do a clockwise yeah so hey everyone um I spent a lot of time

2. 0:25

   thinking about risk management and optimization problems within crypto also invest in a lot of kind of early stage uh technical companies and uh have just been around for a while I guess at this point so hey guys uh my name is ly I started kyber one of the earliest decentralized chains back in 2017 stepped out a couple of years ago and started caliber recently which is a venture Builder um mainly invest in new fs and build uh you know with them um and a lot of them are

3. 0:58

   like you know building defi and also work with n in a couple project everyone I'm Tom I'm a GPA dragonfly uh we're Global crypto Venture fund and uh we're early investors in near 19 triple I'm uh CEO and founder of agoric we do smart contracts to be able to do cross chain uh um orchestration and to me that's the next generation of defi so pretty cool so the topic of the panel is the future of defi in the AI age um I mean it's AI is kind of hot

4. 1:32

   right now D is always been hot minus the past two years um I want to poke to the very basic uh fundamentals do you expect any use cases Beyond just yield navigation to emerge at the intersection of AI and crypto um or do you see this meme trend of goat yield navigation like high level stuff to be the trend for the next 6 8 months I'll throw that to Dar first yeah so I mean I think in general there's a lot of

5. 2:05

   um shock and awe and hoodwinking in most crypto and AI stuff right now it's like generally people gluing together stuff that the AI is not really doing real crypto transactions um arguably you know the telegram Bots like Bon bot and stuff that do transaction identification looking from malicious hacks they're doing much more real machine learning than like 90% of the uh like AI type of stuff um but there is kind of this natural phenomena I think that

6. 2:34

   occurs which is there's almost like you know life imitating art like people have now had this dream of these creatures these metaphorical creatures or the these agents using crypto and doing these things and that has inspired people to start building stuff so I do sort of think a little bit like defi in the Ico boom where there were a lot of projects where like they're like trade

7. 3:00

   without using a centralized exchange and then a lot of them were sort of Half Baked or didn't work or whatever but then eventually I inspired people to find things that worked I think that's the stage we're in for for this so that means there will be other use cases there's just generally kind of I think this like boom and bus cycle we're kind of like on the way to bust right now yeah um that's an interesting point and by the way nothing that we discuss here is financial advice Anders only

8. 3:25

   represent themselves so uh Tom I'll I'll throw this to you um do you see a billion doll D5 protocol emerge in the coming months with AI at its core uh like and what would that look like probably not um I think so far I mean most of the breakout defi uh projects from the past year have been sort of new oneofone use cases like I would put Athena for example in that bucket um or you know hyper liquid for that matter I mean this isn't a a market that people trying to conquer for a

9. 3:57

   while and now um you have players actually being able to conquer it but what you're not seeing is a replacement of the Old Guard you're not seeing a someone out Unis swap Unis swap you're not seeing someone out a a um I don't think AI necessarily changes that Dynamic um but maybe there is some whole new category that AI can unlock that we're just not seeing right now I think if you are going to see a billion dollar outcome that's probably how it's going to be but I don't think we're going to

10. 4:21

    see you know Unis swap plus AI equals you know new UN swap that that seems unlikely to me sounds good um so Dean uh when when we talk about Ai and defi um I mean I personally give this example of agents are basically solvers executing intents um like how do you describe uh the emergence of AI and do you see any like sort of overlap in the coming months with Dy especially in the intent realm so so I see several places where AI is relevant um right now from the

11. 4:55

    perspective of like complex software systems what we see in defi what we see in BL generally are very simple right I mean tens of billions of dollars great but it's a swap it's really really simple right dexes are really valuable but really really simple and so the power of AI to actually you know make some material difference it's going to do well in a more complex environment so I think it's a non-trivial contributor to a next generation of more

12. 5:19

    sophisticated stuff on chain but um you know where I'd like to see it some of the elements I'd like to see are help in analyzing programs that are there security analysis help in producing good robust Rich smart contracts where I can understand the output I can understand what it's going to do and it did something custom for me I said I want a strategy that's going to do the following yield when it crosses this

13. 5:44

    thing then build up a position over there and redeem my rewards every day every day and go off and do something with it right write the program for that and make sure I can understand it right you know so there's a lot to there's a lot of value to be added there but but we're but we're a long way from an AI agent that's doing trading on my behalf onch yeah so L now over to you um whatever you've heard so far it seems

14. 6:07

    like there's a fair amount of skepticism you know while Ai and D5 it may sound like we're forcefully trying to converge them there might be a healthy amount of skepticism from your point of view you incubate a bunch of teams in Asia primarily do you see this as a primary primarily a Silicon Valley phenomenon where you know something gets funded by Mark hry it goes to the moon or do you also see consumer apps or mass Market

15. 6:31

    sort of D5 Founders exploring AI as well um I could say I have seen several use cases of uh you know AI SL machine learning in different like consumer AI application right like for example people use uh you know basically use AI to analyze the user wallet to recommend them like you know hey you should look into like you know similar asset that has some you know that match some

16. 6:54

    certain like properties that you are looking for right or uh recommend like you know what's the next like protocol that they should try to get like better use and things like that um but I think um the recent um experiment with go right the the the me call started by sorry not Financial advice um the the the the the new M call I think it's a really good experiment for a s AI in which you know basically people want um

17. 7:25

    basically like you know like if you look into AI defi there are a lot of things that people want to do in defy but cannot do right because there's human um factor involved like it's not only because of technical risks but also like legal risks right like this for example say decentralize fund management right like how do you decentralize it because end of the day someone still need to make the decision what to buy when to buy uh how to allocate asset right but

18. 7:51

    if you can replace that with like some sort of like AI factors right that can just like automate and like you know can make all the decision right based on some predetermined algorithm then I think that that can enable like a new whole application of of Defi and I think that is the future that I'm looking for and really like looking for people to build that so that you know we can be part of as well awesome um now I want to

19. 8:15

    open it up to more uh I'll say I I said there were a bunch of uses for for for crypto and and defi the other direction and I don't know whether you want to explore it is the opposite side of AI as as partly um defy and blockchain as a enforcement of relationships among AIS right you you long ago you might have heard the phrase right on the internet no one knows whether you're a dog right

20. 8:40

    in web 3 when you're trying to deal with a party and you've got a smart contract that is operating to enforce the terms of your relationship you don't care if the thing on the other side's a super intelligence right you can still do a deal and make a trade even though yeah it's going to win some clever thing that I just don't understand but I can at least participate in the market with it it and so you know as AI moves towards

21. 9:03

    um agent destructuring where my doctor has an agent that is doing analysis of my medical history and discovers or not me but discovers there's something that a specialized agent that could analyze heart information could be a win but that's a private agent that they don't want to share their model with the doctor and I don't want my data shared with that that vendor and how does that all manage and how does this ai go

22. 9:29

    contract with that agency out there or that other AI agent all that should just be able to happen and happen in a trusted structured enforced environment that blockchain could provide it and so there is a future economy where much of the activity is AI agents conducting business on in in a web 3 kind of environment because you know it's a kind of hard for them to have a bank account but they can perfectly well have a crypto Pur and so I think that that's a

23. 9:56

    future that's not 6 months from now and not 8 months from now but it is one of the most important intersections of AI and uh crypto is is is that as a business environment for oh so before I go to my next question U you raised a few interesting points I mean trust and coordination have been challenges for humans for ages um blockchain in particular you have a huge coordination problem um would any of the panelists

24. 10:21

    here uh agree or disagree with uh AI being facilitators for coordination or do you think that's kind kind of A disruption we shouldn't really explore I mean I think it's very good for removing the existing structure right because you might have some existing structure that says you have to take some sequence of steps in order to come to agreement right like a consensus algorithm here you're saying what if we didn't and we asked the third party Oracle here's a question right now obviously the the input to that Oracle

25. 10:55

    could be biased the way you structure the query could be biased as you know from using all the different llms right like 99% of the difference is not even in the weights anymore it's like in the rag front end decoding stuff at the outside which is very custom to each use case but there is a sense in which it gives you a third party thing and to to ly's point we kind of have all these legal structures that have almost prevented us from doing things because there's effectively some

26. 11:24

    some existing algorithm that says yes or no this is an acceptable or unacceptable trade-off but in a world where the agents are doing this amongst themselves that's a way for us to learn oh maybe these are the flaws in our existing rules and this is why this type of thing might be more beneficial so I I think like there's there's some of that where it's like we kind of don't know enough to to be art so on the on the policy piece that you just raised up around

27. 11:49

    like legal uh ambiguity I mean even defi U regulations around defi are kind of ambiguous around the world um so I'm curious how would Regulators view agent to agent agent to agent trade or agent to agent onchain Define interactions like how do you see that evolving over the coming years maybe Tom ly as long as they're not uh transacting in derivatives you know I think it's okay no onion Futures allowed to be traded by yeah AI um you know you don't have uh it

28. 12:20

    depends I guess on how far the liability goes if uh you know an agent is working on your behalf and and and whe it transfers on to you but I mean ultimately I think yeah governments around the world don't even know how to regulate uh open source AI or AI contributions or you know what should or should not be um in sort of the training set for a model and so um I feel like the intersection of crypto and AI is again not six months out not eight months out it's probably you know

29. 12:43

    several years out from uh you know impacting uh sort of policy creation cool um any thoughts about that ly well I mean like I think even like for defi regulation now I don't think any country has any like Clarity on it right I mean like we seek for opinion for like Singapore like you know a couple of other countries as well none of them has any like clear regulation about defi right so like Asian to Asian

30. 13:07

    kind of like regulation I feel like you know it's still going to be like a few years away yeah but AI is moving sufficiently quickly that probably AI will help people to write the regs that will then well I me I I hope I hope that's not true for too soon cuz hallucination on regs sounds is a real problem yeah yeah so we were talking backstage uh um and I jokingly said I want to bring out philosophical disagreements so why don't we um start with some hot takes about

31. 13:38

    the convergence of AI and crypto let's go anticlockwise this time around Dean do you have any hot takes on AI and Ai and defi um you know do you see what's happening right now with goat or some other onchain like hedge fund experiment with AI 16z um what's something you think um is unsustainable there or something that we should continue exploring well a a little bit related to the hallucination that that run just mentioned people will try to apply AI to writing smart

32. 14:10

    contracts and that will be terrible because you will because it's already B you know the what you want to have happen and what uh solidity for example can safely Express from the point of view of of interesting Rich access controls and rights are such that you know you can build stuff that humans just can't understand you take two simple components put them together and you end up with an unsafe system and so having an AI write that stuff that we then can't really be sure of what the

33. 14:42

    properties are is a terribly dangerous place to be analys of these things great or writing things onto infrastructures like agoric maybe move maybe a few others that provide much stronger enforcement of of of of security rules will be a much better place to be and my fear is that people will apply it to these systems that are much less uh that have these security hazards that are too

34. 15:06

    dangerous yeah I I think those are fair points I think um fundamentally crypto Andi are sort of budding heads with traditional legal system for the main reason being that um the legal system has these these two invariants in it that they assume to always be true and crypto I book them both not true which is you know you need an intermediary to transact um or you know publishing code is equivalent to publishing free speech

35. 15:31

    um and oh like this thing doesn't require either of those how can that be true this doesn't work and so every time you see it budding heads with the traditional legal system is because the legal system can't process the fact that like this thing breaks sort of your sort of your core sort of fundamentals and so um you know the way they sort of get around it is by these sort of third party touch points of oh it's a safety issue of oh it's you know the website is a broker thought that obviously makes no

36. 15:53

    sense but they have to find some way to sort of jam it into this these existing definitions I think the only way make meaningful progress is if you're willing to throw out these old definitions and start fresh um and look at things from the perspective of 2024 if that's what does it mean to build or run a model what does it mean to build or run a protocol um and I think until we actually see that this stuff is just going to kind of keep hitting uh um hitting hitting Road

37. 16:18

    Blocks sorry what's the question again just hard takes on like what Ai and Def is all about um I think eventually a lot of the operator roles in several defi protocols now will be replaced by AI mhm I think one is a lot more safer and a lot more automated and more precise remove like all the human error Factor uh and secondly it really like makes like you know all the defi protocols like decentralized in a way right um and I think um uh as I mentioned earlier um

38. 16:52

    it also like make it a lot safer for def5 Builders to launch new initiative if they know that they several AI agent that they can plug in for example for risk analyzing right uh for like liquidity analyzing um and that they can just like plug in and or or even like governance right that they can plug in to sort of like build that def5 protocols so that they don't have like to hold the admin key they don't have like to like contact like you know other parties to help them like with like you know all the Safeguard and things like

39. 17:23

    that um so I'm quite excited for that about that direction yeah yeah I guess I would take maybe like I agree with that in like a longer term Horizon like longer than I think uh the crypto Market attention span Horizon um I think in the short term though I'm actually generally pretty bearish on on um what currently people will call AI for uh onchain systems mainly because the robustness issues are quite high and also fundamentally I I I think like from a a

40. 18:00

    numerical perspective a lot of the machine learning algorithms that are manually tuned will generally do better than like the autogenerated ones from from large language models I think the key where large language models are really good is like I have some unstructured data I don't really know how to compress it into something that's a good embedding that I can given that embedding I can do some numerical math on and I just think

41. 18:24

    like we haven't figured out how to do those things without all of these like poisonous attacks and and trust issues and also the fact that if you know if you you look at like the Llama 3 code base evolved over time you will see that like it's started as like here's a Transformer and here's some light stuff uh that converts your input query to format for the Transformer and here's a an output thing and what you've seen is the code bases have gone from like 80%

42. 18:50

    neural net to like 10% neural net and the rest of the stuff on the outside is growing that stuff has a lot of subjectivity in it and so I I I I think like there's a little cart before the horse still for for for a lot of this stuff I think like a little bit like I was saying with the the goat type of thing where it's like it's not really trading on its own it was just getting air dropped by people and then people started kind of believing this which is cool right it's a that's the it's it's

43. 19:15

    almost like performance art it's like an art form that inspires something you want to build but I I think like if I look at the code bases and then I look at the the marketing the the Gap is quite large and that's just my the fact that you said it instead of HE though means that the marketing worked and uh you believe there you go there you go I mean he him so um staying with the investors for a bit uh

44. 19:39

    tun do you see AI to be embraced by risk management and risk modeling research funds like Gauntlet or do you see that to be a disruptor disruptor where anyone can run their own Gauntlet for like finding the best parameters for uh M of capital the liquidity thresholds like how does that look like I think it if it it's going to look like what you see in normal Finance right which is like very few models that are deployed in production that need X9 of operational

45. 20:13

    stuff generally are like high capacity non-interpretable models and the main reason for that is whenever there's an error like a hack like something goes wrong you if you can't interpret what went wrong and you can't reproduce the exact environment under which the error propag at you have a very hard time of doing emergency uh kind of response now I think you can get to that point but like I said I think it's like the robustness of the the language stuff is

46. 20:39

    like so poor whereas the robustness of like you know the traditional models you would use in in in trading for instance for for for risk in the same way those are much more robust right and like the relative Gap is still high now like do I think in 10 years the Gap will go down for sure but I I think right now it's like it's not you don't get the same statistical guarantees um but you know will will all of that happen sure I just don't think I think it's like again

47. 21:07

    there's a little bit of like one tiny thing worked and everyone thinks like the ocean is going to work tomorrow but it's like a lot of incremental steps to get there so Tom um does AI Aid in capital formation on chain because we've heard from tun there might be angles there for enabling more efficiency uh in the future but from an investment point of view um and capital formation from

48. 21:29

    the ground up does AI Aid that in any way yeah I think like trun was saying I think some of the most maybe directly relevant applications for AI and crypto today is like due diligence and security around different smart contract systems but a lot of those can also just use traditional uh model use like a classic you know sort of supervised machine learning model you don't need like an llm to know hey this this smart contract have an admin key or not um maybe over

49. 21:56

    time again like also just uh the uh uh spectrum of possible inputs is probably probably going to keep growing we'll see you know different VMS different languages different chains and so therefore you'll need something that can actually like um you know take all those as as inputs and tell you more about um what's happening or even something that's more um Bottoms Up of hey just knowing sort of the outputs of this contract um what is it doing I don't think we're there yet and so I

50. 22:20

    don't and also frankly I think maybe the overall level of sophistication of you know Capital Market participants isn't quite that high but over time yes certainly um I think can just help sort of create the right set of guarantees around transacting on chain I think already in the way that you kind of see you know some of the kind of simulation based software do that today I would suspect that'll only get more and more sophisticated where you can say hey I want this set of guarantees to be true about my transaction um and you know

51. 22:44

    hand it over to a a you know bot to sort of take over um now let's move to uh Dean and L uh and take on the protocol sort of research hat you guys know the various dynamics of scaling applications on a layer one better than me so what kind of AI or defi use cases are you most excited about given the discussion we've had here around considerations for security uh regulations do you see any immediate

52. 23:15

    like novel Greenfield ideas uh emerging and what would you be personally excited about so you know background you know one of the one of the things that I've paid attention to for a while is AIC sa kinds of things and so I care a lot about what people end up doing with that but the other side is usability I mean the thing that that that where where the llms catch on more is usability and

53. 23:41

    access to something that used to be opaque to a much more general public and a much less expert public and so you've got you know this to some extent does amplify some risks because you know I'm imagining uh you you see several efforts right now of hey what's a strategy for me to be able to do this thing and get some Dogecoin or what's a you know or or there's this you know stable coin

54. 24:06

    liquidation Market I'd like to participate in it what do I do you know help me help me out here right and and people are working on that or just arrange you know right now we've got this you know connected world but you kind of have to hand carry your assets around so I've got a thing over there get me that nft you figure it out right and there's a lot of people that are presenting that and sometimes it works and sometimes it doesn't um and so I'm

55. 24:31

    excited about those getting more to the point where they do but also back to the you know the whole reason why smart contracts excited me is that they enforce the terms of contract like Arrangement between parties and so having the infrastructures we're building enable people that are um you know that are using these tools to both be protected but also for the other people they're engaging with to you know again I don't care if you're a super intelligence or an idiot with an AI

56. 25:02

    either way you know I'm doing a trade and it's going to be safe things are in escrow if they go wrong I get my money back you know I want those kinds of infrastructures and so I think it will enable more people to be able to participate in a more sophisticated way with the with these networks I think will enable more chain abstraction of some stuff that that that um that needs

57. 25:25

    to be there I think the right way to do that is to generate programs that can then run on you know a platform like aor that can do cross-chain orchestration but there are plenty of ways to do that offchain that go there and so I think that it will provide some some usability that will help raise the overall value of of the crypto ecosystem and raise the access of it into the rest of the the trafi world uh L do you see builders

58. 25:50

    gravitating towards more fundamental ideas U that we're exploring right now in this discussion or is it still a mem meme Market out there uh I mean like I think even for meme like I think I I started like seeing like people like trying to utilize AI to build uh uh more I guess useful product for users this for example like you know like in the world like all the new memes what kind of meme that you should look at but like you know I think you can like I I really see like telegram boss that recommend different kind of like

59. 26:20

    memes for user to look at because based on their previous activities based on their like wet profiles and things like that um I really look forward uh to uh people who can use AI to replace like all the like you know centralized operator right like for like say like you can certainly run an onchain fund uh that is driven and managed by AI U and of course like how to make sure that this AI really runs by a decentralized network but not some centralized

60. 26:47

    operator it's a different question right or like how to verify that this AI is actually built in a fair and like correct way uh like you know it's actually works uh uh you know according to the marketing stuff right um and yeah I mean like I I feel like eventually AI will become you know the assistant to the users to navigate like throughout the defi world and also like can be like

61. 27:11

    this the actual operator of several like def protocols that's the future I'm looking for awesome now we have less than two minutes left I want to give everyone 20 seconds for a closing thought um we have this running joke on Twitter uh where you know people say that crypto people talk more about AI than AI people talk about crypto what do we need to change or what needs to happen in the next let's say

62. 27:37

    one one year two years uh that changes that perception and we get more AI Builders leaning into crypto that so I'll I'll tell you for a fact from all the people I know who I worked with in the past who work in AI literally zero of them ever mentioned crypto but the best time best thing is my friend who's um you know one of the authors on the p paper and also on liberus he he's always like crypto a scam crypto is a scam the only time he ever messaged me and was like Hey Oh my

63. 28:07

    God what what is up with crypto was when bology made that crazy prediction about Bitcoin going 100 care million one million yeah 1 million yeah yeah that thing so that tells you how bad the perception is and I think the only way you flip that is you build something useful that people like actually that takes advantage of all the open Source models I would prefer we view the opportunity as there is there have been hundreds of billions of dollars shoved into giving you free subsidies for cheap

64. 28:38

    compute models whatever like use that to make something interesting for crypto versus like trying to attract AI people because I I don't I think they look at crypto and laugh almost 99.9% of the time um L Tom Dean any thoughts I guess decentralized AI on AGI I I think the two examples where crypto is still beating AI is uh one prediction Market's basically trouncing every single uh you know sophisticated model of this election um maybe that

65. 29:12

    will change over time but right now that's clearly a scenario where the markets beat Ai and um you know maybe there's something to learn there and two being frankly uh worldcoin and sort of pre of humanity being something that is I think going to be increasingly needed um in the future where we don't know if something's going to be you know AI generated or not and I think um uh having something that can basically do Global sub resistance seems like it's

66. 29:36

    going to be valuable I think that the big thing that crypto needs is to provide real user solutions for problems they have the ones that show up in AI that are most interesting that that crypto provides some unique value to are you know there's anol is agentic stuff where you're breaking out different roles and different purposes of models that want to be able to cooperate and you want to enable a framework of that cooperation such that they're not vulnerable to each other and and and

67. 30:01

    we've got a lot of Technology around that parts of that include zero knowledge stuff so I can verify that I'm you know that they're using a heart model to analyze the medical data without disclosing that model to me and various things like that and then the other is you're seeing a lot of evolution of having a bunch of different versions of the same AI configured as this one's a farmer and that one's an economist and that one's a math why

68. 30:23

    don't you simulate an argument for for some voting property right and so you're going to start to see utility for essentially Dow infrastructure for AI agents to be able to come to consensus using some of the kinds of consensus technology that we've all evolved and I think those are both useful places where we could actually bring value and then they go oh I don't I don't care about crypto but I care about this you're helping solve a problem that I actually

69. 30:46

    have thanks so with this we wrap up please give a round of applause to our panelists I hope you guys enjoyed we'll see you for day two tomorrow