Arie Trouw on why AI agents now need proof of action

0:00I'm Ashton Addison from the Crypto Coin Show, and today on Blockchain Interviews, we're back with us Ari Traub, co-founder of XYO. Last time Ari joined us, we dug into XYO Layer 1's major performance upgrades. The chain is getting faster and faster about AI hallucinations, providence, and ensuring accountability in AI as it's growing so quickly, it's hard to keep up. We want

0:25to make sure we're on the right track and it's not leading us astray and where that synergy between blockchain and AI can fit in and much, much more. Ari, welcome back to the show and thanks for taking the time.

Oh, thank you, Ashton. Thanks for having me on the show.

Yeah, you're very welcome. So, I'd love to start off with sort of a high

0:44level on artificial intelligence since we last spoke 2 months ago, it just continues in this exponential growth rate that it is getting faster and faster beyond PhD level. And with that also, I feel like intelligence is being commoditized and the cost of it is sort of going to zero. It's just becoming so accessible even in third world countries.

1:09So, with that happening, what becomes scarce and therefore valuable if AI is giving us access to something at almost no cost?

Well, well, you know, it's kind of like the internet, right? Well, when it first came out, you know, like data was much more accessible and stuff like that. Then we find things that are missing that we need to do. And AI is the same

1:31thing. We have this raw horsepower now where it's like a really sharp knife or a really powerful tool, but still controlling that tool can be difficult. And people have have shown that where you know, if you just kind of like let AI go off even that was demonstrated with the hugging face hack, for example, where it's like you give it a give it a

1:49purpose or a goal and you tell it to have at it without any sort of oversight, it's extraordinarily good at trying to reach that goal, but we have to have tools and harnesses and those sorts of things to actually have it do what we want it to do or to do something useful. A completely uncontrolled or for that reason just a untrained AI will do a lot of work, but get

2:12very little productive work done potentially at the end of the day. It's almost like having a super smart college graduate that doesn't really, you know, know the world go and just crank crank out a bunch of code. If the code doesn't actually work together, you end up with a system which is not very maintainable, it's not very good, stuff like that. So, really figuring out how

2:30to take something as super horse, you know, high horsepower and make it so it's, you know, a tool that can actually do something useful and what people want it to do is the challenge. And so, I think that also redefines to some degree, you know, where humans fit into that equation and the existential crisis or the existential questions people are

2:51trying to ask themselves about, you know, how the workplace is going to evolve from this, you know, what they thought before was the commodity which made them valuable no longer is, but now, well, what is the new commodity that makes them valuable?

Definitely. And with that Hugging Face example, we saw an acquisition this morning from Nvidia for almost 13

3:11billion dollars acquiring Hugging Face. And maybe you can dive into that a little bit more cuz I feel like a lot of people have heard the name and but they're not familiar with it and just about open source AI models and how that is different and maybe important from the oligarchy of LLMs that most people are using.

Well, the primary importance is

3:34sovereignty really. You know, that's one of the things that we talked about in the last show also, you know, sovereignty and providence are things that are very important. But to me, sovereignty is something that's important and one of the things you definitely cannot do with most of these frontier models is you can't run them on your own hardware and can't run them at

3:47home. Or, you know, at an office for example. So, I can't go out and buy a bunch of H200s, set up a data center or if they're just using some servers in my office and you know have my personal cluster that's data secure that's running Claude or ChatGPT's latest models. They do that for a few reasons. They want from a business standpoint they you know they

4:07want me able to make money off of the system. And two also it's it's hard to make something that works generically. We've seen this with Apple versus you know Windows for example. Windows where you have a bunch of different drivers or a bunch of different things. There are some costs to that and you know having your memory at pluggable for example makes it so

4:25it's a little bit slower. And where Apple people put it all on the same board you have just one configuration it's a lot easier to optimize but it doesn't give you that flexibility. So I think an advantage for the frontier model companies is that you know they're running it on a specific stack of hardware that they know they can control they can see the feedback. And also

4:41probably the most important thing for them is they get the data from us using it to be able to train on. And they get that feedback like where everyone is running their own models on their own hardware that doesn't happen which for many people's minds that's a good thing because I don't want my data necessarily to go out there where you know if you

4:58ask the new model of ChatGPT or Claude it knows something private about me that I accidentally you know had in a chat that it learned from. So there's big differences there but really the big difference is the open source ones are able to be downloaded and run locally. Now granted the bigger ones are you know they're going to require a 256 you know

5:19gigabytes of RAM or more and you know H200s a lot of horsepower so not you know average person on a laptop not going to be able to download those anyway and run them from a practical reason but you can if you want to.

Definitely. And on that note of you know the personal information I saw some updates in Claude that it's starting to differentiate and say at least on the

5:43front end that it's not saving your personal confidential information and or you can you can change those settings. Whether that's happening in the back end or not, I don't know if it can be proved.

Yeah, that's difficult to enforce, right? Especially when they have a agent that's running natively on your computer, right?

Yeah.

6:04

and even if you're running it in say, for example, a sandbox or on a separate server, you're typing things into it that might be your company proprietary, might be personal, might be those sorts of things. So, for it's kind of catch-22. For it to be able to do some of the fantastic things it can do, it needs to have access to information. Like, for example, if you wanted to help

6:22manage a calendar, it's going to know your calendar, it's going to know where you are.

Mhm.

and it you know, either don't do that or do it on your own personal stack where you know where the data's going. And so, I think a lot of companies are looking at that. I think they do offer the frontier model companies, they do offer setups where if you're large

6:40enough, and by large enough, I mean quite large enough, they'll set up a enterprise setup for you that runs the frontier models in the data center, which is isolated from the rest of the system and it doesn't trade off of that. But, to your point though, it's you know, they say they don't do this and they say they don't do that. It's it's definitely a won't and not a can't.

6:58And that's one of the things that I think fascinates me about a lot of cryptographic security and blockchain, that sort of thing, where there's a big difference between won't and can't. Like, for example, it's not like the Bitcoin Foundation is saying we won't take Satoshi's funds as well. They're saying we can't because they don't have the key for it. It's And you know, can't

7:20is something which is not negotiable, where won't is something which is negotiable or you know, potentially fallible.

Mhm. Definitely. And what are your thoughts on the frontier models? I've been reading about the recursive improvements where the models are making themselves better, and the open source models seem to be a little bit behind.

7:45And as you're saying, if you want to run the frontier models on the centralized versions, you need a lot of horsepower. You need warehouses of graphics cards. But there's been a discussion around slowing down the frontier models because of safety reasons and other reasons. And maybe the open source won't catch up, but do you see you know, potential issues or a

8:09need to slow down the frontier models that are growing so fast?

Well, I don't think you really can because you can say, "Well, we want to slow down the frontier models." But there's always going to be somebody doing something that's not necessarily in a certain location. Like, you know, for example, CRISPR. I'm sure we can make laws in state or in the United States for that

8:30reason about what can be can or can't be done with CRISPR. But I'm sure somebody finds someplace on Earth to go and do things which are not so great with CRISPR. So, I think the argument is always that the best way for us to defend against misuse of AI is to have our own personal powerful AI that basically can help us defend ourselves against that. And have

8:50its goal be secure the user or make the user happy as opposed to its goal being survival, for example. And so, I think we have to make sure that those frontier models that assuming that they're companies or their structures that we trust. And that's kind of the hard part where he's like, "Well, do we trust OpenAI? Do we trust Anthropic? Do we trust

9:09X? Do we trust all these different companies?" And I think Meta actually came up with a pretty good you know, 1.3 version of their model recently to kind of close the gap to the frontier models as well. So, we have a fourth one there. Now, granted they have some sorted history as far as data retention. So, there's some questions for them there already. But

9:31it's easier to trust those guys than it is to trust somebody who you've never heard of before or somebody who's in a jurisdiction which you can't control at all. And so I think slowing them down is probably a mistake because of the fact that all we're doing is hamstringing ourselves to a large degree. It's kind of like saying, "Well, we don't want advanced

9:48weapons, so we're going to have to slow down our own advanced weapons growth, right?" But so what it doesn't mean that the other guy's going to slow down his weapons growth, right?

Yeah, definitely. And on the notion of trust with the information that we're putting in whether it's personal information or just anything that we're typing in receiving

10:05information back from these LLMs, is there what aspects does does blockchain fit in that and can it improve the trust and that in what XYO is working on in ensuring not just hallucinations, but overall the validity of the conversation and the information that you're working with on AI?

Well, we just recently announced and we're launching this thing called crypto

10:28cards, which is a great demonstration app in partnership with Kate and Kate has this Kate.ai, which is their AI offering along with their the crypto stuff. And what we're doing there is kind of demonstrating exactly what you're talking about. Well, how does that integrate and how do we use cryptographic sovereignty and provenance especially with XYO XYO layer one to

10:53basically make the use of AI a little bit more secure and a little bit better. And what we do there basically is we use transparency. So I've always been a fan of transparency as security where like for example, a troll on social media, if they had to go and put their actual identity next to the thing that they said, they'd probably think twice about what they said, right?

11:15And so for example, if we have AI that's running and that AI is controlled by a certain entity, if you have transparency and an audit log of what it did and who actually did it and controlled that, the A, the repercussions could be even more easily followed through on it you can say, "Well, we know who did this and we can go and you know, make them

11:35stop or ask them to stop doing that or whatever it is." Or we can you know, if it's a bug for example, we can go and repair it. And so really what we do is we focus on those audit logs. So, when a game for example in CryptoKitties happens, what we do is we record all the provenance of the original hand, which model was used, what the prompts were for it, and those

11:55sorts of things. We store those in a data lake and then put those hashes on chain. So, that way later on you can go look at that and you can review exactly what happened and potentially see, you know, if you run that same game with a different model, what would happen and those sorts of things. So, really having that layer of transparency and

12:14the ability to go and see an audit log of what happened is important.

Definitely, I agree and I did see that announcement. Congratulations. I do like Gate and their AI offerings as well. And I'm excited to see the first hands. I have following along with that for the game. With that established as a foundation, how do you see the next steps beyond

12:36that in establishing this trust and validity to more business use cases outside of crypto as well, but things where there's more that's at stake than just a game?

Well, the next steps for applying this to other uses are things like robotics for example. You know, what did the robot actually do? Or you know, anywhere where it matters. And

12:57in the game for example, next step there also could be if the game has high scores and there's a prize for high scores, then well, how do we know this was actually done legitimately or how did that we know that the house actually did not put their their hand in after the fact. So, did the person go and have the opportunity to see the market data in

13:17CryptoKitties prior to playing or did they put their their their hand in prior to the start of the game. So, for example, you have to put a hash in, that's your your play for your hands prior to the game starting, and then the game starts. And so, I can't, you know, at the end of the game say, "Oh, no, I This is the hand that I used that

13:38basically would have won the game, and here it is. I Just trust trust me, I didn't cheat." Where if you had to put the hash in prior to the game starting, like the hand which you reveal has already been locked in, basically. So, you can't change it as opposed to won't change it. So, that really, you know, gets back to the can't versus won't question, where is we really try to add

13:56can't to AI as opposed to won't to AI. And you know, back to a point that I hear all the time is people say, "Well, there's these settings that says that, you know, AI won't do this on your computer, or it won't do this with your data, or you know, it won't, you know, delete your things, or whatever." It's like, "Well, won't won't is kind of scary. Can't is a

14:15lot better."

Yeah, it would be Well, I you know, there are some permissions that you have to allow, but it's sort of like terms and conditions. People just don't read it. They just click allow all, and then eventually it starts deleting things that, you know, you forgot about the terms and conditions from last year. Now it can do whatever it wants,

14:35and it keeps encroaching closer and closer to taking over all of your stuff.

Well, also, it's it's much more easy to use AI power if you give it more ability to do right, cuz what happens is, it does happen to me when I first started using it, where you know, I'm going to say yes to every single prompt, but it's because just once, yes just once. But then, I found

14:57myself basically not really saving much time, because all I'm doing is sitting there reviewing what it does and saying, you know, yes or no, and me basically being the guardian. And so, for me to be able to have a genetic experience as opposed to a prompt-by-prompt experience, I need to have it have enough rights so it can go and do things and try things and those sorts of

15:16things without me being interrupted, especially if I have like 10 of these going at the same time. I just don't have enough bandwidth to be able to go and review all the prompts for every single one of those 10 10 agents. So, getting that right balance of fluidity for your AI usage with security is difficult. But, at the same time I think the transparency of what

15:40actually happened is very important. And they do keep logs inside of most of these tools that you can go and look at, which is nice. Because I've had where one agent complains that, you know, something changed in this repo, for example, because I have a different agent that's doing something there at the same time. But, if they could more easily communicate and see, "Okay, well,

15:56these are the things this guy changed. These are the things that those guys changed." and somehow negotiate what's going on. You can use work trees, for example, to do that. But, and GitHub to a large degree is one of those solutions of how do you actually store things. They hash what's on GitHub and see what happens. And it has a history there as well. So, GitHub is like a web

16:14two version of provenance in many ways.

Mhm. You mentioned the this the difference between, you know, making ensuring the validity of data and then of actions, especially when there's money at stake. You know, when the game is a high-stakes game and there's money to be earned or there's, for example, an AI training program that's out and it's meant for

16:39humans to enter enter information and they get a reward. Maybe it's a small reward. Maybe it's Maybe it's big. How do we prove the authenticity of a making sure they don't cheat? Is there a difference between proving the validity of data and of actions to prove something was done at the right time? Well, the goal in most cases is to try to use the laws of physics to

17:07to prevent cheating. So, for example, if you have a piece of paper and you say, "Well, write your answer on the piece of paper and then put it face down." and, you know, I do the same thing. And then we have, you know, like maybe I changed the answer or maybe I got the separate piece of paper or it's hard to say where. If we use hashes for example and we say okay well, you

17:23know, hash your data and put that you know, on the ledger. That's there forever. I can't the laws of physics are that I can't go back in time and change that. And so I've declared that prior to a certain point of what that answer is. Now, if it's traditional things you say well, you write down your answer, I write down my answer, we both give them to the arbiter

17:43of the game, they can go look at those and they know what the answers are, but in theory they could be compromised or cheated as well and that sort of thing. But the thing which is nice about a system like XYO layer one is that we can basically store data hash and its providence without having the data. So, what happens is you hash it, you keep it, you don't give it

18:03to anybody. You give us the hash, we store that. Now, at the end of the game, you have two choices. Either you reveal your hand and it matches that hash and then you play that hand. Or if you want, kind of like in poker, you can say well, I'm not even going to reveal my hand. I'm going to just muck this these cards and throw away. And at that point in time we don't know

18:20what your hand was, but you just you lose by default because you decided not to reveal. So, the one option that's not available to you is you can't reveal a hand that you didn't play.



So, that basically it uses time time and physics to prevent that. And so, unless you make something which is like physically impossible, it's kind of like

18:38a zero knowledge proof. A zero knowledge proof just uses math to make it so that it's perfect. And XYO layer one basically uses time in many cases to store information about information it doesn't even have. And so, that's what makes it impossible for a person to cheat their their stuff. So, even if I have a log, let's say for example I have

18:58an audit log, I could have you know, 10 GB of audit log that's in a database or a data lake in our case. If I put that hash in there, unless I provide that entire log, I don't I haven't provided the log. So, a person could you say well, I deleted the log and I'm not going to provide it you, but then we assume that they cheated, right? It's kind of like you

19:15know, IRS coming to you and asking you for your receipts and you're like, well, I don't have any receipts at all. It's like, well, that's a problem for you, but if you can prove the receipts, then you know, they're there. And so, especially if you say for example, in tax filing, if I could hash all my receipts, send that hash to the IRS, and then when

19:32they come and ask me for my receipts, I can prove that the receipts I use are this hash. That would carry more weight than it's like, maybe I, you know, had these receipts or I didn't have those receipts. So, it's it's important to use things which humans can't affect to make a can't as opposed to a won't.

Definitely. And in this case, it's it's

19:54a beautiful use case for blockchain in that it is that piece of paper you write on before, but you can't change it. You can't swap it out. And physics and time really puts yeah, blockchain in the perfect place to support AI. But you've said in the past that, you know, not everything should be on blockchain. And that's what the use case of the hash is.

20:17Could you explain the difference in what is important? Is it just the hash of the information that needs to be on the blockchain?

[snorts]

And what else should be on the blockchain and what definitely shouldn't be?

Well, it's definitely important to minimize the size of the blockchain because one of the problems that Ethereum has is everything's on the

20:36blockchain. So, you end up with a very very large shared hard drive which is what everybody has. Every blockchain has that to some degree, but you want to minimize what those are. So, we call them elevated payloads where the elevated payloads actually go on chain and they have to be there and they get validated. But then the non-elevated payloads are things that are in data

20:52lake. So, for example, if I want to take a movie and hash that movie, I can store that in the data lake or I can store it somewhere else and put the hash on chain. You know, the downside of that is you know, it's it's loseable as far as I can I can lose that data. And so, people when they hear the word permanence, for example, like when we

21:10talk about XYO Layer 1 provides permanence, there's two definitions of permanence that I think are important to differentiate. One is it can't be lost, which is often what people think of permanence, and one is it can't be changed.

Mhm.

We focus on the second one, it can't be changed. So, the I've permanently locked in my hand that I played, for

21:30example, is permanence. I've permanently put my hand on Ethereum's blockchain, so I can't lose it is also permanence, but it's a different permanence. So, the permanence of I can't lose it is very expensive. The permanence of I can't change it can be very inexpensive. It's just one hash, basically, for as much data as you want to. So, we really try to tailor XYO

21:52Layer 1 to be as optimal as possible, and which does make it much better for certain use cases, and also makes it not possible, actually, for other use cases. So, if you have to have your data not loseable, that's something which is not really practical for XYO Layer 1, but I would also argue that it's not really practical for anything that's out there,

22:11because things can still get lost.

Mhm. Definitely. Yeah. Especially those private keys from Bitcoin back in 2010. Wish we had those. What are your thoughts on how much focus there is on the synergy of AI and blockchain throughout the blockchain industry? You know, I know some other There's a couple protocols that are focused on decentralized AI and improving AI. I

22:40feel like a lot of the focus is on finance, which is important, of course, and an early use case, but maybe there's not enough discussion going on right now in the understanding how important blockchain is to building up AI properly.

Well, I think it's kind of, you know, an interesting situation where this comes along, every few years, where

22:59people are like, "Oh, you know, AI is a hot word, and blockchain is a hot hot word, and let's put them together and put them in a pitch deck and see what happens, right? So I think often you'll see more projects that are trying to merge them, and in some cases it's just you know, like have a AI bridge or some sort of AI router, for example, that uses blockchain. And it's

23:20well well, how is it better with blockchain in that case? You always have to ask, you know, is it better if you if you combine these two things or is it not better? And I'd say in most cases it may not be better necessarily, it's just buzzy. It's not better. But in many cases, like in an actual layer one case, I think there's a distinct usage for it and a way to make

23:42the system better. Where if I could have a provable history of say robots or autonomous hardware that's that's running around, especially if they can check in with each other, cosign things, and then put those on the blockchain, I can see where these items were. And in a weird sort of way, it's almost like a real-life version of our coin app, where you know, having these things

24:06check in autonomously into the blockchain allows you to see what actually happened in the real world. Because I think it's going to be one of one of the difficulties is going to be well, you know, something bad happened. Was it my robot that did it? Was it your robot that did it? Was it yeah, was on purpose? Was it by accident? And if you know, having audit logs for those

24:26things that are constantly being stored and also constantly being like bookmarked on an XY layer one system, will make it so that we can go back in time and we can at least find some sort of direction for where it is. And so the actual solution for an actual problem is important to have, and so I think some cases where we see that synergy

24:50that's not the case. In some cases it's extremely true to be the case. And so I think there's probably more place for blockchain in AI or in the world now with AI around than there was before just because it's a natural fit.

I agree. I'm looking forward to seeing it more. And I think another major synergy we haven't spoken about yet is not just AI and the data

25:17providence of blockchain, but the payments and the wallets of AI agents or humanoid robots or self-driving cars and things like that. Is there a plan for XYO to extend beyond data to the ensuring the validity of payments or the success in having that infrastructure in there as robots start interacting and paying on our behalf or paying each other?

25:44

Well, we're looking more at facilitating payments as opposed to validating other people's payments. So, for example, one of the things which we built into crypto cards that's not really that easy to see unless you go and look at the blockchain and understand it is we have a an eventing system where basically on our blockchain a person can go and register for an event by signing

26:04a payload with their address. And then that eventing server will start sending those events to you through like a web hook for example. And then it can stop when you stop paying for that. But what you can do is you can say, "Well, I want to get this event for the next say 1,000 blocks." And I'm going to pay 3 XYO 1 to get those events or you know a certain amount per event.

26:25So, the actual payment of microservices and I mean like micro microservices like nano services is something which I think's going to become much more prevalent where a robot or an AI might be like, "Hey, you know, I want to know for the next you know 2 minutes, you know, when does a new block appear on XYO layer 1?" And I want to get those events pretty frequently

26:47for the next 2 minutes and then stop. But I don't want to go and have to create an account with you know a credit card and that sort of thing to be able to use the API for whatever it is. And so, I think real-time ad hoc usage of services that get paid for the via blockchain payments in a very, very efficient way is going to be extremely important for AI and for robots or

27:08robotics because of the fact that they need a very broad spectrum of access to things. And they need to be able to pay for those things efficiently and to authenticate for them efficiently without having to go through all that pomp and ceremony of credit cards and accounts and logins.

Definitely. And micro payments has been a supposed use case of blockchain for

27:31like many, many years, but we haven't really seen it, at least in real-world payments. But now with AI and token usage, we're starting to see, you know, every word is a micro cent. And like that seems like the perfect use case for micro payments. We're still, you know, putting your credit card into the AI and just preloading it, but I feel like there's a perfect use case for micro

27:54payments on blockchain with AI. It's like the number one use case that's now coming to light.

Yeah. I thought it was actually kind of funny when they came out with the AI thing and called it tokens. You know, it's like, "Yeah, how many tokens do you have? How many tokens do I have?" And I'm like, "Well, I've heard the word tokens before. It's it's ERC-20 token."

28:11So, they're kind of modeling that to some degree where it's a token economy. You know, I'm surprised they haven't said yet, "Well, I'll pay you in I have 500 cloud tokens and I'll use that as money to pay you with." or something like that. But there's no way for me to actually transfer tokens from, you know, one account to another account on cloud as far as I know. But

28:31there's nothing saying that I can't, you know, use their tokens as a payment method. But I think that's what you're using blockchain as a way to autonomously give tokens, for example, to an agent for it to use is something which people are going be doing very often in the near future as opposed to saying, "Well, here's my login for my cloud account.

28:53Use as many tokens as you want." That's a little scary where if I can give it an allowance with a blockchain system, it's something which I think people would definitely want to use. But, the token economy is something which I think will will blossom a lot under AI just because AI is so much more fragmented in the way that it uses the world as opposed to human where it's

29:12like, "Well, I don't want to do a micro transaction, you know, for every single bite of food that I eat at a restaurant, for example, or you know, for every minute of the movie that I watch. I want to just get the whole movie." And so, micro transactions have kind of I think failed to this point with blockchain for those two reasons. One, it's just the gas cost for a Bitcoin or

29:30for Ethereum is way too expensive. But, two, it's humans don't think in those, you know, micro steps the way that AI can.

Yeah. No, that makes perfect sense. And, you know, it's 2 months since our last video. I would love to get you back on once more at the end of the year and see what Q4 has in store for the growth of AI. It's just getting faster and faster

29:53as well as blockchain now kicking back up. Not just the development of blockchains and, you know, XY Layer 1 is developing super fast, I understand, using AI. But, also the prices are going back up, so people are getting more interested and more people are looking at different ways that these two technologies can synergize together. What is some of the most important

30:13things that are happening at XY between now and the end of the year that we may have to look forward to?

Well, you got to say something interesting there which I think is very appropriate for thing about AI in the context of XY Layer 1. We use AI in two different ways. One is we use AI to develop into, you know, accelerate how fast we can do really

30:33cool things and make new products with XY Layer 1. So, AI as a tool for us has been fantastic and we really dove into that, I've a bunch of skills and that we use ourselves, we share with our partners and they can use the skills as well. And then also we prioritize AI things. Kind of the crypto cards is that where the house which plays the

30:53hand of crypto cards uses the AI from Gate for example in this case to actually decide well, how should it play that? And so prioritizing AI is the other aspect of it. Well, how do you use AI as a feature in the product that you want to use. And so we're really focusing on both of those. We're focusing on using AI as a tool to accelerate ourselves and we're using AI

31:12as a feature that we can use in our products and in different things and in you know, interlacing those features with hashes and with provenance and sovereignty through XY O layer one.

I'm definitely going to test out these these crypto card first hands. I saw the announcement. I'm I'm subscribed to when the first hands go live. I'll put it in

31:33the show notes below cuz we've mentioned it quite a bit. I appreciate your insights into everything AI. You're such a smart brain and I'm looking forward to talking with you again in the near future. Thank you so much for the time, Ari.

I appreciate it. Thanks so much for having me on.