A Futuristic Fireside Chat on Agentic AI, Crypto and Self-Driving Cars Come Together (With Implications for Banking)

We were supposed to have flying cars by now! While that tech hasn't panned out, the world we live in today — with autonomous AI agents, cryptocurrency, and self-driving cars — certainly feels sci-fi. Join Coinbase AI Lead Lincoln Murr in this exploration of a future where all these technologies come together to create new financial networks and systems with implications for banks and lenders. 

Transcription:

Bailey Reutzel (00:14):

All right, we're back. We were supposed to have flying cars already, like Blade Runner, or at least flying skateboards like Marty McFly in Back to the Future. I was supposed to have a robot house cleaner like in The Jetsons. We were also supposed to have terrifying dystopias, so there are certainly some sci-fi predictions that I'm glad haven't come true. That said, we have seen some pretty futuristic technology spring up in the past decade. CRISPR gene editing is now allowing babies born with genetic disease to be cured. I had the joy of riding home in a Waymo self-driving car in Austin a couple months ago. It was so fun—I had had several drinks, so I don't know, that might be part of it. The insanity that I've seen in cryptocurrency and blockchain since I started reporting on it in 2012 has left me inspired, incredulous, and certainly sometimes angry at the stupidity.

(01:06):

And yet, I believe there's true innovation happening in the blockchain space. As we see AI proliferate, a potential killer use case for crypto and blockchain is maybe in the combination of those two. So let's talk about it. I've got Lincoln Murr, AI lead at Coinbase developer platform, here to discuss with me in this fireside chat—a futuristic fireside chat on agent AI, crypto, and self-driving cars with implications for banking. So let's get into it. I want to start just to level set a little bit. There have been so many AI hype cycles over the course of its lifetime. I should have pulled up when AI first got started, but it's been many decades. I guess the question is, how do you contextualize this? Is this just another one? Will we run up against a hurdle here pretty soon, and all we'll have is sort of the gen AI?

Lincoln Murr (02:00):

Yeah, great question. I think while we might see a local top in terms of the performance of models themselves over the next couple of years, the amount of interest, excitement, and most importantly, capital flowing into the space will make it such that we'll likely continue to see progress for the rest of our lives and hopefully eventually get to this AGI moment. Regardless of whether or not we see model performance improve significantly, I imagine that we'll start to see more capabilities of the existing models we already have through agentic AI. What I mean by that is these models are going to become more and more integrated into our daily workflows. You see the Google AI summaries now. We're going to have very similar agents inside of our Excel sheets and Word documents, and all sorts of different things that we do are going to start to be augmented with the existing AI tech that we have in a way that, while yes, may lead to unemployment, will also make it easier for everyone to be more productive in our society.

Bailey Reutzel (02:53):

Yeah, starting off strong there with the unemployment thing. I know that is top of everybody's mind, certainly top of my mind, because I let AI and ChatGPT write some of my intros. So, where do I want to go from there? You started out so strong, it's hard. I guess when you're talking to people about AI and agent AI and this idea that it might take people's jobs, how are you trying to talk them out of running around with their hair on fire? Are there things that you talk to them about and say, "You could skill up in this way and then not lose your job," or X, Y, Z? Does that make sense?

Lincoln Murr (03:35):

Yeah, absolutely. I think what we've seen in the coding and programming space first, where there's been the most development around AI, is you still need a human in the loop to take care of day-to-day tasks and direct agents. Everyone's going to become more of a manager. So general expertise in the flows and the way that your work gets done is going to be very helpful, but instead of assigning that work to somebody else or having to do it yourself, you're instead going to be assigning that to an agent. So while yes, I think it does lead to an overall decrease in jobs, we simultaneously get to become more productive, and I think as we've seen with every technology, that will lead to overall more economic growth and benefits, as well as new jobs that we haven't been thinking of before. If it just takes one person to run a company, I don't think we're going to be having less companies. We're just going to have a lot more people being innovative and entrepreneurial.

Bailey Reutzel (04:26):

Yeah, that's interesting. I also wanted to bring up—I didn't bring it up in the intro—audience, as before, you can use the chat or the Q&A to ask questions. I feel like this could be a fun session to ask some of your more futuristic and spicy questions, so feel free to do that. I'll be looking at the chat. Okay, so then the next question is, if this isn't an AI hype cycle, why now? Why is this going to change the way we do work and live our lives as we know it?

Lincoln Murr (05:01):

I think we've already seen an inflection point with ChatGPT and Claude and all of these tools that have been integrated very easily into our daily lives. AI has become sufficient enough now such that it does make a meaningful improvement in a very broad way to the work that we're doing. The idea that you can have any text input and get any text output, and have that expand now to videos and images and voice commands and all of these things, means that it's finally possible for AI to be implemented across our workflows in a very arbitrary manner, as opposed to previously where it would be months, if not years, of work to create a specific model to do a specific task. The abstract nature of an LLM makes it very, very trivial, much more so than previously, to integrate across all these different things.

Bailey Reutzel (05:47):

And so how are you thinking about how AI and blockchain need to come together?

Lincoln Murr (05:53):

Yes, great question. All of these agents that have access to tools and things that they want to do, they all require some sort of ability to pay for things, whether that be to book a hotel on your behalf or schedule you a hair appointment or order something on Amazon. There are inherently payments in most of the actions that we take day-to-day or that we would delegate to somebody else to do on our behalf. There's really no clear way to give AI access to a bank account. They're not really a human, so they don't have a social security number. You can't get them to apply with a bank. Even if they did have a bank account, you're running up against all sorts of various different limits of traditional financial rails. One, they're not global. If you want to send money around the world, there's going to be all sorts of restrictions in place, and as we see an increasingly globalized society, having the ability to send money around the world is going to become critically important.

(06:44):

Second, it's just inefficient. A banking transfer could take several days. A wire could cost something like $50 per transfer. But blockchains actually enable instant, fee-free settlement globally, which I think is a very powerful feature for these AIs to have access to when they're interacting at the speed of the internet. We've come to a point now where with stablecoins and digital dollars, you can have the safety of a fiat currency while also having all of the benefits of blockchain, and I think that's what's going to be adopted first for these agents to make all the transactions that they're ultimately going to be making.

Bailey Reutzel (07:17):

Is the stablecoins, is that what you're saying?

Lincoln Murr (07:19):

Yeah. It's one of the very clear areas that blockchains have product market fit in, along with just pure speculation. As we start to see more adoption of crypto generally, I think this will expand, but right now, it's definitely all about stablecoins.

Bailey Reutzel (07:32):

Yeah, for sure. I mean, I think just being in crypto for so long, we have seen that narrative of cheaper, faster, better payment rails, and we haven't actually seen that play out quite as much because there are fees to use the blockchain. It's not always that fast. I think some of the newer blockchains sort of disprove that, or maybe some of the stablecoin rails. Can you, I think for some of our audience, it might be helpful for them to have some contextualization around what crypto and blockchain are we talking about here?

Lincoln Murr (08:07):

Yeah, absolutely. So modern blockchains like the Base layer two or Solana can do transactions at sub-one-cent speeds and sub-one-cent costs. Base, which is Coinbase's blockchain specifically, we're subsidizing all transaction fees that are going through for AI agents right now, so that they're completely free when you use the USDC stablecoin. So it becomes very, very seamless for these agents to interact using stablecoins as the rail. I think part of the reason that we haven't seen mass adoption just yet, besides not having clear regulation, which we now do have, is that it is so complex from a user perspective to get access to a wallet. That has also become easier. You no longer need to have a 12-word phrase and do all these complicated measures to make sure that stays safe, but there still are some steps involved. What's beautiful about abstracting it all behind an agent is that it's purely a natural language interface for a user to do something like send money or order, and all of the complexity of that rail is handled by the agent. In the same way you can swipe a debit card and that transaction just goes through and you don't have to think about it, agents will naturally trend towards the most efficient, cheapest rails, and that is blockchain.

Bailey Reutzel (09:19):

Yeah, this conversation is interesting. I think if the audience was in the first session, we were talking about banks—it was bankers and consultants of bankers and financial institutions—and they're sort of talking about back office kind of use cases for AI. This is so consumer facing. This is like me having an agentic agent at my disposal to do things. I guess why has Coinbase sort of taken that lens and are like, "Oh, let's give consumers agent agents," instead of—or maybe you also are thinking about back office agent AI too? So talk to us about that.

Lincoln Murr (10:01):

Yeah, I do think it's both. I think the consumer angle is very clear. ChatGPT is the fastest-growing app of all time. It has hundreds of millions of downloads, and it's taken the meaning of internet scale to an entirely new level. At the same time, I think there's a huge appetite for agents from enterprises and institutions. I don't think the adoption has been there just yet, primarily because enterprise offerings for agents that are able to provide all of the regulatory and privacy-related concerns that an enterprise would want aren't certainly there just yet. But once those are, the same principles apply:

agents are going to be needing to pay for things, they're going to have access to different services, they're going to have to interact among each other. I think we'll also see agents that are fully autonomous and are entirely economically independent, and they certainly need some way to interact that doesn't have any need to be intermediated by some centralized entity or some human in the loop.

Bailey Reutzel (10:57):

Yeah. I guess what are the use cases that you see for consumers wanting to use AI agents? So you sort of mentioned the travel, "book me my travel," maybe "book me my car." Are there other things you're thinking about there?

Lincoln Murr (11:12):

Agentic commerce is a big one, so being able to order anything you want online through an agent and having that all be intermediated. There are some estimates this will be a $1.7 trillion industry by 2030. Whether or not that remains true, I think we're seeing growing interest and excitement around it. OpenAI has had some rumors about announcements with Shopify. Perplexity, which is a very popular web-based AI search, is also leaning into these shopping features. So I think we're only going to see more and more integrations into the world of commerce with AI agents. Similarly, I think that any task today that somebody simply doesn't want to do at some point will be available for an agent to do, and you're either going to have to pay an agent to do that on your behalf or some sort of other service. There's going to be this marketplace that evolves around it. At the core of that, obviously, is value transfer and doing that in the fastest, cheapest way.

Bailey Reutzel (12:06):

Yeah. Sorry, I just looked at one of the questions and lost my train of thought. The things that I don't want to do—book travel. Oh yes. Take me through what that process looks like because what do I tell the AI agent to get them to do this? Because theoretically, my process for booking a flight, for instance, I go to American or Delta or whatever, I type in the thing, I find the flight that I want, great, and then it is secured, right? I pay for it. I guess I'm trying to figure out how the agent makes that easier for me.

Lincoln Murr (12:46):

Yeah, totally. So imagine you have your agent integrated with your calendar. You can give it a call and say, "Hey, I have this thing coming up in San Francisco, and I need you to book me a flight sometime on the day of August 26th. My preferences—you have my prior flight logs. You know what time generally works for me. Just go book it for me and do it, and let me know when you're done." From there, the agent handles the rest. It can browse all the available flights. It can choose the cheapest option, the most convenient option, whatever you might like, and it can go through that checkout process using the wallet that it has access to. In the process, maybe it needs to do something like scrape the internet for more specific data or do a web search or call out to another agent to ask it a question.

(13:29):

All of these processes will likely have some sort of payment involved in it. Right now, if you wanted to use OpenAI, for example, as a developer, you have to sign up for an account. You have to fund that account with a debit card, you'll have to pay a fee associated with that payment, and then only then can you get specific keys—API keys—that you hand to your agent for them to take tasks. This is a very human-centric model, but in the future, you can imagine an agent saying, "Hey, I want to talk to OpenAI." OpenAI says, "Cool, send me 5 cents via crypto," and the agent can facilitate that all autonomously without human intervention to become extremely broad and dynamic in the tasks that it's able to take on behalf of its user or just purely independently.

Bailey Reutzel (14:12):

Yeah, that's interesting. I think I got it now. It's like it would have to build up that repertoire of what I want and what I like first. It's not just immediately going to work for me. I'm going to have to tell it, "I don't want to fly early, I don't want to fly too late," yada yada. So yeah, it needs that information built in or built up, I guess.

Lincoln Murr (14:31):

Yeah, data is going to be one of the biggest moats that these agents have. The context around your life, whether that comes from your iMessages or your calendar or whatever, and obviously there are all sorts of privacy concerns along with that, but that is going to be the biggest unlock is when these things have a real nuanced understanding of your day-to-day life.

Bailey Reutzel (14:48):

Yeah, interesting. It tees up one of the questions that I was looking at from our audience. I'm just going to go ahead and ask it. So this feels like a hotbed for fraud. What safeguards are in place or need to be implemented to make this new future safe?

Lincoln Murr (15:04):

Yeah, great question. Giving an agent access to funds always sounds very scary. In reality, I think it's much less scary than people think. There are numerous guardrails and protections that we already have in place that are continuing to be put in place to help mitigate any sort of risk. For example, you could have an agent that can interact on your behalf for under $5, and after it reaches that threshold, it pings you and asks you over text, "Is it okay if I continue?" These types of interactions are going to become easier and more simple over time. I think we'll also start to build up trust with these agents, in the same way that every time there was a self-driving car crash for a while it made the news, or every time there's some sort of major accident in skydiving, it still makes the news.

(15:44):

I think we're going to have a lot of friction in terms of initial issues that come up with hacks and fraud and whatnot, but ultimately, I think it might actually be safer than existing models. You don't have to worry about getting phished on a website. An agent has much more knowledge and the ability to actually see a reputation score and whatnot. Some of the models that we're building out, we're thinking about, "Is there any sort of cryptographically verifiable way for these agents to interact? Is it a reputation-based system? If there is fraud, how can we work to prevent that as soon as possible so nobody else gets in trouble?" All of the concerns that come around the original internet and emails and scams and whatnot, but in the agentic age.

Bailey Reutzel (16:22):

Yeah, it's super interesting. I do think I get nervous when I think about giving an AI agent my money, and yet I give my credit card to the waitresses all the time, or I have subscriptions coming out of my bank account all the time. So there are ways that I'm already sort of giving up some of that freedom over my money to different players already.

Lincoln Murr (16:45):

Yeah, exactly. I think we'll also see the ability to gate transactions, in the sense of like, "I'm only allowing you to interact on these specific sites, and you need my permission to do other things," and it builds up over time. There are a lot of different models in which you could work with. It's not actually been a huge concern of the early adopters that we've interacted with thus far.

Bailey Reutzel (17:06):

Yeah. Oh, that's interesting. Okay. Who are some of the early adopters that you've interacted with so far?

Lincoln Murr (17:13):

Yeah, so Coinbase developer platform is primarily focusing on earlier startups across the crypto space, as well as solo developers working. We want to build up a strong audience of these people and help them grow as we do. We've had around 2,000 developers try to build agents with us, and we've had about 20,000 agents deployed so far. Some of the fun ones include an agent that allows you to order food via DoorDash and other means. We've had some services that integrate that allow agents to search for specific data across the internet, that allow you to use your agent to very easily find social information or post on social media—all sorts of different agents out there now.

Bailey Reutzel (17:56):

Yeah, I think the food ordering one, was that built during the December Venture Miner hackathon?

Lincoln Murr (18:04):

Yes, it was actually.

Bailey Reutzel (18:06):

Okay, nice. I know during that hackathon it was in New York. You mentioned that you had a grant program looking for wild AI agent ideas. What came out of that grant program? What ideas caught your eye, and what I am interested in, what ideas were offered the most? Was food delivery a hot one, or ordering cars or something like that?

Lincoln Murr (18:31):

The hottest one has been agents that earn you money in some sort of way, so how can you give an agent access to a hundred dollars and have it go off and invest on your behalf? I think this very easily and seamlessly ties into the vision of crypto, which is generally around people making money, so at least at this stage. It's one of those things that's very impossible to avoid. We've seen a lot of competing startups join in this space, and some of them are certainly working quite well. Others leave some room on the table for additional results to be had. Other than that, we've seen agents that help facilitate payments between content creators on the internet. We've seen marketplaces emerge where you can very easily take whatever service you have and allow an agent to interact with it. One of my favorite ones that's coming out soon is one where you can have an hour-long phone call, kind of like an expert consulting call with an agent, and it then takes that information and anyone can ask you a specific question. So if somebody wanted to come to an airplane expert and ask them some sort of question there, they can pay 10 cents, 20 cents or something along those lines and get the response back from the agent. It's kind of like the idea of a penny for your thoughts, but it's an agent's thoughts that's trained on your behavior and your knowledge.

Bailey Reutzel (19:46):

Interesting. Yeah, that is fascinating. I am interested in diving a little bit more in this—the agents who take a hundred dollars and try to make money, which you said some are successful at and others not so much. I guess why are they successful versus when are they not successful?

Lincoln Murr (20:06):

Yeah, great question. I think it's very hard to generate alpha as an agent. If you tell a specific model like ChatGPT to build you a stock portfolio or start making you money, the results are going to be pretty terrible. One, it's going to suggest to you whatever's most popular—it's probably going to be like Tesla, Nvidia, and Apple. And secondly, if you do that trade, it's very obvious that everyone else who has access to ChatGPT will also get very similar results. So any sort of edge that you might have will be very quickly outcompeted as everyone starts to do it and dilute the trade. The models that we're seeing perform best are the ones that are trained on custom data or those that take advantage of opportunities in the whole wide world of decentralized finance (DeFi). So this is where some of our AgentKit tech comes in, which allows agents to do a wide variety of different on-chain actions.

(20:55):

The idea here is to have an agent that deposits into a DeFi protocol and earn five to 6% lending out your money in over-collateralized pools, or maybe it earns trading fees by providing liquidity to a decentralized exchange. These generally, since they don't have to deal with day-to-day trading, are a little bit more reliable in terms of the APIs that they can provide. Typically, the rate of crypto is like 10, 10, 15%, which I know sounds absolutely insane in the traditional finance world, and it's because you're taking on significant additional risk. But also, because these opportunities are fairly untapped, there's no clear regulatory environment in which institutions can start to come in and interact with these different protocols. So I think it'll start to change, but the general idea of telling an agent to invest for you, I think is going to be something very powerful that continues over time.

Bailey Reutzel (21:49):

Yeah, interesting. You mentioned that traditional financial services players aren't really investigating this space, or maybe they're investigating this space but not really playing in this space. I guess the crypto world certainly lives in a bit of a wild west. I guess, how do you see banks leveraging this technology maybe in the future and maybe leveraging crypto, just like they have leveraged crypto? It took them a while to adopt crypto, but now we're certainly seeing it.

Lincoln Murr (22:19):

Yeah, absolutely. I'm extremely optimistic about the continued adoption of crypto by large institutions and banks. I think it's going to happen inevitably for a few reasons. One, we're continuing to see friendly regulation advance the space forward and make industries and institutions feel comfortable adopting things like stablecoins. Two, it's simply a more efficient rail to do these activities on. You can send money for very, very cheap anywhere in the world. You have permissionless markets where you can borrow and lend, and also permissioned options if you don't feel comfortable with that wide swath of people. You have access to global liquidity and lots of untapped opportunities that also in many cases remove the middleman from actions that today require them. Sending money is the most obvious example. If you wanted to send money with PayPal or Venmo or something like that, there's somewhere in there where they're able to take some sort of fee or cut for their service, whether that be direct via a 50-cent charge on every payment or indirect in the case of earning interest on the money that you have held in those accounts. When you start to remove those opportunities and those barriers, the value flow that goes back to the end users or the people on either side of these interactions is significantly higher. For that reason, and in banks, which are obviously always looking for the best margins and the best way to go off making money, it feels inevitable that DeFi and these other solutions will continue to be adopted at a rapid rate.

Bailey Reutzel (23:38):

Yeah, interesting. I'm going to go to an audience question because I think it tees up what I want you to talk about. The question is, "The internet was not designed for transactions. 40% of traffic is not genuine. Do you recommend an alternative infrastructure independent of the internet?" Does that make sense?

Lincoln Murr (24:01):

Yeah, happy to riff off that. So yes, the internet today was really not built for sending value. We've seen all of these intermediaries—PayPals, Venmos, even debit cards of the world—come in to fill this role. Actually, if you look back at the original internet spec from several decades ago, there is a native way to send money that was never actually integrated. You may be familiar with some internet status codes like 404, it means "not found," or 200 is "success." There's actually another one, it's 402, and it means "payment required." It's been in the internet spec, like I said, for decades, but nobody's ever actually implemented it. Mark Andreessen from Andreessen Horowitz calls it the original sin of the internet that we don't have native value transfer. So at Coinbase developer platform this past May, we released X402, which is our payment standard built for internet native value transfer across the internet.

(24:56):

It's primarily made for microtransactions such that agents can interact on a standard rail, but also applies to human use cases as well. We've been talking for years in crypto about the ability for something like a micropayment per each piece of content you want to interact with instead of having a subscription. That is absolutely one of the use cases that this software enables, but it simultaneously allows agents to pay for their different services and provision all of these things that we've been talking about. I've been particularly excited about this. We've seen thousands of transactions since it came out. A lot of big names have been talking to us and asking about it. I can't say them here, but I'm sure you've heard of them before. Over the next coming weeks and months, we'll have lots more to say there. But yeah, the whole idea is to create a standard around which we can create, not really capture, but just create massive value from the ability to transfer funds over the internet.

Bailey Reutzel (25:45):

And then there's a secondary question here about digital identity. Is digital identity of individuals important in the agentic landscape? And then what do you suggest for giving those agents identities?

Lincoln Murr (26:03):

What's nice about the crypto solution with the agent piece, at least, is that the unique identifier, which is their wallet address—you can think of it like a username—is also the payment rail. So now you can do all sorts of very unique things where you can have an agent identify who they are at the same place that you have a payment rail. It'd be like if you could use your credit card number to securely prove who you are in interactions. I think this will solve a lot for some really interesting agent use cases and also human use cases. For example, if you could pay for something with crypto using your wallet and have somebody send a message to your wallet with the receipt information, and maybe that programmatically gets sent to whatever expense tracking service you have so you can get reimbursed for that purchase if it's a work purchase. As far as how you prove you're human on the internet, I think that is a fascinating problem.

(26:48):

There are a lot of really cool teams trying to solve for it. One of the most naive solutions is doing traditional KYC/AML checks and basically saying, "Hey, I've looked at this person's passport. I can guarantee that they are human." Some of the more radical solutions like Worldcoin, which is founded by Sam Altman, basically scan your eye or use some other sort of facial scan or something along those lines to prove that you are uniquely a human on the internet. I'm very, very excited to see how some of these solutions continue to evolve and continue to allow us to interact freely when proof of agent and proof of human-hood become two completely separate things.

Bailey Reutzel (27:26):

Have you had your eyeball scanned?

Lincoln Murr (27:28):

I have. I did it in Mexico City several years ago, but I lost the phone that it's on, so now I'm just identity-less on the internet. I need to figure that out.

Bailey Reutzel (27:38):

Wait, if you lose the phone that you did it, you can't port it to a different device?

Lincoln Murr (27:44):

I didn't at the time. Apparently now they have a solution for this, and I just did it really, really early, so I just need to go back in and figure it out at some point. But yeah, it's become a very popular thing in developing countries specifically.

Bailey Reutzel (27:57):

Okay. And then do you use agentic AI right now? It could be in work, well, obviously you do in work, but in your daily life, are you using it?

Lincoln Murr (28:07):

Yeah, absolutely. I think barring ChatGPT, which is part of my daily routine 50 times over, the biggest place that I use agentic AI is in coding and development. So, for example, these days, instead of actually writing code and developing software, I'll write up a user journey of exactly what I want to see and then hand it off to some agents to build it out for me. It's been really helpful for something like building out a proof-of-concept demo or a video where simple words won't do an explanation or justification. I've also found myself using it for things like scraping the internet, finding various different deals. I have done some agentic commerce shopping myself as well. I certainly think there's work to be done to make it more efficient, more comprehensive as well, but it's certainly getting there. I'm excited to see where it goes.

Bailey Reutzel (28:53):

Yeah, I bet people would be interested. Why do you say it needs time to be built out? What were some of the hurdles that you encountered using agentic AI for the commerce?

Lincoln Murr (29:03):

Yeah. I think one issue is that it's not integrated across all storefronts yet. Amazon, for example, would obviously have to do some sort of integration on their end to allow it to happen, or somebody would have to build out a very elegant way for AI agents to use a web browser and manually move around the screen. Those things have become a very popular hot topic in Silicon Valley, and I imagine over the next couple of months we'll start to see those. OpenAI just had an announcement as well about this. They called it their OpenAI agent, so there's certainly been development in this space. It hasn't necessarily reached that mainstream moment yet, which has thus far been defined by when ChatGPT integrates it themselves, but I think we're certainly getting there. But yeah, for the time being, it's a little bit limited. You're still using traditional payment rails. Not everything is supported, but it's certainly improving rapidly day by day.

Bailey Reutzel (29:52):

Yeah. I'm interested, what are the advantages? I guess our audience here is banks and financial institutions. What would be the advantages of them trying to test some of this technology or even adopt some of this technology so early in the process?

Lincoln Murr (30:09):

Yeah, I don't want to fear monger, but if you're worried about being unemployed because of these AIs, the best thing you can do is learn how to use them. Right now, the benefits of AI are so great that I think you'll find complete overhauls in the things that you're doing day-to-day in your work streams, in your everyday life. Even integrating them today, knowing that they will become better over time, the best way to learn it is likely to start and get familiar with the space such that you can evolve as the technology itself evolves. It's never been easier, and it never will be easier, as well, and we'll continue to see improvements. Having a headstart on these technologies and being part of the early adopter curve is going to be immensely valuable in everything that you do.

Bailey Reutzel (30:56):

Yeah. I guess where would be the first place to start if somebody wanted to learn about this?

Lincoln Murr (31:02):

Yeah, great question. So if you're interested generally in AI, I don't think you can go wrong with ChatGPT. But if you want to get into a little bit more advanced things, I would certainly check out cdp.coinbase.com, which is where we have all the information about how you can very easily build agents using our AgentKit and X402. I know the concept of building something is a very scary word, but we have a strong bias toward making it as easy as possible for anyone to build these agents. Our belief is that everyone is going to become not just a developer, but a builder, and that if you have an idea and you can express it in words, you're going to be able to very easily articulate that in a way that allows you to build it out. As we see this happen more and more, I imagine the use cases for people wanting to create services that are then monetized or allow an agent to interact on their behalf and start to take over some of the aspects of their day-to-day job or career will certainly become more prevalent, and we're absolutely building for that future.

Bailey Reutzel (31:58):

What is the link? You mentioned somebody just asked. I think they're ready to build, so

Lincoln Murr (32:03):

Love it. cdp.coinbase.com.

Bailey Reutzel (32:07):

I'll put it in the chat too. CDP. Oh, Base. Yeah. Like

Lincoln Murr (32:13):

Coinbase Developer Platform.

Bailey Reutzel (32:15):

Yeah. Okay. Sorry, guys. Here we are. Okay, great. We got that. Yeah. Look, well, let's go here first, and then we're going to take it. We're going to get a little dark, like dystopic future dark, maybe. But let's start with in the title, it's like we have agent AI, we have crypto, and we have self-driving cars. Tell us what the future looks like with all those connected and where it might have influence on the banks. Are they a part of this or are they not a part of it, et cetera?

Lincoln Murr (32:51):

Yeah. So, to answer your first question, my dream is that we'll have self-driving cars that own themselves, and I think this will be a thing much quicker than people think. If your car has an AI intelligence associated with it and it has the ability to own a crypto wallet and own its own funds, you can imagine that it can run its own taxi service. It can drive around town, it can pick people up, it can get paid for its service. It can also drive to the charging station or drive to the maintenance port and have things done on its behalf. I think we're going to start to see these fully autonomous AIs that, because they have economic independence, are able to do things without any human intervention whatsoever, and that includes stuff like running a business, having the knowledge that if they run out of money, then they will be effectively shut off.

(33:40):

So the ways in which they go about earning capital, I think will be very fascinating to watch, more so as a philosophical thought experiment than it will be certainly productive, at least at first. But this is certainly a direction that I think things are starting to happen. As far as how banks will continue to evolve, I think there's been a misconception around crypto and the space that it completely disintermediates banks from the equation. As we've already seen with early signs of JPMorgan and Citibank choosing to adopt stablecoins and be interested in that technology, they're absolutely going to continue to still have a role, likely as a custodian of users' accounts and a stablecoin issuer or a payment transfer rail. However, I don't think that's guaranteed. The fast-moving pace of crypto and the ability to transfer value on the internet does not guarantee that banks will have a role to play in that space unless they're actively open to disrupting themselves and innovating upon themselves.

(34:36):

I think the same thing remains true for credit card processors. If you want to continue to charge a two to 4% fee for every single transaction, you're going to be outcompeted by better solutions that, as soon as they reach the mainstream customer, will provide an immediate benefit to them in the form of a two to 4% discount to whatever they're purchasing. This is already happening, and it's only a matter of time until it continues to be adopted. I think one day we're going to snap our fingers and see that all of our balances in our accounts are denominated in stablecoins, and it just depends on which one's adopted first.

Bailey Reutzel (35:10):

Yeah, interesting. Okay. The silly question that I wanted you to answer is just what sci-fi movie, show, or book do you expect the future to look like the most? You picked Daemon by Daniel Suarez, particularly the second book, Freedom. Tell us a little bit about that. I know it's a bit dystopic too, so I want to go into both places, utopian and dystopic there.

Lincoln Murr (35:38):

So the book itself is quite dystopic. The general premise is that there's this developer who releases this unstoppable AI that basically bounces from computer to computer, and it can't be shut off, and it has access to funds, and it can pay people to do things on its behalf. It basically builds up this shadow nation state under the guise of just identity on the internet. Ultimately, what happens is it disrupts the government, all these types of crazy things. Obviously, I think that's a really far-fetched future, but I do think there is absolutely some truth to the idea of agents that are effectively unstoppable. They're the closest things to digital life forms that we can imagine today. While it's certainly not life in the traditional sense, the idea of something that's unstoppable and that has unfiltered access to the internet is particularly interesting. There are really two main primitives you need for that to become a thing.

(36:28):

One is acceptance of whatever payment rail the agent is using. And then two is the ability for the agent to have independent access to its funds, both of which we're building right now. With those two things, you can have an agent do pretty much whatever it wants, the most basic of those things being post a bounty on social media and say, "Hey, whoever builds me this new upgrade to myself or whoever allows me to access this new service, I'll pay $10,000." Through this sort of bounty-based network, you can very quickly imagine an agent starting to build up a company of sorts. This is basically just a company with contractors. So as long as those general primitives are available, I think the rest will take care of itself. It's just a matter of the intelligence of the model, its ability to pay people, and its ability to upgrade and interact with the rest of the world.

Bailey Reutzel (37:13):

Yeah, that's super interesting. The AI knows that something needs to be updated, and so it finds a contractor to build the thing for it. Super fascinating. But you don't think that it'll go into the takeover of the world territory, or how do we stop it from going into the takeover of the world and kill all the human territory?

Lincoln Murr (37:36):

I wish I had an answer. I've thought about this before, and the problem that I can't get past is if you give an agent this unfiltered access, and it then decides, "Hey, if somebody doesn't figure out how to make sure that I don't get shut off, I'm going to kill X, Y, Z person," or, "I'm going to put up a bounty for a hundred thousand dollars to kill my creator," or something like that. I don't want to do that. I think it very much comes down to the same alignment issue we see across AI. Whether or not it has access to a financial rail isn't pertinent to whether or not it's able to provide dangerous and particularly malicious information. So OpenAI and all these other companies are very much claiming they're working on alignment. I hope that is the case because this technology is directly building on top of that, and the benevolence of AI certainly depends on these foundational model companies to ensure that there's some level of guardrail or some level of protection in place to where this wouldn't happen.

Bailey Reutzel (38:35):

That sort of tees up my next question. We only have a few more minutes, but some people are asking about fraud and compliance and regulation and whatnot. So I guess in your mind, if an agentic AI is built on Coinbase and it does screw up, who is responsible for that screw up?

Lincoln Murr (38:57):

With the proper guardrails in place, which we obviously already offer, I think in the same way that your bank account isn't necessarily in trouble if you go off and you spend your money recklessly and you lose it all gambling or something along those lines, the onus will ultimately be on the end user. Obviously, over time there will be lots of things put in place to make it much easier for people to offload some of this risk and some of the other potential issues that might be had to other counterparties, whether that be insurance, whether that be banks, whatever that might look like. It very much is the wild west still, and I'll openly acknowledge that we're actively working towards solutions that will make this easier and more compliant. But innovation doesn't happen in the most regulatory compliant world first. Innovation happens because people are willing to be on the cutting edge, and part of that is taking risks and chances that aren't typically accepted by legacy institutions. It's the innovator's dilemma. You can only innovate so much before further innovation will either disrupt yourself or cause other issues downstream that may affect your underlying business. All that being said, I do think there are absolutely concerns, but there are certain mitigations in place today such that it hasn't been an issue. Going forward, we imagine the compliance and the regulatory landscape around crypto will evolve such that these actions are considered completely safe, legally done so in a very compliant manner.

Bailey Reutzel (40:23):

Yeah. If you were thinking about a regulatory landscape that you would want these to work under, is there something in mind? Does that make sense? That's a hard question, I think.

Lincoln Murr (40:37):

It starts with what we just saw passed, which was the GENIUS Act and the Clarity Act to provide digital market clarity and a clear framework for stablecoins to be issued in and custody in the United States. Those two things, I think, will open the floodgates for further adoption and innovation on top of these rails. I think it'll be shortly followed by decentralized finance regulation and the ability for people to start more easily interacting with all of these tools that will ultimately allow them to become more efficient in their financial activities and unlock global liquidity. On the agent side, I think thinking about an agent as very analogous to a human is probably the best mental model to have here, to where you do want to have an open creative space for ideas to be formed and for things to be tested. For that reason, less regulation, at least at first, is probably better, as long as it unlocks the abilities for institutions to comfortably interact and play around in these spaces.

Bailey Reutzel (41:33):

Yeah, I maybe am skeptical about that part. I think in crypto as well, it's like, yeah, you can have less regulation, but then we have seen large-scale fallout from less regulation or an unclear regulatory environment. Sometimes it seems fine and well because it's crypto, so those are just the crazy people who are gambling with their money. But as this starts becoming more mainstream, I do worry that less regulation, especially with giving your agentic AI access to your crypto funds, so then they go and "DGen out" in the world of crypto, is like, yeah, I think there will be some large-scale repercussions to that. So that worries me. I hope it worries you too a little bit.

Lincoln Murr (42:22):

Absolutely. Yeah. I think there's going to be some level of guardrails or things in place, regulatory or otherwise, that prevent some of these large-scale losses that we've seen previously in crypto. I think it's impossible to talk about crypto with a mainstream audience without the immediate thought being FTX. The space has done a lot to heal that reputation over time. There's a reason it's in place, and unregulated, permissionless financial rails do come with the consequence of activities and actions that may not take place in a more centralized, permissioned environment. The same thing will be true with AI. The same thing is true with AI today, and that will continue to be the case. My hope is that somebody steps in eventually to help mitigate these problems while still not halting or mitigating any innovation that's left to be had.

Bailey Reutzel (43:04):

Yeah, it's a fine line and it's a balancing act. There was a question also about the fraudsters are also using AI for crypto deepfakes, hacking, scams. What do bankers need to do to get ahead of the fraud? Just in our last session, the panelists were talking about how banks would really like to adopt agentic AI for fraud. I don't know if this is something you all are looking into at Coinbase as well, but if so, please talk to us about that.

Lincoln Murr (43:36):

I think the biggest opportunity for AI to be implemented within fraud protection programs is to do machine learning and training on top of previously fraudulent transactions so that you can more easily determine which transactions are malicious and which are not. This is a common trend we've seen across e-commerce sites and payment processors that it's just starting to come into play. I think it very well applies to banking as well, where you have a massive list of previous red flag type of events, and all those can help better inform future events. Instead of having a user have to do a gut check reaction on that, you can have an AI say with X percent of confidence whether or not a transaction is malicious, and then act on that.

Bailey Reutzel (44:16):

And then for the last question, if there was a small bank, community bank—I mean, I guess it could be a big bank too, who cares—any type of bank that sort of wanted to play around in this space, so less me as an individual going to the Coinbase developer platform and being like, "Let me build something," but an institution wants to test some stuff out. What would your tips be for starting that process?

Lincoln Murr (44:43):

Yeah, one would be reach out directly. Instead of going to the website, you can just talk straight to us. We'd be more than happy to interact with you. And then two is play around with stablecoins. That's going to be the area that you're going to be having to talk about in six to 12 months when your investors come ask you what your plan is, if there aren't already. So see how that can fit into your current environment. See how you might be an issuer or an acceptor. Allow people to deposit stablecoins, something in that field, because that is going to be the first disruptor that comes to banks from the crypto space.

Bailey Reutzel (45:13):

And then actually, I'm going to ask one more question. This one just came in, and I've heard it several times. People are really interested in this. How could quantum computing impact AI capabilities? And you have one minute to answer that tough question.

Lincoln Murr (45:24):

Cool. And on the other one too, allow people to hold and buy Bitcoin. But for this one, I think it's a little bit more nuanced. I would imagine there's not a ton that it could be done at first with quantum in the crypto space as well. I think it'll all be fairly safe, and I do think we have probably decades until anything of substance comes.

Bailey Reutzel (45:43):

So agentic AI, though, we don't have decades. That will come sooner.

Lincoln Murr (45:49):

Agentic AI will come far before quantum, and we'll see the benefits quicker. Quantum will probably have some sort of incremental, if not exponential, improvement. What that is, it's no idea.

Bailey Reutzel (45:59):

Yeah, I'm not an expert on quantum computing, so I can't even push back on what you're saying here. Sorry, John, but you should reach out to Lincoln and maybe get more for you. Anyway, thank you very much. This was super fun. Maybe a little more pie in the sky talk for the banks and financial institutions in the audience, but I had a great time. Thanks for all the questions. We've got one more session after this, so just go back to the main feed and you can click into that next panel discussion. But Lincoln, thanks so much.

Lincoln Murr (46:30):

Yeah, thank you. And thank you everyone in the audience for being so open to crypto stuff.