From Vibe-Coded Demos to Production Money

Transcript

[music] Good afternoon. Uh my name's um Jeff uh co-founder of Curve Grid. Um I used to start my talks with a quote, but it's 2026, so instead I'll start with a tweet. Um here we have a AI generated picture of two rockets. Uh both of them sitting on their launch pads. Um you can see the piping down below. One of them is about to take off into the heavens uh to explore new worlds and new civilizations

and one of them is very likely to explode on the launchpad. Um, but I think the interesting part about this in 2026 is that both of these rockets were very likely as software products created with AI. First, a little bit of background about myself and our company. Um, so I'm the co-founder of Curve Grid. Uh, we've been in business for nine years. We're a blockchain software company based here in Tokyo, Japan.

And our core product is called uh Multibass or at least our core product until now. And it's a blockchain development platform that software engineers use to build blockchain applications faster, easier, and cheaper than doing so from scratch. It's been used by Fortune 100 companies. Uh it's been used by global DeFi platforms. And what we found over time was that our customers were also using it to manage their stable coin and digital asset portfolios. But if you look at this web UI, right, this multibass web UI, it's very much a developer tool. Like if you're if you're a CFO looking at this, it doesn't necessarily make a lot of sense to you.

But indeed, we had CFOs and asset managers and finance folks using our uh software, this this interface uh to do things like complex DeFi transactions on different blockchains, using a combination of hardware wallets and multi- signature wallets. They were using it to run their reporting uh and accounting and uh finance workflows. And so we took all of that. We thought, well, Multibass is great. We'll continue to build and sell that. But we see an opportunity for a new product uh to help companies manage their digital asset treasuries. And so what I'm going to talk to you about today is as our story about building this new product and the

lessons that we learned working with AI to do so. So we started here at the top right. We have uh our existing platform Multibbass. We have customers that are using it for managing their digital assets across chains with uh smart contracts and multisig and hardware wallets and accounting and all these sorts of things. Now the traditional way to go and engineer a new software platform is to come up with an idea, write user stories, draw wireframes in Figma, iterate with customer feedback, uh and then implement it. In 2026 with AIS, all of this is flipped upside down.

And really when I say AI or LLMs or coding agents, what I'm talking about is Cloud Code or OpenAI codecs. And I'm really speaking about the last 10 months. November, December 2025 is when they really started to get good. And so this is all very very new for our entire industry. I'm sure some of you are software engineers, some of you are not, but we all work with software in the space every single day. And so this is the usual route. The new route is you go straight from idea to working UI and iterate very very quickly with customer feedback.

And so that looks a more uh more like this where basically you test the idea while you're explaining it. And for anybody who hasn't actually experienced this, uh it goes something along the lines of you come up with an idea and you put it to the LLM. It builds it for you and then you iterate in real time on what that looks like. You don't just end up with wireframes. You end up with a fully functioning UI that yes is vibe coded in the back end but actually works well enough to really explain your identity your idea in very high fidelity to the point that we would be on calls with customers and we would tell them at the beginning of the call look this is just a mockup. It's it it works. I mean

it looks like it works but it's not really functional. And by the end of the call they were asking us oh you know so how do I access this function and when can I use this and we you know the the the the logical connection in their brain was that this is already production software that they can use. The other interesting thing it did was it took that feedback loop of wireframes and user stories down from uh days or even weeks to minutes to the point where we've been on a call with a customer and they've told us about an idea for one of the a feature that they'd like to see and we've had the AI build it in real time during that conversation and then demonstrated it to them. And again, for anybody who hasn't seen this, this is a

completely new way of building and ideulating with software. Now this isn't unique to us. A lot of companies are doing this are starting to do this. Um but the real question becomes if it's possible to create software extremely quickly and cheaply then what is scarce? And I'll come back to that in a minute. Uh but first I want to take a little bit of a side tour uh to uh talk about an article written by a gentleman by the name of Chris Ly called the rise of industrial software.

and he's got a number of interesting ideas in this article and a number of number of interesting charts and this is one of them. So here you can see your standard four quadrant matrix where we have cost and value on the two axes and imagine for a second that traditional software uh is sitting in the upper right hand quadrant. It's relatively expensive to produce but also produces relatively high value. It's been crafted with care uh and uh and uh understanding with AI. We now have this entirely new class of software which Chris has termed disposable software.

And the idea in this article is that uh AI being applied to software development is similar to uh the application of machinery to producing products during the industrial revolution. Hence the rise of industrial software. And so the idea of disposable software is it's very quick and cheap to produce to the point where for example if you're going on a trip with your family to let's say Hawaii, you could produce an app specifically for that trip that has all of your air tickets, your hotel bookings, what you're going to see, where you're going to eat, and you can be all interacting real time in a message board specifically for that trip. That's the idea of disposable

software, oneoff, cheap to produce uh um you know, single-use software. Just to kind of complete this quadrant because I think it's a it's an important idea from the article. Uh in the high cost, low value, we have kind of the traditional idea of boutique software. And so this is where you would have dev shops creating uh software for potentially one specific use case at a relatively high cost. I think one of the points here is that this kind of boutique software uh is likely to disappear over time.

So I think disposable software has its place and I'll come back to that in a second. But the really interesting piece out of this article for me was that it's not that we just add disposable software. It's that we also add this idea of applying advanced research and development using AI to the to enhance uh the value of traditional software which means that if we put if we use the AI in a uh thoughtful um careful wellapplied way that we can uh drastically increase the value of our traditional software.

Before I move on from this, I wanted to also address one I thought very apt analogy that Chris wrote about in the article which was not only as it uh relates to goods produced in factories uh but also to let's say the industrialization of farming right so the industrial revolution uh also allowed us to at least in the developed world more or less eliminate famine for example we can now produce a lot more goods uh sorry a lot more food on farms efficiently with machines. So we can have a lot more potatoes which means we more or less eliminate famine. It also means we have potato chips. And so I think the potato chips are kind of like disposable

software that came out of the industrial revolution. We can have both. We can have we can draw that link from the industrial revolution all the way to um you know space travel and computers and the internet and AI that we have today. We can also appreciate the fact that potato chips are delicious and we can have both. One more kind of thought on this, you know, for some of you who may not have seen just how fast you can iterate on this. This is a tweet just the other day from uh Toby Lute, who's the CEO of the uh global e-commerce platform Shopify.

And so this is him tweeting about the fact that there's a new Linux uh desktop operating system called Omari that he's been hacking on. And the the unique thing about this operating system is that it has AI agents built into the operating system to the point where if there's a missing feature, you can simply ask it to build that feature for you. And in this case, he's built a kind of free disk space tree graph application and sort of just wield it into existence. So what does all this mean? If code becomes cheap, really what stays valuable? Um, I would contend that what remains valuable to us as software developers or those of us consuming

software is anything that's high friction non-scalable. And so that's that sort of leads to all of the different things around this this sort of central uh code uh piece here. And it starts with distribution, customer relationships, and trust. You know, if there's a million travel apps or a million digital finance investing apps on an app store, you're going to go to the apps that are coming to you from a trusted friend or business colleague or vendor or counterparty. And so those interpersonal relationships and the time spent with our customers matters more than ever. The second is really around uh authority, policy and verification, right? The compliance and government

pieces uh sorry, governance pieces. the compliance and governance pieces that uh give us this trust in the software. And the last is around integrations and context. Is it plugged into the things that uh that I care about and I need to use? Another aspect to this is that similar to how we have started to train ourselves to recognize AI generated text or AI generated images or AI generated video, we also need to train ourselves to recognize uh AI generated user interfaces or software experiences. Just because something looks production ready doesn't mean that it is. And I think three key points here. the the three

things that we really need to consider are what are the operating history or the provenence of this company that we're we're doing business with or that we're buying software from. Um how deeply is it are they considering the edge cases and how uh how tightly is it integrated to the the the things that I care about? And finally, what kind of checks and balances and compliance pieces are in there. The other key piece to consider and I think when we think about AI generated code we just think about the computer code that's been written but there's a lot more to it than that anybody who's spent time in technology understands that uh the concept of supply chain and infrastructure

and you know for example if you're asking your AI agent to uh add a PDF export feature to your financial software probably what it's going to do is go out and uh build something that relies on tens of thousands of third-party packages in a supply chain that's extremely opaque and difficult to audit. And so that's absolutely critical as we're uh you know thinking about this uh that we don't only consider the computer code but we also consider the the infrastructure and the uh the supply chain.

Okay. So, how do we how do we deal with this as software developers and how do we think about this as consumers of these these new kinds of software? Well, what we found as we're building our new platform, it's that when it's so easy and cheap and fast to produce code that we need to spend at least as much time, if not more time, on tests and containment and thinking about these bulkheads and boundaries so that we can limit the blast radius when something goes wrong.

So if you think about uh a barrier or bulkhead uh or a boundary where we let the AI roam free and kind of do as it wishes or where we can move quickly uh is on the left hand side of this. So prototypes uh readonly dashboards experiments test scaffolding or tests themselves. They still need normal engineering. We still need to put time and effort into them but we can move much faster. on the right hand side is anything that touches money, private data or personal information, physical systems, identity and permissions and of course signing keys. So it's this concept of this control bulkhead or boundary when it comes to separating high velocity

software development from real money or personal data. So what does this all mean? A number of the other speakers today have talked about, you know, the future of uh finance and digital assets. And I think everybody who's in this room attending ETH global pragma uh is is a believer in in the future of stable coins and and onchain finance. Uh and there's many many charts to illustrate it. Uh this is one of my favorite uh which my my co-founder uh you know surfaced and and I wanted to share with you as well. So this is from last year but I think directionally still accurate. uh and it's this connection of uh onchain finance to traditional

finance and also to speak about the size and volume of it and so this is a chart that uh illustrates uh the US uh dollar treasury holdings or US treasury holdings of various entities Japan obviously the biggest uh and then you don't have to go that far down the chart uh to number 18 to get to stable coins US the combination of USDC and USDT or uh uh Circle and Tether just after Norway and just before Saudi Arabia. The whole point of of of showing this chart and kind of tying this back to digital assets is that unlike traditional finance and Armani this morning from Backpack actually touched on it a little bit, which is that uh if

there's one lesson that we can take away from traditional finance is that it's really all about risk management. And our industry in digital assets has tended to move much much faster than traditional finance. And I think that's a while I think that's a very positive thing, it also means that we need to be extra careful about risk management and controls uh especially in the face of AI and that's really uh you know what this talk is is all about. I think more than you know maybe any other technology in the past agents really make this control problem uh unavo unavoidable and you know it it's obvious but it's worth repeating that we

should not just be giving uh agents a private key and the ultimate question there really is where does where does this authority live and how does that uh control layer come into play here between discovery permissions approval and the result of what happened. And so you know the key concept here is really around uh intent and authority. human or agent initiates something or makes a decision and at the other end of it a transaction happens and is then reconciled.

Where we still have to spend time and this is where uh you know at at Curve Grid where we're spending our time right now is on this control layer around managing identity whether it's human identity or agentic identity policy limits uh approvals uh and verifications and checks. So I uh I talked earlier about our new product uh going from Multibbass, our blockchain development platform to what we're branding Neo uh which is this control layer.

So this is uh just a view of uh of the web UI of Neo still in beta and uh to be uh uh released uh next month as a MVP. And it's again really a digital asset management platform for businesses to manage their corporate treasuries uh giving them the ability to transfer stable coins, earn yield on digital assets and put their capital to work in a controlled and compliant way. So I'll just leave you with this thought that interfaces uh can be beautiful uh they can be disposable software can be disposable but it can also be handcrafted and uh uh carefully thought out um but authority can never be uh

left to the side. Authority is critical and uh we can't make that disposable. Thank you. [applause] >> [music]