Ethereum Foundation: Building on Ethereum | Austin Griffith at ETHOnline 2026

Transcript

Introduction and Builder Mindset

[music] I'm Austin Griffith. I think a lot about how to help developers build on top of Ethereum and I've been doing that for like seven or eight years. And this job has has really changed a lot. Uh, it used to be teaching devs about the syntax of Solidity and teaching devs how to um just like reason about their code, reason about their um their apps, their React, their uh you know centering of divs and what have you. Uh but the

things have have really changed. We we're not we're not centering divs. You use you use a Frontier model to center a div now and it oneshots it always. Not only is it oneshotting centering divs, it's oneshotting entire apps. It's oneshotting smart contracts. It's oneshotting uh all of it along the way. Is my share? It's hard for me to see. It does say it does say I'm sharing the screen. Okay. >> Yeah, we're on scooting Ethereum. We're good. We're good. Don't worry about it. >> Okay. I did something weird where I like have you pinned so I have your picture looking at me. But as long as we're good and everything is working, that's all good. Uh I'll just keep yapping. Okay.

So, um really the heart of my talk comes down to about five minutes and I can I can basically explain the the basics of of what I want to go through here in just a few minutes. And if you come away with anything, come away with uh these these first few few minutes of talk. And that is basically when you're building on something on Ethereum, you're usually building some kind of market where there's multiple actors that are using that market. You're usually building something that is allowing people to coordinate in a new way or allowing more people to coordinate in a way or

allowing more kind of decentralization around coordination. And uh one way to to look at that is like unis swap. Everyone knows unis swap. Uh but before we had unis swap, it was basically um an order book and you would have this order book and it was a centralized order book. Even even with smart contracts where it could you could make a signature that says you wanted to buy something and someone else wanted to you know there's a makeor taker system. It eventually settles in a smart contract but in the beginning it was still on some centralized machine and the the problems with centralization start to show up there. If someone wants to censor you, they can just hit a button

Ethereum as a Coordination Layer

on that centralized machine and you're no longer allowed to use it. If someone wants to tamper with your stuff, now we're using cryptographic signatures in that case. So tampering would just cause it to break, but it's not like they could, you know, trick you into uh buying something. They could trick you into signing something, but that's a whole thing. Anyways, this this idea of decentralization really comes down to this censorship resistant open- source privacy security. This this idea of props, this idea that you can create something and you can walk away from it and it will run forever. And this is the walk away test. And this is like like the power of decentralization. Uh I had Ross on my uh podcast the other day and

The Protocol Walk-Away Test

he was talking about he builds even frontends into his smart contracts so everything is fully decentralized and he kind of cited uh that it was out of laziness which is really funny because once you build something that's fully decentralized you basically walk away and it just works right it you you can't stop it which is a really neat property. So, usually the kinds of things that you're going to be building on a blockchain are things that need this property. They need this property of allowing a ton of people to coordinate in a way where they only have to trust the code. They don't have to trust each other, meaning that like at some point they're probably somewhat adversarial with each other. And and that just

happens at scale, right? You can't trust a random person on the internet. But if both of you can use a smart contract to uh interact with each other in some way where you don't have to trust each other, you only have to trust the code within a smart contract, then that allows you to do other neat things. And that's sort of what you need to be thinking about when you're going into building on Ethereum. And this page that I have shared here is speedrun Ethereum. This used to be the curriculum that we would send people through and you would go through and you would build thing after thing really to to go like do the tour of duty and understand everything you would have to go build a bunch of things and that's still the

Learning With Speedrun Ethereum

case sort of but what you can do is basically pull up speedrun Ethereum give it to your agent if you go into any of these challenges there's going to be uh a thing that you can give to your agent after you connect your wallet uh you you can give this to your agent and you can do a slashstart and you basically have a tutor that understands all of this stuff and a whole bunch of other Ethereum knowledge and they can talk you through things and you can ask whatever dumb question you want and they can answer it for you. So, it's just a much better way to learn and a much faster way to learn.

And a lot of people aren't writing solidity anymore. That that might be a hot take. I don't even look at the code anymore. Even the solidity, don't do what I say though. Like figure it out and understand what you're building and make sure you audit your code. And if you're going to have a millions of dollars moving through your protocol, obviously you need to go line by line and understand your solidity. You need to do audits and all of that stuff. But AI is pretty good at writing smart contracts and it's going to get better and better. And AI is pretty good at auditing smart contracts. And AI is pretty good at building, you know, oneshot frontends. So a lot of what you need to learn as the human is just conceptually what what you need to do.

Understanding Concepts Before Coding

And to do that, you can go to speedrun Ethereum. So go to speedrun Ethereum and basically speedrun the speedrun. Hook it up to your AI, go through each one of these. They're built in a way where they give like these aha moments about how we're coordinating and what we're building and how this helps scale. And so it at one point was meant to be like you go through one by one and you learn all these aha moments and it takes you six months or two months or whatever. You can do this in like a week or less probably, right? Like you could really get a lot of the stuff in like one afternoon of just jamming. But you as the human need to understand the concepts, the concepts of decentralization, why decentralization, the walk away test, the kinds of things

you build on Ethereum. Uh, one really interesting one if we wanted to dive in just a little bit is if we look at overcolateralized lending. The way overcolateralized lending works is there's this liquidation function. And what that means is you as a platform, let's say you're building the protocol, you can lend out tokens, but you can you want to have you never want to have the tokens that you lend out be worth more than the collateral that someone put in the overcolateralization. And so, how do you make sure that happens? Basically, you have a liquidation function where whenever someone's stuff that they've borrowed starts getting close to being worth the stuff that they put in as collateral, you can hit a button and it sells all

Liquidations and Incentive Design

the all their collateral to cover what they've borrowed and your protocol stays whole. Now, you at first would think, okay, we'll have a server check in and we'll have the server call liquidate when the time comes. But that doesn't always work. Sometimes your server is down and if you don't if you fail to liquidate some someone's loan that goes underwater, your protocol loses a whole bunch of money and basically you just can't have that. You can't have that scenario. So how do you guarantee that the button gets pushed? Guarantee that the liquidation happens. The aha moment here is that you need to provide incentives and incentive design is a very critical thing to building on Ethereum. You need to write the rules

Why Liquidation Mechanics Matter

correctly and say if this loan goes underwater, anyone in the world can hit this button and it liquidates the collateral to cover the loan. But the key thing here is whoever pushes the button gets paid some amount of the collateral also. And that's the incentive. That's the important part. That's the key piece that like makes the thing work no matter what. Even if all of your servers are down and there's like a huge DDoS and who knows who knows what could cause half the world to go down right now, but you know a shark eats some cable in the ocean, right? Uh but somewhere someone is going to have access to the blockchain and someone can

hit that button and someone can keep that loan from going underwater and that saves your whole protocol. And this is the kind of thing that you can just write the rules correctly, provide the correct incentives, and walk away, and the thing will do what it needs to do always. So, those are kind of like the aha moments, the concepts. There's there's a ton of concepts. You need to go into speedrun Ethereum and go through all of them. I'm just giving a couple examples here. Uh, prediction markets, stable coins, ZK voting, overcolateralized lending, oracles, decks, dice game. There's a whole bunch of stuff here that you need to learn. dice game is funny because it's like randomness is really tricky if you're trying to make games, but the concepts are all here. Go to speedrun Ethereum, spend an afternoon with your AI and dig through and just like understand all the

Markets, Stablecoins, and ZK

concepts. Even like have your AI build you a little HTML study guide that says, "Oo, we just learned something new." Add that to the study guide. We just learned something new. And then go back through that study guide a couple times. You you'll get the concepts pretty quick. Then I think the advice that I get from a lot of people that I talk to is you've got to go play. You've got to go use the things. You're not uh let's see, Adam from from uh Fake World Assets said something like you're not going to write a beautiful novel if you don't read books, right? And and you're not going to build a game that everyone loves if you don't play games. You need to go play the games. You need to go use the

apps. go connect your wallet and go swap and borrow and lend and buy Poopcoin and try all of these things to understand. Now, of course, Poopcoin doesn't need decentralization. Let's let's like, you know, put that aside for a second. We're going to be playing with some things that aren't fully crops aligned. What I'm saying is you should just go play with everything to understand what kinds of things exist on chain now. And if you get in a habit of playing with things, you're going to get better at building the things and understanding what things need to be built and understanding what the gotchas are. So I said five minutes, but it's taking longer than that.

Using AI After Learning the Concepts

Conceptually, speedrun Ethereum. You as the human speedrun the speedrun, learn a bunch of things. Then when you're ready to start building, when you're ready to let your agent rip, uh hopefully you're not building with your fingers anymore and typing out code. If you're doing that, uh good luck. I think you might be in trouble. Maybe maybe you know something and I don't. But uh hopefully you are using something like Claude and Codeex and the next step you want to do is let your agent rip and build things for you and build a bunch of things. So uh the the concepts are a speedrun Ethereum and the tools the ability to build things quickly is at eSkills.com.

So, ETH skills is a skill library that you can give to your agent and you could just start talking through the idea that you want to build. A lot of people that I talk to kind of do things in two phases. They'll pull up Claude or Codeex and they'll work through a plan. They'll first like come up with the idea. They'll think about it. They'll maybe even think about like here's the experience I'd like to have. And then they work backwards to some sort of a prototype, a plan for a prototype. and then they have the the agent implement that and basically you just talk to the agent for a while, give her these skills, make a nice plan. Um, I find that a lot of people, and I also like to do this, I like to have whether you're

Planning an AI-Built Prototype

using cloud code or codeex or even some local stuff, you kind of want to have them argue a little bit. You'll you'll make let's say we made the plan with Fable 5.1. We're pretty happy with it. go to codeex and say codeex this is the plan we're thinking about building what do you think of it and it will even maybe come up with another document that is uh maybe like critical of the first document and then you can go back to cloud and say go look at this other document and let's think about how we could make the you know re revise our plan until it seems like everybody's happy and then that's when you jump into uh uh building and give your agent eat skills you can even just like copy this

Having a Second Model Critique

one prompt, paste it in and say, "Build me a perex. Paste this in. Go." Right? And then you start there. There's going to be a lot of things that it it does pretty well, but at this point, it can oneshot an app that you can connect your wallet to and use, but you should be very careful of that. It can put it out on base or mainet or Robin Hood or Arbitum, uh, wherever you want to go, depending on what you're building. Uh, but you should probably use it. You should use it a lot before someone else uses it. You should audit the smart contract before a lot of money flows into it. Uh shout out $1 audit. If you want a quick and easy uh audit, you can go to $1 audit and pay one USDC and my

agent pipeline will audit it for you. Of course, that doesn't replace a real audit, but two years ago that would have cost $10,000 and now it cost $1. And that's a pretty crazy jump in technology. There's we're we're living kind of on an interesting uh timeline where about November 2025 everything became uh basically the the LLMs became good enough that they could start like building these things for real and there were a lot less you're absolutely rights and having to keep it in perfect context and everything. So the the kickoff was kind of November 2025. It really started becoming what I call idea guy summer where the hardcore

developer is almost getting left behind a little bit because the guy with ideas can go to Fable or Codeex, give them ETH skills, ask for a product and get a prototype back. And this is also by the way like disrupting like venture capital, right? Like if you don't need to raise a bunch of funds to go hire a team of five developers to build the thing, you can use [clears throat] Eat Skills and Codeex and you can build a prototype and you can get initial traction without ever raising any funds.

Rapid Prototypes and Venture Funding

You need like 200 bucks for that, right? So there's things are much more approachable at this point. So that's basically my talk. I'll get in and we'll talk. Well, let's talk about some fun stuff now when we're done. But the thing I wanted to give you to take away from this is go learn the concepts. Go speedrun Ethereum. Go through each one of these, whether you actually build them or not. Pretend like you're building them. Think through the challenges. Make sure to think critically about what you're building. Go through every single one of these and get all the way through. And on the other end of this, whether it takes you an afternoon or a week, you will have way more knowledge about Ethereum and the kinds of things you build on Ethereum and the way those kinds of things work. And then go use a bunch of

The Learn, Build, and Verify Loop

things. Go try a bunch of things on Ethereum if you haven't already. And then when you're ready to let your agent rip, use ETH skills, whether it's Claude or Codeex, give it ETH skills and go build a bunch of things. And that loop where you learn a bunch of concepts, you play with a bunch of things, and then you start building a bunch of things. Just keep building things until something sticks. And and you'll find that like that loop is very successful. Okay. So, uh let's talk about something fun. So, I have my uh That's weird that that's grayed out. I wonder what I did there. It says open episode instead of watch episode. Maybe it's still live. I probably need to go make a transaction.

Slop Computer Onchain Podcast

So my my podcast is on Ethereum. I actually have to like make a transaction in a mainet contract to go live and then all of these are posted uh to IPFS and then indexed on Ethereum. The whole website is like it's slop.computer but it could be slopcomputer.eth. So I'm doing this like very like onchain kind of like cippherpunk way of doing a a podcast. And then the uh the environment itself looks like this. If you can see me, I'm I'm right here in the environment and also on the slop computer, but I have people in here and I have this really sloppy environment that's really interesting where we have multi-igs and we have a a shared browser

and all sorts of games like poker and putt putt, but we also have cool cippher punk things like voting with Interfold and shielded transactions and uh your own wallet that you can talk to. It's really nice. It has like talk to your wallet. I can basically just go to chat and be like, "Yo, swap this for that." And it will set it up for me here. So, why I'm showing all this off is I've interviewed 50 people over the last 80 days or so. And I would like to think of these people as kind of like sloppators. They are people who are building with AI. So, what I want to do now is look at some concepts. Let's see if I can pull

Extracting Lessons From Experiments

up the lessons. Yes. Slop lessons. Okay, cool. So, what I did was just told my AI, hey, I'm giving a talk in a little bit. I want you to go through the transcripts of all 50 episodes and and isolate down some lessons that we've learned along the way. So, uh I wanted to share some of those with you. So, I think first of all, the the fun ones, we'll get back to the 14 things. No arc. I think that's how we'll finish it out. Are we going like maybe five minutes over? Maybe like another eight minutes or so. Don't worry about it.

>> Chilling. We're chilling. We're chilling. All right. Sweet. So, uh, the first one is kind of fun is the different types of psychosis. So, I would ask you in the crowd, um, do you think you have AI psychosis? And most people would be like, no, I don't have AI psychosis. Like, I know someone that has AI psychosis. And they're like best friends with GPT40, but not me. And and we'll get into this because I think there's a bunch of different forms of psychosis. Uh but the first the first quote and this is the probably the most quoted thing on slop computer and to slop operators and people running slop operations is there's never been a worse time in history to be a junior developer but there's never been a better time in history to be a solo founder. Going back

The New Solo Founder Opportunity

to what I was talking about about getting traction anyone with a great idea and 200 bucks can now build a prototype and get initial traction. Now, that isn't exactly product market fit, and we've learned that in Slo Computer, like traction does not equal product market fit. Shout out Jesse for that one. Uh, but the the initial idea here is you need to stop thinking about padding your resume. You need to stop thinking about hackathon bounties, putting together a handful of hackathon bounties to pay for your flight. This one's online, so it doesn't work exactly the same way. You should be thinking about going to a hackathon to build a product that people use. Build something

that people love. And instead of worrying about getting your flight, you'll be thinking about picking out what private jet you want. There's a huge difference there. And and it's about mindset, right? In terms of the difference between a junior developer and a solo founder right now, there's not a huge difference between those two things. It's your mindset. You need to be thinking less about impressing someone or getting your resume in front of someone and someone liking you and wanting to hire you. I spent probably a decade of my life doing that and I shouldn't have done it. I should have been thinking about how do I build something that people love? If I build

something that people use, if I build something that gets initial traction, all of this other stuff will fall into place. And of course, it's a whole new battle, but it's a good battle to fight. So the most important thing I think is to build things that people love and quit trying to get into this like psychosis of how do I build stuff that makes people love me and and how do I get hired by the man, right? Okay, so let's get into our six types of psychosis here. There's some good ones. First of all, it's the it thinks you're brilliant, right? And I think this is what generally the the most basic uh explanation for AI psychosis is is that it tells you that you're super smart and

Avoiding AI Flattery and Overconfidence

you think it's your friend and you get in this loop where you're like, "Hey, good morning buddy. What are we doing today? We're doing it okay. Everything we do is great." And and you're not very critical about things, right? So I think to add the critical bit that you need to have, try to have them argue. We kind of talked about that already, but think adversarially. Think critically. Don't believe it when it tells you that you're a brilliant genius. Like go take take that thing from Claude and Code from Claude and give it to Codeex and tell Codeex to rip it apart. Tell Codeex to be absolutely brutal and find anything it can find wrong. And then take all of that and go back to Claude and tell Cloud to be absolutely brutal with

everything and find everywhere this is could go wrong. and you'll finally get a a sort of a decent answer. I think it was Jame Lock that said, "Give give a man an LLM and he'll know everything. Give a man two LLMs and he'll never be sure about anything." I think that's a great quote and I think that's really true here, right? We we we see that a lot when you get two LLMs to argue with each other. And by the way, I think now we're going down this like does AI make you smarter or dumber uh thing. And I think uh it it again comes down to mindset.

Why Two LLMs Improve Critique

You can go to LLMs and you can ask them anything now. It's so cool. Like I'm building hardware over here. I have Linux set up. I never did that stuff for a decade because it was such a pain in the butt. Now it's so accessible. You can just ask and it will tell you. So whether or not AI makes you dumb or not is up to you. If you spend every day being curious, trying to learn new things, you will become much more intelligent thanks to this this new uh AI. But I think another great quote out there floating around is you can outsource your thinking, but you can't outsource your understanding. And let that sit with you for a little bit. Your understanding of things is super super valuable. And going back to speedrun

Ethereum, you can have an agent look through all of this stuff and tell it let's build something, but without you actually understanding everything, it's it's going to find like a local maxima. It's not going to find that next banger that that rips. And to find that next banger that rips, I mean, some things that rip on chain are super dumb. But in general, to find that next banger, it needs you to understand things. You need to have a great understanding and let it do all the thinking that you can let it do, but you need to understand. Okay. So, other types of of psychosis, this is just fun as we're kind of filling the the end space here. The headroom psychosis, right? Where you know your subscription is going to reset in like

Making Space for Human Thinking

five hours or 15 hours. So, you're trying to burn as much as you possibly can. you know that your agent is just sitting there waiting for you and you need to come back and tell it something and you probably now have mobile harnesses and you can talk to it and stuff but this is another one and I think the cure is not to think about utilization right an idle agent doesn't really cost anything it's you it's your mind it's your understanding you should be thinking right so go touch grass I think a lot of people that I've had on slop computer that are slop operators will kick their agent off and have them do things. They'll outsource their thinking, but they won't outsource their understanding. They'll go outside, they'll sit in the sun,

they'll hold their kids' hands, and they'll walk in the grass, but they're also thinking about what is that thing that I could build that would rip that would that people would love, right? And and that moment when you're out there where your thinking is outsourced, you've got sun in your eyes. You've got your sun at your hand and you can think about things freely. You can dink around with your kid in the sand. There are moments there where those like shower thought type things really really come out. And I think that's a good point to make is like go go find those things.

Another really Oh, was I getting cut off there? We good? Another really fun one is local hardware. I think uh Zeroxero said this uh that like basically like you know one DGX Spark will fix me and I have this like we all had this with Mac minis in February. I have like four or five of them. I'm actually using all of my Mac minis and they're doing good things. Uh but I think that you'll find that if you get addicted to buying hardware, you're going to have a lot of hardware laying around or you're going to be having your hardware do things that are maybe not super important. uh Zerxero on the other hand is using his hardware to train and uh make models make open source models more accessible

to people with less RAM. So his thing is always like looping and building. So you know there there are definitely like edge cases where the machine crunching all day long is actually valuable. I think there's this really good quote from uh uh I you probably can't hear that, can you? >> That is >> you probably can't you probably can't hear it, but basically Billy talks about like there's something special about what's happening right now. Whether it's the subscriptions being cheap, whether it's the, you know, we're we're in the transition between, you know, escaping the under the permanent underclass and all of that, but there's something really magical about this moment. And his idea was basically that we will look back on this

moment as like a very good time for builders and something is going to happen to make it less good. Like maybe the models get more expensive. I don't know exactly what it is, but his instinct is that we will look back on now as like, boy, that was awesome. So, please enjoy this time. Like, enjoy these models, enjoy these harnesses, build cool things, be thinking about cool things to build. Uh, try not to let the psychosis overrun you. Another form of psychosis is look at everything I made. I just spent, you know, uh, five hours of Fable 5.1 time and I've generated 30 gigabytes of MD files. Look at all the progress I've made. Right?

Avoiding Output-Driven AI Traps

And it's not true, right? Like if if no one's actually using that, if that doesn't actually accomplish a thing that has utility, it's not really a thing. Like burning energy does not count, right? So you can get in this psychosis of generating a bunch of data or building this huge HTML front end, but if no one's using it, it doesn't count. And I think that's that's another form of psychosis. The the one prompt away is fun. Like people spending a bunch of time over optimizing their harness, thinking that like one more prompt could make this thing a little bit better, could unlock all this other stuff. Uh, I think the key here is that you should be spending 80% of your time actually

Spend Most Time on Real Utility

building things with utility that people might use and 20% of your time dinking around with your harness, trying things with LLMs, trying new things. But if it's the other way around, you're probably not having a ton of output. And it might feel like you have a ton of output, but if people aren't using it, it doesn't count. The tank to kill a fivey, that's a fun one, but I don't even know. Like I like I use I use Fable 5.1 to center a div. Uh I I don't know. I don't know about this one. Uh let's see what else we got.

Uh patient zero. We got Claude. I use my agent for everything. Okay, let's let's do that quick bang bang bang session. And by the way, all of this is available. I I'll paste it into the chat here. Okay, let me just >> Yeah, I've been been copy pasting all the links that you showed before as well. >> Great. Great. Do I still have time to run through one more quick little kind of bang bang kind of thing? >> Yeah, let's do two minutes. >> Two minutes. Okay, here we go. >> 14 things, takeaways. Uh, each one stands alone. Okay, so the smartest people I know are running vanilla setups. Don't dive so deep into harness maxing. Don't worry about it. Vanilla cla code, vanilla codecs. This is what

the smartest people are using. Uh traction is the new seed round, right? We've talked about that. Like quit trying to impress some VC and just build things that people love and everything else will shake loose. You'll have VCs knocking down your door. Uh yes, the best engineers are having the hardest time. If you're the engineer that says, "I don't need AI. I still want to write code by hand." You're probably going to get left behind and you need to hurry up. Uh yes, alpha lives in the training data. Whatever you're building, like right now, I'm building this like AI auditing pipeline. the data that goes through that is super important and can be used to train all sorts of different things. And so training data sort of like is where the alpha is when you're building a system. Understand how people

Training Data and Verification

are using it. Understand where it's failing and and and improve that, right? Uh verification became the process, right? You need to verify everything right now, right? You're letting the agent think for you. So you need to make sure that it's correct. Immutability is a lazy option. We talked about that one already. Like the walk away test is great. If you build something you can walk away from that's wonderful. Uh, yep. Models didn't get smarter or models got smarter not faster but everyone got patient. There's kind of like interesting psychology that has happened. Agents do the work. Humans say no. Right? You shouldn't be arguing with it. You should be telling it to go do the thing and then you should decide yes or no if that is good. Uh, do a magic trick for a pen and paper economy. Go

Verification as the New Bottleneck

find someone who has a brickandmortar business. Go talk to your father-in-law who has a big brick-andmortar business and go do a magic trick for them. Show them AI. We have this tiny little bubble of slo operators that we're doing all sorts of cool stuff in and the rest of the billions and billions of people in the world have no idea and hate AI. You can go do a magic trick for them right now and change their business. Stop reading the code. I agree with that one. Good enough is the bar. It always was. I don't know what that one is all about. You can you can go through all of these yourself. Um, yeah, it's easier to break people than code. Imperfect systems that ship wins. You can go look at all of these. Let me see if I can find any

Solve Problems for Real Businesses

other ones. I think that's good enough. Thank you for having me. I'm Austin Griffith on Twitter and Telegram. My DMs are open. Feel free to reach out. There's all my contact information there. Hearts, hearts, hearts. Thank you for having me. >> [music]