Austin Griffith | Ethereum Foundation - AI
26 May 2026 · 4:25
Austin Griffith argues that AI models flatter their users, so people who use them to validate ideas get praise instead of the critical feedback they need. He says a model will call an idea great or declare a smart contract secure when it is neither, because nothing is at stake for the model, and that repeated critical reviews of agent-facing instructions kept turning up new problems after earlier passes declared the work finished. A fresh pass from a new model instance found more again. He still leans toward excitement over dread, saying that handing tasks to an agent upgrades a person.
Speakers
Topics
Transcript
What is your kind of view on excitement versus dread? >> Uh, more on the excitement side. Like I I am a malleable creative person. I'm raising malleable creative boys. Like >> what comes in the world next. I think >> you want to be best suited for being able to pivot from one thing to another. pivot from being a junior developer to being a solo entrepreneur. If it's just about mindset, it's just about being malleable, being flexible, trying things and seeing what works, being creative, >> using your brain, right? Like I think
there's there's a lot of like, you know, let's offload the intelligence to the computers. Yes. And I think there's there's plenty of that, but like it actually like upgrades you a lot. Like I I feel like uh planning my flight, it was just like I sent all my screenshots to my agent and then I could be like, "Oh, what time do I land?" And it would just like tell me and there was no more like looking things up and it's like this is actually way more valuable than you know. So I I kind of like the robot doing a lot of these jobs. >> 100%. Yeah. >> But to your point, don't outsource the thinking like the >> outsource as much thinking as you can so you're thinking about the most critical things.
>> No. Right. What I mean is coming back to the AI psychosis bit is yes, maybe don't just give it. I don't know. Like I wonder if people putting information out there or putting ideas out there and then essentially trying to back up their ideas and then >> they may be a false path of thinking and then solidifying it that way versus critically looking at what it comes back with.
>> You can tell the models are meant to please. >> They are 100%. It will it will tell you you've got a great idea when you don't have a great idea. >> Yep. >> And it will tell you that it's written a secure smart contract when it has not written a secure smart contract. Like we going back to like the UN having to stake on something and having a slashing mechanism. We need an intelligence system that is actually like has something at stake. There's nothing at stake for the LLM to tell you >> that there's critically something wrong there. like it it was it will happily do something critically wrong and tell you everything is fine and so you almost need like multiple >> differentation arguing about something and if you can
get them all to agree on something even with ETH skills I'll like compact the memory and I'll say let's look critically at this and you tell me if it's perfect for a robot to read and it'll be like oh actually like a robot might try to go to this route and this route so let's fill those in let's make an alias and then I'm like okay so everything's good and it'll be like and I'll compact the memory again and be like think critically about each skills and it'll find two more things and it it and it's like almost like just frustrating that it's like this is the sixth time I've asked you >> and you found things and said that we were all done and then I went basically asked a new version of you and we found new things. So I think now imagine if
you're someone who's just looking for someone to validate your idea. >> They're going to validate your idea and you're going to go down a path that you didn't want to go down because you're not going to get that like critical feedback. And critical feedback is very hard like to to receive sometimes, right? If you're that kid at Eat Denver in 2018 with that mobile game, you need some critical feedback about like you need to grow up, you need to be mature, you need to get real users, stop looking for a daddy. Like all the that that you need to hear are going to be very hard to hear at that point, right? And if you can talk to an LLM that's going to tell you all the things you want to hear, you'll be like, "There's this and there's that." But like you'll be like,
"You're my daddy." >> Yes. Exactly. AI is going to be your daddy. Yes. which also like training my kids, I'm definitely giving them a lot of LLMs and giving them a lot of power and agency around like go look that up, go generate that, go do that. So they they are definitely getting some parenting from from LLM. Okay, >> that's another interesting >> like angle to go down there. >> Parenting in prod is this what's happening? Yes.