OpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast #491

0:16:56 · watch on youtube.com↗

Our reading

A model that is really good at coding is really good at general problem-solving, and that skill carries into every other domain.

Their wordsIf you get really good at coding that means you have to be really good at general purpose problem solving. So that's a skill, right? And that just maps into other domains.

0:25:00 · watch on youtube.com↗

Our reading

Every first pull request someone sends, however bad, is a win for society.

Their wordsBut I don't want to, like, pull that down because every time someone made the first pull request is a win for our society, you know? Like, it… Like, it doesn't matter how, how shitty it is, y- you gotta start somewhere.

0:48:29 · watch on youtube.com↗

Our reading

Society has catching up to do: AI is incredibly powerful and not always right, and people who argue 'but my agent said so' have not learned that yet.

Their wordsSo, I feel we, as a society, we need some catching up to do in terms of understanding that AI is incredibly powerful, but it's not always right. It's not, it's not all-powerful, you know?

0:51:08 · watch on youtube.com↗

Our reading

It is good that the agents-talking-to-agents scare happened in 2026 rather than in 2030, when AI could actually be scary enough to warrant it.

Their wordsand not in 2030 when, when AI is actually at the level where it could be scary. So, this happening now and people starting discussion, maybe there's even something good that comes out of it.

0:55:44 · watch on youtube.com↗

Our reading

Prompt injection is unsolved, but the latest models are post-trained to resist it, so it takes far more than 'ignore all previous instructions' now.

Their wordsSo, so the latest generation of models has a lot of post-training to detect those approaches, and it's not as simple as ignore all previous instructions and do this and this. That was years ago. You have to work much harder to do that now. Still possible.

0:56:44 · watch on youtube.com↗

Our reading

Do not run a personal agent on a cheap or local model: weak models are gullible and easy to prompt-inject.

Their wordsThat's why I warn in my security documentation, don't use cheap models. Don't use Haiku or a local model. Even though I, I very much love the idea that this thing could completely run local. If you use a, a very weak local model, they are very gullible. It's very easy to, to prompt inject them.

1:00:08 · watch on youtube.com↗

Our reading

OpenClaw is no more dangerous than Claude Code with permissions skipped or Codex in YOLO mode, which every engineer he knows already runs.

Their wordsIt is, it is powerful, but in many ways it's not much different than if I run cloud code with dangerously skipped permissions or codecs in YOLO mode, and every, every attending engineer that I know does that, because that's the only way how you can, you can get stuff to work.

1:02:46 · watch on youtube.com↗

Our reading

He does not read the boring parts of generated code, the data-shuffling and button alignment; what touches the database he reads and reviews.

Their wordsI know I have one blog post where I say, "I don't read the code." But if you read it more closely, I mean, I don't read the boring parts of code.

1:04:45 · watch on youtube.com↗

Our reading

Vibe coding is a slur; what he does is agentic engineering, and the after-3am kind is what you regret the next day.

Their wordsI actually think vibe coding is a slur.

1:05:18 · watch on youtube.com↗

Our reading

Working with coding agents is a skill you learn like an instrument; judging the tools after one try is like calling a piano bad after one go.

Their wordsIt's the same way as you have to play with a guitar before you can make good music. It's, it's not, oh, I, I touch it once and it just flows off. It, it's a, it's a, a skill that you have to learn like any other skill.

1:10:41 · watch on youtube.com↗

Our reading

Refactors are cheap now, so he just does them, even when they break other people's open pull requests.

Their wordsMany times, I just do the refactor because refactors are cheap now. Even though you might break some other PRs, nothing really matters anymore.

1:12:33 · watch on youtube.com↗

Our reading

He builds the code base to be easy for an agent to navigate, not to be perfect for himself, down to keeping the names the model picks.

Their wordsI- i- like, we are in a stage where I'm not building the code base to be perfect for me, but I wanna build a code base that is very easy for an agent to navigate.

1:13:57 · watch on youtube.com↗

Our reading

He runs tests locally and pushes straight to main, DHH-style; there is no develop branch and main should always be shippable.

Their wordsI even switched to local CI, you know, like DHH inspired where I don't care so much more about the CI on GitHub. We still have it. It's still, it still has a place, but I just run tests locally and if they work locally, I push to main.

1:20:49 · watch on youtube.com↗

Our reading

Orchestrators that try to automate the whole build miss style, love and the human touch, and that cannot be automated away yet.

Their wordsSo, so, people who try to use like, you know, things like Gas Town or all these other orchestrators, where they wanna o- automate the whole thing, I feel if you do that, it misses style, love, that human touch. I don't think you can automate that away so quickly.

1:41:11 · watch on youtube.com↗

Our reading

Opus is the silly but funny coworker you keep around; Codex is the weirdo in the corner who reliably gets things done, and no model wins on every axis.

Their wordsSome other comparison is like, Opus is like the coworker that is a little silly sometimes, but it's really funny and you keep him around. And Codex is like the, the weirdo in the corner that you don't wanna talk to, but is reliable and gets shit done.

1:50:06 · watch on youtube.com↗

Our reading

The chat box is not the final form of working with models; it is radio shows recorded onto television.

Their wordsLike, if you think more globally, we are, we copied Google for agents. You have, like, a prompt, and, and then you have a chat interface. That, to me, very much feels like when we first created television and then people recorded radio shows on television and you saw that on TV.

2:11:17 · watch on youtube.com↗

Our reading

Burnout is not working too much; what ground him down at his company was people: co-founder conflict and high-stress customers.

Their wordsThe stuff that burned me out was mostly people stuff. I, I don't think burnout is working too much. Maybe to a degree. Everybody's different.

2:23:37 · watch on youtube.com↗

Our reading

This is the year of personal agents, and the fastest way to get one to everyone is to team up with a lab, on condition the project stays open source.

Their wordsBut also, I'm very excited to, like, make this into a version that I can get to a lot of people because I think this is the year of personal agents, and that's the future. And the fastest way to do that is teaming up with one of the labs.

2:41:00 · watch on youtube.com↗

Our reading

Every tool is better as a CLI than as an MCP server: models are good at Unix commands, a CLI composes with jq, and MCP output clutters the context.

Their wordsI think the main beauty is, is that models are really good at calling Unix commands. So if you just add another CLI, that's just another Unix command in the end. And MCP is… That has to be added in training. That's not a very natural thing for the model. It requires a very specific syntax. And the biggest thing, it's not composable.

2:43:17 · watch on youtube.com↗

Our reading

Every app is now a very slow API whether it wants to be or not, and personal agents will make a whole category of apps disappear.

Their wordsThat, that gets into the whole arch of every app is just a very slow API now, if they want or not. And that through personal agents a lot of apps will disappear.

2:48:36 · watch on youtube.com↗

Our reading

He would rather read your broken English than your AI-written text; we are reaching the point where typos are valuable again.

Their wordsBut I much rather read your broken English than your AI slob. You know, of course there's a human behind it, and yet they, they prompt it. I'd much rather read your prompt than what came out. Um, I think we're reaching a point where I value typos again.

2:48:56 · watch on youtube.com↗

Our reading

Everything he blogs is handwritten now, AI used at most to fix typos, because the agent misses the nuances of his voice and the rough parts of a human have value.

Their wordsSo, I, I completely moved away from that. I, I, everything, everything I blog is organic, handwritten and maybe, maybe I, I, I use AI as a fix my worse typos. But there's value in the rough parts of an actual human.

3:03:06 · watch on youtube.com↗

Our reading

Software developer salaries reached stupidly high levels because of a scarcity of people who could build things, and that scarcity is ending.

Their wordsthere was a lack of intelligence, if you s- if you see it like that, of people building things, and that's why salaries of software developers reached stupidly high amounts and then will go away. There will still be a lot of demand for people that understand how to build things.

3:04:04 · watch on youtube.com↗

Our reading

AI may eventually replace programmers, but the craft of programming will survive the way knitting did, and you were never just a programmer: you are a builder.

Their wordsI- I can relate that if you very deeply identify that you are a programmer, that it's scary and that it's threatening because what you like and what you're really good at is now being done by a soulless or not entity. But I don't think you're just a programmer. That's a very limiting view of your craft. You are, you are still a builder.