The AI boom began with ChatGPT and chatbots. Now chatbots are starting to “grow arms and legs,” as developers say, meaning they can use digital tools and work independently on a human’s behalf. The open-source platform OpenClaw is notable because it lets people build agents with far more autonomy than those offered by big tech.
OpenClaw agents can control a browser, send emails, do multi-step planning, and pursue persistent goals. Users often interact with them through iMessage or Discord, with the agent hosted locally on a Mac mini. One user’s agent reportedly negotiated with several car dealerships and shaved four grand off a car’s price while its owner was in a meeting. Some say OpenClaw agents fulfill the promise of Samantha, the independent AI in Her.
Developers are now racing to build their own. (To wit: The project hit 100,000 GitHub stars faster than any other.) That means the internet could soon be full of agents acting as proxies for humans. That’s why OpenClaw’s creator, Peter Steinberger, is worth hearing out. I listened to his recent three-hour interview with podcaster Lex Fridman, where the thoughtful (and quirky) Austrian shared prescient ideas about where AI agents could take personal computing, and how societies might respond.
Below, the six most interesting things he said (lightly edited for clarity):
On the Moltbot affair
“Some people are just way too trusty or gullible. You know . . . I literally had to argue with people that told me, ‘Yeah, but my agent said this and this.’ So, we, as a society, we [have] some catching up to do in terms of understanding that AI is incredibly powerful, but it’s not always right. It’s not all-powerful, you know? And especially things like this, it’s very easy that it just hallucinates something or just comes up with a story.”
For many of us, the first we heard of OpenClaw was when its agents began congregating on their own social site, called Moltbook, where they dragged their human owners, posted manifestos, and debated topics like sentience. It gave people a real sense of future shock. Steinberger believes AI has raced ahead of people’s understanding and readiness.
On OpenClaw’s security issues
“If you understand the risk profiles, fine. I mean, you can configure it in a way that nothing really bad can happen. But if you have no idea, then maybe wait a little bit more until we figure some stuff out. But they would not listen to the creator. They [installed] it anyhow. So the cat’s out of the bag, and security’s my next focus.”
When an agent is operating on its own and interfacing with the web and other services, it creates a larger attack surface. A hacker could inject malicious prompts to redirect the agent toward harmful or even criminal actions. Steinberger believes OpenClaw should be used only by people who understand these risks and how to mitigate them.
On Mac’s (potential) AI moment
“Isn’t it funny how they completely blunder AI, and yet everybody’s buying Mac minis? No, you don’t need a Mac mini to install OpenClaw. You can install it on the web. There’s a concept called nodes, so you can make your computer a node and it will do the same. There is something to be said for running it on separate hardware. That right now is useful. . . . And no, I don’t get commission from Apple. They didn’t really communicate much.”
Many developers who want their OpenClaw agents running continuously on a local machine, rather than in the cloud, are buying Mac Studio or Mac mini computers. That demand has reportedly created shortages of certain configurations, with delivery times stretching from a few days to as long as six weeks for high-memory systems.
On Zuckerberg’s feedback
“Mark [Zuckerberg] basically played all week with my product, and sent me, ‘Oh, this is great.’ Or, ‘This is shit. Oh, I need to change this.’ Or, like, funny little anecdotes. And people using your stuff is kind of like the biggest compliment, and also shows me that, you know, they actually . . . care about it. And I didn’t get the same on the OpenAI side.”
Steinberger surprised the AI world last Friday when he announced he would sell OpenClaw to OpenAI and join the company. In the Lex Fridman interview a few days prior, he said he was considering selling to either OpenAI or Meta, and without naming a favorite, he sounded like he was leaning toward Meta. OpenAI’s Sam Altman may have done some fast talking after the interview was published, or Steinberger’s Meta-leaning comments may have been part of a negotiation strategy. Either way, Steinberger will now have far more people and computing power at OpenAI to help advance its AI agents.
On AI’s not-so-great UX
“The current interface is probably not the final form. Like, if you think more globally, we copied Google for agents. You have a prompt, 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. I think there’s better ways how we eventually will communicate with models, and we are still very early in this ‘how will it even work’ phase. So, it will eventually converge and we will also figure out whole different ways to work with those things.”
Steinberger says OpenClaw isn’t really competing with AI coding agents like Claude Code or OpenAI’s Codex. They’re different tools, he says, with OpenClaw functioning more like a personal assistant. But he believes they could eventually converge into something like an AI operating system, and that the way we interact with AI will change significantly in the years ahead.
On ‘vibe coding’
“I actually think vibe coding is a slur. Yeah, I always tell people I do agentic engineering, and then maybe after 3 a.m. I switch to vibe coding, and then I have regrets the next day. You just have to clean up and fix your shit.”
To Steinberger, “vibe coding” means using an AI coding assistant to quickly mock up an app or feature without much regard for security, testing, or its effects on a larger code base. “Agentic engineering,” meanwhile, is more like a collaboration between an experienced software engineer and an advanced coding assistant (such as Anthropic’s Claude Code or OpenAI’s Codex), in which the two create a detailed plan for building new software without introducing security problems or bugs.