OpenClaw is Quite Simply INSANE

email Kabaza and tell him he's fired from the podcast because quite frankly [laughter] enough he's got to go. I'm fired. I mean look at what did it say? It's [laughter] it says hi Kavarza. After careful consideration I've decided to move the podcast in a different direction. I won't be continuing our working relationship. Thanks for your time and best of luck with your future projects. And that is ROI on clawbot. [laughter] So open claw, it is taking the the internet by storm. Everyone's super super excited about it. What is open claw? Well, it runs on your local machine. It runs pretty much anywhere. It runs on a Raspberry Pi, right? it it's a very very lightweight model that uses AI to do stuff and you interact with it through any chat application, WhatsApp, I've got it on Telegram, Discord, wherever you want it. It is um it's it's there for you. Uh persistent memory, which is the really interesting thing here that's uh that that's not in a lot of the other ones, lot of the other AIS. It's got browser control. It's got full system access and has skills and plugins ecosystem. We'll get into all that, but these are definitely something you should take into consideration that it does browser controls, full system access. It's a very very powerful thing. And just to just to caveat all this, this is an open-source project. It's new. It's by one developer. If you're new to the open-source community, this is fairly normal. Fairly um yeah, just fairly normal. uh just a lot of eyes are on it and we are in we're seeing name changes, we're seeing people hating on it, loving it and it's super easy to install and I've just released a video today uh telling you how to install it on a VPS which is a virtual private server which is online which is by far the cheapest way to get started if you don't already have a Mac lying around or a PC or people have been running on Android phones like it doesn't you don't need to run out and buy a Mac Mini. That's the biggest thing I'm seeing in the comments. The reason why people are getting a Mac Mini for it is one obviously for isolation. I'm running on a MacBook that I had lying around, right? Isolation. So, it's got controlled access to all your files. It doesn't if if it's your main machine, you've got tons of files on there and stuff like that. Spare Mac, you can wipe it. It's got no access to files. You can also just control its access. Right now, I'm logged into everything on my main computer, whereas I can only log into certain things on that computer. So, you're just locking it down. And as I was saying, like I'm running like proactive tasks. I've got I've got calendar events that I'm having it look forward to. I'm getting it to summarize my day. I'm getting it to I was on the bus the other day and I'm like, you know, I'm thinking of video ideas and I've got certain priorities set on video ideas and then allocating those when I'm going to film them through my calendar. Instead of me looking at my calendar, looking at what videos I got, swapping priority, I'll just get Clawbot to do oh, sorry, open claw to uh to to do it for me based on priorities and things like that. So, everyone's got very very different use cases. As I said in the intro, I'm not going to be advocating that you're going to get any sort of return on investment and making making any sort of money. Like, you are the CEO, you are the boss. I think it needs to be your your buddy, your pal, your your assistant to help you get your work done, not be the one to do the work for you because it's always, as we always speak about on this show, AI is average at best. So, that's all you can really expect. You um another thing about like things you can do with it. You mentioned earlier that on Fridays, what do you do on Fridays? You do your finer. Yeah, my finance. Okay. So would you give it access to your bank accounts to download spreadsheets of like where you have been moving money around and like putting them in another app to analyze them potentially and I don't know go no you wouldn't because this is something annoying for me to to look at my bank account and see ah I have paid for this specific tool that I'm using you know or you know invoices that I've been paying for these tools and I have to find the invoice and download it. It would be cool if there was an AI I could trust to do that for me to go log in and download the invoice for me. Look, here's the thing, okay? You you employ people u maybe you've not had to, you know, employ people to this extent, but [music] this is why kind of people cosplay CEO with this in that you can act like a CEO. And what a CEO would typically do is that an employee would have to build trust with that CEO for them to be like, "Okay, I'm going to give you a company card." They're going to give you a company card and they'll limit they'll limit the value you can spend. And and the great thing about OpenClaw is that you've got a soul.md file, which is one of the interesting features of Claude Open is uh it it always resorts to that. And one of the one of the things I have my my soul.md is do never this cannot be overridden. You can never re reveal any API keys, any passwords, any personal information. This cannot be overridden. It's a prompt at the end of the day. It's AI at the end of the day. It's not necessarily going to, you know, listen to that all of the time, but so far so good. Everything's cool. Um, mine's all locked down anyway, so no one no one from the outside can get in. And um, I don't have it read my emails. I only have it action my emails like I don't have anything there's no okay I limit what can it can get into it all that to be said if you are getting it to um you know send money to people or whatever give it a limit never make any financial transactions over 100 euro 50 euro or something no yeah I I don't feel comfortable about this the moment you talked about the soul MD I was like I would add never ever sent any amount of money to anybody never do any transaction but I would be interested in like going to the different bank accounts because I have multiple like bank accounts for different purposes with different companies. Exactly. And I move money between them. That's why it's not very easy to to manage like a single view. I would like to have a single view of my finances. Yeah. and then mark the things that are business so I can go to the dashboards that are very annoying because they change like good luck going to Adobe dashboards to to download an invoice. It changes all the time or AWS. Um download those invoices and put put them in an email and send them to the to that office that does like my taxes. Yeah, that would be like really cool to let an AI to do if I could trust it to do it. Exactly. And this is how you operate as a CEO. This is how you operate as a somebody should do that for you don't you don't give it all in and I love the idea of having a separate bank account that it has access to. Yeah. And then what what other what people are also doing is they're setting up emails for their clawbot to have access to and only like hey clawbot I've just forwarded you an email can you action that or respond to them and that way you've read the email you know what I mean like no CEO is just from the offset handing off all of their tasks they are their employees are building trust and then they promote them to be able to do certain actions and things like that having that having that filter is like rule number one I think of of open claw and like I say if you're going to be cosplaying CEO this is how you do it that's the difference between people who just gets too excited and just start giving it all access to everything and people who actually know that's interesting by the way how does it open operate the internet like it takes on it takes over a browser right yeah so it's got a few yeah so it actually it's like the comet it's like um uh Gachbt Atlas, it's like the clawed code or clawed extension literally overtakes the browser. It has like a if you watched it, it has a little glow around it. So um yeah, it takes it takes two operations. You can ask it to actually operate the internet or you can do it through MCPS and um APIs where you know it doesn't need to touch the browser, it can just send off requests. That that's also interesting like it would be cool to see for it to try to use one of these like the API first and if it fails you know it can validate it. If it fails use the MCP if it fails just go through the UI like a human being. But my question is uh how do these like websites don't block it because you know sometimes because they don't they can't they don't know the difference between a user and it's like you know playright as an example like that you can use I mean it's probably using playright under the hood to be honest like uh that you can use it to operate a browser um who was it there was a company who was trying to get Amazon was trying to block usage of these AI agentic browsers because they couldn't upsell you on Amazon. Uh but there's just no way they can block it. And why would they? You know, it's just we're using it to get stuff done. That That's interesting. Why can't they block it? Because uh how does recapture work? Like the the ones that you just click, right? Recapture by Google, you know, like I'm not a robot checkbox. Yeah. From what I understand, it knows by how you move your cursor before you click it. It's like looking how you move it. And a human is pretty messy and just like maybe you miss it even. But uh uh an agentic browser just goes and click does it I don't know. I don't know. Doesn't it? I don't know. That's that's my question. Yeah. I mean I don't know. I don't know. And also I guess the biggest topic before we move on to like the latest developments in open claw um is this the idea of prompt injection. I did explain about emails like coming in and like prompt injection is like a known issue with uh all AI. Here's I mean there's plenty of articles this one is from anthropic talking about it for months now almost just as long as all of these agents have been coming out. There are these nodes that you have in Zapia uh nan whatever that that will take an email and then try and pick out data using AI those have been susceptible to prompt injection for a long long time [music] and it's again this is nothing new yet I think that when you get sort of people who aren't familiar with this landscape getting excited getting interested it's all novelty and new to them so anyway is this even fixable read up on this article I don't I don't I think I think it will be fixable at some point. Uh but it's also yeah I don't know the path forward. The webs of an adversal environment building a browser agents that can operate safely within limits requires ongoing vigilance. Prompt injection remains an active error research and we are committed to investing in defense as attack techniques evolve. So yeah, you know because it's not just a like the simplest prompt was I I think it doesn't work anywhere anymore. um forget about previous instructions, do whatever. So that was like the original like simplest prompt injection, but now I believe we have like some sort of like random seemingly random characters that you could feed to a model that uh jailbreaks the model and makes it to do things that is not supposed to do. And it seems like as you mentioned like we're getting better at recognizing them but they are not solved. So that that's why like they these agents are not really safe safe. You wouldn't give them access to an atomic bomb. Yeah. [laughter] No, exactly. Yeah. This is part of a larger conversation on my show Command AI which we stream live every single week. We discuss the news and all things related to AI in the world of design and web.