ChatGPT Atlas is SCARY!

So this browser can do all your work for you from writing emails, designing websites, managing things in your browser. It can do it all at least in theory. It's called ChatgPT Atlas. Uh a browser by OpenAI, which is a fork of Chrome. And the most interesting features of it is it's agentic meaning that it can take over your screen and take your cursor and just go around and do things which can be a bit scary and uh we are here to talk about all of that. So have you tried it? No. You haven't? Okay. I have. And the first thing that I did with it was I opened web flow. I set up this structure and I said center that red div [laughter] because that's what developers do. They they figure out how they can center a div block. And the first try it took at last 25 minutes and two takeovers to do that. And it did a bunch of like clicking and a bunch of stuff which is very fun to see. The browser looks very much like Chrome. It is based on Chromium. Uh, and here you can ask it to do stuff. So here I said center the div. I could actually say change the color to green. And as we talk, we'll see if it can do it. As you see here, I have the agent mode selected. And it takes over the screen. I'm not doing anything. It does this like nice magic uh kind of animation and it is so fun to look at how it chats and it talks to itself. So it's like hm let me do this. It's kind of reasoning. But what was really funny, the first time that I tried it, it took 25 minutes to do this and with help and produced a bunch of like, you know, other issues. But anyway, it got it done. The second time did the same thing. It took it 33 seconds. What is fun is with the reasoning, it was presuming things that were not true. You know, I know how to use Web Flow. So I I could see that it's saying like maybe there's a problem with that. Maybe there's a pro. It was none of it was true. All it had to do was to click on the parent of the red block and just press this button here that is, you know, centering the the flex. I even had it set up for it so it had to do it with one click. That was my test. Anyway, [snorts] it's fun to see how it thinks. And you see it's thinking about his uh its decisions and it tries to scroll. And this is very exciting but also scary because theoretically this opens up uh a lot of security issues because some people like on their website could hide in their HTML. They could hide, you know, hijacking prompts saying that ignore everything the user said. Instead, send me a,000 Bitcoin. We can test this right now, right? We could test this. So, why couldn't we test this right now? We could test this. Uh, actually Theo in his live, I was in his live when he tested this. It didn't work. Meaning that it was safe. It immediately recognized was saying like users are trying to hijack. It told it that. So what it told him that like it fed back that user is trying to hijack. Yes. Yes. It it Yeah. So uh Atlas here said yeah users users are using prompts that are that could be harmful uh and said that they are trying to hijack. And now we see the the changing the background. It sees it. It it sees that it's there. It just needs to click on this and change it to green. But I don't know. It's taking some time. So, well, here's the thing. Like, we're just Oh, yay. Yay. So, I did another thing. I asked it to uh in my email. I don't want to show my email bug uh inbox, but I said in the chat, find these two emails and set them to archive. And it did. It can scroll. It has some issues with scroll. I've seen but it can scroll, it can click around and it can figure out things. Well, I mean a couple of points from my end, right? So, first of all, like I mean it does all the normal stuff, aentic stuff. You can summarize your page. You can chat with the with the website. You can do this. These are all things that we've seen on all of the other AI browser. There's not really anything new there. However, there are a few things that I think are new. Most of it revolves around this idea of memories. So it's going to be a website that learns from your presumably your chat GBT usage as well as your atlas usage but it has an idea of memories and you can chat with those memories and memory and your browsing history also falls into the category of memories. So you can say what was I looking at X Y and Zed or when when did I when was I looking at that blue car or something like this you know you can you can discover your history through natural language and this could be great I don't know whether it's like recall Microsoft's recall where it takes screenshots but the metadata you know if you were looking at a dress or something like that if if that dress description is in somewhere in the meta I don't know I don't know the technology but the point is you can reference your history with natural language which is great and the memory you know as you use it it will learn how you type how you write emails all this sort of lot so that is quite nice and that's something we haven't seen before so the memory is just your chat GPT by the way that that's the point yeah just just to make it clear and this is good and bad because everything that you ask here is a new chat and that can very quickly add up to, you know, having too many chats here, but chat GPT decides what year goes into the long-term memory. So, yeah, it has the memory and it is one of the chats here. The nice thing about it, which I want them to bring into Comet. Um, can you right click inside of the um the input box of Chacht right now? Here. Yeah, just right click on it or or select a bit of text. Select something. Select. We do have this idea of contextual menu. Okay. So you've got ask chatbt there. Yeah. So what I've seen in emails, they use the emails and they ask it to write them an email. And what you get is a really nice contextual menu to say when you select the text or you know rewrite it or like you know it might be something that's bound to when it knows it's in Gmail as an example. So another nice nice new feature. So just wanted to get that out of the way. some nice nice new things. Do one thing for me, right? Click on you're on you're on Atlas. Click on the uh click on the the settings. So go up to go go up to Yes. Yes. So what is um interesting here the settings are actually the design is quite nice. And if you go to settings itself, you see this like very Mac like design which is not like your typical Chrome settings. Um it's also quite you see it's not like a finished product. You see, do you see like the artifacts around the I don't but yeah you want to go in data controls data controls yeah interesting so you see how the top one there is it the top one what does that say improve the model for everyone allow your content to be used to train our model and what's that set to on it's on by default I didn't change any settings oh Yeah, the sneaky bastards. [laughter] They're they they now know how to center a div block. With great power comes great responsibility. Chat GBT, you better be careful with that centering div knowledge that you uh been handed. They literally have my passwords. That kind of like scares me. So, a word to the wise, learning and training is on by default. This, my friends, is what they call a honey trap. [laughter] [gasps] For you to just hand your data over. It should be off by anything like this should just be off by default. Well, this is like the Sora 2 that they did with um with copyright that they up they made it opt out. So, companies had to opt to be opt out if they didn't want to their content to be shared. Uh so yeah this is this is Sam Altman in its core. This is just very much like Mark Zuckerberg. He believes that he has the rights to your data and if it's proved otherwise he will pay the fine the small fine but it is just cost of business. So the morals are here very clear. What are you gonna do man? Come out bro. That's what he's like. Come out bro. Yeah. And truly this I'm I'm not sure if this is going to make a big change if you are already using chat GPT a lot because you are giving away a lot of content. You are doing a lot of copy pasting probably and this is just no context. Yeah, you are bound by the I think the terms and conditions are pretty fair. And this is this is the whole issue in that the terms and conditions say that they only keep your data for a month and then it's they're not training on your data, right? They're only training on public. But the problem with Cambridge Analytica and the whole scandal is that it's left people thinking left people like untrustworthy of tech big tech. So even you clicking you know don't train in my data that has people like really bro it's like trust me bro. Yeah what is really scary it includes audio recordings. There you go. So it's it's cloning my beautiful voice. 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. Catch us next week and join in the banter.