ChatGPT Atlas: The "Side Quest" AI

It's not OpenAI/Atlas. It's chatgptatlas. So, it's a it's a chat GPT product which is an interesting choice. Yeah, it is interesting. When they released this, it kind of got me like back in the room of like so I've been using Comet for months as you saw just then. Centering a div ch doing very very simple things. It takes ages and this is the problem they had in the actual presentation. It takes vastly longer for the AI to do things than for you to do them quite frankly. But what this what this is, it's a side quest tool. So what I did, and I actually released a video on this on LinkedIn, I had my accountancy software on one tab with all the unresolved transactions. I had my emails on another tab and I told my agent to go through each of those items and give me a list of all the emails what relates to that transaction. This is a task I do every Friday and it takes ages. I can now Friday mornings I can just set that off in the background while I toddle off and do extra work and then a couple of hours later I could I mean it's not going to take a couple of hours but when I've got the time to check back it would have done most of the work for me because a lot of the a lot of the PDS and transactions are in Gmail right so that all being said this is a long horizon task background task feature it shouldn't be front and center of your workflow at all you literally staring at that color box waiting for it to change it to green. It's like just do it just do it yourself. Yeah, exactly. I'm just having fun with it. And I know these tasks will take me maximum like 20 seconds to do. Uh I just want to see how it does it. So it took it two minutes to do it. But I'm actually surprised that it did it well this time. Like adding a color variable. These are semi- new in Web Flow. So it knew what button to click. Obviously the it looks at the HTML and but it knows what button to click. Do you have Web Flow AI? No. Exactly. Okay. Because when that comes out and you have Web Flow AI panel open. Yeah. What the AI will probably do is ask the AI in Web Flow to make the box dream. That would be so fun because it's like why should we spend tokens while the other AI can spend the tokens exactly for me that that's that is going so well well is this not like the optimum idea of sub agents like there is a agent other agents hey these are smaller tasks you can do them just go yeah that's this is exactly what a sub agent is it's someone who's or a skill you as someone who's specialist in that field, it's like look, you you go and do it. And that's actually the distinct difference between skills and sub agents. Um sub agents will work at something and iterate on whatever. It's like it's it's an agent, right? Whereas a skill is like it's it's a reference. It's it's not going to continually. It's still going to be the original agent that's working on it just referencing the instructions from another agent. So you know um so the top agent will be like sometime in the future maybe uh the sorry what I'm the top agent you're I I want to say like you are the top dog so the the higher level agent than this would be on a brow on now this is on a browser level but maybe we'll have it on the Mac and on the operating system level and then we'll have us, you know, just talking to our operating system and saying like go do these things and these like agents having sub agents and them having other sub aents. That's at least the idea. But I think we are pretty far from this uh future. I I I believe these AI tools because of their current architecture, it's not that easy to scale them to that level. Well, saying all that, I mean, like I think the the the winner here will be a potentially might even be Google, right? In that they've released the agent API which is a ondevice agent or LLM that can run without thing is this is constantly sending images back and forth to the server, right? And we're there's a lot of latency involved there. So really this is a task that should be done natively on your local machine. This needs it's bizarre that we would need to send anything like this off really. Do you know like these the idea of remote LLMs is I think will be an antiquated subject because it will be like well this is so easy it's so simple now LLM are so fast or computers are so fast now you might as well just do most of it you know really you'd want to go maybe it's tool use you could have it go off into the internet and find up to date data but it could be that you send stuff off and get up to be able to get up to date data but anyway I think that yeah this is sort of should be should be done locally But we said at the beginning of the the this the um the show or I said like it's having me rethink what I do in the browser versus what I do on my computer because I had notion calendar and notion uh mail maybe no just notion just notion notion calendar I had and I'm like I'm missing out on a lot of opportunities if I'm having a local if I'm having an app running, why not put it into the browser? And so I've quit notion calendar and I've now got Google calendar opened up in a tab because if I'm able to then the more I'm able to move now into my browser, the more I'm able to leverage the agentic functionality of comet. And so I'm just slowly rethinking again what apps do I actually need? What benefits from being an app or what benefits from being another browser tab? And so I'm in this kind of world now and that's where I said to you like what determines what do you determine as whether you should be using something in the browser versus what what should you be using as an actual app? Uh yeah. Do you have like other use cases for this? I would, you know, I would I would open up the CMS thing and then if you're putting data into the CMS, if we're in the context of workflow, yeah, you know, um but then you potentially just use MCP for that anyway. But if if you're living in the world of browser, if you have a one article open, you could thing is you can transform data as well. So it could be that you're reading um you know you're on a on a blog post and you want to break that blog post up into individual blog posts as an example. So divide you know what I mean? So you can transform and reinterpret the data you're receiving. I think again it's just being a bit uh uh progressive with thinking and and I'm sure you'll find some some use cases for it. I bet it would be great for this podcast and um we could probably open up notion in one of the tabs and then say just make some notes. That's a good idea. Instead of reading the notes and summarizing the page in the chat, just saying can you make notes on this article? Put it in put it in the notion page. No, whilst you're off doing this, that and the other, you know, because again, it's going to be slow. Like it's just it's not going to be it's it's not your primary focus. It's your side quest. Yeah. While you're scrolling Tik Tok. while you're scrolling Tik Tok, you know, while you're out touching some grass. 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.