5 Ways AutoGPT is Already Being Used

0:00The show you're about to hear was originally released as a video on Thursday, April 13th. In it, we dive deeper into AutoGPT, including looking at five use cases that are actually being implemented right now. Those use cases range from starting a business to figuring out how to build an app to doing a task list that does the tasks itself.

0:28What's going on, guys? Welcome back to the AI breakdown. down. I'm actually on the road right now. As you can probably tell, it's a different background than normal, but there's so much going on that I wanted to do a show today anyways. So what we're going to do today is go a little bit deeper on auto GPT. This is a theme we started to talk about in yesterday's video, but I want to go, as I said, a little bit deeper and talk about in more depth some of the emergent use cases that we're already starting to see. So for those of you who are just catching up now, AutoGPT, what makes it different from the other GPT you may have heard of, ChatGPT?

1:03Well, one, it can conduct internet searches. Two, it has short-term and long-term memory management. And three, it can generate other AI agents to accomplish tasks. So basically, AutoGPT produces AI agents capable of figuring out how to solve particular problems. A couple implementations of this arose about a week ago. That's how fresh this is. Baby AGI is An open-source AI platform inspired by various cognitive development that aims to train and evaluate various AI agents in a simulated environment. The platform focuses on reinforcement learning, language learning, and cognitive development, allowing AI agents to learn and perform complex tasks.

1:42Baby AGI operates in an infinite loop, constantly pulling tasks from a task list, executing them, enriching the results, and creating new tasks based on the objective and the results of the previous task. The workflow consists of four main steps, tax execution, result enrichment, task creation, and task prioritization. What you're seeing here, even in just this description, is that we're talking about actual AI agents who figure out how to solve the problems that you assign to them. This is different, obviously, than sort of the human-operated chat GPT, ask it a question, get an answer, and maybe you can weave together different answers to produce something, but this is really doing that weaving together all by itself.

2:22Now, another implementation of something very similar is AutoGPT. That's obviously the one that's lending its name to all of this, given that it was the one trending on Twitter and seems to be the term that people are latching onto. AutoGPT from Sig Gravitas, who was the developer who first put the repo out, AutoGPT is an experimental open source application showcasing the capabilities of the GPT-4 language model. The program driven by GPT-4 chains together LLM thoughts, quote unquote, to autonomously achieve whatever goal you set. It pushes the boundaries of what's possible with AI. Obviously, the key part here is this idea of anonymously achieving whatever goal you set.

3:01So that's the background. Like I said, these things are very, very young, but that hasn't stopped developers from diving right in to try to explore them. So let's look at a few of the use cases that people are excited about. And we'll start with the starting a business use case. So this comes from Graham Fleming. He writes, I have an auto GPT from Sig Gravitas currently developing an e-commerce business. It has decided to browse the internet for business ideas, saving its findings to files to reference later on. So I'm going to pause this little GIF video right here. Let's go right there. Try to pause it.

3:37There we go. So e-commerce GPT thoughts. As an e-commerce AI, one of our primary goals is to increase our net worth. In order to do this effectively, I need to leverage my ability to browse websites, read and write files, and search the internet for information. I should also consider delegating certain tasks to GPI agents to further optimize my time and resources. Reasoning. My goal is to increase net worth, which requires me to make smart decisions and act efficiently. I need to leverage different resources available to me to achieve this. It then shows its plan. Browse websites for e-commerce and business ideas that it can increase net worth.

4:13So basically this agent is going out and figuring out what business idea to pursue. Save important findings to files for later reference. Utilize GPI agents for delegated tasks. Continuously review and analyze my actions to identify areas of common and refine my approach. Evaluate and improve efficiency of current processes. Now, of course, this is basically exactly what a human who had decided to go start a business to make money would do. They would browse websites to look for ideas. They would save their findings to files to reference later. They would look into what resources were available to them to delegate certain parts of the tasks.

4:51Very, very human-like, which I think is why people are so captivated by this new implementation of the technology. So that's one example of AutoGPT. Next up, building apps and fixing coding problems specifically. So this one comes from Varun Maya. He writes, AutoGPT was trying to create an app for me, recognized I don't have Node, Googled how to install Node, found a Stack Overflow article with Link, downloaded it, extracted it, and then spawned the server for me. My contribution, I watched. So there's actually a couple things going on here. First, AutoGPT has been assigned to create an app. But the person who assigned it as such or assigned it to do this didn't have the proper infrastructure, right?

5:33They didn't have Node. So what did AutoGPT, excuse me, do? Well, it figured out how to install Node. It found an article that had the link. It downloaded Node, extracted it, and then spawned the server as it needed. So it was not going to let, in other words, a problem such as not having Node stop it from completing its task. It figured out how to solve the intermediate problem to get to the long-term objective. Next up on our list of implementations, and again, this is less than a week old, guys, the do-anything machine. a to-do list that does itself. This one is capturing a lot of people's attention, I think, for kind of obvious reasons.

6:13So Garrett Scott writes, over the weekend, I finished the to-do list that does itself. Every time you add a task, a GPT-4 agent is spawned to complete it. It already has the context it needs on you and your company and has access to your apps. It's called the do-anything machine. So again, let's look at the examples that they put here. Task one, find the best person at Walmart for us to sell to, add them to our Notion CRM, and send an outreach email. Number two, create a memo about our first in product in Notion and email it to Ryan Cooter and ask him if he has anyone in mind who would be a good fit to run it.

6:48Number three, make a web app that sends mass email and Slack Chris a tutorial on how to host it and launch it. The point of this is that this is the same exact type of to-do list that you might write, right? These are things, These are tasks that the person or the company involved wants accomplished. But instead of actually doing them, you are effectively, your to-do list is the prompt for the AI agent to go figure out how to do it itself. Hugely powerful. In fact, Garrett says that they've had to turn off signups. Current users will be enough to find bugs and improve UX. Then we'll roll out next weekish.

7:24And then later, by popular demand, they added a wait list. So that is the do-everything machine. Next up, we have the content creator. This is one that I mentioned in yesterday's video as well, but I'll go a little bit more in depth today. So this idea is to use a GPT agent, an auto GPT, to change the way that content gets created for the internet. So in this case, the author, JB here, is asking the GPT agent to read about recent events and prepare a podcast outline. He uses All In, the podcast with Chamath Palapatiya, Jason Calacanis, and David Sachs, and among others. And it did five searches and 15 web browsers.

8:07It found five topics podcasts on recent news. It made accurate references, and it wrote a cold open. So this is sort of the whole process of building a podcast, right? Interestingly, what it outputted was a current task, a draft output, suggestion for next tasks, excuse me, a task list, and reasoning. This is a potentially game-changing approach to creating content. Obviously, one of the first use cases that people have flocked to for LLMs, and it seems likely to be an early use case for this type of autonomous agent as well. And fifth and finally, let's talk about AgentGPT, which is effectively a way to more easily spin up an AutoGPT for yourself.

8:55Awesome here writes, introducing AgentGPT, an attempt at AutoGPT directly in the browser. Now, what's worth noting here is that AutoGPT is basically just used by developers so far because you have to install a bunch of different tools and kits to actually make it work. You have to have access to a few different APIs. Agent GPT attempts to short circuit that right so you can see here in the little gif explanation or example rather demo that the name of the of the agent GPT that they're working on is hustle GPT its goal is to create a new startup with only$100 of funding and then it goes off and does the task so this is a little bit different than a single use case this is an attempt to improve the infrastructure for lots of different people to go figure out their own use cases.

9:41This is the type of thing that is happening left and right with these new tools where they come out and someone figures out what's needed for them to be implemented in a way that is much more user-friendly and potentially less technical, and that comes out next. Really, really fascinating stuff going on here in this space. Now, a couple more that are theoretical, right? They haven't come to the fore yet, but are someone imagining what these types of AutoGPT could do. Greg Eisenberg writes about three different examples. One, he says a customer service rep, AutoGPT could understand customer inquiries, provide support, and even suggest up sales.

10:21It would mean an assistant that was available 24-7, speaks in every language, etc., etc. Number two, social media manager can be used to manage social media accounts for business based on goals of retweets, likes, and even sales. Number three, financial advisor, AutoGPT could make it a breeze to invest your money, saying it can analyze financial data and provide recommendations on how to stay ahead of the curve. Now, these are obviously a little bit kind of farther out examples. I think a lot of the first tests that we're seeing here are self-contained, but it shows just how fast people are thinking about what these sort of agents could do.

10:57Now, one really interesting last piece of this story for today is folks are already comparing these different implementations. So Lauren Marie here writes, today we tried AutoGPT, AgentGPT, and BabyGPI. Thoughts and insights below. They think that AutoGPT is the best overall. It's most difficult to set up, like we were just talking about. It needs basic programming language. There isn't really a user interface. But it seems closest, they say, to accomplishing running autonomous agents for complex tasks. She says that it communicates with itself. She says that it runs analysis on its own errors to diagnose issues, which is super cool.

11:35It critiques its own strategies and analyzes potential challenges and pitfalls. Seems well-equipped with logic to execute on the goals you give it. So basically, there is a higher burden and a higher barrier, but once you get through that, it does a better job of actually figuring out how to accomplish the task that you've set for it. Now, they say, Lauren Marie here says that the next best at completing tasks is baby AGI. It gives a detailed task list to reach goals. It directs you to where we need to go, but she says that they haven't been able to get it to execute, but perhaps that's on their end.

12:08They're maybe not prompting it correctly. Lauren says that it seems to have sound insights, but it is this implementation problem, how to get it to actually act on the task list it creates. The third, Agent GPT, perhaps not unexpectedly given what it's trying to do based on what we were just discussing. It has the best user interface. It's the best overall user experience, but there are some challenges. Lauren says it claims it's executing tasks, but nothing happens. And it says that it can scrape accounts reasonably well, but nothing happened after. The task that they used to model this was to grow Twitter and gave it this profile.

12:45Didn't read it correctly, saying it thinks that Lauren here is a beauty influencer. So it didn't scrape the right profiles, but it did scrape some. Overall, I think that it's clear that these implementations of this type of autonomous GPT type technology, these autonomous AI agents are still very nascent, right? These aren't necessarily fully production ready. You're not going to turn these into sort of business tools in a full form quite yet. But we're talking about a technology that's less than a week old in many cases. It has absolutely captured everyone's attention though. And I think that these early examples, the content creator, the do-anything machine, this fixing coding problem app builder, and this business starter are clear examples of where this might be going.

13:30Anyways, guys, I hope that gives you even more of an insight into what's being built with AutoGPT and gets your brains going on what you might want to do with it as well. Thanks for watching, and until next time, peace.

13:50Thank you.