The AI Daily Brief: Artificial Intelligence News and Analysis - What is AutoGPT and Why is Everyone Talking About It?

Episode Date: April 12, 2023

Less than a week old, AutoGPT has exploded onto the scene. AutoGPTs are AIs that, once assigned a task, figure out how to complete that task, including generating other AIs. They have access to the internet, longer term memory, and thus can be used in much different ways. In just a week, AutoGPTs have become the most active GitHub repositories.

The show you're about to hear originally came out as a YouTube video on Wednesday, April 12th. The discussion is all about the rise of Auto-GPT, what it is, where it came from, why people are so excited and what use cases they're already exploring. Welcome back to the AI breakdown. Today we are talking about Auto-GPT. If you have been anywhere on Twitter or around AI conversation circles, Auto-GPT has been all that anyone is talking about. So what we're going to do on today's video is talk about, what auto-GPT is, where it came from, why people are paying attention, and what they're already starting to build with it. And let's start at this high level with a thread from Sully. So Sully writes, by 2025, autonomous AI agents will be in every aspect of life. They've been out for a week and already the possibilities are mind-blowing. But what are they? How do we just figure them out

and what's next for us? So what Sully explains an AI agent as is something that is an AI system that when given a task, it runs in a loop until the task is solved. Another way to say it is that AI gets assigned a goal. It figures out what it needs to do to accomplish that goal and then spawns more AI to do it. Now, this is a concept that isn't new necessarily, but it's only in the last couple months that it's really come to the fore as something that's viable. So one version of this or one early explanatory version of this was baby AGI. And then another one, which is the one that's really caught people's attention is AutoGPT. Now, just about a week ago, a user named Sig Gravitas released a repo on GitHub called Auto GPD that demonstrates a self-running AI agent that could

write its own code and heal itself if it had any errors. This has blown open the doors of this development field and tons of people. Basically, everyone in AI is talking about it. So in terms of understanding what makes AutoGPT different than ChatGPT, just as a for example. Linus here describes it as AutoGPT having Internet access. It has long and short term memory, which is different, and it operates without human input. In a nutshell, he writes, Auto GPT allows you to set a goal and then get it done. Ethan here explains a couple other ways in which it's different from ChatGPT. AutoGPT is connected to Google via the search API, which means it can go conduct its own searches. That is, again, different than chat GPD, which is trained

on websites and other linguistic sources up to a certain period of time. Auto GPD is also connected to Pine Cone, which gives it long-term memory of a sort. It's able to read websites and break information into chunks. It can create files. It is self-directed once it is given a goal or a task. Nathan also explains the main features here saying that Auto-GPTs, one, assign tasks and goals to be worked on automatically until completed. Two, they chain together multiple GPT fours to collaborate on the tasks. Three, they have internet access and the ability to read and write files. And four, they have memory to know what has been done before.

So this is a whole different level of capacity than something like chat GPT does. ChatGPT is a human mediated and moderated experience where we are using AI to accomplish something. This is really a version of us asking AI to go figure out a problem how to solve it and then to just let it go on its own, right? It is something that a ton of people are building on right now. And so let's look at some of the examples that people have spun up in just, again, the last week or so. Omar Perra shows that one of the examples is an AI agent that autonomously does sales prospecting on its own. This one is powered by baby AGI. Linus points to another example of doing product research. So this is one from Chilzilla,

where he runs an AI agent that conducts product research and writes a summary on the best headphones. Really basic use cases so far, but still pretty remarkable. J.B. here writes another use case that is reading about recent events and preparing a podcast outline. I don't know if this is something that I should be scared of or diving into, right, as a way to augment my work. But J.B. gives the all-in podcast example. That's, of course, the podcast with Jason Kalakanis and Chamath Palahapitia and David Sacks. So his example is with five searches and 15 web browsers, auto-GPT research agent prepares a five-topic podcast on recent news with accurate references and a cold open. And again, the point here isn't that a person sat down to chat GPT to go figure out what

those news searches were or what those news topics were. And they didn't sit down at chat GPT to write this intro. This was auto GPT being given the prompt to read about recent events and prepare a podcast outline and just figuring out how to do it on its own. Wildly, wildly different. Sully here again, we're back to Sully writes, still not convinced. He pretended to be a fake shoe company and gave it a simple objective to do market research for waterproofing shoes, get the top five competitors, and give me a report on their pros and cons. Now, he basically says that this worked incredibly well. He says, first, it went straight to Google to find the top five waterproof shoe reviews,

and once it found leaks, once it found links, it created questions for itself, such as what are the pros and cons of each shoe, what are the pros and cons of each top five waterproof shoe, top five waterproof shoes for men. These, again, were not created by Sully. They were generated by the AI itself in the, the context of the larger task. Sully goes on and saying it continued to analyze the various sites with a combination of Googling, Googling, updating its queries, until it was happy with the results. Here's an example, he says, of when it thought critically. It knew that some reviews

could be based biased to fake, so it had to validate the reviewer. Again, Sully didn't program this AI to say or to determine what type of review was fake or biased. The AI figured out that that was important on its own. Sully says it even spawned its own sub-agent to carry out a task of analyzing the websites. Now, there were times that it got stuck because there was no text file, and then it had to figure out how to fix the issue, which it did all by itself. The result, Sully says, a pretty detailed report of the top five waterproof shoe companies with their pros and cons and a nice conclusion summarizing the report. Oh, and it only took eight minutes at a cost of 10 cents. He says this is a pretty basic example too, entirely unoptimized. I don't think that it takes too much

imagination to understand how if this is the type of thing that people can do with this technology after literally one week of it being available, that it could blow the doors off of a huge number of applications. Over in crypto, of course, we already have people talking about how there could be disruptive implications for automating trades as a part of it. But overall, regardless of the specific use cases, it's very clear that this particular subset of AI has really captured developer attention. Siki Chen writes that the top three trending repos on GitHub are all self-promoting primitive AGI projects in this frame of reference, right? It's baby AGI, it's auto-GPT, and Jarvis by Microsoft. These are just the beginning.

of this set of types of technologies. And in fact, they are, the underlying technologies are evolving themselves. It's not just that people are using auto-G-T and baby AGI. They're building on top of them. They're forking them and building on top of them. This is DSNR saying about how their co-founder recently built teenage AGI inspired by baby AGI, which has different attributes and properties, including recalling infinite memory, thinking before it speaks, and not losing memory after.

being shut down. Now, there will be a lot more discussions around not only the use case side of this, but to what extent it pushes us farther down the path to AGI. It's no surprise that this is happening at the exact same time as governments are trying to quickly catch up and figure out if they need to have some sort of regulatory body or licensing body for AI, but it also shows just how difficult that process is going to be given how fast things are moving. So anyways, that is a brief overview of auto GPT, what it's doing, where it came from, how people are using it so far. I'm sure that within a couple days we're going to need to do an update to see how people are building on this. But for now, that is the AI breakdown. Until next time, peace.