The Latest on AutoGPT and BabyAGI: Semi-Autonomous Specialized Agents (SASAs)

0:00The AI breakdown you're about to hear originally came out as a YouTube video on Thursday, April 27th. In it, we look at the latest in AutoGPT and Baby AGI, which is semi-autonomous specialized agents, a more discreet and useful implementation.

0:20One month in, AutoGPT and Baby AGI have been through a full hype cycle from peak excitement to, hey, are these things even really useful, to now people finally understanding where their utility lies, at least in the short term. And it appears it's semi-autonomous specialized agents. What's going on, guys? Welcome back to another AI Breakdown. Today, we are talking about seemingly everyone's favorite topic, auto-GPTs. Now, auto-GPTs and their related technology, Baby AGI, are about a month old. You actually have Yohei here who created Baby AGI saying, happy one-month birthday. And this is just from a couple days ago.

0:58Now, there's been a lot of ink spilled on what these are and why they're so interesting. There's a million articles like this one from TechCrunch right here. What is AutoGPT and why does it matter? So there are a few key things. One, unlike ChatGPT, although that's changing right now, AutoGPTs can search the internet. A second difference is that they have more memory so they can engage in complex tasks. And then that's really the third piece is when you assign an auto GPT a task, theoretically, what it does is it figures out all the steps that are needed to complete that task and then can even spin up other AIs that can help them complete that task.

1:36At least that's the goal. Now, this is a topic that has captured tons of attention. I did this video 13 days ago, five ways auto GPTs are already being used, and people were just so excited about these use cases. It was things like the do-everything machine and the self-executing task list. And it really was a group of people who very quickly saw the implications of this technology and raced out to see how much they could do. However, over the last week or so, there has been a shift in narrative, let's call it. And I think Rachel Woods captures it here. She says the example is auto-GPT. Yes, impressive proof of concept.

2:12No, not actually useful yet. That nuance is very difficult to communicate. And I've seen lots and lots of pieces like this. But of course, that doesn't mean that people aren't excited about it. Omar Parra, an AI product lead at Meta, says auto GPT hype is unreal. 85K plus stars on GitHub. AI autonomously completing tasks based on a... Overhyped? Maybe. More wow than useful? Yes. A peek into the future? 100%. A catalyst for innovation? No doubt. So this is roughly the state of the conversation coming into today. Well, when I signed onto Twitter this morning, I saw this thread from Nate Chan, who's been playing around a lot with auto-GPTs and baby AGI.

2:54And I think he did a great job of pointing to where the tinkerers who have been experimenting are starting to actually see where the short-term utility and usefulness of auto-GPTs might come from. Nate writes, Today, auto-GPTs are thought of as general autonomous agents, a tool that can do anything and even thought of as a precursor to AGI. When you build something that can do everything, turns out it can't do anything well or at all. But I think we're starting to notice a shift in the auto GPT space. Specialized agents are being built, and we're now seeing more useful and narrowly focused autonomous agents hit the market.

3:32P.S. This is not a criticism towards the OG auto GPT creators and projects. Now, Nate from there gives a few different examples that he's seen of these specialized agents. One, he suggests, is autonomous agents for research. Nate writes, research can't rely on AI-generated content, so this agent browses the internet for what you're researching and only extracts relevant info from trustworthy sources. He points to a project called AOMNI, which launched on April 17th and is an agent specifically designed for research. Another specialized auto-GPT he points to is one for medical research that can call medical APIs and cite sources.

4:08Another was a specialized agent for podcast preparation that could research recent events, prepare a podcast outline, lay out topics and discussion points, or even write a cold open. Another is a specialized agent to control your web browser. And Nate says there's some nuance here, writing, Though some see this as a do-everything agent, its abilities are limited to certain web actions. I think we'll move to seeing this as a specialized agent simply to automate tasks in the browser. order a pizza, book a flight, open tabs for 10 Craigslist posts of couches near me under$100, that sort of thing.

4:41Nate concludes his thread, There's a ton of greenfield opportunity right now in building for specific types of tasks and industries by using a robust chain of GPT calls optionally to fine-tune models with narrowly focused prompting, all to get one category of things done and done very well. Yohei, the creator of BabyIGI, actually responded to the thread, saying, I agree, I think something like this is what we'll see. First, specialized agents that work for one person, then specialized agents as a service, then slightly more generalized agents, then increasingly generalized agents. So this is a theory of how this might evolve from something more general like we saw people try initially, but starting from the standpoint of these highly specialized agents which seem to be coming to the fore now.

5:26Now in this thread, I asked Yohei basically what types of specialized tasks he was seeing there be a lot of value in, especially as compared to a more human-mediated chat GPT style experience. His answer was, for me, I have a couple of what I'd call semi-autonomous specialized agents that run in the background that would take me maybe 20 prompts to do in chat GPT. Having it automatically kick off from CRM activity saves me the time it would have taken to do that in chat GPT. So effectively, I'm kind of imagining a string of tasks which an auto GPT can put all together without him having to mediate each one of those tasks along the way.

6:03I followed up and said, how routinized and task-based are they versus creative and production-based? His answer was some of both. The easy stuff is pinging APIs and scraping websites and summarizing, but you can do some creative stuff. For example, if you have a podcast about AI and you see layoff news from big tech, you could ask, and then he tweeted a picture of him doing this. His prompt was, You're generating content for a blog about the future of AI. Generate a discussion specifically referencing this news in a way that is relevant to the audience. News. Dropbox lays off 500 employees, 16 % of staff.

6:41The discussion idea that came back from the AutoGPT started, The news that Dropbox has laid off 500 employees has sparked debates about the future of AI. On the one hand, AI has the potential to automate many processes, making them more efficient. On the other, the automation of certain processes could lead to job losses in the future. This news from Dropbox serves as a reminder of the potential effects of AI on the job market. Nate Chan responded to all of this and said, That is awesome. Semi-autonomous specialized agents. I think abbreviating to Sasa has legs. I've got my Sasa's running. Give me five minutes.

7:13So as you can see, the cool thing here is that people are taking this technology, which again, was so hyped and so exciting when it first came out a month ago, and then maybe had the patina wear off it just a little bit. But now people are figuring out how to hone in on where its actual utility lies, and they keep building these really interesting tools. I mentioned Aomni before, which is an agent specifically designed for research, and I took it for a test spin earlier today. I did a couple of tests. The first, I asked it to find the most relevant and discuss news stories in AI. It brought back a bunch of interesting things, certainly not about today, but you could see how this would be a really helpful, useful tool.

7:50And then I also asked it to write a crypto podcast outline. For that task, it lays out its subtasks, including search, find popular topics in the crypto world to include in the podcast. Browse, extract information from the top search results. Browse, extract information from the second search result. You see where this is going. Process, combine the information from popular topics and recent news. Process, create a podcast outline based on the combined information. Now, it's a very expansive suggested podcast that it came back with, but of course, I didn't define it strictly at all. I could definitely see integrating something like Aomni into my research flows.

8:25Another tool that a lot of people have been excited about is this new implementation of AutoGPT for mobile. It's effectively a Telegram bot that allows you to use your own version of AutoGPT right inside Telegram. Anais Calio put it together on GitHub, and of course, I had to try it out. So here we are in Telegram, and I ask, I'd like some help with some research. NLW agent says, sure, what topic are you researching? I say, AI. NLW agent says, great, artificial intelligence is a vast field. Could you please narrow it down? Then I said, yep, I'm doing a podcast about auto-GPT. NLW agent started to write about the basic steps for implementing auto-GPT.

9:02Then I tried to define the task a little bit closer. I said I'd like to find three potential topics for a podcast on AI. It gave me AI and the future of work, AI and healthcare, AI and society. Now you'll notice this is much more like a chat GPT style experience in Telegram versus an auto GPT type experience. And so eventually I asked, are you an auto GPT? And NLW agent said, no, I'm not an auto GPT. I'm an AI language model that uses machine learning and natural language processing algorithms to generate responses to user input. I'm more similar to chat GPT than auto GPT. Now, I just started to try this right before I filmed the video.

9:35I'm sure there's a lot more to do. And frankly, having ChatGPT sitting in my Telegram, that's my own little NLW agent, is still very cool. So I think to sum up, what's cool to see is that people are starting to hone in on the use cases that make AutoGPTs actually useful for them, that actually fit into their workflows. And this is kind of what you'd expect. Things start like a toy and with people understanding the broadest possible implications, and then they get narrowed to where they can actually drive and provide value, whatever those use cases might be. However, to the extent that they are competing in some ways with a more human-mediated experience of just using an LLM like ChatGPT, they now even have some extra competition.

10:14As I was having these conversations earlier this morning, I started to notice that people were talking about the fact that ChatGPT default mode now includes browsing mode enabled for many plus users. It's still in alpha and hasn't been rolled out to everyone, but this is ChatGPT allowing you to also access the internet. Now we're ignoring all of the AI safety implications of all of these AI tools being hooked right into the entire data of the internet. But if we do hold that aside, it's an extraordinarily exciting time to see how these agents can change our workflows, make us more efficient, and allow us to do more of whatever it is we want to do.

10:49I like this direction for AutoGPT of these semi-autonomous specialized agents, and it's certainly where I'm going to be experimenting in the near future. All right, guys, that's it for another AI Breakdown. Please subscribe to this channel if you haven't yet. Go check out the podcast. And until next time, peace.