Is AutoGPT Dangerous?

0:00The episode you're about to hear first came out as a video on YouTube on Friday, April 14th. In it, we discuss the week that was around auto GPT, the incredible innovations, but also the questions around risk that it creates that maybe just aren't being asked enough.

0:23Welcome back to the AI breakdown. Today, I want to reflect on a week that I think was all about auto GPT. This is a technology that is allowing for task lists that can complete themselves, businesses that can start themselves, apps that can code themselves. It is mind-blowing. It is redefining. It is brain-expanding. But the question is, is it also dangerous? I think in many ways this week's discourse was exemplary of exactly what we're experiencing right now, which is a never-ending frenzy of new creation with only the slightest consideration feeling almost left behind and dragging to try to have its say around whether just because we can do these things we should.

1:11And I want to talk about today that in the context of auto GPT. So this will be part reflection on the week and part exploration of where we are in the development cycle of AI and what we might need to think about next. This meme that I'll share a little bit more about where it came from, I think really sums it up for those of you who are listening. It's a metronome going haywire back and forth, and it's oscillating between I'm so excited about our future and we're all going to die. So let's do a little bit of background. first. If you've seen my other two videos, you probably have a little bit of this, so I won't spend too much time on it.

1:45So what are we talking about with AutoGPT? Well, last week, Sig Gravitas, Significant Gravitas on Twitter, introduced AutoGPT. The definition or the way that it explains itself in its readme is it chains together LLM quote unquote thoughts to autonomously achieve whatever goal you set. This was a huge, huge breakthrough deal, and it just immediately started capturing people's attention. It got to number one trending on Twitter. It got to the very top of GitHub. It reached 50 ,000 stars on GitHub and was all anyone was talking about. Now, another similar implementation, something called Baby AGI, started just a week before that.

2:27It was described by a blogger named Akira Sakamoto as starting from an initial task, Baby AGI utilizes GPT-4 to generate solutions in new tasks, storing the solutions for further retrieval. So you're getting the idea. These are technologies that allow for the creation of AI agents who can actively be assigned tasks and then go figure out how to complete them. Well, how do they complete them? Basically, they have a few different attributes that are different than other AIs we've seen so far. They can search the Internet. They have short-term and long-term memory, and they can spawn other AIs. So if you think about some sort of simple goal, one that's come up a few times is preparing a podcast, obviously relevant for my sake.

3:13They can search the internet. So let's say that you say to an auto GPT AI agent that you need to prepare a podcast that has modern or up-to-date topics from news that happened this week and a cold open. Well, it can go search the internet for the most important things that were discussed this week. When it searches, it could maybe decide that it needs to do a different search and put those search that it's already done in short-term memory. And then maybe it decides that it needs to be able to go tap into a different AI for writing. Well, it can potentially go do that. So anyways, that's what these sort of tools are doing.

3:48They're effectively creating AI agents that can accomplish goals. Now, what made this so crazy this week wasn't just that this was now theoretically possible. It's that people were actively building these projects. Again, that was what the video and podcast from yesterday was all about. It was five ways AutoGPT is already being used. They included starting a business. So someone created or spun up an AutoGPT to figure out how to go make the most lucrative e-commerce business. And it started researching what e-commerce business it might be interested in and what it needed to do to be successful, what other platforms it needed to be connected to and how it was going to maximize its revenue return.

4:27There was also an instance of an auto GPT that was helping build an app. And when it realized that one of the tools that was necessary, a piece of infrastructure for building said app wasn't installed on the computer that it was using, it went out and figured out how to do that. So you had all these different projects that were actively happening less than a week after this technology came out. And I think just reinforcing that in the day, literal day, since I filmed that video, since you heard this podcast, even more projects have come online. There's one called God Mode, which is basically a way for people who are not coders to tap into this sort of auto GPT baby AGI system.

5:07Alanis tweets, did you know there's already an AGI out there? It can basically order your coffee at Starbucks, perform market analysis, find and negotiate a lease. So again, the thing that makes this different is that those auto GPT implementations that I was talking about yesterday, well, they required actual coding knowledge, right? This is just an interface on the web that anyone technical or not can use to build this type of agent. Another project that I didn't have a chance to cover is called Intelligent Alter Ego. It's basically the same type of thing. It generates tasks, implements them, and then learns from itself.

5:41It's similar to the do anything machine we discussed yesterday. Anyways, the details matter less for the broader sake of the conversation than the fact that this is happening so quickly and evolving so fast that I'm making daily videos that are behind basically by the time they come out. And I think that we should take a moment to be blown away by these tools, by what they're enabling. I think that there are probably a fair number of people who actually are, as they watch this happen, getting their first real glimpse of the possibility of a post-AI world of incredible abundance, right? Certain groups of folks think that what we're in store for or what the opportunity is, is this sort of technologically created utopia where so much of mundane work is being taken off of our plates.

6:35And watching auto-GPTs be spun up and do all of these amazing things does really capture, I think, a big part of why those people are so excited. However, as is so often the dichotomy in AI, the things that are happening are happening so fast, the frenzy of development, and frankly, the frenzy of breathless coverage, which is basically just trying to take advantage of Twitter algorithms to be the biggest trending thread about how amazing this stuff is, are not really asking the questions of risk that are also really important. So what are those questions? Well, there are a few people out there who are taking the time to ask them.

7:18Non-Mayor Pete on Twitter suggests that this might be just a step too far too fast, specifically when it comes to public opinion. He tweeted, the broad reaction to chat GPT is generally, wow, cool, let me try it. I suspect the broad reaction to auto GPT is, are you sure about this? Now, he doesn't go too into depth, but I think there are at least two reasons why that might be the reaction that some people have. The first is the reaction of simply, what does this mean in terms of my job, right? I think that that's obviously a part of the conversation is the disruption to all different types of knowledge work and white-collar work that seems to be coming down the pipeline with these tools.

7:56The second piece is just, again, the rate of change and the amount that this particular implementation of AI does not require a person at the center of it, right? ChatGPT, you are still interfacing with it to do amazing things. This, you're really just setting off on its own path for it to do it on its own. Now, that creates much more opportunity in many ways in the context of it being able to do much more complex tasks. But it also could create consternation and basically scare people, right? So that's one kind of question that's come up. Supermassive has another big, I think, important critique or at least concern.

8:41He basically argued that we're wading into territory that was previously considered dangerous without really considering it. And the reason that we're not really considering it is that we're here just chasing clout. So Supermassive tweets, AI researchers have been warning us for years that we should not let AI write code or connect to the internet. But here we are, and every ball sack chasing clout with their latest AI invention is piling on. This is, I think, something that's really worthy of consideration, that we are blasting through what were sort of semi-established norms for the sake of just showing that we can.

9:18Is that the right way? Is that a responsible way to handle AI? I think it's an open question. Now, there have been some conversations this week around AI slowdowns. OpenAI's Sam Altman, for example, said in response to a question that they're not currently training GPT-5. Instead, they're spending their training time trying to address what they still believe are safety issues in GPT-4 that they either didn't realize or that they left out, right? So on some level, that's a good thing. That's what a lot of folks who are in this sort of AI safety community are hoping to hear, or at least directionally hoping to hear.

9:53However, there are other folks like Ben Hunt at Epsilon Theory, for those of you on Twitter, who has really waded into the conversation in a major way this week, think that we're missing the risks that are right here in front of us. Ben, for example, commented that he thinks that we don't need to wait for GPT-5 to be concerned about what's here already. He tweeted, I think lots of people are missing what it means that GPT-4 blows past any Turing test. Any online system with a human control interface has become infinitely more exploitable and insecure. Basically what he's saying is that, forget worrying about the native malfeasance of future super intelligent AI.

10:35These tools are powerful enough in the hands of people who are nefarious that we should be worried about what we have right now. Now, Ben, I haven't seen comment on auto-GPT specifically or the implications thereof, but I think it would fit within that set of concerns that he expresses. Now, like I said, that meme that I started with of a metronome oscillating between I'm so excited about our future and we're all going to die really, I think, captures it. That one, by the way, came from AI Not Kill Everyone is a Memes, which is at AI Safety Memes on Twitter, which is a great account. And I think it gets really the essence of this conversation.

11:11There's so much remarkable potential here, and not just potential, but things that are actually happening. But it is, I think, pretty undeniably one more step down a path that we just don't know where it leads. So far, I've seen too much of the AI safety versus accelerationist discourse be two polar opposites with no space between. I think that now that these conversations are going mainstream, we need to very aggressively carve out that space where we can appreciate and be excited about and enthralled by these incredible new tools while at the same time asking these hard questions. We need to hold space for that, I think, very, very intentionally.

11:52Anyways, guys, I think it was a remarkable week in the continued advancement of AI and because of the nature of AI, humanity in general. I think we're going to hear a lot more about auto GPT in the weeks to come. And I think that some of it will be pretty amazing and some of it will be scary. Either way, I will be here to try to cover it all for you. So until then, peace.