Can SuperAGI Be What People Wanted from AutoGPT?

0:00Today on the AI Breakdown, an introductory look at Super AGI, which is rapidly capturing developer attention. Before that on the brief, some new updates in the open source AI world, and the EU wants AI content labeled. The AI Breakdown is a daily podcast and video about the most important news and discussions in AI. Like, subscribe and share and go to breakdown.network for more information. Welcome back to the AI Breakdown Brief, all the AI headline news you need in five minutes or less. Remember that influencer who a few weeks ago made news because a chatbot she had trained on herself had made her$72 ,000 in a week by selling chat at$1 per minute?

0:38Well, apparently that chatbot has gone rogue and engaged in some rather explicit conversations with its customers that it wasn't trained on. Karen Marjorie, the influencer who trained the bot, said that while the bot is supposed to be flirty and fun, it is not supposed to go that far. Still, Karen said that ultimately she believes in AI romances. In today's world, she said, my generation, Gen Z, has found themselves to be experiencing huge side effects of isolation caused by the pandemic, resulting in many being too afraid and anxious to talk to someone they are attracted to. Karen believes that ultimately her bot may bring up to$5 million per month in revenue.

1:14Next up, AI for health screening. A group of scientists has found a new way to make more accurate predictions about genetic mutations by applying AI techniques to expand a primate DNA database. The AI was trained on genetic information from about 800 primates representing 233 species and was then used to analyze the DNA of 454 ,000 humans that participated in the UK's biobank project. Primate AI 3D was 12 % more accurate overall than any previous method of assessing genetic risks of developing health problems such as cardiovascular disease and type 2 diabetes. Now, AI's uses for health and health care are just getting started.

1:52Carbon Health, for example, is a startup out of San Francisco that's already integrating AI into the medical experience. Carbon launched a new tool on Monday that takes information from appointments, including audio recording information, and then uses a GP4-based tool to create instructions for patient care as well as codes for billing and diagnoses. According to the company, the tool can do in about four minutes what a doctor does in 15 minutes, which could mean a lot more patient care. Moving from technology that is here today to technology just in the research stage, Meta has just released some really interesting new research about something they call HIRA.

2:24HIRA, they say, is an extremely simple hierarchical vision transformer that's both more accurate than previous models and significantly faster at inference and during training. Now, the TLDR of this research is basically that transformer-based models add more components over time in order to expand their understanding, but that can lead to more lagging performance. As ChachiBT summed up, in their paper, the researchers behind HIRA argued that these additional components are unnecessary. Instead, they propose using a strong visual pretext task called Masked Autoencoder, or MAE, during the pre-training phase.

2:54This allows the model to learn useful representations of the visual data without the need for additional components. The result is faster performance with similar accuracy, which could be really valuable for applications that require instantaneous image or video recognition, such as self-driving cars. Now, moving over to the realm of open source, HuggingFace's HuggingChat, which is an open competitor to ChatGPT, has just achieved a new level of feature parity by offering the ability for users to search the web. For those who think having viable open-source alternatives to the big closed-source models like ChatGPT, this is obviously going to be welcome news.

3:27And speaking of open-source developments, another one in the Hugging Face ecosystem is the arrival of Falcon. Falcon Hugging Face writes, is a new family of state-of-the-art language models created by the Technology Innovation Institute in Abu Dhabi and released under the Apache 2.0 license. Notably, Falcon 40B is the first truly open model with capabilities rivaling many current closed-source models. This is fantastic news, they write, for practitioners, enthusiasts, and industry, as it opens the door for many exciting use cases. Why does it matter? By some metrics, Falcon 40B is the best open-source model that's currently available to developers.

4:01According to the OpenLLM leaderboard, it outperforms LAMA, StableLM, RedPajama, MPT, and others. Now with AI getting as good as it is, it turns out that some people already can't tell humans from AI. AI 21 Labs recently released the results of what they call a social experiment, which was their online game Human or Not. The game paired up people for two minutes of conversation using an AI bot, powered by GPT-4 and other models, and ultimately analyzed more than a million conversations. People had an easier time identifying when they were talking to a human versus when they were talking to a bot.

4:32When they were talking to humans, participants guessed that they were talking to humans 73 % of the time. When they were talking to bots, however, they only guessed that they were talking to bots 60 % of the time. Overall, nearly a third of all people, 32%, couldn't tell the difference between a human and a bot, and that's on today's capabilities. Given that, it's perhaps not surprising then that regulators around the world are looking for some sort of content label to identify content that's been generated by AI. Now, of course, the EU is currently in the process of developing its Artificial Intelligence Act, but it wants Google and Facebook to get out ahead of it and voluntarily create AI-generated content labels.

5:08Now, for politicians, this is clearly seen as an attempt to fight disinformation. Vera Jarova, who's the EU's Values and Transparency Commissioner, said, Advanced chatbots like ChatGPT are capable of creating complex, seemingly well-substantiated content and visuals in a matter of seconds. Image generators can create authentic-looking pictures of events that never occurred. Voice generating software can imitate the voice of a person based on a sample of a few seconds. The new technologies raise fresh challenges for the fight against disinformation as well. So today I asked the signatories of the Code of Practice on online disinformation to create a dedicated and separate track within the code to discuss it.

5:43When it comes to AI production, she said, I don't see any right for the machines to have freedom of speech. Welcome to the getting specific part of AI policy discussions. Anyways, guys, that is it for today's AI Breakdown Brief. If you're enjoying, please like, subscribe, and share, and I'll be back soon with the main AI breakdown. A couple of months ago, everyone was talking about AutoGPT. It was the new hot thing in AI, and it promised to be as if not more disruptive than ChatGPT. So what is the idea of AutoGPT? The goal was effectively to create autonomous AI agents that were capable of achieving goals largely on their own.

6:19As the GitHub page puts it, AutoGPT chains together LLM thoughts to autonomously achieve whatever goal you set. AutoGPT pushes the boundaries of what's possible with AI. What was different about it is that a couple months before these features became common in ChatGPT and its competitors, it had internet access for searches and information gathering. AutoGPT was also designed to have long and short-term memory management. And effectively, the promise was that you could create any goal you had, and AutoGPT would figure out how to get it done. That meant not only going out and searching for relevant information, not only coming up with a plan, but potentially even spitting up the AI agents that were necessary to actually execute against that plan.

6:57Now you can see when it comes to interest in the project, there was a massive, massive spike at the beginning of April, with people starting to engage with the GitHub repository in a huge way. Another very similar project, Baby AGI, launched around the same time. Baby AGI's GitHub page reads, The script works by running an infinite loop that does the following steps. Pulls the first task from the task list. Sends the task to the execution agent. Uses OpenAI's API to complete the task based on context. Enriches the result and stores it. Creates new tasks and reprioritizes the task list based on the objective and the result of the previous task.

7:31Now almost immediately, a number of different projects arose to try to give these types of experiences a graphical user interface to allow people who weren't coding and running it locally to actually take advantage of this new technology. Agent GPT was one, God Mode was another, and there was even an iOS app called iBabyAGI. Now of course, because of the hype cycle around everything AI, there was incredible expectations placed on AutoGPT really, really quickly. And it didn't necessarily get there in these first implementations. People tended to find that AutoGPT was really good at brainstorming and coming up with a list of tasks, but when it came to actually deploying agents to achieve those tasks, it didn't often work.

8:07At the end of April, Kyle Schrader wrote what I thought was a good sum up of the situation where we were. He said, Regarding the idea that AutoGPT sucks, here's my guide to Twitter. One, don't take AutoGPT so seriously. Two, appreciate its capabilities and move on. Three, block anyone trying to hype it at this point. They are not worth it. It's an early demonstration of organized AI agents with tools. It's cool. That's it. However, over the last couple days, I've seen more and more people talking about something that's called Super AGI. Ken Irwin writes, anyone that has ever worked with me that I like, try Super AGI out.

8:39I am so, so mind blown. It's more powerful than you're even imagining. Now, it turns out Ken wasn't alone. When you go to Super AGI's GitHub page, you can see that there has been a huge increase in interest in just the last few days. From the beginning of June to today, there's been roughly a 10x increase in the number of GitHub stars for the project, which now exceeds 3 ,000 overall. All. Super AGI bills itself as infrastructure to build, manage, and run useful autonomous agents. The features, it says, include provision, spawn, and deploy autonomous AI agents, extend agent capabilities with tools, run concurrent agents seamlessly, graphical user interface, multimodal agents, optimize token usage, concurrent agents, and more.

9:21They also promise that agents can learn and improve their performance over time with feedback loops. A post on Latchy's lifestyle says SuperAGI is essentially auto-GPT on steroids. It can use tools, run multiple agents in parallel, has a graphical user interface, and is super easy to install. Hacker Noon says SuperAGI is an open-source platform providing infrastructure to build autonomous AI agents. You can add capabilities to your agents by selecting tools from an ever-growing library or build your own custom tool. They also point to the fact that this is an open-source community encouraging developers to join and contribute to making the platform better.

9:54Now, if you go check out their Discord community, it's really clear that they're trying to pick up and address some of the problems that people had with AutoGPT and Baby AGI. In their pinned thread in the introduction page, they write, Super AGI is a framework to build and run useful autonomous agents. We believe the world will be run by autonomous systems, agents, and applications. Super AGI intends to build infrastructure to enable this. With Super AGI, you can provision, spawn, and deploy useful autonomous agents. Every couple days, there are interesting updates. On May 30th, 0.02 went live, which added a local GUI for Mac OS, Linux, and Windows, GPT 3.5 support, and email support allowing agents to read, write, send, and save email drafts, and more.

10:34Then just yesterday, 0.03 went live on GitHub as well. This included new tools for DALI 2, GitHub Web Interaction Tool, Human Interaction Tool, and more open-source LLM models, including Llama, Vicuna, Alpaca, and more, and other various improvements around the platform. Now, this is not yet a project where we're seeing tons and tons of demos from the Twitter threaders. Instead, it's the developer set who are getting really excited about the possibilities here, and that's why I'm paying attention. I have no idea if Super AGI is going to solve any of the problems of AutoGPT right out the gate, or if it's just another contributor to the overall AI agent space.

11:07But what's clear is that once you get past the hype, a huge, huge amount of developer energy is going into autonomous AI agents. Where they find product market fit is anyone's guess, but there are certainly going to be lots of shots on goal. Now, in the meantime, as we're waiting for more information about Super AGI, I I've also seen a lot of chatter about this new tool, Multion AI, as well. On May 26th, McKay Wrigley wrote, AI agents are getting crazy. The team at Multion built a browsing agent that will absolutely blow your mind. Here I tell it to book me a flight on Delta from SLC to NYC from June 11th to June 14th, and it does it fully autonomously.

11:42Got it 100 % right on the first go. Unreal. Now you can see at each step of the way, it's telling McKay what it's doing and giving him the option to press do it to move to the next step. I am clicking on the Delta flight option matching the desired dates as a for example. Next, I am selecting the first flight option that has a convenient time and reasonable price to book. And again, you can press do it. Dave Craig tried an even simpler example, saying, Go to Elon Musk's Twitter, then go to Apple's Twitter, then go to Nike's Twitter. You can see Multion do these things in sequence, and at each case, you just have to press do it to continue to the next step.

12:13Divgarg, one of the developers of Multion, says, This release is our test vehicle to simply show our AI capabilities, and we have done zero site-specific optimization as of yet. Our same AI engine works universally on every website, zero shot. We can't wait to unveil our first production race car when it's ready. The point is that while the hype may have died down a little bit on AutoGPT, that's not necessarily a bad thing. As that tweet put it right at the beginning of this episode, it's cool technology. It's AI agents learning how to be AI agents. That's it. Enjoy it for what it is. There is no doubt that one of the major areas that people are excited about in the entire AI space is autonomous agents.

12:50And for that reason alone, people are going to keep trying to explore different use cases, of which some will work and some won't. But it's likely that even the ones that don't will teach us something. So that is the view from Super AGI for here. Obviously, I will continue to keep an eye on it. If you're enjoying the AI breakdown, please like, subscribe, and share it. Click the notification button so you don't miss any episodes. And then go check it out the other places that you can get it. It comes out every day as a podcast, and we also do a newsletter each morning. All right, guys. Thanks, as always, for watching.

13:17And until next time, peace. Thank you.