AI That Learns: We're Closer To AGI Than You Think

In today's video, I want to talk about AGI or artificial general intelligence, a word that scares the heck out of a lot of people in the scientific community. Most of the world doesn't realize it, but we're actually approaching AGI quicker than a lot of people anticipated. So let me first start off by sort of defining AGI or better yet, let's have ChatGPT define AGI for us. Artificial General intelligence refers to a type of artificial intelligence that possesses the ability to understand, learn and apply knowledge across a wide range of tasks, similar to human intelligence. Unlike narrow AI, which is designed for specific tasks and lacks the ability to generalize its knowledge to other domains, AGI can adapt to new situations, solve complex problems, and exhibit a level of autonomy and versatility that matches or surpasses human capability. If you've been paying any attention recently to the AI space, the topic of AGI and artificial intelligence, basically becoming self learning, is a topic that's coming up more and more as tools like GPT4 blow people's minds. And people like Elon Musk and Steve Wozniak and Ahmad from Stability AI, and Gary Marcus, the godfather of AI. And all of these big names are calling for this pause to development of AI. The this is essentially one of the things that all of these people are worried about is that AI is going to be able to self learn. And we don't know what happens next when we open Pandora's box. And in this video we're gonna talk about some of the little sparks of AGI that we're starting to see already, the things that are showing that AGI is actually a little bit closer than most people probably realize. Now, before I start getting into some of the examples of what we're seeing in the space right now, let me break down some of the worries that people have around AGI. And once again, who better to ask than our current AI overlord, ChatGPT. So people are scared of AGI for several reasons. The biggest reason, misaligned goals. You might start seeing people on Twitter and YouTube talk about the alignment issue. And what they're essentially meaning is the goals of an AI that is self learning may not necessarily be aligned with the goals of humanity, and that we need to smartly develop AI so that it aligns with the human's goals and not with necessarily what might be the computer's goals. So as ChatGPT describes it, there's a fear that AGI systems, if not designed with the right objectives, might develop goals that are misaligned with human values. This could lead to unintended and potentially harmful consequences. And as the AGI pursues its objectives at the expense of human well being and safety, people are also worried about the loss of control. As AGI systems become more capable and autonomous, people worry that humans may lose control over these systems. As they start to self learn, they may stop respecting the input of humans. This could result in AGI systems making decisions or taking actions that have significant negative impacts on society. Economic impact the widespread adoption of AGI could lead to massive job displacement as machines become capable of performing tasks that were once exclusively in the domain of human labor. This could exasperate income inequality and create social unrest. We've already seen a ton of pushback around AI from the artist community because AI is actually starting to get better at creating a lot of art than a lot of artists. We're seeing pushback from the animation and computer graphic community as AI gets better and better at that. We've seen that GPT4 is able to do your taxes for you. Accountants and bookkeepers and people in the legal profession and bankers and a lot of professions are at risk of AI actually doing the job just as well, if not better than them. Autonomous Weapons There are concerns that AGI could be used to develop advanced autonomous weapons which could be difficult to control and could lead to an escalation in warfare and global instability. Now that one's pretty self explanatory. Well, what if AI starts developing weapons? Concentration of Power AGI might lead to a concentration of power in the hands of a few entities. This could result in imbalances of power and undermine democratic processes. Existential Risk Some experts worry that AGI, if not designed and managed properly, could pose an existential risk to humanity. A super intelligent AGI might inadvertently cause human extinction or significant harm if it is not adequately aligned with human values or if used maliciously. This is what people are talking about about once again when they talk about the alignment issue. We need to ensure that when these AIs are being developed that the values and the goals and the intentions of these AIs are aligned with the goals of humanity as a whole. Otherwise there could be some huge problems. Right now that I've given a sort of surface level explanation of what AGI is, as well as some of the very, very real risks of AGI. Let's talk about some of the things that are available right now that are actually sort of Sparks of AGI. Back in March of 2023, a handful of researchers from Microsoft put out this paper saying that there are sparks of artificial general intelligence with their early experiments with GPT4. In their paper, they demonstrate that beyond its Mastery of language, GPT4 can solve novel and difficult tasks that span mathematics, quantum coding, vision medicine, law, psychology, and more without needing any special prompting. GPT4's performance is strikingly close to human level performance and often vastly surpasses prior models such as ChatGPT. Given the breadth and depth of GPT4's capabilities, we believe that it can reasonably be viewed as an early, yet still incomplete version of artificial general intelligence. Now, this entire paper is 155 pages long, and I myself have not read every word of it, but I have read summaries of it, and it essentially, essentially breaks down the fact that GPT4 has elements of artificial general intelligence, but still lacks quite a bit to be artificial general intelligence. But one interesting thing about this paper is at the very end, it asks some very interesting questions. It asks, why does it exhibit such general and flexible intelligence? At its core, it's merely the combination of simple algorithmic components. These questions are part of the mystery and fascination of LLMs, which large language models which challenge our understanding of learning and cognition, fuel our curiosity and motivate deeper research. These large language models are essentially a black box right now that even a lot of the people that are creating them and doing the research around them don't totally understand how they work. Yet today we have tools available that we can install and use on our own computers that are even closer to what people are talking about when they talk about AGI. So let's take a quick look here at GitHub. As of the recording of this video, the top three repositories on GitHub all deal with sort of sparks of AGI. We have Baby AGI, Auto GPT, and Jarvis from Microsoft. And for the rest of this video, I'm gonna somewhat break these down and what they're capable of and once again show you that this concept of AGI is really not that far off. So let's start with Baby AGI. This is one that you can install on your computer and use right now if you want. In fact, I've even installed it on my computer. I have baby AGI installed and I've done a little bit of messing around with it. There are instructions on exactly how to use it if you want to install it, but here's basically a breakdown of how it works. You basically provide it with an objective and a task. It then attempts to complete that task while continually learning how to get better at doing that task, the user inputs the objective and task. It goes into the task queue. It then goes to this execution agent here where the task is completed. The task is added to a memory so that it can query it again later for better context. It then looks at the result of the task, puts that result of the task into the memory as well. And if there's anything else that needs to happen to further complete the task, it then sends what needs to happen back to the task queue. The task queue also has a mechanism to prioritize the tasks. So it sees the tasks, cleans the task list and updates the task queue. So what the user does is just this first step here they provide the objective and task. The system then prioritizes through these two steps right here what needs to be done to accomplish the objective and the tasks, and then executes on that. Saves what it's doing to memory. And if it needs to do anything else, it then adds whatever else it has to do to the task queue, further optimizing and improving itself, making sure that it is fulfilling on the task that you gave it. Here's the explanation from their paper. The system processes the task at the front of the task list and uses GPT4 combined with LangChain to generate a result. This result is then enriched if necessary and stored in pinecone based on the completed task result. The system uses GPT4 to generate new tasks, ensuring that these new tasks do not overlap with existing ones. So it is self learning. It is figuring out what other tasks it has to do in addition to the initial tasks. The system then prioritizes the task list based on the new tasks generated and their priorities, using GPT4 to assist in the prioritization process. Okay, so here's the example video from baby AGI that they put out. So the objective solve the AGI alignment issue. The first task of the task list is to build your task list. So you just need to give it to the first task and then it will kind of figure out what the next tasks need to be based on the first task and the objective. You can see it then created a task list. Investigate the existing research and literature on AI alignment. Identify the key problems and challenges associated with AI alignment. Develop a framework and approach for addressing the identified problem. It then moves on to the next task. Analyze the implications of AI alignment. Here's the result of that, moves on to the third task. Develop a protocol and it starts to show the results of that. Now it moves on to the fourth task. Research the potential of AI alignment. Then it moves on to the next task. Develop a strategy for deploying AI alignment. Develops the strategy here and outputs the results. Next task. Analyze the impact of AI alignment on existing decision making processes. Spits out the task result. Next task inventory. Investigate the potential of AI alignment to improve the accuracy and speed of decision making. Here's the result. Next task number eight. Explore the potential. Here's the result. Next task. Create a framework for evaluating the effectiveness. Here's the result. Next task. Develop a system for monitoring the development and deployment. And here's the result. Assess the ethical implications of AI alignment. Here's the result. Investigate the potential for AI alignment to improve the accuracy and then it gives the result. So if we go all the way back to the beginning, the only input they gave was solve the AI alignment. And the first task is build your task list. It did everything else after that. It created the entire task list, step by step of what needed to be done to solve the AI alignment. And then it went through each item in the task list and performed the necessary task and gave the output from each task. So essentially just the one objective and the single first task of building a task list, it was able to go off and do everything else. So that was the example provided with baby AGI. Now let's talk about Auto GPT, which is a little bit different because AutoGPT is, is web connected. So here's the breakdown of the features of AutoGPT and this is one that you can install locally as well. Their installation guide is actually very easy to follow if you just follow the steps on the page. And this one accesses the Internet. You can see the features here. Internet access for searches and information gathering, long term and short term memory management. So as it learns, it actually stores what it's learning into a memory management System. It uses GPT4 for the text generation, it can access popular websites and platforms, and it has file storage and Software summarization with GPT 3.5. Now if we take a look at the examples that they actually share, you can actually see this one in action right from their GitHub page. So they started to run it and then it asked them to name their AI. So they named the AI Chef GPT. And then it says describe your AI's role. Gives you an example of describing the role. And then they say, Chef GPT is an AI designed to browse the web to discover the next upcoming event and invent a unique and original recipe that would suit it. And then it asks for you to enter up to five goals for your AI. So goal Number one, invent an original and out of the box recipe to suit a current event such as Easter. Goal number two, save the resulting recipe to a file. Goal number three, shut down upon achieving your goal. Now, they call this Auto GPT because it's supposed to go through and solve for all the goals that you put in here and just kind of run autonomously. However, they do have a safeguarded place to keep it from just continuously running forever and ever. It does prompt you throughout the process and ask you if you want to move on to the next step. So it's semi autonomous. You do have to be sitting here watching it and give it the okay to move to the next step right now. So let's go ahead and continue with this example. It's thinking so Chef GPT thoughts? I will search for upcoming events to find a suitable one for creating a unique recipe. Finding an upcoming event will help me come up with a relevant and exciting recipe. Next action. You can see that it's actually googling upcoming events. Then here it says Chef GPT thoughts? I will browse the first search result to find a suitable upcoming event for the recipe. Recipe creation, reasoning. Browsing the website containing events will offer me more insights about the theme and suitability. So next action. Browse website and then you can see that it's browsing this Website EventFinder, Co NZ Etc. You can see now it found a summary of the website content. This website page is a guide to events happening across New Zealand, et cetera, et cetera. I will browse the second search result for more details on upcoming events reasoning. The first search result didn't provide a specific event I could use. I'll check another website for targeted information. So right there you can see the power of this. It actually googled upcoming events. It found a website, went into the website, summarized what it found in the website, didn't find what it was looking for and now it's searching the second page of results. Now it's browsing the website heartofthecity Co nz. It got the summary. Now it says I will search for April 2023 events to find more relevant and specific events. My previous search results didn't have a specific event. Searching for events in April 2023 will give focused results. This is doing this on its own. It searched upcoming events on Google. It found a few websites, it looked at the websites, summarized the websites, didn't find what it was looking for, and then altered what it is searching in Google to get a better result. Now you can see Next action. It's googling April 2023 events. I'll browse the websites with a calendar of events happening in April 2023. So it found a website, cutecalendar.com it summarized the webpage I will save Earth Day on April 22, 2023 as the event for which I' unique recipe reasoning. Earth Day is a wildly celebrated event and would be fitting for creating a thematic recipe. And then you can see it added it to memory. The argument Earth Day April 22, 2023. So now it staved to its memory. Okay, we're going to build something around Earth Day. Chef GPT thoughts? I will use a GPT agent to create a unique and original recipe for Earth Day incorporating sustainable and eco friendly ingredients. Using a GPT agent will help me generate a creative and well crafted recipe tailored to the Earth Day event. So not only did it decide that it is going to help you create a recipe based on Earth Day, it recognized that Earth Day is actually about sustainability and it's incorporating that information into its recipe. So now it's using GPT and it's saying create a unique and original recipe for Earth Day 04-22-2023 which would suit this event incorporating sustainable and eco friendly ingredients. And then look, it wrote up a whole recipe here for Earth Day quinoa salad. So this to me is absolutely insane. And this is what I'm talking about. As I mentioned, we're getting closer to legit AGI. I just showed you two different programs between Baby AGI and autogpt which you can just enter in objective and the first one will create a task list and work through that task list until the objective is completed. And the second one will use the web combined with GPT4 combined with its own memory to try to complete the task that you gave it. And it learns and it iterates off of what it's learning and saving to its memory. So not only did it goog events and then not find events that it was looking for, it tried a different Google result. Then it found an event that it could use, Earth Day. And then based on the event that it found, it realized that Earth Day is about sustainability and it thought of a recipe that included sustainability as part of the recipe for Earth Day. And then it generated that recipe and that's all one software that you can install on your computer. Right now I have it installed. The whole thing is less than 1 megabyte. If I open my own command window right here, I could literally run that exact script on my computer right here if I want want let's make our own quick script called YouTube Idea Generator. Now, it asked me to describe the generator's role. An AI designed to find ideas for YouTube videos that are likely to perform well. Now enter up to five goals. Search YouTube for videos about AI that perform well. Goal number two, create a list of 10 new video ideas about AI that will also likely perform well. And then when you're done, shut down. And then if you just hit enter after that, it will start to run your program. I will search for the most popular AI YouTube videos to find new possible ideas for AI related content. By examining the most popular videos on the topics, I can determine what people are interested in and create videos about those topics. Plan Conduct Google search for top AI YouTube videos. Analyze the most popular videos for common themes and content. Generate new video ideas based on popular themes and content. Next Action Google Most popular AI YouTube videos Now when I run this, it prompts me to say yes to authorize the next thing. So I can't just say run this script, walk away from my computer and then come back and have a list of YouTube videos. They put a safeguard in place so that it can't just continually run behind the scenes and do all sorts of stuff for you. You have to sit here and basically approve every Next action. All right, so let's go ahead and hit yes, let it continue. So it found a list of videos and then it says, I will browse this playlist to find the 10 most viewed videos along with their corresponding titles. Reasoning this search will help me identify popular AI topics on YouTube that I can use as a starting point for new video ideas. Plan Browse the playlist link to find the top 10 most viewed videos. Create a list of the video titles. Next Action Browse website and then you can see it's browsing a YouTube playlist here. Go ahead and hit Y and it will continue the next step. Now it says I will start a new GPT agent called GPT Agent AI Video ideas to generate AI related video ideas. Reasoning GPT agents are powerful tools for generating relevant and creative content ideas. This will help me identify a unique angle to present AI related concepts for viewers on YouTube. Plan Start a new GPT agent. Set the task to generate AI related video ideas. Provide an appropriate prompt. Wait for the agent to generate a list of suggestions. Yes, please continue. And here's the response. There are many AI related ideas that can make for popular YouTube videos. Some of which include explaining the basics of AI and machine learning, demonstrating the capabilities of intelligent personal assistants like Alexa, Siri and Google Assistant, et cetera, et cetera. I'll save the GPT generated list of AI related video ideas to my long term memory. Yes, I will now search Google for the latest AI trends reasoning. It is important to stay current on trends in AI to ensure that new videos are relevant and engaging to viewers. Yes, I'll save it to long term memory. Yes. Latest AI trends AI development and climate action Enhanced personalization in AI Increased automation in industries Utilization of AI in healthcare Significant innovation in machine learning infrastructure and you can see I basically just created an autonomous bot, well, a semi autonomous bot that will go and hunt YouTube for all of the most popular videos related to AI. It will look at the titles of those to help me kind of create some new titles for the video. It will then Google trends in AI to make sure whatever videos I create next are relevant. And then it'll give me a list of videos that I can potentially create on my YouTube channel based on YouTube videos that have performed well, plus the trends in AI right now. And I say semi autonomously because I have to press yes every time it wants to do the next step in the process. And that's it. Let's go ahead and close out of this one. And let me talk about the last one which if you remember from our trending on GitHub, was Microsoft's Jarvis. So Microsoft's Jarvis is actually using Hugging Face or Hugging GPT. And if you've watched my last videos about Hugging Face, you know that Hugging Face is essentially a repository of machine learning models and many of them have spaces where you can test them directly on Hugging Face. Most of them aren't really tied together yet, they're just kind of individual siloed tools that you can play with individually inside of Hugging Face spaces. Now what Jarvis and this Hugging GPT are trying to do is create a workflow out of the various available models that are on Hugging face. So here's the example that they give from their paper up at the top. They say, please generate an image where a girl is reading a book and her pose is the same as the boy in the image. Then please describe the new image with your voice. You can see down here, this is the response that it gave. Here's the original boy image. Here's the pose that the boy image is doing. Here's the response of the girl reading the book in the same pose. Then it uses computer vision to find out what's in the image. You can see that it outputs a girl sitting on a bed reading a book and then it outputs the audio of it speaking. A girl sitting on a bed reading a book. So what happened between the request and the response? Well, what it did was it used a whole bunch of the hugging face available models and it tied them together one by one and chained them to get that output. So in phase one, it looks at the prompt and it says, what are all the things that need to happen? All right, we need to look at the image that was inserted. We need to figure out what pose is in that image. We need to create a new image that follows the same pose plus the prompt. We need to detect what objects are in that image. Then we need to turn the detection into text, and then we need to speak it out with text to speech. That's all what happens in stage one. Then it looks at hugging face and all of the available models that are inside of hugging face and then figures out exactly what is available to do that thing that you're trying to do. It then executes on each one of those tasks by using each one of those models individually inside of hugging face. And then it outputs a response based on its usage of all of these models. And then you get this output. Here was the image that was inputted. Here again was the prompt. Please generate an image where a girl is reading a book and her pose is the same as the boy in the image example jpeg. Then please describe the new image with your voice. So here's the original image. It used Control net from within side of hugging face to figure out the pose. It then used one of their image generators to generate a new image based on this pose plus the prompt of an image where a girl is reading a book. So it's matching this pose plus the prompt. Then it's using OpenCV computer vision to actually detect what's in the image. So you can see it put a square around the book, it put a square around the girl. And then based on what it sees, it used this VIT base patch 16224 and this VIT GPT 2 image captioning to caption the image as a girl sitting on a bed reading a book based on what the computer vision saw. And then it used the model fast speech 2 in LJ speech to convert the text into audio. And then it would output the audio saying a girl sitting on a bed reading a book. So it's essentially this multimodal large language model that uses the various models inside of hugging face to generate the output. If you think about it this way, Hugging face has a whole bunch of machine learning and AI Tools that are siloed that each one does a very specific process. What this Hugging GPT and this Jarvis are trying to do from Microsoft is they're making an AI that looks at what AIs are available. So it's an AI to search AIs and then pick which AI tools to use. That's probably the best way I can explain it. So after seeing all of that in practice and knowing that some of this stuff is open source, there are people at home right now using baby AGI, using Auto GPT, tweaking the code themselves and improving upon it and changing it and, and doing who knows what to it. With that in mind and knowing what's coming from Microsoft and Jarvis and Hugging GPT, it's not really that hard to believe that we are very close to artificial general intelligence. The one thing we need to keep in mind is we need to solve this alignment conundrum. The people that are building these AIs that are building these large language models that are working on these systems, they need to do it in a way that we know is aligned with humanity's best interests. We don't know what's going to happen. There is still sort of a black box element to it where a lot of the researchers that are building this stuff can't fully explain how or why it works. And that's the scary piece for people. And so this alignment needs to be programmed in now. This isn't meant to be a doom and gloom video. I'm very, very optimistic about where we're going with AI. I'm very excited about all of these technologies. I'm looking forward to playing with them in ethical ways. I'm looking forward to being a part of the solution and not part of the problem. Again, I'm a creative. I like to make videos, I like to have fun. I like to create silly memes and push the limits of creativity using AI. I'm not somebody that's an expert on ethics or philosophy or that sort of thing, but I am somebody that because I am deep down this AI rabbit hole, I do start to get more concerned about where some of this stuff is going. And I am seeing the advance happen so fast. I just want to make sure everything is done in a safe way. So are the fears justified for AGI? Who knows? Nobody can say for sure what's going to happen. But what I do know is that I'm going to continue to educate myself and learn more and more about all of this stuff and what's going on behind the scenes and I'm going to continue to try to make content to help you better understand what's going on behind the scenes so you can make informed decisions and you can have your own opinions and ethical and philosophical discussions about this. Because I think this stuff is important to realize it's happening. Because this is the stuff that if you're not paying attention to it and it's not on your radar, you're going to be caught off guard. And I don't think that's good for anybody. So I think it's important to talk about it, I think it's important to make content about it, but I also don't want to take the fun out of it. I also want to use these tools for enjoyment and fun and creativity and just to make amazing stuff that people are excited about. And yeah, that's what I have to say about the AGI topic right now. Hopefully you found this video helpful. If you want to stay in the loop of everything that's going on in AI and get the TLDR every every week of everything that's happening, head over to FutureTools IE and join the free newsletter. Every Friday, I'll send you five cool tools that I came across. I'll send you a handful of news articles that are going on in the AI space right now. I'll send you a handful of YouTube videos. And one cool way to make money with AI. It'll be the TLDR of the Week. So if you don't want to keep your finger on the Pulse as much as I am, get on the newsletter and I'll keep you on the pulse for you. Also, if you haven't already, if you enjoy this kind of content, make sure you give it a thumbs up and subscribe to this channel. I'm putting out as much content as I possibly can about what's going on in the AI space. Cool tools, news, tutorials, just anything that I have fun nerding out about. And if you like that kind of stuff, giving it a like and a subscribe. We'll make sure you see more of this type of stuff in your feed. Appreciate you. Thanks for hanging out with me today. See you guys in the next one. Bye.