AI Agents - BabyAGI, AutoGPT (AgentGPT) - Research, Tools, Repos

hey everyone welcome to Star morph where we talk about artificial intelligence and web development I'm really excited to show you this video today where we're going to go deep into AI autonomous agents because not only is chat GPT and large language models exploding in all of these new use cases and GitHub repos coming out but we're seeing a lot of really interesting stuff with this concept of an AI agent that can perform tasks interact with this environment and learn through reinforcement learning so we're going to talk about tools like Auto GPT baby AGI Yarvis agent llm and a lot more we're going to look into research how you can get started using these tools building them on your own and just a whole overview of agents and the related Concepts so before we get started with the rest of the presentation I figured why not get an agent working for us in the background while we do it because that's what it's all about is you get it going and it's working for you so this is Agent GPT this is available on the star website I've just deployed this open source GitHub repo so we'll also go into how to deploy these and we can use this similar to chat gbt we can give our agent a name and then instead of a prompt we can give it a task so I can say write a next.js web page with a list of card components and now we can deploy this agent and it's going to iterate on that task and think about how to accomplish this task here we go it's generating so we'll let that run while we go into our presentation all right so I've labeled here a few things that are potential benefits of the AI framework the agent framework and we'll go deeper into these but the key one is kind of what we have going on right now which is that you can get these agents to run continually tasks uh run continually on your tasks so you don't need to write a complex prompt and then get one output write another complex input you get another output it's able to chain events which is really helpful for scaling um let's continue on so what are we going to go over in this video first we can we're going to talk about some of the concepts of agents and then um we'll I'll show you some tools that you can use like the one we're going where you have running right now to use AI agents yourself with tools that are available then we'll go into a lot of the academic research that has happened regarding agents reinforcement learning uh llm Improvement and relevant topics that are synthesizing into the products that we're seeing coming out and then for those of us who want to dive into the code a bit we will go into how to deploy some of these amazing GitHub open source repos it's such an exciting time for open source development right now so you can get your own agent tools set up on your website and then finally for people who want to take it further than that and do their own custom development for agents we'll go into some Frameworks like Lane chain and reinforcement learning Frameworks to be able to code your own agents okay so why do we care about agents in the first place and I think it's a solution to some issues that I've been noticing uh with these large language models now Stormer has a for a few chat Bots live that get a lot of visitors and um I've seen I've learned a lot about the interaction and I come from a UI ux background so I'm thinking a lot about how do we interact with these Bots and you know what are the limitations and what are the the strong points so I just want to explore this for a second because this is something I'm seeing is a real um a real throughput um Block in using these llms and that is not the power of the llm but the humans ability to query the bot and be able to extract value from the llm so this right here is my visualization of a super intelligent computer and this this is a Swiss army knife that has an overwhelming amount of tools so imagine that you were gonna go talk to an omniscient computer or alternatively imagine you're going into a room and you and Elon Musk is there Albert Einstein is there and Obama's there what do you say you know you see them in a coffee shop what do you it's hard to come up with what exactly to say to them because there's potentially so much value that could be had in that conversation it's hard to know where to start and I think that's an issue that a lot of people have with llms I see a lot of people show up to an llm and say hello and so that brings up the current solution to that which is engineering specific prompts that detail out what we want the llm to do and this is great because we can direct the powerful tool in a specific direction that we want to take it but it also takes a lot of work so you know for an end user this might be something that we need to see abstracted more and packaged into a product that the user doesn't have to see this it just does this for them okay so summarizing a little bit some of the issues and the limitations with our current UI ux and using llms it takes time and effort to write detailed prompt instructions and not everyone knows about prompt engineering yet and has the the time or awareness to write out very specific prompts or know what what to write the second thing is a lot of people talk about chat GPT being able to interact with external tools get up-to-date information on the Internet or custom data that is relevant to their business and that's something where we're seeing agents start to be a solution for that that's you know chat gbt plugins and chat Bots certainly can solve that as well but a lot of Agents have apis for web scraping or searching the web um to be able to get real-time data like Auto GPT can show you live what website it is currently scraping to get information from Okay the third one is yes we can produce amazing content and get great Knowledge from large language models but also how do we utilize that information and how do we save it for later use to make use in our business and our personal lives and we're seeing that with these agents that have a file storage mechanism you can start to accumulate documents of code or you know whatever you're writing documents um in assets and I think that's really important being able to walk away and have something to show for it after having the chat and then just highlighting back on the knowledge having specific knowledge you know with embeddings um or another form of bringing your company data in through an API um I think I haven't seen as much with an agent yet but I think that that would make an agent really interesting to see it with um embeddings and more specific data so I think that's agents could help with that um relevancy issue as well all right so how do we start to solve some of these limitations with chatting with Bots and of course people have been studying agency and reinforcement learning for many many years so I'm not you know claiming to say anything new here um but something that's been really relevant recently with these large language models is the reinforcement learning feedback loop and there's two kind of models for that here super basic one is the human interaction with a large language model so that's kind of what we've already been talking about how do humans interact currently with the chatbot UI u x wise and I think this is a relationship that the world is going through rapidly right now we're learning how humans interact with large language models and this brings up a whole other issue or a whole other area rather of the ethics and alignment of large language models how do we build large language models that are useful for humans you know and not um harmful to humans and I think that's a very relevant question right now in software development um so I think this human computer interaction between humans and llms is extremely interesting area right now where it gets even crazier is when you have the llm starting to talk to other llms and I believe this is how gpt3 was able to get a lot better into gbt4 was through a mechanism called instruct GPT which I believe was based off of the self-instruct GPT paper and this is where the llm can basically revise its own answers and learn from its own um its own responses and create a loop where it can become more intelligent from its own feedback and you don't even need human input to increase the intelligence of the system and uh so that's a very powerful Loop and then this is a recent tool that someone released that is an interesting exploration into some of these ideas of human and llm interaction and prompting the user to engage with the bot by giving the bot a large personality that reaches out to the user so they're inclined to connect with the bot as well and then this last section here is just breaking down some of the parts of an agent uh the lane chain agents have a concept of a tool that they can utilize in their tasks so things like bringing in external apis searching the web I think you could consider the ability to save files as a tool and then I want to see some more embeddings mixed in with agents I think would be super cool and these tools allow them to accomplish specific tasks which when you have a reinforcement learning mechanism can start to create a loop to improve its own abilities more before we go into that I want to mention that I'm compiling a lot of these resources into this blog post so let me just run through this really quickly so first off there are some links here to start just using agents if you want to get started so this is a baby AGI UI that I built on Star morph I'm going to go more into that later in this video the next thing is the agent GPT that we're using that's still running probably maybe it finished it's still running yeah it's been just coding so that's pretty cool hopefully not running up too much of an API bill but okay so we have some tools you can use then we have a ton of academic research because there's so much cool stuff on agents some of this is more abstract this isn't even really artificial intelligence it's just the concept of agency um some of these are way more targeted specifically to AI agents like let's see okay camel is a really cool one and then this is a really cool one too and I'm going to come back to this to dive deeper into it I'll come back to that one in a little bit um but yes we have a lot of research papers here that you can check out if you want to dive really deep into this stuff we also have some programming Frameworks and then we have some demos this is that Samantha chatbot I just brought up earlier and then we have some GitHub repos uh so you can deploy and use these tools yourself so that is the blog post this will be in the description if you have any tools that you think should be added to this list please leave a comment below and I'll add it to this list I'm also thinking about making this into a GitHub repo so people can add to it and uh maybe even a website if we if people find it useful okay so going back to the presentation this is Microsoft Yarvis which they don't necessarily describe as an agent but has a lot of the characteristics of an agent it has a it's a system that Microsoft is creating where the tool the AI tool can plan tasks and then dynamically figure out which machine learning model on a hugging face is most appropriate for that task then it will use that model to accomplish the task and give the response so this is an example of how you can start to piece some of these puzzle pieces together to create a really powerful agent that not only uses one machine learning model one you know gbt4 but is able to dynamically select from potentially thousands of models that I believe are on hugging face so uh the fact that Microsoft is you know thinking in this kind of agent framework I think it shows a lot okay this is that paper that I said I was going to come back to I hope I'm not bopping around too much I wanted to give you guys a broad overview of a lot of different areas of this and then allow you to decide which is the most appropriate to dive deep into so this is supposed to be you know a breath um search algorithm search over the topic so this paper here they did something really cool where they created agents in like an RPG video game and um then those characters were able to interact with their environment and walk around the space and chat with each other um so this is an example not only of an AI agent but of a group of AI agents collaborating in an environment and I think that's an extremely powerful concept um and I made a little sketch here let me see if I can find it I'm not sure if this is going to load from what I had before no it didn't save it okay but my idea was that my idea with this is imagine if instead of a video game you have an agent who is a coder and then they pass the code to an agent that's a debugger and then they pass the code that's an agent that's to a deployer and then they pass that to the marketing agent so basically what I see here is a civilization of AI agents that are performing tasks and even though we're super early on in seeing the products that are of reaching consumers with agents I think this conceptually is extremely crazy and um you know who knows in a week or two weeks we might see an agent that just breaks through and is 10x of what we've seen so far Auto GPT is already one of the fastest growing if not the fastest growing GitHub repo of all time so clearly this concept of Agents is really striking people um and that's why I wanted to go over all of this okay so that is the end of the presentation I have now I'm gonna let's go back into the agent GPT and check how this tool is doing and then we'll work on deploying our own version of one of these um Auto GPT or rather agent tools so let's check in here let's stop this agent and just take a look at everything it's done so it started off with some basic tasks that it broke down my initial task into different categories so gather requirements make a project skeleton and then make a card component and put it in the web page and then you can see as the agent as an agent it wants a bunch of tasks and I think it's it wants a bunch of information I think it's going to find that itself but it's able to think through this process until it's able to complete the task so now it's starting to write some code looks like the card component nice it's mapping over this list of card components okay let's just skip down here and see how good it got it by the end I like what I'm seeing handle search we have a search okay sort order create a function handle sorting to sort the card components based on their titles in ascending or descending order so it just kind of loops and keeps on improving and keeps on improving and some of these tools you can set how many iterations it does because otherwise it will just keep running forever but pretty cool that it went past what I even asked for and started to write a search mechanism and you can see here here's some of the code card list card component so it's already done what I've asked it for you can also with this tool export an image or copy the text to get the code yeah so that's here's our next JS page right here it's not perfect that's got cut off there it's certainly not perfect but it you can see that the agent is able to work through a task so it's still very early on and personally I prefer using gpt4 stock or one of my chat Bots um to the agent model currently but I think conceptually we're really close to this being extremely useful and so this is the agent gbt which again is a browser version of Auto GPT so Auto gbt is usually through command line and this just makes it a little easier to interface with I would absolutely show you how to deploy this one open source repo but it takes a few environment variables and I think it'll be easier for non-developers to deploy another tool so let's go ahead and do that so instead of deploying the agent GPT alternative to Auto GPT is baby AGI another AI agent tool that recently came out and someone built this amazing uh UI for it so there's now UI for it as well as the command line interface and you gotta love it they have the one click for cell deploy so I'll put this in the description it's also on the blog post and all you have to do is go to this repo and it looks like it's actually hooked up with pinecon oh so there might be an embedding in this as well I haven't hooked it up with Pinecone yet but let's go ahead and deploy it all right so we just go to the repo and we click this deploy button and then we can name baby AGI video I'm just calling it video because I'm making it in this video all right and this is going to take about a minute I believe to build and then we're going to have our UI to use baby AGI and this is similar to the Asian GPT tool that we were just using we can set the number of iterations so if we don't want it to Loop forever it will actually finish that's working how's our first cell doing I think it's almost done there we go there we go so now we have our own baby AGI app on our versel URL and you're able to go into your settings domain and add your website here so I could add uh you know video test.starmorph.com and then just go into my domain registrar and add okay I'll just do this so now I would just go into my GoDaddy or my name cheap wherever I have my domain and then I would add a new cname record with this and then um we're gonna have that baby AGI UI available at that domain and you do have to put in your openai key down here so that's how you can set up your own AGI tool baby AGI tool and the last thing I'll go over in this video is um or maybe two more things one more is going over some of the Frameworks so if you want to not only deploy your own open source repo which are down here some of the options um but you also want to really dive deep into coding with this stuff here's a few resources for that one of them are link chain agents and they just came up with this thing today that is a new tool chat agent structured tool chat multi-input tools so a lot of development happening with agents in Lane chain is you know one of the Hot Topics in terms of building agents right now so there's a lot of Open Source Code and a lot of cool functionality that can be done here I also ran into this framework which is a reinforcement learning framework that was made by openai so in terms of learning more about how these agents are learning and improving themselves I think this is a great resource but and another reinforcement learning library related to open AI is stable baselines which is an improved version at least according to I'm not sure if they're both officially from openai it's a this is a fork apparently um of the open AI Baseline so another framework to learn about reinforcement learning and then let me just mention as well that um this one's pretty cool this is an open source repo that could also be deployed and this one is similar to agent GPT except for instead of using gpt4 it's using native um it's using a locally hosted model and I think it's actually saying that this is running in the browser which is pretty awesome that's really new functionality and that's why this only is going to work in one of the newest versions of Google Chrome but this is pretty amazing being able to run large language models in the browser the people that came out with this technology to run in the browser also just made it so you can run large language models on your phone so basically we're going to start seeing these everywhere continuing to grow at amazing speed but this is another cool tool I wanted to highlight no open AI API cost for running this one although it did take a lot of uh GPU power on my machine when I ran it or memory on my machine I think so that's another cool tool camel is also doing some amazing stuff with like multiple agents in an environment and there's a camel uh based Auto GPT there's also a demo from camel um and some research papers from camel about agents they have this research paper they also have their website which has their research their demo some YouTube videos so they're doing some really interesting work with agents as well and I think that is a good starting point um for this video please let me know if you think I should add anything else to this list I hope this video was a helpful overview into agents and some of the some of the exciting things that they're bringing to the space and the amazing tools that are coming out um and yeah thank you so much for watching if you found this video helpful please do like and subscribe and I will see you in the next video

Channel: StarMorph AI

Views: 4,481

Keywords: autogpt, babyagi, ai agents, autonomous ai, autogpt tutorial, agentllm, agentgpt, reinforcement learning, large language model, microsoft jarvis, machine learning research, autonomous ai agents, langchain agents, openai api, openai, artificial intelligence, RLHF, self instruct, simulcra, babyagiui, autogpt example, automate openai

Id: SYG3j6elkDs

Length: 24min 15sec (1455 seconds)

Published: Wed May 03 2023