AutoGPT & BabyAGI: autonomous AI Agents for LLMs explained

hello Community today we're gonna talk about Auto GPT and baby AGI let me show you why we have here this very famous GitHub repository now about Auto GPT and as the author writes it an experimental open source attempt to make GPT fully autonomous well the solution is quite easy so what we have here is we have internet access for searches but this Internet access is over an API an API you have to pay for then we have some long-term and short-term memory management in the form of some vector representational vector spaces I will show you an in-depth view in a moment we have GPT 4 as our main intelligence in the whole system and then we have via the interface functionality of length chain access to popular websites and the platforms where we have data and then we have if you want summarization with GPT 3.5 3.5 towable and whatever you have so beautiful and as you already noticed and this is now something very positive that here on the GitHub right next to experimental open source application there is that you need an open AI API key and you need to set up a paid account with all your credit data for this system to work so this is one of the first the very first API that you have to provide your credit data for you will have a free trial and in my understanding it depends on the region where we are I've seen free trial of three dollars I've seen uh subscriber telling me they have 18 dollars a free trial so wherever you are you have a free trial but be aware whenever the free trial is finished you have your credit card over there and therefore one of the most important buttons to press here is the limits you can set monthly spend limits here for example with open UI and I can highly recommend this that you say Okay per month 20 US dollars for this or 50 US dollars but remember this is just one of the first API you have to pay for so depending on the complexity of the task you asked for you have this 5 times or 10 times so choose your monthly threshold wisely okay and here we go now at first it is always great if you read this system can act autonomously without requiring a human this is what you are dreaming of a system has your credit card data and it goes shopping well it's not that beautiful but have a look interesting fact about Auto GPT is that it is currently not integrated in the Lang chain ecosystem so the interface functionality between GPT 4 and all external Services is not a length chain but it is done manually with code written by the authors and talking about code just very short you have of course to specify your open AI API key as I told you you can run on different Cloud platforms like Azure and if you run on Asia for example you have to have all your different keys and whatever so keep every documentation of your Cloud compute accounts ready and so let's have what is the easiest view on the whole system architecture well on the left hand side we have gpt4 maybe you have also GPT 3.5 or whatever you work with for Microsoft open AI and remember we have this beautiful 4K prompt where 4096 tokens is our input Corridor to GPT form and for gpt3 it is cheaper but GPT 4 is of course if you want a better operational system with a higher accuracy and then you have on the other side here all professional external Services they have a professional API where you can put in your Professional Credit Card and you can pay your professional money for the Professional Service that those microservice service provider or real service provider whatever you need from the internet and then in between you have simply memory but this memory has a special form if you have seen my last video you know that here the text for example sentences are converted to mathematical objects to Vector representation in a specific Vector space recall this a vector embedding and if two sentences have a similar semantic meaning this takes care that those sentences are close in this constructed Vector space and if we take the mathematical operation of a cosine similarity onto Vector objects we have now with the cosine similarity an instrument to find sentences that are about the same topic this is if you want a vector store a vector database with a cosine functionality this is somewhere in the cloud you pay the cloud service provider you pay the software service provider you'll pay you know what it means so those are your main building blocks those are the main chunks that you have in your system architecture now let's look at the flow since we do not have here as I told you Lang Jane as an interface functionality everything is here Simply Connected with the internet and with a manual written code example is specifically fitted here for auto GPT beautiful so let's start there was a funny publication where they said that the main user in America is about 16 years old if you are asking about Auto GPT so I don't know if this is the case worldwide but it's a nice idea to have here 16 year old teenager and he has now a question here and he uses Auto GPT so there's one input with one question and it goes into gpt4 beautiful then gpd4 has here the possibility to ask to Interlink with other gpts well you know all the tokens that come out here are here input tokens you have to pay for input tokens then your question here is calculated you have to pay for that then you have output tokens your pay for output tokens then those output tokens are input token here you pay for input tokens and you get here this beautiful circle of life and then GPT sometimes decides okay now I have enough information now I can have an interface to the world now I need resources and those resources are here for example external Services some of them are very costly some of them are medium costly and maybe some of them are free now chat GPT or gpt4 or whatever you have when you watch this video we have here now the control cycle number three sorry there's nothing here it's just I couldn't find a curve here but chat TPT gpt4 the decides only this AI which external service it would need to fulfill here question number one so it goes there accesses API you have your API key a service provides the service and you get an answer and those answers are now fed back into a database why because the answer can be quite lengthy and you know we have here if a 4K token length so you store all the different answers hops and then something beautiful is happening all those sentences are transformed to Vector objects Vector representation in a very specific Vector space and then you have the cosine and given that you have now that you know what you're looking for ninety percent of the answers that we are fed back here either you delete it because you just need 10 of your answer because in those 10 there is the factual answer to your question so 90 lost or you store everything here in the database and you have higher database cost depending if you are working with gigabyte terabytes whatever amount of data and just to show you here about one year ago I made a video about the ipcc report 3600 pages and used exactly here as 3D visualization here of a sentence burden padding so let me show you how does it look like you can have different correlation you can have different topological analysis you can have a cluster analysis so whatever you want to do here you can do amazing things you can have I don't know you can Define how many thematic clusters you want to have you see here in 3600 pages I had close to 100 schematic clusters and you can zoom in you can have a detailed analysis of addressing so this is something we're familiar with two years ago one year ago I had all detailed videos if you want to code this for yourself and coming back now this result here this reduced set of information but where the information is incorporated this tiny result is now fed back into our 4K token or whenever you watch it 16k token length or maybe even 32k token length so GPT is now has now the first iteration to its job now chat GPT or gpt4 may decide that it needs now another run that it needs now additional data so you see what is going on and it's on and on and on and this is why I told you please do think about that you define threshold for your financial apis where you have your credit card data because if you run this and if you're coding and if you're deep and focused and concentrated and working believe me if you look up up after 10 hours of coding you would not believe the bill amount of resources that you have to pay for so always good to have some threshold installed and just to make sure the only body who decides in this whole system is the intelligence of this AI system and if this AI system is an llm that gpt4 for example this gpt4 decides all processes that run here and all processes that could be very expensive talking about expensive let's talk about security since gpt4 is here steering which web server or which address it connects to there is now unfortunately in the last days happened that people understood how the system works and they set up some apis with let's call it some malware in general there's a let's reinstruct formation where there is some prompt interaction or whatever you can think of so whenever jet GPT or gpd4 Auto connects now to a link for whatever reason maybe it was given the address in another service then you could have that maybe if there's a direct feedback linked to chip to gpt4 that you could have here some problems or a more disastrous phenomenon I've heard of at first they play here that they are free service but there's something going on I do not want to describe to you so security here is a real real odd topic that has on this system not been addressed that you really can recommend this is a safe system please keep this in mind and talking about risk and security I've read here in the news Auto GPD lets you set it forget it and rake in the passive income that O2 GPT creates for you whenever I read this somewhere in the news you know what happening to me exactly this this yellow warning sign appears and I hope whenever you read something like this that you just have to set it up give it your credit card data forget it and then do you really believe that you rig in the passive income what are the costs that you pay for this so please be careful recommendation if you want to see here the full code development there's a YouTube channel Sophia young this is your handle have a look at this channel if you like it give it a like subscribe to her and you see here in the background the output of Auto GPT I'm not gonna let you read it because you have to go to her Channel and have a watch beautiful next object baby AGI there's a reason I put it second although baby AGI is taking Silicon Valley by storm Yuppie so okay so what is it it is very similar it is a task driven autonomous agent we have our key players we have tbt4 we have again a vector store Vector database codes and similarity but now we have a lang chain interface functionality with a task to autonomously create and perform tasks based on an objective sounds familiar not for example attractive might be create and run a small content marketing business let me tell you the intelligence always is only in gpt4 with its 4K token length beautiful so this is the GitHub repository of baby ATI and they have a warning baby AGI is still in its infancy and thus we are still determining its direction and the steps to get there so this is not even experimental this is Alpha prototype experimental times 2. beautiful how does it work easy instead of a single task a single question you ask the system we have a list of tasks this is one of the main differences then also you have open EI API pay for this you have you enrich the result by external Services you know what this means API stores it in a vector database with a cosine similarity you know that it sounds similar or you know this then it creates new task and re-prioritizes the task based on the list now this is something interesting because it has now an additional functionality it learns in each step and therefore maybe on the task list it finds out after Task 1 task 2 I have to shift down a little bit and I have to insert another task because I have to find out about a specific code segment first so nice idea as I told you baby ATI uses the length chain framework so we have all our code segments that we love and use and we can use it here and as I told you since we have a list of tasks we have now three functions some people say there are three agents or one agent and three functions whatever wording you like just the content is important you have an execution agent task creation agent and you have a prioritization agent so this is the text to it please read it I'm going to show you the same now in a visualization by the order so here is his documentation have a look at this it's really interesting to read so we have an input the first input is an objective and a task if you want and this task here goes to the task execution brings back a result a salt goes to the task generator to the second one creates now a new task task list goes to the task prioritization the task prioritization reorganizes the different tasks and Define what is the next task to execute interestingly you have here a security agent that might emerge in the next week I hope so and it's a very good idea to integrate your security of course as you know we have a vector store outside of all of this where we can store the tax document PDF information Finance data string data whatever you have in the mathematical representation of a vector in a vector space this is it simple but let me show you this now our third time and this third time the same thing now in code now we know we have three agents or one agent with three functions the first one is execution so we come here we have task execution then task generation and task prioritization when I show you now that this tasks are very simple to understand if you remember that I told you a gpt4 you have the prompt and you have something like in the fine tuning methodology toolbox you have something like instruct fine tuning where you provide a template with clear instruction and you're not going to believe it can exactly this here is kind of happening so you tell now gpt4 you are an AI who performs one task based on the following objective create a smaller medium-sized Enterprise for marketing or whatever take into account these previously completed tasks then you have whatever you have a marketing study or have nothing whatever and then you tell tbd4 your task is to ask response so you see you could type this into the prompt on your free chat GPT or your chat GPT plus there's no difference this is just the automated version so you say Hey you are known here so if you're role playing these are the tasks you have to do this is your input this is the output I expect you have an instruction you have a template beautiful Next Step equation we have now the same thing you tell now gpt4 you are now a task creation AI that uses the result of the execution agent of the source here our execution chain to create new tasks with the following objectives we know the objective the last completed task has the result so you integrate if you already have a result here this result was based on the task description your last as prescription there is an incomplete task the rest of your list based on the result create new task to be completed by the AI system which is GPT 4 itself the Do Not Harbor overlap with incomplete task return the task as an array so you see again you could just type this manually into chatgpt and you will get the same results you have here an instruction based fine-tuning prompt when ICL if you want then here because we cannot fine tune here chat GPT or gpt4 this is also an ICL and this is what we get back and the last one is as I told you the task prioritization and again you tell gpt4 hey you are a transportation AI task with cleaning the formatting off and re-prioritizing the following task ABC consider the ultimate objective of your team create a marketing small medium-sized business do not remove any tasks return the task as a number at least like first task second tasks for task start the task list with number so you have a template you could type this in manually to chat GPT or you pay here for gpd4 Epi whatever it is you see this is in context learning on a gpt4 system and this is so beautiful to do and you understand now what those three agents or at One agent with three function is doing and those three agents you see here those are my circles Merry-Go-Round circles those I'm sorry my three agents here so you see we have gpt4 here our 16 year old Samsung teenager gives a list of tasks with an objective and the first objective is create here my new business and first task start now beautiful so goes into a 4K funnel to gpt4 it's an ICL command if you want we cannot change the weights of gbt4 to learn anything internally via fine tuning so we got a return result back and this intense communication with GPT 4 here is one of the characteristics here of this specific baby AGI implementing the standard interface functionality of Lang chain Frameworks beautiful the rest is the same you have external Services you have your vector store you have your cosine similarity or maybe you use a topological similarity or you use anything else but the main structure for your understanding this is what's going on these are the main constituents in your system architecture beautiful this is it more or less the nice thing about baby AGI is there's a very short script to install you have seven points and you are up and running even if you do not know how to program at all so you just have your jupyter notebook for example your collab notebook use a git clone and you have here the baby AGI git command you change into the Clone repository you install the python requirements the other python libraries that you need to execute this then you need to put in all your API Keys you know where we have a page with your credit card you have here beautiful all your top secret API keys as I told you can set an objective create a business in marketing whatever and then you just type in Python bbagi hooked python return and the system runs this is here the beauty of baby AGI and this is why it's so as you've shown in my first slide on baby AGI took a Silicon Valley by storm the nice thing reading the documentation by the order is that he is aware of the key risk associated with his method and he clearly addresses data privacy and security issues so yeah risk of data privacy security breaches accessed by unauthorized individuals you remember I showed you whenever gpt4 decides to link to an external internet address you can't be sure on to what address the system connects to so the idea to create something like a security agent I think I would highly welcome this additional security layer here and baby AGI let me give you an idea and some voices the chief technology officer of Ernst young said in an interview with Venture beat this is the article about entropy Auto GPT is fascinating but he admits that he doesn't yet understand the details of how it actually works now you as a subscriber here of this channel now you know how it works so you know more than your Chief technology officer for Ernst young about it what about Nvidia in Nvidia AI scientist tweets I see auto GPT as a fun experiment as the authors point out too but nothing more I'm not sure that it is nothing more it has some really interesting aspects but of course for the financial data I told you how you can protect yourself against unforeseen Financial expenses when you work with auto GPT Nvidia scientists goes on prototypes are not meant to be production ready yes of course it's really early in the development stage don't let Media fool you most of the cool demonstration projects that you see online are heavily cherry-picked of course of course this is true but nevertheless and if you like more the sector of here stability eye image and so so on we have here the challenges becomes that we don't really know what an error looks like with Auto gbt well I can tell you if you are the father of a teenager of a 16 year old teenager and you get your credit card bill from your credit card that your son uses and you suddenly have 500 US dollar here from any internet sources I can tell you you know how an error looks like you know how it feels this error because you get an emotion about this so and he goes on that auto GPT currently fails in 15 to 30 percent of the time in reasoning so whenever you use Auto GPT if these are the error rates one sort of an error rate well it depends it is very early in the development cycle but how we can reduce this error rate almost to zero with security agents there's a lot of work ahead of us but I think from the technical perspective it is a highly interesting system that is coming into life and we have to be really aware and we have to know how the system functions and we have to know how we have to control this system and if Government really wants to regulate it we have to help here all legal politicians or whatever you call them to come up with solutions that are useful to implement and that really will not hold back the Innovation cycle that we are currently in this is it for me I hope you enjoyed it I hope there was some information for you as I told you I like to share my knowledge that I have and I hope to see you in my next videos

Channel: code_your_own_AI

Views: 4,668

Keywords: AI, AGI, GPT, GPT-4, Autonomous Agents, LangChain, BabyAGI, AutoGPT

Id: 5cnk5zpf1EA

Length: 29min 36sec (1776 seconds)

Published: Sat Apr 22 2023