Fine-tune ChatGPT? Buy Embeddings /OpenAI? What are Embeddings? My own ChatGPT? | Visual Q+A

Key Takeaways

hello Community Welcome to our q a about how to outsmart Microsoft paid version of chat GPD so here we go I got a lot of questions for you I know about half of them is hey I just bought embeddings from I don't know out my eye cohere how can I now build my own chat topt okay let's look at this I hope at first you ask Siri how to build a chat gbt and then you come to me and say hey code for AI so what I do now with my new embeddings okay let's start you have knowledge written down in books a lot of books and some of those books you put in a treasure chest and you say okay these are my training data for the system and then if you accompanying you lock this training chest and uh Ida I don't know for example if there's gold inside or some diamonds inside it's a locked system and then to be really transparent you say okay I put this lock system this treasure just in a black box and this black box will do the work for me and every customer that approaches me will only have a black box to interact with right so where are those embeddings you have an input and an output from this black box the input and I've chosen you an example by opening I hook very clearly has in its documentation please read the copyright reference as how it is done you say okay here yeah application this is my open API key this is where you have the connects to my credit card and then you have here your input your text string goes here so here's your your question what you want to know the information that you ask for great and you send this into the model and then you get an output of the model and if you bought embeddings if if you really just bought embeddings this is the the result you get back you get back a vector attends a high dimensional tensor a vector object representation of a vector space and then and then comes the beautiful part and at this moment some viewer turned to me and say hey what now okay so welcome to my course if you want to create an AI model you can sell to everybody on this planet like Microsoft did so we have of course if you want to sell it to everybody we have to include all their knowledge days on the internet we copy the internet done of course since every day millions and billions of new ideas and documents are generated around the globe you have to have a cutoff date for your model and here I can take here the documentation by Beta of mei.com docs guides embeddings water embeddings you see here you have a clear for version one a cutoff day say August 2020. stop no data behind August 2020. or the newer version version 2. you see here September 2021 and it is not a coincidence that it's about a year between those models because training those models with all of their input data all after training data cost Millions if not tens of millions if you run it on let's say Microsoft azure so you see here this is a cutoff date and you just put all of the knowledge you could legally allowed respecting the copyright copy from the internet beautiful and then you have a model and this model produces a vector embedding and here here this is what you bought of all training data so just to be absolutely clear the training data is all the knowledge the company put here in this data chest this treasure chest and locked it up for us and the model is here this beautiful black box that for example opening I trained on Microsoft Azure Cloud infrastructure and they have now a model and they say okay if you pay us you can use it to generate your embeddings so a model the Black Box produces a vector embedded what the hell is a Bitcoin battery I just showed you of all the training there beautiful so now you say okay so I have now my vector embeddings and I have all that data I have Vector embeddings what I do now with my input vector embeddings and how I build a system out of this beautiful beautiful I really learned a lot based according to the question of my users but it is important that everybody understands it so if you have here your chest and here all here your Digital Data now with the model what is the main purpose of the model the model needs to structure to bring order in this chaos of information overflow so the model structured input data what is an embedding now have a look here at this car park isn't it beautiful now imagine you are looking for your car where you parked your car in this car park you remember there are these beautiful huge signs for example A4 and you remember your Park directly under the sign A4 so this indicator this Vector embedding of the position of your core gives you the location of your car yes it is this easy this is a vector embedding this is a positional indicator giving you a direction where you want to go where your course where your data is beautiful now what does the model do the model takes all the data from the treasure chest and now it has to bring order in this Digital Data so it says okay all books all articles all Facebook comments about politics I put here first line here are my politics everything about Finance I put here in the third row everything about science science can go here art literature painting whatever might go here so what you see what's happening it is more or less a segmentation of the available space you say okay to bring an order like in a library you say you have local places local spheres local Vector spaces where you assign the data to beautiful so now imagine this is you you are driving a blue car and this blue car is at this position and you have an input query the car symbolizes your input query and you say hey show me the information about politics and you have no idea how this parking lot is organized so you get an embedding you got like the A4 sign you get a vector showing you the way where you have to go to where to locate your data and this arrow and you say hey I know I know firm's cool error they're not two-dimensional ecladium plane yes I know there's an X and A Y chord in it it indicates here the direction of this indicator I'm surprised of you because you see if you take this symbol now and we have here our input string that we put into our Black Box this is our blue car and out comes and now you know what is an embedding and embedding is a vector representation in a this particular created Vector space and this Vector space was created here by our black box so this black box took more or less here all the Digital Data and ordered it in a particular way so it knows exactly if you have a question hey show me data about I don't know about Picasso it knows exactly where Picasso is and it tells you hey you're the machine code of your question the embedding of your input data is X Y set beautiful now you see here that there I I didn't put any symbol here in the black box but to be a little bit provocative LMS and this black box is an llm like chat GPT for example 50 of the work people invested in them is to clean data because imagine in this data chest here in this treasury chest here we have all Facebook comments and you can imagine that some people on Facebook are emotional some people have some very polarizing speech some people are very emotional maybe write something they do not really mean but there are some words you do not want to have or communicate to other people so here you have a huge filter a filter plane here that's going on just to absorb everything that is connected if you want to call it hate speech to get rid of this and it is amazing the resources and the capacity in the manner and the compute time that run into the system just to say it clean the data from some human polarizing speech without this my goodness life would be so much easier so just for you to imagine this huge filter is not indicated at all I just wanted to show you now that you have embedding and you understand what are embeddings indicator to a particular region in Vector space where the system placed all the information that you ask for in your input string now you have to path to go there and you see there are some other system representation you choose your floating points and you say okay what are those figures now imagine you have a two-dimensional car park somewhere on the field and then you go to the city and you have a 3D multi-story car park this is more or less you just add layers Dimensions to the model three-dimensional four-dimensional five dimension you can color code I don't know the right side it's red on the blue side and the left side is blue so you have further distinction here that you can incorporate in a particular Vector representation a tensor so you see easy you go higher dimensional embeddings and now you might say hey great now I know what a model does I know what embeddings are and I tell you no you have you have an idea now but only in theory why first if you were a client from one of those companies you do not know what's in this hidden away and closed up Treasure Chest you have no idea what are the original data that the system was trained on and the system can just do something where it has the data if you are in a particular topic where there were no documents or no data files available and is not represented here in the training data set for the model you will get a very a very particular answer because it does not have segmented the knowledge comparted the knowledge in a way that will give you an answer that is satisfying you do not know the model if you do not know this black box how this black box operates what kind of Technology this black box uses how much man machine human interface bias whatever is generated because of the architecture and the quality of the input data you do not know nothing about the system and as I told you you do not know the relevance for your input since you neither know the input data nor the processing unit whatever comes out of this system if it is relevant or not I can not give you a a trustworthiness index an internship between zero and one and you see this is where other systems come in play and if you ask me can I build my own GPT the answer is no because you do not have the resources of Microsoft azure but you can do something more clever a smarter system look here this is my two-dimensional open error cardboard and now you see here there here almost at the Horizon we have a specific topic and this is quantum field Theory this is something in physics or mathematics or whatever you have and because the car park has a limited size chat GPT has to run on a supercomputer from Microsoft Cloud this here and if you have 10 million topics of 100 million topics the signs for some outlying topics maybe there are just three cars that characterize a Quantum field Theory or three spaces here so if I have now a very dedicated input and I ask specific about this topic although the machine has 100 million other topics this topic it has only three car spot spaces available so you see where I go from something that is everything for everybody like chat TPT which is amazing but it needs a cloud supercomputer to something that is highly specialized like we humans somebody has a specialization somebody is gifted somebody is a painter somebody is a Creator somebody is you know what I mean so this system here now for example these three car places what I call Quantum field Theory and physics it is not anymore a large language model because it is not all circumventing afferencing it is a highly specialized Focus system on one or two topic only therefore it does not qualify as a llm we call this a neural search model if you want but you know what if I built this system only on one topic and I use the same car park here like before and now all cars here let's say we have 10 000 cars here in our picture are all specific detailed data on a highly detailed level all about qwifty so I have now really A specialized system for all details for every aspect of this topic where in an llm I just had three core places here somewhere at the end of The Horizon now if I have a specialized system I can use all my place here that I have available on my course here for this particular topic and this is really a deep dive into this topic if you want to build such a system in this video I'll show you how to build it easy and if you want to know about the difference here because jet GPT you will not guess what GPD stands for and Bert they are more or less the whole is a Transformer infrastructure Transformer invented by Google and the encoder stack here on the left side is bird and this is what my specialized Focus system would be where I have some transfer learning where I can continue learning every month and this GPT like that GPT further development that is based on GPD you see this is here the decoder part of the Transformer and if you have a decode a stack of 24 or 36 layers or whatever how many layers you have Transformer blocks you have operational in GPT so you see this are more or less very similar architectures for your AI either you go with a system an llm system like GPT that you have a supercompute center run by Microsoft yes you have to pay Microsoft if you want to use it or you will have a very specialized system also based on a Transformer model not an llm that is everything to everybody that is highly specialized just for me this system I can run on my home PC yeah wait so to summarize those alternative system if you ask me can you build your own GPT no I mean if you are multi-billionaire yes you just buy a cloud compute supercomputer and then you can run it there but normal people normal companies normal universities no way this costs too much runs on a super computer center because it has to include almost a whole knowledge that we have it is every topic including for everybody it is a black box system we do not know what the companies who trained the system put in there how they trained it their parameter it's a black box it's a black box regarding the data content and it is a black box regarding the model architecture and as you've seen the data I showed you before it's hard to update because you have to run the whole system and if you say how to fine-tune GPT for example because you freeze 90 of the very early layers and you just update 10 of the decoder uh stack architecture yes I know it is a way currently if you fine-tune this but if you compare this to the other model this model is a highly delicated specialized model it's not an llm but it's a neural search model that operates something very close to an llm but it is absolutely focused on one topic like Quantum field Theory physics here for me it is a professional focused system I have more than I don't think 20 000 publication only scientific publication I don't have to clean the data for hate speech because I'd only take pre-reviewed journals it has a very narrow Focus but data from all over the world all r d Department I think for the last 10 years I have every publication in this particular topic published and I can run a neural search on this I have it in my home I have it behind several firewalls cyber security if you are a company and you have very sensitive data you do not want to have it in a cloud although clouds are very safe today but if you insist that you want to have it in-house in a dedicated entity within your company you have hid the advantage of this system but again this includes everything for everybody and it's a highly dedicated specialized professional focused system so this is an alternative to Jet GPT and chat is really really the correct term it just chats with you it says blah blah blah blah blah blah and and here this system here does not chat at all with you the output is purely professional hardcore data full stop no hey how are you do you want your output in a in a special form do you want whatever so Black Box system so all the viewers who asked me hey what about my embeddings what I do now with my embeddings you have to think a little bit different you need to perform a task on an llm we call it a downstream task like a search you are searching for some information Therefore your input query is transformed into an embedding and text Vector embeddings measure for example the closeness the similarity the relatedness my goodness of strings of text and of course you know we do this with the cosine similarity function in a defined Vector space where we have even maybe a metric so you search then for your specific information you have your input string and you get a result back from the Black Box and for example the results this the information ranked by relevance to your query string beautiful if you want to have very good documentation again here please check beta of myeye.com docs guides embedding's use cases I think they have an excellent literature where they show you what you get back and how to operate this now an llm because it's so huge and it includes everything for everybody it has to run on a supercomputer Center so what it does and this is really the task that it performs it is convert your input text to a vector embedding this takes time this takes compute time it Compares this Vector representation normally in the form of a tensor against all other Vector representation it has stored in the last run of the training run of the systems those huge data of all the Digital internet or main part of the internet has to has a data storage somewhere at the supercomputer Center either it's cloud or Azure or wherever you are so you have to pay for data storage and then the performance as I showed you the cosine similarity function if you combine your input against all the other input on the internet that has been digitalized and included into a vector representation of an llm my goodness this is a main part here you compare once one tensor to I don't know 100 billion tensors and then you have an output and you have to convert this output from code to human text and all this has to be executed in a supercomputer and damn it takes time it takes energy it takes a lot of resources CPU GPU TPU everything on the network everything you somebody paid for this to build the hall to build all the infrastructure and you use this and of course you have to pay for this so yeah if you have chat GPT then this output for example you can say hey give me the output a search result in a form a particular form give me a Japanese haiku or I don't know a French poem all those things cost compute time I mean Heaven's Sake and then of course yeah I see example can you give to me the results the form of a speech that Winston Churchill would have done ma yes this system can do this because let's call it the the neural interconnectivity or if you want a simple weighted template is available for this but yes on a supercomputer if you execute this you have to pay for this and up until now in the beta version or in the free version in the research version of well however you want to call this free version Microsoft paid for this because they provided you the infrastructure for free if it came from the marketing department of Microsoft or not is not a question but yeah it always did cost huge amount of money so have a look again open my.com API pricing they have very detailed information on pricing and they show you exactly what it costs if you have an input text if you compare this if you have done the storage if you have compute time if you have conversion if you have further they have very simple formulas you can really calculate the pricing for your specific task okay this was it short overview short Q and A for you I hope there was something in for you and if you have any other question do not hesitate to ask me until then I see you in my next video