How Does ChatGPT Work? (Explained Simply)

Key Takeaways

how does chat GPT or other large language models or llms actually work no hype no scams just AI llms explain the way I do everything else now before I get started llms rely on huge data sets to function effectively and our everyday online activities generate a lot of personal data and that's why I use Pia VPN I'm headed down to Virginia Beach in a couple hours to see this girl it's kind of a casual ual thing and that's why I always bring protection that's right Pia VPN so click the link in the description below we go to Pia vpn.com McBeth to get 83% off plus 4 months free look I'm old enough to remember the 1980s when if you're playing games and put the wrong 5 and a qu inch into that slot you get a virus that could spread to everybody else you played with so I'm very careful before I connect to some girls Hotpot especially if I'm not committed so think of P VPN as protection but for your computer P VPN creates a tunnel between your computer and the public internet that hackers and Rogue government agents can't penetrate in just a couple of clicks you can select from several different servers in multiple locations I just started that Lance Armstrong 30 for30 Show on Netflix and when I'm at this girl's house I don't want to deal with that Netflix household error so I'm going to put paa VPN on our Xbox and connect back up to Maryland that might solve the problem if you don't have paa VPN you are sticking your Wi-Fi antenna into God knows where so click the link in the description below or go to P vpn.com McBeth to get 83% off plus 4 months free now let's talk about chat GPT this video is probably going to get shared a lot so allow me to introduce myself my name is Ryan McBeth I'm a software engineer and director of integration for The texasa Med lab which uses AI powered drones to deliver whole blood to wounded soldiers without relying on GPS we actually don't use large language models in our software at the Texas AA Med lab mainly because the battlefield is a dynamic place llms like chat GPT aren't actually capable of reasoning or figuring out their way around problems but large language models are good at predicting the most likely token a word or a letter that comes next in a sequence and their answers are limited only to the training data that they've seen it's like they're in a box this means that large language models are constrained to the boundaries of what we know think of it like this um you've used autocomplete you know how you type in Google and you go flights too and in my case it might complete to Miami well there are a couple of ways you can get something like that to work the first way would be an alphabetical text list like let's say you live in Maryland like I do you type in the letter M and the software looks up a list of all US states and displays everything with M and then you hit the letter a and that list gets reduced to main Maryland to Massachusetts then you hit R and you just get Maryland now for a list of 50 states that's fine but what if you're Google and you want to autocomplete the term flights 2 around the world well it would make sense to autocomplete the most popular place that people are searching for flights now there might be also uh other factors like if a person's searching for skiing a lot you might autocomplete the word Utah after flights 2 but in this case I'm just going to give a basic example that uses training data to display a frequently typed destination so how does an llm know what to suggest well it all starts with a neural network think of it as a digital brain inspired by our own llms are trained on massive amounts of text Data learning patterns in context of language one key feature of llms is called Transformer architecture uses something called attention mechanisms to focus on different parts of input text understanding the context better but I'll get to that in a moment the first thing that happens when you type in an input prompt like flights 2 is that flights 2 is split into tokens now if you want to get technical a token is usually a word and it can be parts of words or sometimes even characters but usually it's just a word so now we have two tokens flights and two each token is then converted into something called a high-dimensional vector now don't get wrapped around the axle about the term High dimensional Vector we use vectors all the time like um let's say you're looking at a location on a map you're using a two-dimensional Vector right X and Y and in a three-dimensional Vector we know where to put the camera in a firstperson game you have x y and z a high-dimensional vector is essentially a list of numbers that represents data in a space face with many dimensions and in the context of language models each word or token is mapped to a high dimensional Vector this is also called embedding which I'm going to use going forward these embeddings capture semantic meanings and relationships between words based on the model's training so the token flight might actually have an embedding that looks like this now this example has 10 Dimensions but typically applications like Chad GP have 768 or 1,24 Dimensions this is a tradeoff between performance and efficiency and if that sounds like a heck of a lot of Dimensions language is incredibly complex but using a high number of Dimensions we can capture the nuances and relationships between words more effectively it also represents contextual information as in what kind of flight are we talking about are we talking about an airplane flight or are we talking about flight as in fighter flight like what an animal might do when you spook it are we talking about a flight as in a drink sampler other dimensions might represent syntactic Bridges is this a flight of PLS a flight of birds a flight of wine other embeddings also explore relationships with other words for example if I say haste makes you're probably going to say waste haste makes waste and I could guess that you were going to say waste because your brain has all this training data that you've been Gathering all your life you've heard people say haste makes waste plenty of times now is it possible that someone might say haste makes plutonium I mean plutonium is a word but it probably doesn't appear very much after the words haste makes even though it is a valid word to use in a sentence so if the prompt was being asked to make a rap song then maybe the next word might be based embraced taste paste uh face uh they all rhyme so those words might have higher values to appear so I'm sure you kind of get the idea here other dimensions cover things like sentiment Associated word order and so on then you enter another type of transform called context aware embeddings through something called a self attention mechanism this allows the model to weigh the importance of each token in context with other tokens for flights 2 the model will determine how much attention to pay to flights and how much attention to pay to two when predicting the next word this generates an output that is most applicable to The Prompt input so let's talk about this output in order to choose the next word the model can either pick the word with a highest probability directly and this is called greedy decoding or sample from a probability distribution to introduce some kind of variability and in this case the model predicts Miami as the next token and all this information is transferred again into text and printed on screen now remember how I said that llms aren't capable of actual reasoning that's because these vectors are coming from values that have been pre-trained in a case like chat GPT the values have been pre-trained on basically everything that has ever appeared on the internet in fact the acronym GPT means generative pre-trained Transformer and it's that pre-training part that makes this behavor the way we think an AI is supposed to behave if you trained your AI on samples of rap music in order to generate Rhymes that AI will probably be pretty good at generating Rhymes but it won't be so good at writing software code large language models like chat GPT really can't reason in fact I would be reluctant to even call them AIS at all llms are really more like Advanced databases let's say you have a database and you search all the time times that a user is logged into a website the database only knows about the data that it has in the database if the database ever got temporarily disconnected from the logging server it wouldn't be able to display the logins it didn't know about and it wouldn't be able to guess at them either this is important because people who are grifters or who don't actually know anything about AI claim that one day AI will be able to generalize the new data it hasn't seen yet as if a computer could generalize when you logged in to a website but people who say that AI will be generalizing things one day they can't ever seem to explain how it will be able to do that and if you want to get technical if the same prompt is used unless conditions change the answer will always be Miami large language models like chat GPT are incredibly sophisticated in how they process and generate text they use high-dimensional vectors and Transformer architectures to predict the next word in a sequence based on extensive pre-training data but remember these models are not capable of true reasoning or understanding they can't generalize beyond their training data that's in that box they're powerful tools but they can have their limitations always be mindful of what they can and cannot do and if someone says otherwise they're probably trying to sell you something all right look if you really want me to sell you something head on over to Bunker branding get a live laugh launch for a patriot missile shirt uh you can also grab one of my Ryan McBeth inaction figures from the knif hand company even comes with its own trading card with my stats on the back thank you guys so much for watching it's me Captain Bannon of the documentary Team Yankee when I'm not kicking Kami butts I'm wearing T-shirts from Ryan MCB Beth available at Bunker branding knife hands high Mars landmines Patriot and even my favorite the toe missile mush we want t-shirt too take a hike come so come on down to Bunker branding and take a stand for what's really important about America capitalism