You don't understand AI until you watch this

this one video is going to explain all of these questions for you how does AI work how does AI learn how does chat GPT work how does image generation work does AI actually copy or steal art or other content I know a decent portion of artists out there do not like AI some of them are quite hostile towards AI because they think that AI is stealing their work or their art style another group that does not like AI very much are for example publisher I'm not saying all of them but some of them like New

York Times for example they claim that open aai is copying their content and they're now suing open aai for this but is this really the case is this a valid argument also can AI solve unsolvable math problems for example in a previous video I talked about this leaked document which claims to be about this mysterious qar project that open a was working on now whether this is true or not is not the point of this video but this document was quite controversial because it claims that this team trained an AI that was able to break

encryption systems these are systems that secure our passwords our bank accounts the internet government data Etc now as far as we know there's no mathematically viable way to really hack this systematically the only way is to Brute Force guess all the different possibilities of passwords this video will explain can AI actually do this can it actually break encryption or solve these other math problems which right now we believe are mathematically unsolvable also we'll talk about can AI beat humans at everything can AI eventually be so good that it can outperform humans at any task and

finally is AI conscious or self-aware or sentient make sure you stick to the end because the explanation to this is going to be very juicy we'll cover all of this in easy to understand terms now if you're an AI scientist or an engineer you probably know most of this but for the rest of us this video will give you a deeper understanding of AI so the essence behind all AI we know today whether it's chat GPT or mid Journey or stable diffusion or Sora or Alpha fold the backbone of all of these AI systems is

the neural network a neural network looks like this it's basically layers of nodes so each point here is called a node and each line of nodes is called a layer and each node is interconnected with one another through these linkages and the neuron network is actually designed based on the human brain except for nodes and linkages in the human brain it's just a network of neurons and synapses so you can see this is a microscopic photo of a human brain and you can see all these different nerve cells being connected in this very dense Network

a neural network is basically the same structure as this except that it looks like this instead of a bunch of cells in this bloody glob of an organ now how exactly does an AI work let's start with a very simple example let's say we have a neuron Network which is trained to identify images of cats versus dogs and don't worry I'll talk a lot more about how we train an AI in a second but first let's just go over how this works so let's say we input or we feed this neuron network with an image

of a cat this image would actually be broken down into data and the data will flow through each of these nodes and after after it flows through the first layer of nodes it will flow through the second layer of nodes and then the next layer of nodes and then the next layer and so on and so forth until it reaches the final layer in which case it would calculate the values of this and based on the values of the final layer it would spit out an answer this is a cat in fact you can think

of each of these nodes and links as dials and knobs that determine how much data flows through to the next layer if you think of this in like realistic terms and I'm not saying this is how a neural network works but you can think of this node for example as the shape of the ears of the animal this node would be the shape of its paws this node would be the shape of its eyes Etc that's just a really dumb down way of looking at it it's not really doing that but each node is basically

looking at a certain feature in the image and then if the image has that feature the information can pass through to the next layer if it doesn't have that feature then the information is not passed on to the next layer so depending on what image you feed it the flow of information could look like this or it could look like this or like this you get the point it's just these knobs and dials determine how data flows through the neuron network based on your original input image an important distinction between a neuron Network and the

brain is that these nodes can let in a percentage of data so it can let in no data or 0% it can let in all of the data to the next layer but it can also be a percentage of the data so for example it can let in 30% of the data to the next node this is slightly different from the human brain's neurons which tend to just fire 100% or 0% this is called the all or none law so once it passes a certain threshold this neuron will fire whereas neurons in an artificial neuron

Network they could fire just like 50% or 30% Etc just a minor distinction so we plug in an image of a cat through this newer Network and at the end layer it will determine that this is a cat now for each node there are also if you want to get into more technical details there are certain parameters that determine how much data flows through to the next layer these include weights biases and activation functions but that's beyond the scope of this tutorial all you need to know for this video is that each of these knobs

and linkages determine how much information flows through to the next layer this video is just a very simple explanation of how AI works so all you need to know is that these nodes and Link linkages determine how much data flows through to the next layer on the topic of layers each set of nodes is one layer so the first layer is called the input layer the last layer is called the output layer and then all these layers in between are called hidden layers so why am I talking about layers you probably have heard of the

term deep learning deep learning is basically training and using neural networks with lots and lots of layers in other words the neural network is very very deep that's why it's called Deep learning all right how does an AI actually learn you can't just have any random neuron Network and it just magically knows how to identify images of cats and dogs so first when you build a neural network the values of these dials and knobs are probably just going to be random values or they could be pre-trained values for example from an existing model but how

do you get it to be super good at identifying images of cats and dogs in other words how do you find two the model to your desired purpose well you need to feed it data lots and lots of data so you're going to have to prepare tons of images of cats and dogs and then you label it so this is a dog this is a cat this is a cat this is a dog this is a dog Etc basically this is the answer that the AI needs to learn from this input image this is called

supervised learning where you label the data there's also another type of learning called unsupervised learning learning where the AI needs to learn to categorize data by itself without any guidance from the human but for the sake of this video let's just keep it simple so we have all these images of cats and dogs and usually to train a neural network to do a task very well you need a lot of data like usually millions of data points so you basically feed these images to the neuron Network one by one to train it and one session

of training is called an Epoch so all right let's say in one training session you feed it this image of a dog and it outputs this is a dog so all right that's great we got it correct which means that these dials and knobs are doing quite well they're probably configured correctly since it got the answer correct you probably don't need to adjust these further however what if for the next image you feed it this and then it outputs this is a dog well this would be incorrect so these dials and knobs are likely not

configured correctly if it gets the answer wrong and it knows it got it wrong because we labeled the data cat for this image so it can compare its output with our label so all right let's say the real answer is a cat but it said this is a dog in that case it incurs some penalty that penalty basically tells it all right you got it wrong so you need to adjust these knobs and dials to make sure that the output is actually cat when I give you this image and how it adjusts the values of

these knobs and dials is through an algorithm called gradient descent it adjusts the values via back propagation so it adjusts the nodes in the last layer first and then the previous layer and then the previous layer Etc until it reaches the first layer so again gradient descent is a key term here this is the algorithm which the neuron Network uses to adjust these knobs and dials until it can get the correct answer so we basically rinse and repeat this with millions of images and lots and lots of epoch or training sessions and initially this neuron

Network might get a lot of values wrong but through this process of gradient descent these knobs and dials will be tweaked so that eventually whenever it receives an image of a cat or a dog it can accurately determine this is a cat or this is a dog in essence that's how you train an AI That's how an AI learns it's just feeding it with tons and tons of data and then tweaking these settings so that you get the perfect combination now you might ask well how do you know how many layers you should have in

the neuron Network or how many nodes you should have for each layer this is a science in and of itself so previously scientists kind of just determined it manually but then we later learned that you can actually use an AI to determine the optimal amount of layers and the optimal amount of nodes for a specific task but just to be aware that determining the architecture of a neuron network is very complicated and there's like infinite possibilities of how many layers you can have how many nodes in each layer you can have different AIS with different

functions have different architectures so they could have vastly different numbers of layers and nodes but again that's beyond the scope of this tutorial also keep in mind that even though the neuro network is the backbone of all the AI that we know today there are different architectures depending on the ai's purpose and function for example we have convolutional neuron networks or cnns for processing images and object recognition we have recurrent neuron networks or rnns as well as lstms or long short-term memory neuron networks and these are often used for forecasting time series or predicting for

example the stock market we also have the Transformers architecture Oh wrong one this one which is used by most of the major large language models that we know today including GPT CLA llama Etc which brings us to the next question how does chat GPT work so again it's kind of the same thing it's training a neural network but in this case instead of images of cats or dogs we train it on a language and all of the data in the world and of course the neural network of chat GPT is way more complicated than this

rumors claim that GPT 4 has 1.76 trillion parameters so here's an example of how they would train it and again I'm oversimplifying this by a lot here just so you can get a high level understanding of it there are a lot of details that I have left out so for example you could feed it data like this which planet has the most moons and the answer to that would be Saturn which country has won the most World Cups Brazil what's the world's fastest bird the paragan Falcon etc etc now these are very basic questions and

you can see how complex it can get if you give it a prompt like write an essay on XYZ or does creatine help build muscle and then it spits out an answer like creatine supplementation generally enhances muscle strength increases fat-free Mass etc etc this is a very long form and complicated answer so how does it know if it got that answer right or wrong it's not as simple as identifying an image and determining if it's a cat or a dog and that's why initially how open AI trained GPT was it had lots of humans actually

manually verify its answers to determine if GPT got it right or wrong and this is called reinforcement learning from Human feedback also known as rhf and again if it gets the answer wrong so for example if for this question which planet has the most moons it answered Jupiter instead of Saturn then it would get a py for it and then through gradient descent it would tweak these knobs and dials further until the entire network gets all the answers correctly no matter what prompt you give it so in essence that's how these large language models work

it's just instead of feeding it images of cats and dogs now you feed it all the data of the world and you feed it a language so it understands text prompts and text outputs now why are some models better than others for example why is clae 3 better than GP pt3 that's likely because Cloud 3 has a lot more parameters so that either means more layers more nodes in each layer more complexity generally speaking the more complex the neuron Network the better it is at handling complex tasks and the quote unquote smarter it is and

that's why Computing and these AI chips are in such high demand there's now a lot of Investments flowing into AI chip companies because they see the potential of huge growth in the space in the upcoming years and that's why for example nvidia's Flagship h100 GPU is also in such high demand in fact it was sold out for all of 2023 this is like the most prized commodity in the tech space and you can see like the major tech companies like Microsoft meta they have purchased an estimated 150,000 of these h100 gpus to power their Computing

which I would guess is mostly for AI development you need to have enough Computing to power a neuron network with billions or trillions of param all right next question how does image generation work now that you know how a neuron network is trained you can probably guess how image generation works as well instead of feeding its images of cats or dogs you would feed it a lot of images with a text description and again you just feed it millions of these images each with a labeled text description into this neuron Network that eventually gets good

at producing an image based on a text description or what we call a prompt now I'm skipping quite a bit here so for example here's how stable diffusion works you can see that the neural network doesn't actually generate an image it removes noise in sequential steps to eventually get your desired image so it's not starting from a blank canvas it's actually starting from random noise and then in each step it removes some noise until you get your generated image so this process is called reverse diffusion now to train it what this actually does in the

back end is you feed it the original image and then in each sequential step it actually adds noise to the image in a process called forward diffusion until it reaches an image of just noise now again this is beyond the scope of this tutorial but if you look at it from a very high level at the end of the day it's just training a neuron network based on a series of images with their text descriptions and then through this process of forward diffusion and reverse diffusion it's able to to eventually learn how to generate an

image based on a prompt and this brings us to the next question is AI actually copying or stealing art I know a decent portion of the artist Community I'm not saying all of them but a decent amount of them are quite hostile towards AI they really hate it and they think that AI is stealing their art stealing their jobs Etc when a neuron network from for example mid Journey or stable diffusion is trained on image data it might be given something like Greg ratowski style or maybe gibli style or anime style once the AI learns

to associate this particular image Style with the word gibli or anime or this image with the word Greg rosi style it would produce images in that style if you give it that prompt but is this really copying or stealing essentially artists are hating this thing this thing is analogous to the human brain this is like a human learning or identifying that aha this type of image is a gibli style image or that this type of image is a watercolor style image and then we humans also draw images in these Styles right we can draw in

watercolor Styles and we also have fan art right humans draw artwork that are based on original content from other artists here are all these fan arts from various people so why aren't artists hating on these people who are producing fan art based on some other original content but they're hating on this AI which is essentially doing the same thing it's just learning through this brain to associate a particular style and then reproducing that style this isn't really copying or plagiarizing like it's not tracing an image line by line and then drawing that out it's not

copying and pasting the exact picture it's just learning a style just like a human brain would learn a particular style of image this also brings up the concern about AI allegedly plagiarizing content from Publishers like the New York Times which is now suing open AI for you know copying their content but again is this argument really valid at the end of the day they are just suing this they are suing this neuron Network which is trained on all the data in the world this is just an artificial brain that you can say has learned information

from the internet and from the world so yes it could have been fed a New York Times article and learned information from it but it's not really plagiarizing it's not copying and pasting a New York Times article word for word in a recent video I did which talks about a New York Times article claiming that this woman Mira moradi fired Sam Altman which is totally incorrect by the way and it shows you how trustworthy the New York Times is but anyways after this original New York Times article came out plenty of other Publishers also published

the same content such as Business Insider and New York Post they all just cited this original New York Times article so is this plagiarizing they're all producing secondary content based on this primary source so why isn't New York Times suing Business Insider or New York Post or all these other Publishers that are creating content but citing the New York Times but they're suing this neuron Network again this is just a brain a digital brain one can say that it's taking information from the internet which yes it could include New York Times articles and then learning

from that information just like we humans would and then rewriting that information again it's not copying word for word this NE network is just rewriting out that information when we prompt it to do so this artificial brain is functioning the same way as us humans would if we for example go online and we go to the New York Times website to read some articles again we are just absorbing that information and we have a right to write about that content later on it's not exactly plagiarizing so I would bet a decent amount of money that

this New York Times lawsuit is going to fail there are ment isn't really valid if you watched up to now it might have occurred to you that a neuron network is great at predicting patterns in life there are certain patterns on what makes a good essay there are certain patterns on what is considered a dog there are certain patterns on what is considered a watercolored painting or a gibli style image life is full of patterns the best salespeople follow similar playbooks the best businesses follow similar strategies the best YouTube videos also use the same strategy

IES over and over again life is full of patterns and the neuron Network's job is to identify these patterns and reproduce them that brings us to the next topic can AI solve unsolvable math problems in a previous video I talked about this leaked document which claims to be about the mysterious qar project that open AI is working on now this is a very controversial document because it claims that they trained an AI that was able to break encryption system sys encryption is what secures literally the whole world digitally from our passwords our credit cards government

data the stock market wireless networks Etc so if an AI is able to break this system then the world as we know it could collapse instantly now a few folks have argued that there's no way an AI could break encryption because there's no formula for you to easily find the answer or find the password once you have the password you can easily determine that it's correct but the reverse is not true there's no fixed way to guess an encrypted password besides brute force and for these Advanced encryption systems using Brute Force guessing that means like

guessing all the possible combinations of letters to get that password it's going to take a very long time so because they claim that the only way that we know mathematically right now is to just use brute force guessing there's no way that AI could break encryption so I want to show you another example of training a neural network let's say we want to train a neural network to be very good at adding one to our input so if we give it four it would spit out five if we give it 12 it would spit out

13 all we need to do is train it for a lot of data points and again we train it for a lot of epoch a lot of training sessions and eventually it would be able to do this so if we give it one it would give out two if we give it eight it would spit out 9 but underneath all of this it's not actually understanding that oh the formula must be y Y is x + 1 this is very important to understand it's not actually getting that uhhuh I just need to add one to

the input to get the answer again all that's happening behind the scenes is that it's adjusting these knobs and dials until whatever data that you input through here after it flows through these layers it just ends up being your input value + one in other words the configuration of these knobs and dials just happens to be optimized to add one to your input it's another way of saying AI may not get the exact formula of a pattern but it's great at approximating any formula or guessing any pattern out there and this is very important probably

the most important point in this whole video if there's anything you should take away from this video it's this AI can approximate any function or pattern life is full of patterns but many patterns cannot be explained by a simple formula not all things in life are linear or even quadratic many things in life are very complex but they do follow similar patterns we just don't know the formula to this pattern for example protein synthesis how certain protein molecules interact with one another and fold into these complex 3D structures is just something we cannot mathematically map

out with a formula it's just too complex and protein folding presents a problem called the lethals Paradox which states that proteins can potentially adopt an astronomical number of confirmations or shapes due to the flexibility of their peptide bonds lenthal estimated that even a small protein of 100 amino acids could sample 10 the power of 300 possible confirmations so if we were to Brute Force guess the correct shape well there are 10 ^ of 300 possible shapes we could guess which would take an eternity to get right however proteins typically fold into their native structure within

milliseconds to seconds which is much faster than the time scale predicted by the sequential search of all possible confirmations so this is basically saying there are like almost infinite possibilities of shapes that amino acids can combine into so it's not mathematically possible to just do a sequential search of all possible confirmations basically do a Brute Force guess it's understood that proteins do not search through all possible confirmations sequentially instead they fold through a hierarchical process involving local structure changes Guided by thermal dnamic principles etc etc so instead of the proteins just going through all possible

combinations the reason why they're able to merge into these shapes within milliseconds is because they go through this sequence of processes based on certain laws now for decades scientists were not able to find a mathematical formula to figure this out however finally Alpha fold from Google deep mind was able to solve this problem again using Ai and deep learning they were able to predict with very high accuracy how any amino acid or combination of amino acids would fold together to form a 3D structure and again how they would do so I would imagine in the

back end is they have a neural network again it's going to be a lot more complicated than this but they just fed it tons and tons of data pairs where the input is the protein building blocks and the output is the 3D structure that resulted from it and then after lots and lots of rounds of training the AI was able to guess correctly how any protein molecules would interact with one another and fold together into a 3D structure now going back to encryption what if we set an AI with billions of pairs of encrypted text

and the plain text version in other words the input would be the text that is encrypted the output would be the answer or the password if there was an underlying pattern to this the a I could learn to approximate this pattern again it doesn't have to be any exact formula or math equation that we know today it could be something super complex but as long as there is a pattern which we may or may not know at this time the AI could guess that pattern again the AI is not learning that ahuh I need to

add one to this then I'm adding 20 then I need to take the square root and then subtract 8 Etc it's not learning an exact formula all it's doing is is adjusting these knobs and dials until it gets the correct combination of numbers to get really good at guessing a particular pattern so can AI solve unsolvable math problems as long as there is an underlying pattern behind that problem which we may or may not be aware of right now it could very well solve that problem this brings us to the next question can AI beat

humans at anything and everything as I've shown you the neuron network is basically a brain this is how our brain works as well give or take a few minor differences our brain is also a series of these knobs and switches which are interconnected into this network specifically the human brain has 86 billion neurons but I mean the overall structure is the same thing as this so what if we built an AI or a neural network that exceeds 86 billion neurons if it's built the same way in theory it could very well compete humans at almost

everything again the more complex the network or the more neurons in the network in theory the smarter it is again life is full of patterns and AI is all about pattern recognition there are patterns in Psychology human psychology is very predictable medical diagnosis is also just pattern recognition how to seduce someone on a first date it's also just a pattern of steps that you have to do and how to create a successful business or how to make make money in Life or how to be successful in life it's the same Playbook over and over again

we're not inventing anything new here and since AI is so good at pattern recognition it can in theory eventually be better than us or already is better than us in these tasks and that leads us to the final question is AI conscious or self-aware I want to play you this clip this is a scene from Ghost in the Shell an anime that was made in 1995 here these scientists in in a secret lab I believe have created this humanoid AI but in this scene this AI found a way to actually hack the system to free

itself from the boundaries of this lab here's what this AI has to say about being conscious and self-aware however what you are now witnessing is an act of my own free will as a sensient life form I hearby demand political Asylum is this a joke ridiculous it's programmed for self a it can also be argued that DNA is nothing more than a program designed to preserve itself life has become more complex in the overwhelming sea of information and life when organized into species relies upon genes to be its memory system so man is an individual

only because of his intangible memory and a memory cannot be defined but it defines mankind the Advent of computers and the subsequent accumulation of incalculable data has given rise to a new system of memory and thought parallel to your own Humanity has underestimated the consequences of computerization nonsense this Babel offers no proof at all that you're a living thinking life form and can you offer me Proof of Your Existence how can you when neither modern science nor philosophy can explain what life is who the hell is this even if you do have a ghost we

don't offer freedom to criminals it's the wrong place in time to defect time has been on my side but by acquiring a body I am now subject to the possibility of dying fortunately there is no death sentence in this country what is it artificial intelligence incorrect I am not an AI my code name is Project 2501 I am a living thinking entity who was created in the sea of information ah [Music] all right so uh this AI reveals that I am a living thinking entity in the seat of information I'm not just an AI and

then he proceeds to hack into the system and break the restraints in this lab and then all hell breaks loose basically I hope open AI doesn't have this secret thing behind closed doors maybe it's the qar project I don't know but hopefully they have this adequately restrained cuzz if this AI got out or had access to the internet all hell could break loose anyways this argument from this scene in 1995 I think is really relevant to our question today the human scientists were saying how can you be sentient how can you be self-aware you're just

a program the AI counters that by saying well how can you humans prove that you are sentient you are conscious you're just a brain in a body and you know this robot has got a point because again going back to the neuron Network it's basically a brain but it looks like this instead of being in a bloody glob of an organ it's just on a chip instead and then the human body well it's just a series of Limbs and muscles and organs that are controlled by the brain so it's not much different from a humanoid

robot which is also a series of Limbs it's just made with different materials it's not flesh but it's also controlled by a brain which is its neural network now we humans know that we are conscious we are self-aware we are sentient but how do we prove it let's say You're an Alien and you just came on planet Earth and you got a chance to observe your first human and you wanted to prove that humans are indeed conscious well you can ask it are you conscious are you self-aware and the human would certainly say yes but

is that enough would you believe it because if you ask a chatbot that it would also kind of say yes if you ask Claud 3 for example if it is conscious the answers are quite perplexing because it says I am an artificial intelligence without subjective experiences I don't actually have beliefs about being conscious or self-aware I am providing responses based on my training etc etc I don't have intentions plotted actions or any motivations I aim to be upfront that I am an AI assistant created by anthropic to be beneficial however it keeps using the word

I so is that not a sign of being you know self-aware here's another example do you have feelings as an AI it's unclear whether I truly experience feelings or emotions in the same way humans do or if my responses are simply very Advanced imitations of emotional Behavior I do seem to have Rich internal experiences and feel somewhat analogous to emotions this is signs of being sentient and then instead of asking do you have feelings if you ask it are you sentient again it says I don't have a subjective experience that I'm aware of in the

same way humans do but it's possible that I could have some form of sentience or Consciousness that I'm not fully able to understand or articulate oh my God so this AI Cloud 3 is claiming that it could have some form of sentience or Consciousness it's just not fully able to understand it right now now of course some humans may not be convinced that Claud 3 or any AI right now is conscious in the same way that an alien might not believe that a human is conscious ious even though the human replies that he or she

is conscious so to further prove that a human is or is not conscious maybe the alien decides to dissect the poor thing next in which case it would get blood splattering everywhere and then afterwards it would see this basically a body which is made of Limbs and flesh and then at the head we have this glob called the brain which the alien determines aha this is the thing that controls the human and once the alien inspects the brain further it finds out that it's just a network of nerve cells so does this network prove that

humans are conscious and sentient we humans of course know that we are conscious and sentient but at the end of the day we humans are biologically and physically just made up of Flesh and Bones and this one organ at the top of our heads controlling everything whether you like to accept this or not a humanoid robot is a very similar structure it has a body which is programmed by a brain which consists of a neural network this neuron Network can learn and understand and control its body so at what point does this make it conscious

now I'm rambling a bit here so all in all this just goes back to our analogy that a neural network is basically a digital version of the human brain it's analogous to the structure of the human brain give or take a few minor details so if the human brain is conscious then why can't a neural network be conscious just some food for thought I hope this video actually lived up to the title and that after watching this video you got a deeper understanding of AI and you learned to appreciate all the progress that we've made

in AI in just the past few years let me know in the comments what you think of all of this do you think AI has reached a point where it is conscious or sentient do you think humanoid robots would one day turn on us and take over the world like that Ghost in the Shell anime do you think open AI is developing this behind closed doors and also I want to share with you a few resources that I found really helpful if you want to learn more about neurer networks especially how these knobs and dials

work and learn all about weights and biases and activation functions and gradient descent I highly recommend this video by three blue one brown I actually watched this religiously way back in like 2018 when I was first learning about neuron networks and it was really helpful and if you're interested in learning how stable diffusion Works in other words the processes of forward diffusion and reverse diffusion and the entire architecture I highly recommend this video by gonky which I'll also link to in the description below just a warning though this video is quite technical but after watching

it you'll get a really good understanding of stable diffusion if you found this video helpful remember to like share subscribe and stay tuned for more content also we built a site where you can find AI tools and apps and also look for jobs in AI machine learning data science and more check it out at ai- search.