But How Does ChatGPT Actually Work?
Вставка
- Опубліковано 16 чер 2024
- Check out my Linktree alternative / 'Link in Bio' for Bitcoiners: bitcoiner.bio
You’ll learn how ChatGPT works and this will provide many benefits, such as helping you to use the model more effectively, evaluate its outputs more critically, and staying informed about the latest developments in the field so you are better prepared to take advantage of new opportunities.
ChatGPT is a type of natural language processing model (NLP) known as a Generative Pretrained Transformer (GPT) developed by OpenAI. These are the two big terms we will focus on in this video. On top of that you will also get a base understanding of common Machine Learning techniques like supervised learning, and reinforcement learning, which were used to make ChatGPT as good as it is.
Use ChatGPT here: chat.openai.com/
My profile: bitcoiner.bio/tillmusshoff
Follow me on Twitter: / tillmusshoff
My Lightning Address: ⚡️till@getalby.com
My Discord server: / discord
Instagram: / tillmusshoff
Join my channel membership to support my work:
/ @tillmusshoff
My Camera: amzn.to/3YMo5wx
My Lens: amzn.to/3IgBC8y
My Microphone: amzn.to/3SdHdkC
My Lighting: amzn.to/3ELnof5
Video on my second channel about building a blog website with ChatGPT: • I've Let AI Code A Com...
Further Sources:
Scale AI, OpenAI's Greg Brockman: The Future of LLMs, Foundation & Generative Models (DALL·E 2 & GPT-3): • OpenAI's Greg Brockman...
I made a few mistakes in this video that I want to clarify. Thank you for pointing them out Jan Tuitman & BigBadWolf!
1) The 7 steps I initially talk about are an example of how natural language can be processed. I somehow directly linked that to the GPT Encoder. This is NOT what happens in ChatGPT. The GPT Encoder finds its own structure and patterns in the data it gets provided. And this makes sense, because otherwise the output would be as bad as the sentence "I be learn ChatGPT".
2) The breakthrough of the Transformer is that its sequence processing is highly parallelizable. I touch on that fact but give it way too little importance.
3) Before the model gets fine-tuned like I described, it is first trained by predicting removed words on unsupervised data.
I hope you can still enjoy the video. I believe there is still lots to take out of it.
" that its sequence processing is highly parallelizable" means, that it can process the 40 Sentences of my Text parallel?
Another thing I didn't quite unterstand is the reinforced learning. Your AI can learn shooting better in adapting the movement and realize that when moving faster, the kill count gets higher, for example. But How does that look like for chatgpt? It writes an Answer and then compares this answer to what? And what is the reward?
Nonethess a very good video! The first I found that really went into describing how chatGPT works. :)
@@aerobraking couldnt bother to check but they probably used back propagation (like its almost certain they did) for learning
and for the "parallelizable" part I think its because the tranformer read the text step by step with differents models so it can be scaled very easely to be able to read large text input and be run effeciently during the learning step and when its used afterward
yeah I was about to comment on the third point, you should probably pin your comment
otherwise great video 👌
@@aerobraking they first train a reward model based on human labeling that then is used to 'rate' the output during reinforced learning
Please make this ERRATA comment a standard procedure. Many youtubers are adopting this and it really helps reduce misinformation! Thanks for your hard work and sharing your thoughts and knowledge with us!
My first though is: now NPCs in videogames will seem way more human and real conversations will be possible. That means that, for example, really good detective mystery videogames are possible now
Well yes but no you still need batshit insane crazy calculation power to generate these kind of outputs so we are still far from it
I would rather not play such a game, because it should be ultra boring. Game designers put enormous effort to keep the balance that allows the player feel the challenge but not stuck. And to make it work the player should have very limited space of options.
With AI-powered NPC it would be too easy both go the wrong way and lose the ability to complete quest at all (as in real life, where lots of crimes remain unsolved, and detectives work remain useless) or complete it too fast and obvious by guessing the way the game designers haven't thought
@@user-oc6rn4px5t Yes, except if you tell the AI not to do those things. It's fine if you don't play, you don't have to.
Ai dungeon was a cool website that implemented this concept as a textadventure
@@ibozz9187 im using chatGPT as a textadventure rn and its amazing lol
Removing suffixes will only work in languages with simple grammatical structure like English. In some languages a single word can encode meaning, gender and time, which are kind of important.
I'm a native hebrew speaker, and beside it being a bit fancy at times and forgetting to put ה (the) at the start of words on occasion, it performed pretty well. Although, hebrew does have insanely consistent verb structure, so it might suck in a language like Russian.
The transformer architecture was originally designed for sequence to sequence transformation like language translation, therefore should be able to understand that other language have unique language features (like gendered words) if it was trained I’m a dataset from a different language
I suspect future models will instead break down words into multiple tokens, one that's the base words, and others which are modifiers to represent "this is past tense" and stuff like that.
Depending on what part of the processing is happening, these modifier tokens will likely be able to ignored sometimes and not others. Like when talking about current events, past and present tense are often interchangeable, but future tense means a prediction is being made, which is not a fact and shouldn't be treated as one.
I wonder if NLMs will struggle with the difference between the genders humans have and linguistic genders, which are not the same but use the same terms most of the time.
The fact that "I be learn chatgpt now" is several layers of efficiency higher than the original is super interesting to me
English language needs a new efficiency update
Thanks Till. One of the few videos that actually tackles the technology behing Chatbots.
Thanks for the video man, easy to follow and understand.
you are so underrated...gonna be huge.
Thanks a lot!
He probably just used chatgpt for the content. Jk
@@tillmusshoff loved your content.
It truly is informative and simple enough to understand.
My personal opinion is probably you need to work on your thumbnail.
@@rightdefensive3492 probably. All the information in the video is wrong as fuck if you know how things actually work. He’s confused and mixed up the details of how different types of NLP work, which seems to be the kind of mistake ChatGPT would make.
Since my last video was received so positively I immediately wanted to follow up on it. So here it is! I hope you enjoy the video! Consider subscribing if you want to see more of my future content on your feed :)
Are you a polymath already? You inspired me.
No, my knowledge is too shallow and I am a humble beginner in many things I pursue. Great to hear you‘ve been inspired! ❤️
I love you man , i never learned this much in my whole CS degree 😂
Hi Till, great content, however I believe that GPT is decoder-only model and doesn't use any Transformer encoder in its structure. On the other hand, BERT is the encoder-only model that uses a stack of Transformer encoder blocks
Great content and well presented mate. Looks like you are trained by some model as well 😅
😎 Very good explanation! Danke man!
I'm amazed how one person can be a game developer, ml specialist and a youtuber. Very inspiring
Thanks for the informative video.
It was easy to understand.
At 5:03 you mention that this specific model does not train itself anymore.
At the end of the video, you talk about how the AI learns on its own by trial and error.
So what is true?
Is ChatGPT evolving as time progresses or it’s performance will remain the same till it has been updated?
Both is true. The AI trains by trial and error (Reinforcement Learning) as part of the training phase. Once the model is used like we use it there is no further training though. So to improve it OpenAI trains a new model or takes the existing one and continues training with it. We don't see that on a day to day basis though. GPT-4 will be the next iteration and we don't know the release yet afaik.
Thanks Tim for the explanation. Makes sense now.
Thanks for the info this is a great intro course to GPT
Great video! I'm curious what tool you used to make the presentation part?
I edit everything in Final Cut Pro
Really insightful video. Well explained
The fewer words to be more efficient instantly made me think of that Office episode where Kevin thought he was a genius for talking in that manner.
Great content and very informative as well, And you could have given credit to Google for Transformers.
Thanks for your clear explanation
V interesting. I am learning Mandarin and reducing and simplifying the sentence is very similar to translating from Mandarin to English especially with regards to the verbs. It makes sense as Chinese characters function in the same way as tokens.
Chatgpt can actually communicate in different languages I tested it out
Ngl, "I be learn chatgpt now" sounds so much more efficient and cool than "I am learning ChatGPT now".
Well done! 👏 Thanks!
In case it helps in the future, it's not "play-garism", it's "play-jarism". Thanks for the informative video!
Helpful. I be learn pronunciation.
Grow up clown 🤡
Thanks for a great video!👏
Nice video! Is the code of your reinforcement learning game shared somewhere?
No, sorry. I don‘t even know if I have access to it anymore honestly. It‘s using Unity + ml-agents extension.
This video is very useful for revision. Thank you ❤️
Except this video is complete misinformation, I wish youtube didn’t remove dislikes
@@MrFramhow so?
Superb video!
This explanation is great for an overview.. But I can't help but feel that industry-talking-point leaves a little out when it comes to the complexity and intelligence we see these models display. As if the experts who came up with this way of describing things, specifically did so to aid in the "moving of the bar" on what constitutes "intelligence".. They made the job of skeptics easier, but that doesn't do anything to explain what we're actually seeing with our own eyes.
If we merely assume it's only behaving as a language-model of patterns that it recalls from, and predicts the next words.. Something about it's output just doesn't seem to line up with that.. I mean, it produces results that are seemingly more than the sum of it's parts! If you ask it a difficult word-based math or logic puzzle it cannot solve on it's own, then as a teacher might, describe a method the AI could use to figure out a solution.. It CAN figure it out, at least it's done so when I've tested it.. Even more interesting, if you give it a similar word problem, with different numbers, it will then use that method you taught it, to solve it. Sadly it won't remember that for future conversations, but it does seem to be capable of learning, WITHIN a conversation. And when I've gone back and forth with it on complex topics, it's answers prove to me that it DOES possess some level of understanding of them, it can even draw whole new conclusions based on novel information I've given it.. Despite it's mistakes, it can be corrected, and I think a LOT of people are still underestimating what this kind of AI is capable of. It's not perfect for everything yet, but give it time.
When learning, the model essentially tries to find the best set of parameters to get the best results with the largest variety of different inputs. In this way, it can generalize different types of recurring problems, such as learning to calculate instead of learning the results of a set of predetermined math problems, which allows it to find solutions to problems that it has never seen before during training. In this way, it can also learn to execute a set of instructions...
chatgpt doesn't learn for your inputs, the only thing it does that could resemble learning is that it takes into account current and previous prompts as well as its own responses (all in the current thread only) to generate new responses.
it is not that surprising that it can do so much, because it is just a big neural network, just like our brain.
i thik that in the third step in nlp (stop words) you need also to delete i and am
Amazing video, super high quality. Subscribed in an instant!
very nice, please make more videos thanks
Amazing video thank you
Elaborate on what platforms to implement like python java ect.
"I be learning chat GPT"
that's proper Jamaican for you right here
9:16 But how do you program a model to follow the reward model? What exactly is a “reward” and how do you get a model to “want” the +1 as opposed to not wanting it at all? And how do you program it to tell it that it needs to avoid defeat?
Just program it to get maximum rewards. It might be as simple as using a recursive function in which you send inputs(the weight of the particular input), and a map where weights are mapped to the actions. Now, it's a bit like the "maximum sum of node values" problem where the program iterates over a tree, except here, it is a set of inputs. The input could be the key and the decision it takes will be the value. But take this with a grain of salt- I do not know much about AI and ML technology; I just know how to code. So this might not exactly be how they do that.
@@aapp776 Interesting thanks! Yeah that whole field is so complex yet so fascinating.
Really well explained! Like your Videos a lot ^^
but where in what technology this data is store in what of type database example - if we turn on interact and turn of and turn on this will be the same data result ??
Sir bring Playlist on Artificial intelligence..
that will rise up your channel 💯%
Love the Unity game example. Kudos on the recent success of your content : )
What background music did you use?
What is the music in the background? I really like it
Is your thesis publicly available somewhere?
I actually don‘t know, it‘s also in german only and just Bachelor / undergraduate thesis. There are probably more interesting and prolific things to read 😂
Wish people would use the full terms rather than abbreviated version before it comes into common usage like GPT or NLP.
I did so in the beginning. Generative Pretrained Transformer is just not a very handy term. It‘s long, sounds complicated and doesn‘t give you a good idea what it is exactly.
is chatgdp developed by backend languages like python and java?
Please REMOVE the background music. annoying. distracting. shame because you explain it very well.
This was awesome! Content but also I am very impressed by your visuals! The text parts were beautiful. I am just starting my channel on AI. Any advice on how to obtain B roll footage? Like the dog clip and the cyber clips?
Thank you! I have a storyblocks subscription for that :)
Thanks for the quick reply! Will check it out
Anyone know what the music playing at 4:00 is?
How much data in TB has the model been trained on?
I asked this question ChatGPT. This is the answer:
"The chatGPT model is a variant of the GPT-3 language model, which was trained on a dataset of billions of web pages. However, the exact size of the dataset and the amount of data it represents in terabytes (TB) is not publicly disclosed by OpenAI.
GPT-3 is a large and powerful language model that uses machine learning techniques to generate human-like text. It was trained using a combination of supervised and unsupervised learning, which means it was fed a large dataset of text and learned to predict the next word in a sequence based on the context of the words that came before it.
As a result, GPT-3 is able to generate text that is highly coherent and resembles human writing, making it a useful tool for a wide range of natural language processing tasks."
At 9:30 you show a graph. Was wondering what the X and Y axis were???
The Renaissance(Rebirth) was a period where Europe left the dark ages, and where classical art and science reappeared. To call for a 2nd one, would to me require that we were in a dark age compared to an enlightened past. I dont think that is the case, and if so, you should expand upon why you think so? To me this is more of like another 'industrial revolution' or a 'golden age': In regards to the Renaissance Man, that was a very limited option back then, as you had to be both wealthy and extremely intelligent to succeed in more than a single field. Specialist have only become more dominant as globalization has progressed and as fields of science have diverted more from one another. I like your unique idea, but I think you need to expand more upon it, to make a case for it.
Good points. I think my main idea is that through the democratization of tools and education through the internet more generalists can be created. Yes specialism has increased through globalisation, but you can individually counter that trend if you want to. And I believe there are good reasons to do so. I chose the renaissance as the period, because the key inventions and their impact are comparable to today. Another golden age is an alternative way to put it to still get the message across for sure.
good explanation
Nice video, I actually was thinking about RL since their model has 17B params, no human can tweak those manually, it has to be done by the machine. Only thing remaining for them to remove the limitation of pre-trained. I think if it can be tweaked to learn on the go, its gonna be pretty sweet, maybe for chatGPT 2.0 which uses GPT 4
I asked it if I should use a needle valve for a reverse activated sludge stream. It told me that needle valves are very good in reverse activated sludge streams. That's a big no no.....
Did you submit feedback to it with why that is the case?
Im binging your videos. This is no4. I notice if I talk kindly to chat gpt it mirrors it too.
Where can i read your paper on proximal police learning?
Not available online and it‘s in german anyway.
Is it good to learn a code now?
It's one of those inventions that you know will be the basis of the future. Imagine this getting implanted to your brain..
Great video ! What’s the music please ?
Something from Chopin!? Dvořák!? Haydn!? 8:13
Nice video
I love technology, but I also realize that sometimes the journey getting somewhere is more important than the end result and I wonder what's going to happen to humans who have so much of their thinking being done by robots? Remember the original Time Machine? Books falling apart, people drowning and nobody caring, etc. What incentive are people going to have to learn a skill if everything they learn could be done by a machine in 30 seconds? At least by trying to master a skill you might end up building other skills and come out a better person. I was a troubled teen and getting into programming at 19 gave me somewhere productive to focus my energy. It helped me in various ways beyond just writing software.
Very interesting, I asked few topics that were barely documented on google and gave me a decent response, really interesting how well it scraped the internet for information
It does not scrape the internet for information but it's been trained on internet text.
@@StevenAkinyemi thats how the text is obtained, it scrapes the internet and arranges the answers, training on internet text would be too difficult since there are various question
Great overview. Subscribed!
I was about to search that
Perfect timing then 😁
I don't think that "about" is a wasted word there. "Learning" and "learning about" are very different.
So chat GPTs natural tongue is pirate speech?
Here be pirates
To be clear, ChatGPT does not follow these exact 7 steps. They are just an example for what NLP can contain. ChatGPT tries to find structure itself in the insane amount of data it got provided. I have not been clear enough on this point in the video unfortunately.
You've accomplished learning how to ride a bike until it feels natural [to fall on your face a few times]!
Learnt something today 💪🏼
Is google's Bard also based on a transformer?
Plot twist: this video was created using ChatGPT.
Excellent! Thank you! :)
I built a 'Link in Bio' - a Linktree alternative for Bitcoiners. Check it out here: bitcoiner.bio 🧡
Impressive video, what is your native language? There is still improvement possible with your English pronunciation. Keep up the good work! 😊
Oh for sure, I am German 😁
It's obvious from his name. 😉😃😃
thanks a lot
Can you talk about CHAT GPT - MODO DAN?
I'm glad I started using GPT-3 back in 2020, it hasn't changed much, and it's still 100 billion parameters as before.
Still, ChatGPT is impressive and i'm excited.
Yes ChatGPT makes it more accessible for most. I'm excited for GPT-4 😁
what's with the spanish guitar background music 🤔
More detail required on how the transformer works. At the moment your video suggests "and then something magic happens".
I agree. Looking back I don‘t like this video of mine anymore. It could have been much better. Might do an updated version at some point.
who knows even your script is written by ChatGPT.....
Interesting and helpful video. I'm doing a bachelor's degree in computer science right now. Did you do a bachelor's degree or also a master's degree?
Bachelors only 😁
That was an awesome overview. Is it possible for you to create a more technical video where you dive into the nitty gritty details of the implementation? How does ChatGPT even remember the context of the previous questions you asked the model? And are all the previous questions/answers part of the input to the encoder? So many things I’d love to know 😁
No I don‘t think so. There are people who have a way deeper understand of this. I am not suited for that task (yet).
"Remembering" works quite simply - the entire previous dialogue is fed to the neural network input along with the last message, so. Much more interesting is how they taught it to remember it's name (Assistant) and say that it's is just a language model, when you say that it's not
I don't think the previous questions are directly encoded but become a latent part of the weights in the net.
Wow you did an excellent job explaining! Love this video! please do more!
Great video!
Too be honest it's really great for writers you can ask it to make a script and ideas 💡!
Don't you think we have enough mediocre scripts and ideas already?
I mean I've seen it write songs and they are awful.
@@redmed10 A useful tool is to ask it what it would do with an idea, and then not do that. Basically, you use it as a DIRECTION for your idea, to see what someone else might do with your same idea. Which can be pretty helpful. I sometimes ask my friends how they might use an idea of mine if they wrote it, this is just an even faster way to do that.
Is this something that the criminals in the future could use to crack passwords and break into systems? I think it's crucial that the cybersecurity industry, immediately get as much information as they can about AI. AI could give a hacker, new ideas about attacks that are more lateral than any would come up with.
Would there be a day where we can get these models running in our local machines that bypasses the need for OpenAI to be the gate keepers?
Can you provide us an ultra-minimalistic Python version of ChatGPT for us as a single script and just as a function method instead of a class method? I sometimes can only learn by looking at whatever the working code. Thanks!
Derivative regurgitating the redundancies of others
Doesn't matter how much it makes sense, it will never make sense
Really well explained 🙌
Niiice ❤
Chatgpt can improve to a level where we will no need for human job for diagnose a disease, solving a murder ....that will be sad news for some people 😞
I asked chat gpt about scientific differences between white and black people and it refused to give me am answer - it did however give me long life lessons which I did not appreciate...
"It is not appropriate to make generalizations about people based on their race. There is no scientific basis for the belief that one racial group is superior or inferior to another. People of different races can have a wide range of physical and mental characteristics, and it is important to remember that individuals are not defined by their race.
It is also important to recognize that race is a social construct and does not have any inherent biological meaning. The concept of race was created to justify the exploitation and oppression of certain groups of people, and it has no scientific basis.
It is important to treat all people with respect and dignity, regardless of their race or ethnicity. Making assumptions about someone based on their race can lead to stereotypes and discrimination, and it is important to strive for a more inclusive and understanding society."
This makes me think that I cannot trust the answers as they are "bias" towards a particular way of thinking and does not give an accurate scientific answer - this will not make the world a better place if narrative is pushed above facts and truth .....
Can I get your thesis?
okay but what is a reward for chatgpt?
That's a cute dog
I don't comment often but you explained everything so simply and clearly. I've been working on an AI project with some similar ideas in mind.
That‘s fantastic to hear! Thanks a lot! And good luck with your project 💪
@@tillmusshoff Thanks :)
This feels more like a video about the general GTP architechture, not about the ChatGTP and how that differs from older models. Kinda missleading title :/
Chat GPT is fine tuned by human feedback
Why does it care about the reward?
I have such a hard time watching videos with technical explanations with music included in the background. Uggghhh!! The music at 4:25 is fine but before that it’s so distracting…