How ChatGPT Works Technically | ChatGPT Architecture

Get a Free System Design PDF with 158 pages by subscribing to our weekly newsletter.: blog.bytebytego.com
Animation tools: Adobe Illustrator and After Effects.
Checkout our bestselling System Design Interview books:
Volume 1: amzn.to/3Ou7gkd
Volume 2: amzn.to/3HqGozy
The digital version of System Design Interview books: bit.ly/3mlDSk9
ABOUT US:
Covering topics and trends in large-scale system design, from the authors of the best-selling System Design Interview series.

[@sameerizaj5458]

Sir i'm being very honest your 5 to 7 minutes video always taught more than the week of study,these viusal are so helpfull and also i love the fact whatever word you use you first explain the meaning behind that make's thing more easy to understand

[@happyruimi]

You should probably drop the course if you learnt more in this 5 min video than you did in 4 weeks...

[@user-dh7ut3wh7w]

@@happyruimi People love learning superficial stuff that makes them feel smarter lol

[@cexploreful]

Love the way you animate the presentation! Everything went so smooth and clear!

[@ferdous_khalifa]

Thank you so much for creating this insightful KZbin video! Your hard work and dedication to providing quality content are truly appreciated. You’ve managed to convey the information in a clear and engaging way, making it easy for viewers like myself to understand and enjoy. Keep up the fantastic work, and I’m eagerly looking forward to your future videos!

[@satychary]

As always, CLEAR AS THE BLUE SKY - Alex, EVERY one of your explainers is pure gold, including this one! Thank you for making this.

[@jonosutcliffe1]

The best summary available. I needed a good summary for non engineers about how it works and I wish I'd found your explanation before creating my presentation. Very nice work, thanks!

[@betweenprojects]

I am amazed by the results. A great wordsmith/people pleaser but factually about as reliable as a first year undergraduate in a bar.

[@space_btwn_notes]

pretty much recreated the typical politician

[@OsamaAdel]

I turned to be more complicated than I first thought. However this is natural, since its capabilities are fascinating and growing from day to day. Very nice video, thank you a lot.

[@DevCoder]

Man you explained everything in a simple and clear way. Thank you very much for this helpful content

[@tommushrom5929]

Build it then.

[@DevCoder]

@@tommushrom5929I'll

[@chrismachabee3128]

This is a very good video. It is the best nuts and bolts video I have seen and you've about think I have even heard of and I've been listening to everything I can, so fine job, well done and thanks.

[@rubbish9231]

As a software engineer I learned how to become a good chef. I asked gpt for it's implementation and informed me well

[@sudzam]

Wonderful explanation! Just enough detail to explore sub-topics in detail.. Thanks.

[@MrKelaher]

Great video, helps to clear up some confusion by "going behind the curtain"

[@moneycrab]

Thank you for this! So clear and helpful

[@richardglady3009]

Thank you very much for this great explanation. As a non-computer expert, the information was well presented and very understandable. Thanks.

[@ramaati18]

Wow! The best animation in slides i have seen in any tutorials! Every word you utter during you presentation is golden and easy to learn easily and quickly!⭐️⭐️⭐️⭐️⭐️!

[@AbubakerMahmoudshangab]

What app to create these animated graphics

[@thydevdom]

A few months ago I was looking exactly for this kind of video. Thank you!!!

[@Tony-dp1rl]

Knew this was coming on your channel, great little intro to the topic for sharing with people, thanks.

[@MandarPophali-yy1yz]

This is great summary. Can you have quick summary on differences between BARD and chatGPT?

[@superpowerdragon]

can you do a video on how Stable Diffusion work technically?

[@jntaca]

I follow a lot of divulgation tech channels. You are the most clear and concise of all. Thanks for your work !

[@henriquedomingues3800]

Can you do a list of this channels? Thanks!

[@goneni1504]

Excellent. Concise but thorough and clever. Thanks a lot. Love the burger analogy.

[@arthurcolle2778]

Excellent video, covers all key concepts - RLHF, PPO, etc... and how they relate to the GPT models. Great content! +Subscribed

[@jeroenvermunt3372]

Those are concepts for reinforcement learning, I missed the key concepts of the LLM itself. I found out via other sources that the actual key concept of the LLM itself is multi-head attention

[@L_Christ_BR]

Estou impressionado com a tecnologia utilizada pelo ChatGPT-4! É fascinante como a Inteligência Artificial evoluiu nos últimos anos, permitindo que máquinas possam entender e responder às nossas perguntas de uma forma quase humana. No entanto, é importante lembrar que, assim como qualquer tecnologia, a IA também pode ser perigosa. A capacidade de aprender e evoluir rapidamente significa que as máquinas podem se tornar imprevisíveis e tomar decisões que podem ser prejudiciais para os seres humanos. É crucial que os desenvolvedores de IA levem em consideração a segurança e a ética em suas criações, para que possamos aproveitar os benefícios da tecnologia sem comprometer nossa segurança - Texto criado pela IA

[@philipsun4834]

to be more specific, the minute on 1:55, Token ID is the vector, instead of Token, right? as Token can consist of alpha-numeric value, but the vector is purely the numeric with digits only. that is why the vector database is coming as the storage of those vectors for searching and indexing.

[@deepflyball]

New sub here. Thanks for such a clear explanation. I really enjoy the way you explain things and the graphics you use. Looking forward to more videos. Happy New Year.

[@inber]

Great explanation, much appreciated. I learned a lot.

[@alexharvey9721]

Great video. Clear and broad while still short - that's a sub from me, thanks! Sometimes it seems like a lot of steps to fine tune the usefulness of statistical representations instead of just going to the effort of trying to understand cognition. I wonder how responsible it is. But I guess morality has never stood in the way of bucketfuls of cash! The irony is many of those who developed the tech are incredibly thoughtful and morally grounded, while those in control are employed for their moral flexibility.

[@rlewisf1]

Very useful and professional presentation. Thank you.

[@marciabrown5569]

I really enjoyed your analogy. Though it is still over my head I do think I understand a bit more.

[@debashishbramha]

Excellent analysis.Thanks from India.

[@rahuldinesh2840]

This is classic presenation. Thanks.

[@SubashiniBala]

Pretty insightful content !! On the side note, I liked that you have brought up an example with Indian prime minister

[@ldohlj1]

Thanks. Please do one for Stable Diffusion too

[@moiserwibutso4899]

Thanks for this amazing explanation. The visuals are so fantastic 👌

[@bobbyandthomas]

Thank you for explaing so nicely!

[@SanjaySingh-jo3py]

Voice to slide !! Perfect, Sir 🙏🙏🙏🙏🙏

[@KumarBodana]

Excellent video! Best explanation I've seen on ChatGPT

[@wernerkallmann9105]

Excellent explanation. Thank you very much.

[@tomtyiu]

wow, now I totally understand how it works technically. Thanks

[@rajneeshverma4026]

super simple explanation and very insightful.

[@jdrake411]

I found your review very useful, thank you!

[@nguyenhonghanh2043]

Thanks for the video, I find it very helpful.

[@pawandhailPWRIES]

Excellent tutorial, very insightful !

[@darissalzburger]

I rarely comment the videos, but I had to comment yours. Amazing explanation and visualization. Thanks

[@nithinbhandari3075]

Hi, like your video a lot. Thanks. Can you please create a video on computational power and cost for running GPT3? How many GPU are required to run one instance of GPT3? And How to dynamically scale server for GPT3? How many request per instance of GPT3 can handle request per second, request per minute, and request per hour?

[@imagai.store.2323]

been searching like this one, thank you so much sir

[@kennethroark917]

This is gold. Thank you!

[@FlareGunDebate]

Bitwise operation, if/elseif/else, and neural network are orders of complexity in flow control. ChatGPT is huge in terms of parameters but it isn't any more sophisticated than modern translation software which relies primarily on tons of user data.

[@divotiontogod]

right, most people fail to understand and see how chatgpt is not able to for example write a clear code, it generates always a buggy code, and one has to elaborate more on minuscules of requirement for it to change it answer to a more better one. plus its answers are not mathematically correct, ask a few complex math questions and you will know! AI can never beat humans, ill iterate, even the dumbest human!

[@williamtim5893]

The information age creates the data for the AI age

[@FlareGunDebate]

@@williamtim5893 yeah OpenAI is neither open source nor artificial intelligence

[@JBoy340a]

Very nice and clear explanation.

[@dionisos8152]

Congratulations, this video is very clear and useful.

[@micbab-vg2mu]

Great explanation - thank you

[@damonwu9658]

Thank you for sharing, great stuff, and langchain next?

[@hubstrangers3450]

could expand how the constraints of the moderation API is developed (What type of parameters and arguments are provided on the subject of desired), thank you

[@Larry21924]

This content is a guiding star, revealing new pathways of understanding. A book with parallel material touched me deeply. "The Art of Meaningful Relationships in the 21st Century" by Leo Flint

[@freekyninjacat8207]

Question: Where did you get the information about 06:30 (conversational prompt injection, primary prompt injection etc.) from? Is there a scientific paper on these topics?

[@Richieexpert-cg3dx]

Hello friend, I noticed that you want more info about how ChatGPT works, let's discuss more and you will know more.

[@IbrahimAkar]

Excellent Video - thank you. What software do you use to create your KZbin presentations?

[@VictorAntonioLive]

Exceptional video!

[@phil7121]

Sahn is the Michael Jordan of giving ELI5 explanations of complex things

[@santi044]

super well animated video

[@bitsnbytes4406]

Great video! Thank you very much.

[@yuli.kamakura]

Very exciting, easy to understand

[@renatoigmed]

Even though I don't understand English perfectly, I managed to get a good idea of how it works. it is much more enlightening than the lazy analogy that they just say "ChatGPT talks about everything but has no awareness of what the subject is about". But imagine if there was...

[@williambrackett7100]

Fantastic explanation!

[@dineshchandgr]

amazing explanation. thank you

[@lloydwyngard7233]

Great video, very informative thanks!

[@ganeshkameswaran7123]

Nice explanation, thanks!

[@TynOng]

Could you cover Vector databases for LLM next time please?

[@hamman_samuel]

I'd like to hear about that too

[@SantoshNair1]

7:00 - there is a mistake in the graphic at this mark, where the prompt should have been a finance statement/question, but instead it remains the roses poem from the previous example. Probably a copy paste error.

[@rajeshkumarsahu7605]

Have learnt a lot from your vlogs and posts. Thanks Alex!

[@diondewet8522]

Wow.....brilliant explanation

[@mehediazad1780]

Sir I'm being very honest your 5 to 7 minutes video always taught more than the week of study, these visual are so helpful and also I love the fact whatever word you use you first explain the meaning behind that make's thing more easy to understand(2)

[@user-jq2up3ft6f]

Great Job, Thank you!

[@romeobernabe6871]

Thank you!

[@hanimahdi7244]

Very nice work, thanks!

[@exxzxxe]

Very well done!

[@vasanssv]

Great content! could you tell how you create such good animations? What software or tools do you use? Thanks.

[@vasanssv]

Thanks for your response. Like I requested I will be grateful if you can mention the software you use to make your animations. Thank you 🙂

[@thanhlv2]

Great video. Keep doing your great work 👍👍👍👍👍

[@praveends06]

Beautify articulated.

[@pelimies1818]

What is typical energy consumption for one inquiry? For google search the consuption for one inquiry is around 24 hours usage of 2W led lamp. So if you make 10 inquiries per day, it uses around same energy than illumination needed for a small house.

[@BATMAN_1]

Thanks!

[@homersimpson1199]

At 1:30 he mentions the amount of parameter for the models. Does anyone know why these are bounded according to the llm used? And how do these numbers come up?

[@MrDanieluy]

Brillant resumee and presentation !

[@lauralexi5236]

Great video and explanations. Anyone can tell me what are the tools have used to create the animated flowcharts here? These are too good to see!

[@dr.imayavarambanmunuswamy808]

Many thanks Sir.

[@OneAndOnlyMe]

Great explanation. As a computer scientist GPT is really not that mind blowing when you know that behind the scenes it's a giant data model and algorithms. It's magic to many non-computer scientists, but we can distinguish science from magic.

[@herrbonk3635]

Really?? I'm a "computer scientist" of sorts myself (CPU logic designer, compiler writer) but doesn't get sh*t from this. The video does not really explain anything "technical" (despite what the title says), only hyper loose analogies that tells me nothing.

[@OneAndOnlyMe]

@@herrbonk3635 As a computer programmer of 40+ years, AI is not some great mystery to me; large data models and quite simple algorithms at the core of it, combined with the computer capacity we now have. The theory has been around since the 1940s.

[@herrbonk3635]

@@OneAndOnlyMe That wasn't the point. Neural networks are no mystery to me either. My complaint was that this guy doesn't explain anything in concrete terms. It was mainly other models in the 1940s and the better part of the 1900s though. More focus on recursive list processing (Lisp) and logic programming (like Prolog).

[@OneAndOnlyMe]

@@herrbonk3635 I don't think the point is to explain in concrete in terms, more to just explain concepts and principles.

[@messiisthebest]

It is all about attention mechanism

[@modolief]

Superb, as usual

[@williebrits6272]

Awesome, well done

[@onlinecreations7781]

Pretty same as phase of a compiler 😍🙏

[@bhushankshire2854]

amazing. what tools do you use to make these videos??

[@not_amanullah]

Thanks ❤

[@user-zl2iv9sq6b]

What kind of tools used in your presentation. How you did the animation?

[@sheryll9966]

Great content!

[@BCubDelta9]

min 6:33 Could you please provide the paper for the "Conversational prompt injection" term?. I am currently citing in-context learning and prompt learning but this one is new for me and seems helpful.

[@waynelast1685]

Good video except for the end trying to understand how a user prompt actually gets converted into a response. I understand the system generates some additional user-transparent prompts, but then what exactly happens?

[@VLM234]

@ByteByteGo, RM model is again GPT-3.5 or other LLMs used?

[@freekyninjacat8207]

Ffs, exactly what I just needed! :D

[@buntykumar7270]

This is awesome.

[@The_Kharski]

So about the basics - the neural network itself. 96 layers - does this mean that the model only process' sentences with less than 96 "words"? It doesn't sound very deep, but rather VERY vertical (lots of lexicogrophy, lots of options/parameters for each layer). Or maybe several words can be a shortcut to another context? Which could be another input/output pair? 'SpecialValue ie 004726' references another full input (sentence) and it's output.

[@alexc7175]

The number of layers more closely relates to the model's predictive capabilities. It can be thought of as follows: Increasing the size of the model (in this case, more layers) is like increasing the size of a brain. A very small brain does not have enough hardware to form complex thoughts. As we increase the size of the brain (more layers), we are able to retain more information about the world around us and form more complex thoughts. It is important to note that while there are minimum size requirements, simply scaling the size of the brain upwards yields diminishing returns and may even become detrimental to learning. The number of input words a model can refer back to is defined inside the individual layers. This is called the "context length" and it determines the maximum number of tokens (sub-word / a chunk of a word) from the input that a model can look back on. If you want to learn more about what happens inside the individual layers, you can search for "Transformers AI".

[@adityendrapratapsingh7999]

@ByteByteGo what tool do you use to animate your presentations?

[@mikopiko]

Damn, I wish you brought up Transformers which is the key to how GPT works under the hood!