Large Language Models explained briefly

[@MinuteEarth]

3B1B explains LLM with RLHF for the CHM!

[@3blue1brown]

I'll go ahead and update the title to "3LRC"

[@Four_Kay]

@@3blue1brown wow really love your content

[@daylight8296]

just got a research internship because of you, final question was “explain neural nets like i’m 5”, used your explanation from a video with number recognition. Interviewer was super impressed with how concise i was with my response, thanks for making me an expert now blue :)

[@50Steaks68]

Ok this is glorious

[@ricotaline]

That's amazing ahah, well done

[@lilposs98]

congrats! how did you answer tho

[@karimaboucham3694]

2 vids in 12 days we are truely blessed

[@Dohan06]

Ja!

[@georegio]

Truly

[@srivatsajoshi4028]

3 now

[@iskda1585]

indeed

[@Bortnm]

This should be the first video someone watches before watching all your other explainers. Its a great foundation.

[@FuZZbaLLbee]

Just the right amount of abstraction for a short video. Also nice that it contains a bit of the final training part where humans rate answers of the model, to tweak to model. I believe you where still going to make a video about that in your series

[@DiThi]

There's a fantastic video about that part (and about what happens when it goes wrong), it's "the true story of how GPT-2 became maximally lewd" by rational animations.

[@FuZZbaLLbee]

@@DiThijust watched the video and it is indeed a great story 😀 Hopefully something like this doesn’t happen when programming for robots

[@notohkae]

this video was very useful. i loved your deep learning series, but this was a really nice summary and i definitely think it should be at the start of the deep learning playlist, as an intro

[@hugoperhammer]

I have an old friend of mine working in the field of curating data for training AI models. These videos has made me understand a lot more of what his company does than anything he's ever said to me.

[@SendyTheEndless]

I know someone who used to curate cars at night.

[@robertthallium6883]

I saw someone on TV pass a 40 kurig bowel movement back in the 00s

[@cosmicwitness3390]

@@robertthallium6883 Beautiful

[@aquajosei5759]

I am currently studying AI, and it really excites me that blue is talking about this interesting things on Transformers.

[@jamesevans2507]

You mean a meticulously scripted explainer video by an expert, featuring sophisticated animations, was more clear than casual chitchat you have with your friends? You don't say.

[@floodingchen]

One suggestion I have for the Computer History Museum is that they really should bring more computers back to operating condition now that the Living Computer Museum is closed. I totally understand that some of the older ones are prohibitively expensive to bring back, but there are so many more recent computers that they can make interactive. What difference is there compared to a museum for hardware/industrial design without interfacing those computers?

[@DiThi]

Fun fact: if you keep asking the transformer to complete the AI assistant after it has ended its answer, it will then invent text from the user, impersonating them. At least that's the case with most models. It will be trying to predict what the user is going to say, the same way that the answer from the assistant is what the LLM "thinks" the assistant would say.

[@columbus8myhw]

It's as if the machine is playing the "character" of an assistant, rather than being an assistant. It's also why it's so hard for it to admit when it doesn't know something. It's playing improv, trying to perform the character of someone who _does_ know the answer.

[@technologicalMayhem]

I think it's quite interesting that ultimately something like ChatGPT appears to be a helpful assistant that can answer questions. But in reality its just a piece of software that is really good a completing incomplete conversations. And when then just happen to give it a conversation where the user asks something and now the response has to be filled in. But we present it as thought the response is the assistant even thought it's not really.

[@pixelforg]

Honestly, i dont even see AGI any time soon, how can it even come when the LLMs dont understand shit?

[@jogadorjnc]

@@columbus8myhw Ironically enough, that's how humans often work too, we all play the roles we think we are supposed to be playing

[@fieldrequired283]

​@@pixelforg It seems unlikely that a real AGI is going to be a large language model specifically, though it's possible (likely, even) that advancements in tech from that field would inform the design and function of such a hypothetical AGI.

[@HelloWorld-ve3if]

The computer history museum is one of my all time favorites just from the memories of going there with my grandma as a kid. Loved the simple explanation too!

[@BrianOxleyTexan]

I work directly with a client in this field. This is one of the best explanations I've heard.

[@unholycrusader69]

<a href="#" class="seekto" data-time="468">7:48</a> Never thought in a million years that I would ever see Kendrick Lamar in a 3B1B video

[@klungusxyz]

I bet it's up there as a demonstration of deepfake tech

[@zmaj12321]

Makes sense though, that's an AI music video

[@SubatomicPlanets]

This is great! I often find it hard to understand your videos because I'm so bad at math (I still watch them though...). But this video is perfect!

[@ntruter42]

Me: This is genuinely the best simple explanation I've seen for LLM's. I feel like I can already teach LLM's to someone else now. Re-worded with ChatGPT: This is by far the clearest and simplest explanation of LLMs I've come across. I feel confident enough to teach others about LLMs now.

[@jaimeortega8939]

Ace content. Have you ever considered doing a few on image generation? Would love to see those explainers from you

[@offYears]

For anyone who hasn't checked out the Computer History Museum youtube channel, I would highly recommend for valuable primary sources/interviews with pioneers in the field! I love Gary Starkweather's lecture on the 'Birth of Laser Printing'!

[@EggZu_]

<a href="#" class="seekto" data-time="260">4:20</a> it's incredible they've been working on large language models for over a hundred million years

[@MMARavid]

I truly love these ML/DL videos. I'm new to studying the field so these are so helpful to my understanding. Please continue making em!

[@five-toedslothbear4051]

Awesome! I’ve been trying to figure out how to help a friend watch your videos on machine learning and stuff like that without getting lost in the math. You do a great job of explaining both the concepts and the math, and I would say something like ignore the heavy math, look at the animations and listen to what Grant is saying about them. Then you made this video. I thank you very much.

[@myronalcock4716]

Very useful, excellent visualizations, even better than the usual high standard. More beginner friendly video's please. Thank you!

[@imthestein]

OMG, I love that museum and I took my kid there earlier this year! Their favorite section was the old videogames that you could play lol

[@Wraithfighter]

I think this is fine if all you're talking about is the strict "how this works" in a short, digestible format, which sounds like what the CHM was asking for (been there before, fully agreed, its a great museum), although emphasis of "GenAI doesn't actually know anything" might be warranted. Terms like "training" and, well, "Artificial Intelligence" can give the wrong impression that GenAI is actually intelligent, so contextualizing those terms is important. However, I think that the exhibit as a whole would be incomplete without talking about the controversies with the technology, ranging from the rampant theft of people's writing and art being used to train the algorithms, to the obscene power and water usage of the tech, to the problems of hallucinations caused when the AI tries to mix up its responses and creates some nonsense instead, and including how, because GenAI is largely trained off of mass internet scraping, people's sarcastic or nonsense shitposts are being interpreted as factual statements with entirely predictable results. GenAI is a fascinating piece of technology, but its not one without some very serious flaws, and those flaws do need to be addressed in a proper museum exhibit about the tech. And... yeah, I'm a bit concerned that the CHM, beholden to Silicon Valley for funding and support, might try to gloss over those bits in the effort of keeping their backers happy :(.

[@lugiagaurdien773]

This is entirely useful for beginners! I have many friends and family members who would genuinly benefit from watching this video.

[@3blue1brown]

That's one of my big aims. After putting out the videos on Transformers earlier this year, some viewers gave the feedback that when they shared the videos with friends in their lives curious to know more about the topic, sometimes those friends found the videos a bit heavy or confusing. Hopefully, this offers something more helpful to share.

[@mostinho7]

@@3blue1browncan we still have technical videos made for engineers pls

[@jmg9509]

So relaxing to listen to whilst learning something fascinating ❤️

[@marksteers3424]

Grant, this seems to be a good primer for those who want to understand the principles without having to be a degree level scientist. It might well pique their interest in related topics. I have enjoyed your videos for many years now. 🥳It implies but does not state that an LLM does not understand in the way that humans do. When you give the figures about 100 million years of billions of calculations per second, maybe put into context how far human behaviour is beyond what computers can do. Most humans would be able to complete a response in seconds.

[@rorolonglegs4594]

Perfect to show my friends and family!! Thankyou Grant

[@hrishikeshharitas2456]

These deep learning videos are awesome, although less popular, please cover the ever mystical Graph Neural Networks next.

[@PAiWExHD]

Great video. <a href="#" class="seekto" data-time="287">4:47</a> afaik RLHF does not use human annotation for reinforcement learning on the base model. Instead, Human Feedback is used to align a reward model for the RL process on the base model.

[@polygonz3113]

That last part about it's amazing performance is due to all of us. Human decisions, expressed in language, convey an emergent intelligence to this system.

[@walts555]

I've been studying this topic since 2023, and this is the best short video covering the most important aspects of LLM's without unnecessary jargon or misleading statements. Well done!

[@SpencerTwiddy]

<a href="#" class="seekto" data-time="267">4:27</a> - Since this feat only took on the order of a year, the trainers must actually be doing quintillions (10^18) of operations per second🤯

[@ShaharHarshuv]

The idea of attention is brilliant. It basically invents a langunage with little semantic ambivalency, so things "aren't lost in translation

[@jewpcabra666]

amazing video ! Will definitely be sharing with allll the people I talk to often who say unhinged stuff about llms

[@abhisheksoni9774]

Your animations are as always superb 👌👌👌

[@Visitor______________________v]

This shorter more compact video is certainly useful for a higher level understanding of the whole thing I still haven't watched your longer videos about the topic, because they're long, but hey, it's kind of my problem

[@josephlabs]

This was a great summary. Essentially Transformers help us encode words and their meaning into a high dimensional space. The crazy thing is that when scaled high enough it starts to understand the pattern of language and can decently accurate the next word in a sentence. I remember doing a project in college about summarizing newspapers and now it's everywhere, pretty crazy.

[@maurosobreira8695]

Grant: The intersection of an artist and a Mathematician!

[@kyledmorgan]

This deserves a thumbs up!

[@aleksisofevil]

“a large language model is a sophisticated mathematical function that predicts what word comes next for any piece of text” - well phrased sir, that’s a very satisfying definition for a novice like me.

[@jmg9509]

If someone were to dumb it down even further: “it’s a glorified a autocomplete”.

[@duarpeto]

It is a good simple definition, be careful though, it only works if you're aware it is a simplification. For example, there are a lot of people who read that and think "it's just like the autocomplete on my phone keyboard, but a bit better" when that's very wrong. It is so complex we don't understand how it works, but this mathematical function presumably computes the meaning and context of what was said so far as well as the direction of where the conversation is going, refining this information through its learned knowledge of the world into a representation from which we can extract the most appropriate next characters for the response. There is nothing simple about this and we don't know what the limits are for what different versions of it can do.

[@amadif1793]

Always a good day when 3blue1brown uploads

[@trchri]

I like how you showed the “bank vs bank” pictures and then used that to show what the model decided to use.

[@darrennew8211]

As a 60+ YO computer programmer, I highly recommend visiting the museum.

[@chrisalex82]

bro is on the grind 🙏

[@FranciscoPires-y4g]

"[...] what word comes next, for any piece of text" - going to steal that one :)

[@BjornMoren]

I think that in the first half of the video you explain things on a level that is very easy to understand. The second half only someone fairly familiar with LLMs will understand and get something out of. So if this was aimed at the general public, I would have excluded some stuff from the second half and instead fleshed out just a few key ideas in more detail and using simple language.

[@doyletheCat]

I wish there was similar LLM appendix video, but that explained things with a little MORE formality. I love the intuition of these videos, but it’s hard for me to transfer my knowledge when I actually try to make my own LLMs

[@blobman_dev]

PLS a video about the probability of getting the most liked comment (forgot to say thanks, I love all your videos and lot of my friends does my favorite ones are those about AI and also the one about solving the gaussian interval, pls continue to do more)

[@dino-tech]

That would be hilarious

[@wertzuio2127]

You were in Munich some time ago and I heard you hung out with some people from WARR. Was never that jealous before! 😂

[@xassix]

Can we have a video explaining how hallucinations happen? As fascinating as the topic of LLMs is I think it is also very important to understand the limitations of this technology.

[@BooleanDisorder]

I would love to see your take on the manifold hypothesis.

[@Spatakioutis]

never clicked on a video so fast

[@gakpo_era]

You know it's gonna be a beautiful day when Grant drops us a video out of nowhere

[@MiScusi69]

Excellent video!

[@helmutalexanderrubiowilson6835]

The GOAT with two new videos.... i will need some cofee.

[@r.s.e.9846]

🙂Nice! (instruct-GPT, reward model)

[@shashwatbhatnagar659]

Got much awaited video ❤

[@asilverman1949]

You’re my favorite teacher

[@r0cky_2010]

It’s always a great day when 3Blue1Brown uploads!

[@Inspirator_AG112]

*Suggestion...* Make this video, but for image recognition.

[@yoram_snir]

Very nice. Definitely giving a lot of credit to Google in this one.

[@roncarlson8535]

Might be useful to explain what a vector is for the layman audience at CHM.

[@pier-alexbergeron5960]

Excellent video❤❤

[@Nobody-df4is]

Ooohhh. I had no idea. This is amazing. Thanks for the vid. (I probably sound like a bot. But I can assure you I'm not. I'm just a bit regarded but fascinated by LLMs. Haha. It is amazing to me how many people react to these chatbots on a emotional level when you realize what they actually are.)

[@harambe_andy5134]

This is awesome!

[@mithunraaj4097]

I love these videossss, More of these pleasee

[@sagarjamwal6182]

Thankyou very much for the vid haha

[@nvnrmchl]

Wow this is my current job rn. I think it’s a beautiful process that totally won’t accelerate AI too quickly Do you think the legacy of AI will have a negative or super negative impact on human history?

[@bjorntorlarsson]

Reply from random guy here. I think stuff like ChatGPT are fantastically useful as they are already! Maybe it's a good thing if their capabilities don't accelerate much further. That the limitation is human generated digital data. Huge benefits from that already. Might take a generation to implement them.

[@murirokcs5518]

AMAZING. I have a question about System Prompts and user prompts ( a common parameter for making (agents)) How does it affect the LLM? What makes some part of the prompt more important than the other (system vs user prompt)

[@Vytautas4Xfiles]

Thank you!

[@tawabullas5058]

Thanks for this awesome video.

[@YMandarin]

I think the real magic is encoding words, meaning and context as numbers in a vector

[@Sydra.]

We need more videos about AI architectures and cognitive science! MOAR!

[@jonclement]

Nice! Though I wouldn't say, "feed in all but the LAST word" and the model would predict...but "feed in all ..." let the model predict the NEXT word (Explicit LAST can be confusing because it doesn't just predict the last word of sentences). (I also got to wrap my head around 1 Billion computations per second...as a reference point ___).

[@JithuNair95]

I've never been this early to a 3B1B video! ❤

[@DominikJaniec]

nice. thank you!

[@KingGrio]

Hello ! I know this is maybe not within your scope but... could you do a short presentation on the maths of economics ? Particularly: could you underline the key assumptions made when we try to predict the economy and why predictions can be so wildly wrong and contradictory, depending on who you ask. And perhaps tell us how/why someone like Jim Simons got rich. My current understanding of economics is actually very low, but when I listen around me all I hear about is how economists are supposed to be savants, but then we always get surprised by a major crash, or things turn out to be the opposite of what was announced. What's the deal with that ?

[@mutalibgozalov7208]

If I'd need a team for a business or anything I would just visit your videos and contact commenter ) pretty sure brilliant people

[@hrishikeshharitas2456]

IMO would've loved this even more before the main series. It would've served as a nice breadth-first explanation of this concept to prevent a loss of context (pun intended) in the deeper videos

[@ib9rt]

There are lots of AI generated posts appearing on places like FB, but you can tell almost immediately they were not written by a real person. The text is excessively longwinded, flowery and repetitive, using two paragraphs to say what could be said in two sentences. So the next useful function for AI will be to implement a tool that recognizes and filters out all AI generated content, leaving only the content generated by real humans in the output. This will probably lead to an AI war, where LLMs compete to produce more human-like and less robotic output, so they can get past the AI-blocking filters.

[@DennisDavisEdu]

Wow, the museum piece at <a href="#" class="seekto" data-time="10">0:10</a> looks like the teletype machine I used in high school to learn BASIC. Now it's safely behind glass on display as an ancient relic. Wow.

[@facundosolar5356]

sharing this to my grandma bc she thinks chatgpt is skynet

[@waarschijn]

ChatGPT is not, but there are other AI that are more like Skynet (ChatGPT does have goals but it's too weak for that behavior to be noticeable except in a few famous cases)

[@DakshPuniadpga]

A great iniative in CS AI related education.

[@MRL200]

Thanks!

[@uvaisahmad2225]

Just perfect ❤

[@acarril]

Santiago (Chile) mentioned! Thank you for these videos Grant, they're an incredible resource.

[@niteeshjoyal9168]

Loved this

[@ZennExile]

So it's a very elaborate Copy Pasta Machine? I think I'm pickin up what yer puttin down.

[@columbus8myhw]

This is good but I thought the description of "feedforward" was unclear.

[@maxmyzer9172]

I am hoping for a Mamba video, someday

[@paultapping9510]

it just occured to me that that gpu shortage a few years ago that many attributed to cryptomining at the time, may in fact, have been due to companies like anthropic, openai and google.

[@baroianandrei7883]

Simple and awsome

[@mewsick1646]

amazing video

[@Vinzmannn]

And that's why more data won't make much of a difference anymore. You're just getting a better mean of all the sentences ever spoken.

[@soumyadipgoswami1721]

Hi grant❤ love your work

[@saipr96]

Now I know reasonably and intuitively how it works. All this time is was a magic for me😅

[@Merlin_The_Magician]

3Blue thank you much love from va

[@Benjo0900]

New 3b1b video :D

[@BuFu1O1]

my guy...build a physical typing machine that moves the keyboard keys when it communicates tokens back to you. A token should be mapped to the character so that it would be easier to move the keyboard's keys.. In this way, eveybody will understand that's some ghost inside the machine that writes back at you.

[@shenghaipeng]

are you making this video based on the book"Understanding Natural Language Processing on Semantic Web", the sentence examples in your video are very similar to this book.

[@_Solaal_]

It has to be the first time I see "1s ago" on a vid

[@bagool20]

lol same! I was excited that I could be the first oe to comment.