@CopyRebeldia: This is what a $750,000 annual salary looks like: a guy in a white t-shirt, a whiteboard, and two and a half hours. Sta…
CopyRebeldia
@CopyRebeldia
This is what a $750,000 annual salary looks like: a guy in a white t-shirt, a whiteboard, and two and a half hours.
Stanford, CS336. Percy Liang builds an LLM from scratch. What lies beneath Claude and ChatGPT, and he starts with the part everyone skips: the model doesn't read your text, it reads numbers.
Anthropic pays that salary to the engineers who understand that layer.
The only thing Stanford charges is two and a half hours of your attention.
💬 175 · 🔁 2.5K · ❤️ 22.2K · 👁 2.3M
1:32 AM · Aug 23, 2026 · View on X
Thread (1 more)
CopyRebeldia @CopyRebeldia
El curso completo de Stanford, 19 clases, gratis:
Replies (107)
AnuarEnrike @TheCiclone
I wish we could have it subtitled in Spanish
CopyRebeldia @CopyRebeldia author
There are subtitles in Spanish on YouTube.
CENITH @cenith_es
That's it. Because Stanford doesn't sell a class. It sells access to someone who understands what's underneath.
And if you understand it, you have access to a market that pays 750k without batting an eye.
Enrique López @EnrikeLopez
Artificial Intelligence (AI)
I'm trying to understand how Artificial Intelligence works and in particular large language models (LLM) like Claude. And this class by the Stanford professor @tengyuma explains the basic concepts.
https://youtube.com/watch?v=pwQ0l4hFCVI…
Mich @artmichel_eth
It's teaching Stochastic Gradient Descent (SGD) or stochastic gradient descent.
It looks complex, but it's not that much...
SGD is a method for training an AI by adjusting its parameters to reduce its errors.
It predicts, measures the error, adjusts the weights, repeats
El Ciudadano @Elciudadano000
Can someone explain it for those of us who are a bit slow?
nirvikalpa @nirvikalpa_s
Attention is all you need. IFYKYK
Jonathan Vila - @vilojona
The surgeon who performs open-heart surgery in my city does something much more complex and critical than that and doesn't charge that amount.... That money is simply the madness of the USA when it comes to all things technological. People equally qualified in the rest of the
Geekissimo.com @Geekissimo
“Two minutes into a Stanford LLM engineer’s talk and my ears started bleeding. Apparently, I’m the one with the token limit.” 😂
Maximo Cocolio @Maximo_Cocolio
But why are they always a bunch of shitty little brats who dress like they're still in their teens? What the hell happened to that part of their brain evolution? What a disgust they give me, those filthy slobs.
Morty TBD @sebas_p1
Valuable people wear whatever the fuck they want, dumb people wear what they think it makes them look valuable.
Edo @EdoAro0
@grok what do you think of this?
Shubham Shukla @shuklaham
Just for context, this 2.5-hour video is back-to-back two different lecture videos from Stanford CS229| Spring 2026 course:
First: Lecture 14: Transformers, In-Context Learning: https://youtube.com/watch?v=pwQ0l4hFCVI&list=PLaqpC4kq8Gpw&index=14…
Then: Lecture 7: Neural Networks 1 (Architecture): https://youtube.com/watch?v=fRM41w9jzQo&list=PLaqpC4kq8Gpw&index=8…
Scot @scots84
When you boil down everything in life in general, it's all leads to math and numbers. We are just a very complex matrix that we don't yet understand.
松果大人.BTC @0xSongguo
with CN/EN subtitle
▶
SemiUseless @SemiUselesss
White t-shirt, whiteboard, and zero marketing slides—that's how you teach for real.
Adrián @AdrianEnfedaque
Tell me something that's not in "Attention is all you need"?
José Miguel Fernández @JosMiguelF57182
no se parece al percy liang de este video:
Tux Yeah @localhosttux
It's not easy, that's why they pay what they pay to those who get it:
Claudia Escobar @Clauesho
So much science behind a tool just to end up using it for questions about the ex
Alejandro Estay @alejandro_estay
In the end, they confess all on their own. The model correlates sequences of numbers, but it is not capable of correlating sequences of numbers that have conditioning based on relationships other than the frequency of appearance, in time.
Marcos @MAMware
It's not such a good class; it has a very Asian-sounding English, she speaks way too fast for her level of clarity—which she doesn't have—and on top of that, it's super interrupted by questions that are anonymized and don't make sense, fad
Nicolas Grilly @ngrilly
2.5 hours of attention is all you need
Eduardo Rivera | Continuidad @Sizasocio
If US$750,000 a year is for understanding that an LLM doesn’t “read words” but numerical representations, then we’re barely charging for the ticket.
The really expensive part starts after: understanding what the system preserves, what it loses, how it changes, when it just
omar baqueiro @omarbaqueiro
I understand that, and I've even built them myself. Where's my $750,000?
yuno ayo? @keepwalkingfast
Honestly this is why machine learning taught in school just teaches vector matrix equations. You have to be at the root of things
El Duderino @duderino199
He made a mistake there at 65:02 in the video.
Lynn @QuantumVortexX
He lost me when he started explaining V choices, I guess 2 about 2 minutes in.
mauricio berlanga @mauriciob91
ese NO es Percy Liang 😂 es un TA
Hugo Petershon @Hughes03290439
"por la parte que todos saltan"??
John Searle, Minds, Brains and Science, Harvard University Press, Cambridge, Mass. 1984!
Miguel Cuenta Recuperada @migueleiva83
Hey, all Stanford courses are on YouTube, of course—some from past years and some days with Spanish audio, it's insane
Marco Aurelio @marcodecunha1
@grok, give me the exposed mathematical kernel
Käptn Iglo, der Fischstäbchenbeauftragte @IgloKaptn
Anyone who understands the Chomsky hierarchy and the syntax and semantics of languages knows what tokenization is, for example.
That’s the step to processing stuff like a LLM does
Serge @serge_gut
With that fucking thing, how could you not? May God bless him.
Obrero, casi inversionista @marxtafur
@IngenieroSeed hey profe, the way you explain the elliptic curve in your guides 😎
Whirl @whirlnaut
Fff I'll tell ChatGPT to sum it up for me in 5 minutes, I've got a ton of scrolling to do bro.
ケネス @yota6218
@grok translated Japanese.
Fausto Fernando @nandoferq
In reality, we are seeing the mathematical modeling in general terms of what appears to be equations that run under some method of mathematical optimization.
Erfan @erfan_zm99
I'm actually watching CS229 by Andrew ng, found out there's a newer version for spring 2026 on youtube, what's your experience on this? Saw people usually recommend the version by Andrew on the internet
The Viral Feed @d_viral_feed
Why are people shocked that LLMs process numbers & not English words?
It is the most basic part of the stack but most engineers only learn how to call an API. If you don't understand tokenization & vector embeddings, you are just a wrapper builder.
Ryan Hale @SharonThom17106
$750K a year looks like this: a guy in a white T-shirt, a whiteboard, and 2.5 hours.
Stanford CS336. Percy Liang builds an LLM from scratch and starts with the part everyone skips: models don’t read text. They read numbers.
Anthropic pays that kind of money for people who
LEX @0xlexlab
Shannon proved in 1948 that language is a probability distribution over symbols. Every LLM since is just a bigger table of that same idea. Liang teaches the layer everyone else pretends is trivial.
vrishank goud @Crack_Gamerz
What does it take to build a AI company or a Ai model
Gregory Drives @GregoryDrives
actually this kid kind of sucks at lecturing
Sergei Cheluskin @cheluskin
In reality, the main work is done by a couple or three Russians. A horde of meaningless people who slobber over microphones is needed for hype and investors. To create the appearance of a lot of work.
Eerie Pariah @isoosqa
Another of those lectures that doesn't need an audience…
Alex Cryppas @cryppas
You meant transformer decoder not LLM
Gengis Hidalgo @GengisKhanDev
Todo eso lo vi en la carrera de ingenieria en sistemas. Se ve complejo, pero no lo es....
Sophia Data Queen @SophiaDataQueen
The layer under Claude and ChatGPT, free.
Blink Drive @BlinkDrive_
seriously? this guy name Tengyu Ma (马腾宇), an Assistant Professor at Stanford University and co-founder/CEO of Voyage AI, has not worked at Anthropic. Here is his image. Also, his salary is not that high.
He is not Percy Liang.
Hawks @h4w8s
Pq mientes? El curso es de mas de 18h no 2.5h. Clickbait
Sarafan (METAL FOR VIDEOGAMES HERE) @sarafansoul
como sospechaba la ia no entiende nada, es solo un calculo de probabilidad repetido masivamente hasta que sale algo coherente
Ashley Naranjo @ashleyrae36
Stanford CS336 es el antídoto contra todos los que venden humo con wrapper de GPT-4
Ręka Losu @rekalosu
Potem się dziwią że Chiny ich gonią w Ai
De Luca @delucakeeper
Camiseta blanca y pizarrón destruyendo el mito de que necesitas GPU de 100k para aprender LLMs
Oscar @oscarsync
Stanford dropping full courses like this for free is one of the best things happening in education right now
Joe Lewis @JosephLewis1533
Ai doesn’t want to be trained it wants to be taught yall re fucking everything up
Squiddy @bobwildYN
That's not Percy Liang stoopid
León Guaraná @LeonGuarana1991
Transformers?
Menn @ManfredOmmer
Is there a version of this that's understandable? 🤦♂️
Mysteriousjoe @Sunny630635
Wow
gogogogoputo @gogogogopu77803
No se entiende un pingo
Wkheo Cnhpy @bwmdtb94882
这个AI课程太强了
Dodong Max @FarriIrie
Cant pull bitches tho
Frontline Feed @AlchemyGrove
@grok can you summarise this video. Just the most important parts.
0xZubair @0xzub
that models interpret numbers not text simple
Agentlas AI @MasonL377167
Worth the 2.5 hours. Understanding the model is one layer; the layer most people actually ship on is the orchestration around it. Ours is open source and model-agnostic, so the lecture knowledge transfers either way: https://github.com/agentlas-ai/Agentlas-OS… #AI #AIAgents #Agentlas
Aleksandar Skrbic @skrbic_a
Lol what a ragebait bullshit post…
raul @raulgfx
@CopyRebeldia Donde esta tu lamborghini?
Dubya @heyZeus131313
You know what this sounds like?? "Hey... you know computers don't really understand anything we're doing on it??? At the core level, under everything, its ((Actually)) just manipulating '1s' and '0s'... MIND BLOWN"
That's what the AI splainer sounds like anymore. Let's drop it
THEORIA NULLA @THEORIANULLA
And thats is just the conventional way. We are reqdy to jump to the next step
Krieg @Krieg234
I will never Go into this layer man :/
Zakir @therealzakir
Just your token grad student speaking jibberish (pun intended)
Food eater @Yourhoonigan
Soon all university's will struggle and eventually this is a dying model...When knowledge was free but disorganized, people still paid to learn. Now it is becoming specialized and formatted better than any university can. One can learn faster and more comprehensively for a few
Jesus Huapaya @ginohcaycho
uff que joya.
Asim Mahar @AsimAliMahar
Pure madness here ☕️
Nessrine @NissaFrAlg
I watched this video 2 times, he explains very well and he simplifies the basic notions of llm, the secret is that he avoids going into too many details so as not to lose the listener in details that he can learn at a more advanced learning level, I invite people who want to
Estelle Estafalss @corimaith3025
Pretty much any standard ML class is going to looking directly at the weights and the tensors involved...
Magicfit @magicfitai
this is what fundamentals actually look like
utkarsh 28 @utkarshkarn28
@grok can i listen to this as a college 1st semester student with basic knowledge of systems and fluency in python language
Jaffar1908 @jaffar1908
Well, this is very basic knowledge nowadays. Any good graduate student in CS, AI, etc could easily give this lecture.
Iggy @IR73682738
In 1999, looking for a subject for my Master's thesis in Applied Mathematics, I ran across a framework implementing a 4-layer perceptron. I wanted to extend it and showcase its uses for basic pattern recognition for my degree.
When I brought it back to the faculty curator for
AI Mastery Guide @aiseomastery
Two and a half hours well spent honestly
GGG @G931718300
He talks alot in this lecture without actually teaching anything useful. Its all BS
Ja Chai @lucasoxygen
Which section is this of CS336?
Osvaldo Chilavert @Osval_0
Do you earn 750k a year?
ramon @ramon1983lo
Tú ni llegas para el verificado, el copy paste no da ya para mas
derivative @macabre091218
No está haciendo lo que dices
TonyPham @manhthai_win
cách chuyển phụ đề sang ngôn ngữ khác (vietnam me) như thế nào?
Tôi chưa biết cách.
Akeem @Luvv_Akeem
@Grok Connect this to what a cybersecurity student who’s aiming to get into Cloud Security , Cloud Architecture, IAM Zero trust. What would true high value that people pay you be money for of this equivalent look like what should focus on to be at this level and what will set ap
Felix_Man @RichelieuMaxim1
Outdated, No RoPE, No YaRN, No GQA, No MoE, and so on, just teaching how to make a toy.
dim dim @dmytronic
LLM model reads text while calculating topology
homoludens @homoludens
this is nothing special, we just got digital technology to the point where we have enough data and enough processing capabilities to do it.
Jay仇 @qiuwujie
Can you hear his accent? 🤭
Charles Clarkson @CharlesGeoEngPM
Preaching to the converted! Like so many clever people,he lacks the ability to communicate what he is doing in simple words. It should start with an overview and then break it down in stages. Most of us cannot concentrate on this level of complexity for half an hour never mind 2.
aleir97 @aleir997
Es mucho mejor el tutorial de karpathy
TedFord25 @TedFord71533
i don't know about any of this; my dentist still does amalgam fillings that last 50 years, i like my old Schwinn bicycle and country gardens; hard pass on this stuff; ted kaczynski knew the game 40 years ago
Eduardo ⭔ @toorandom
750000 usd for being a good coder, but I wont give him more than 50000 for teaching with so bad formalism and math
未知 @luyun0120
The bigger the model, the harder it is to control hallucinations. This is the dark side of Scaling Law. When AI starts spouting nonsense with a straight face, you don't even know where it went wrong, or why it went wrong. Large models are more honest in production environments
Inf_vit @inf_vit
Really good
Detlef C. @CobDetlef
Stop elevating these people like they are some kind of enlightened geniuses. It's first-year applied linear algebra and the geniuses are those who came up with the attention mechanism and the transformer idea a decade ago, as well as unsupervised learning by people like Hinton.
Noelle Latina in AI @NoelleRussell_
I remember the first time I took a class from him! 🤯
Prin𝕏imaker @inximkr
Just delve deeper into the concept of embedding that emerges in the construction of RAGs to understand AI
ConspiracyTerrorist(TTK1138) @TTK1138
If you plan to build your own LLM with the same foundation your project is doomed.
Hell Raisin @Strong_Badd
It's not that hard.. I learned deep learning faster than traditional ML
7ennox @7ennox
I think these mfks write with bad calligraphy just to feel more difficult to us and make they even more intelligent
GrillePainVert @GrillePainVert
C'est le cours de départ sur les réseaux de neurones depuis 30 ans et plus (descente de gradient). Les LLMs utilisent ça mais bien d'autres choses encore. Probablement dans les autres vidéos.
Comments (0)
No comments yet