Show HN: Visual intuitive explanations of LLM concepts (LLM University)
36 comments
> We've just published a lot of original, visual, and intuitive explanations of concepts to introduce people to large language models.
Kinda frustrating that the main link dumps me onto what reads like a university syllabus, and nothing original, visual, or intuitive.
If I click through the sections in order, there are 5 "preamble" sections describing logistical and other meta-information about the course. All text.
The first pedagogical image I see this this, which tbh doesn't make any sense to me: https://files.readme.io/329efd5-image.png
"Where would you put the word apple?"
The image alone doesn't work without reading the supporting text very closely. I also have to have a pretty sophisticated understanding to get the idea that I can represent words as points in a plane.
Representing the words as icons is fundamentally confusing, too, I think. After all, maybe I say the word "apple" should go in "d" because it has at least two senses: a fruit and a machine.
Oh, sorry, you failed your first quiz!
"You can't fail the quiz, you're not being graded." Then why call it a quiz? Why use classroom metaphors unless you want students to fall back on classroom behaviors?
Of course, you know the #1 student classroom behavior: not reading the syllabus.
But if I have no trouble with that level of abstraction, what's with the cutesty way of describing the problem?
Get rid of all this chocolate-covered broccoli. Just say and show what you mean.
Computers like numbers. Vectors are lists of numbers. Vectors come with concepts like length and distance. We want to transform words into vectors so that words we think of as similar are close together as vectors.
There are many ways to translate words into vectors. Here are 5-10 examples of how we might do that. What are some pros/cons? What relationship(s) do they make clear or obscure?
Get them thinking about what it means to embed things and why we'd want to embed words one way vs. another. That'll pay dividends. Having them remember "where the apple icon goes" isn't going to be something they'll benefit from reflecting on in any future experience.
Kinda frustrating that the main link dumps me onto what reads like a university syllabus, and nothing original, visual, or intuitive.
If I click through the sections in order, there are 5 "preamble" sections describing logistical and other meta-information about the course. All text.
The first pedagogical image I see this this, which tbh doesn't make any sense to me: https://files.readme.io/329efd5-image.png
"Where would you put the word apple?"
The image alone doesn't work without reading the supporting text very closely. I also have to have a pretty sophisticated understanding to get the idea that I can represent words as points in a plane.
Representing the words as icons is fundamentally confusing, too, I think. After all, maybe I say the word "apple" should go in "d" because it has at least two senses: a fruit and a machine.
Oh, sorry, you failed your first quiz!
"You can't fail the quiz, you're not being graded." Then why call it a quiz? Why use classroom metaphors unless you want students to fall back on classroom behaviors?
Of course, you know the #1 student classroom behavior: not reading the syllabus.
But if I have no trouble with that level of abstraction, what's with the cutesty way of describing the problem?
Get rid of all this chocolate-covered broccoli. Just say and show what you mean.
Computers like numbers. Vectors are lists of numbers. Vectors come with concepts like length and distance. We want to transform words into vectors so that words we think of as similar are close together as vectors.
There are many ways to translate words into vectors. Here are 5-10 examples of how we might do that. What are some pros/cons? What relationship(s) do they make clear or obscure?
Get them thinking about what it means to embed things and why we'd want to embed words one way vs. another. That'll pay dividends. Having them remember "where the apple icon goes" isn't going to be something they'll benefit from reflecting on in any future experience.
I read and watched almost all the modules and for me it as it is, perfectly accomplishes the intention of the course as stated by the op.
Your suggestion may work for other intents (like having a Schaum's Outline of LLM's) and I would also love to have that additional material (maybe yourself could provide it as it seems you have a clear idea)
Your suggestion may work for other intents (like having a Schaum's Outline of LLM's) and I would also love to have that additional material (maybe yourself could provide it as it seems you have a clear idea)
I'm not proposing anything WRT the depth or scope of the material.
If the premise of the material is that phrases like "dot product" can be used freely or with minimal explanation then images like the "place the apple quiz" make even less sense. For that person, not much more needs to be said than "We want to represent words as vectors so we can do linear algebra with them. If the representations preserve structure we care about then here are some cool things that happen: (examples of good and bad embeddings)."
Then go deep, having given them an adequate advanced organizer.
If the premise of the material is that phrases like "dot product" can be used freely or with minimal explanation then images like the "place the apple quiz" make even less sense. For that person, not much more needs to be said than "We want to represent words as vectors so we can do linear algebra with them. If the representations preserve structure we care about then here are some cool things that happen: (examples of good and bad embeddings)."
Then go deep, having given them an adequate advanced organizer.
The landing page is technically the course overview. I'd love to hear what you think would've made it more engaging for you. We can probably pull up some of the visuals to it as a preview. Let me see what we can do on that front.
I know this course is meant to be content marketing for your company, and I don't mean this in a derogatory sense--you're doing content marketing right by providing high quality information to an audience who could be interested in your services--but it's a bit odd that going to https://llm.university unceremoniously drops me in the middle of what looks to be your product's documentation. That's not wrong necessarily, but it all adds up to arriving at a site and having a feeling of "what am I looking at?".
Like the grandparent comment mentioned, the pitch is "visual, intuitive explanations", but I don't see that on the landing page. I'm looking for a way to get to the start of your content, but the top and left hand menus don't help and are, if anything, confusing until I realize that I'm now inside of a larger set of documentation unrelated to the course.
Below the fold we see a "Let's get started!", but the link I see, Structure of the Course" doesn't sound like getting started. It sounds like more front matter. From the nav menu I see that after that I still won't get to the content, but instead a page about the instructors. Do I really need to read blurbs of the instructors before I get to the meat of the course?
It just feels like too much wrapping paper and packaging to get to the good stuff--and it really does seem like good stuff! And I think the way that you've embedded this course into the rest of your documentation prevents you from presenting it in a structure that is more familiar and easy to navigate (e.g. an 'About' link at the top that talks about the instructors and Cohere).
It might be frustrating to put a lot of time and effort into high quality materials, only for people to not want to spend a few minutes looking around, but from the audience perspective, there's a sea of LLM-related content out there. I want to quickly determine if this is worth adding to my already-too-long list of LLM related bookmarks of things I want to read.
Like the grandparent comment mentioned, the pitch is "visual, intuitive explanations", but I don't see that on the landing page. I'm looking for a way to get to the start of your content, but the top and left hand menus don't help and are, if anything, confusing until I realize that I'm now inside of a larger set of documentation unrelated to the course.
Below the fold we see a "Let's get started!", but the link I see, Structure of the Course" doesn't sound like getting started. It sounds like more front matter. From the nav menu I see that after that I still won't get to the content, but instead a page about the instructors. Do I really need to read blurbs of the instructors before I get to the meat of the course?
It just feels like too much wrapping paper and packaging to get to the good stuff--and it really does seem like good stuff! And I think the way that you've embedded this course into the rest of your documentation prevents you from presenting it in a structure that is more familiar and easy to navigate (e.g. an 'About' link at the top that talks about the instructors and Cohere).
It might be frustrating to put a lot of time and effort into high quality materials, only for people to not want to spend a few minutes looking around, but from the audience perspective, there's a sea of LLM-related content out there. I want to quickly determine if this is worth adding to my already-too-long list of LLM related bookmarks of things I want to read.
I understand that, of course.
Is your goal to make it feel like a typical university course or like something else?
If like a typical university course, start with a syllabus and a course description and all the logistics.
If like something else then the first 10 seconds of the experience should make people go "Oh, this is different."
What's happening in the first 1 second, 30 seconds, 1 minute, 10 minutes, etc. that are reflective of the rest of my experience? That will serve as an advanced organizer for what's to follow?
The very first graphic I see is labeled as a "quiz" and requires me to read a bunch of surrounding text to make sense of it.
That's the vibe: a promise of something visual and intuitive, first consummated by a long syllabus and a quiz.
Is your goal to make it feel like a typical university course or like something else?
If like a typical university course, start with a syllabus and a course description and all the logistics.
If like something else then the first 10 seconds of the experience should make people go "Oh, this is different."
What's happening in the first 1 second, 30 seconds, 1 minute, 10 minutes, etc. that are reflective of the rest of my experience? That will serve as an advanced organizer for what's to follow?
The very first graphic I see is labeled as a "quiz" and requires me to read a bunch of surrounding text to make sense of it.
That's the vibe: a promise of something visual and intuitive, first consummated by a long syllabus and a quiz.
The goal is to make the materials as accessible as possible. So we're definitely not limited to the structure of a typical university course and are happy to iterate on it.
I appreciate you elaborating on your feedback. Thank you.
I appreciate you elaborating on your feedback. Thank you.
The landing page should probably be more of a marketing page explaining why the course is worth checking out, with low information density and large visuals
start with a visual hors d'oeuvre
That’s just you. I find the apple thing obvious on first sight and cements an intuition that talk of vectors does not or differently. Why choose?
IDK, with only GP's comment and the image for context, and some earlier familiarity with the concept of embeddings, I'd put the apple... somewhere near the middle of the picture, nowhere near any of the lettered placement points. To me, the apple fits about equally to all the clusters, excepting bottom-left one, where I couldn't figure out the relationship in few seconds.
EDIT: in some sense, the whole idea and usefulness of embedding comes from it working like the inverse of this kind of "intelligence"/"logic" tests - tests that ask you to which, out of several groups, a new symbol belongs. Usually there are couple competing answers, but the test has you guess the one that's the Right One. Embedding is about subverting this - it's about telling the test giver, "you know what, it actually belongs to all of them", and adding enough dimensions to the problem space that you can have all the groups be far away, from each other, and the new thing close to all of them - to each along a different dimension.
EDIT: in some sense, the whole idea and usefulness of embedding comes from it working like the inverse of this kind of "intelligence"/"logic" tests - tests that ask you to which, out of several groups, a new symbol belongs. Usually there are couple competing answers, but the test has you guess the one that's the Right One. Embedding is about subverting this - it's about telling the test giver, "you know what, it actually belongs to all of them", and adding enough dimensions to the problem space that you can have all the groups be far away, from each other, and the new thing close to all of them - to each along a different dimension.
Jay, I liked your tutorial on Transformer models. Helped me a lot when I read it in 2020. One of the best resources on a topic then. Thanks for your work! Fingers crossed for your new endeavour.
Thank you so much (and others for your kind messages). Glad you found them useful! Writing is the best way for me to learn, I find.
This looks like a pretty great resource and I'm looking forward to checking it out. My only ask is that since it's the type of site I'd probably be looking at for quite a while it'd be nice if it had a dark mode.
You know what would be helpful? A little tag or something at the beginning of each section that says about how long it's going to take.
From what I've seen so far, it looks awesome. I'm excited to dive in. Thanks!
From what I've seen so far, it looks awesome. I'm excited to dive in. Thanks!
This is pretty excellent material, even just spending 10 minutes I have learned more than most random blog posts in the past few months.
I'm not sure how much is actually known to write about, but what I'd like to see explained is how transformer-based LLMs/AI really work - not at the mechanistic level of the architecture, but in terms of what they learn (some type of world model ? details, not hand waving!) and how do they utilize this when processing various types of input ?
What type of representations are being used internally in these models ? We've got token embeddings going in, and it seems like some type of semantic embeddings internally perhaps, but exactly what ? OTOH it's outputting words (tokens) with only a linear layer between the last transformer block and the softmax, so what does that say about the representations at that last transformer block ?
What type of representations are being used internally in these models ? We've got token embeddings going in, and it seems like some type of semantic embeddings internally perhaps, but exactly what ? OTOH it's outputting words (tokens) with only a linear layer between the last transformer block and the softmax, so what does that say about the representations at that last transformer block ?
This is a field I find fascinating. It's generally the research field of Machine Learning Interpretability. The BlackboxNLP workshop is one of the main places for investigating this and is a very popular academic workshop https://blackboxnlp.github.io/
One of the most interesting presentations in the last session of the workshop is this talk by David Bau titled "Direct Model Editing and Mechanistic Interpretability". David and his team locate exact information in the model, and edit it. So for example they edit the location of the Eiffel Tower to be in Rome. So whenever the model generates anything involving location (e.g., the view from the top of the tower), it actually describes Rome
Talk: https://www.youtube.com/watch?v=I1ELSZNFeHc
Paper: https://rome.baulab.info/
Follow-up work: https://memit.baulab.info/
There is also work on "Probing" the representation vectors inside the model and investigating what information is encoded at the various layers. One early Transformer Explainability paper (BERT Rediscovers the Classical NLP Pipeline https://arxiv.org/abs/1905.05950) found that "the model represents the steps of the traditional NLP pipeline in an interpretable and localizable way: POS tagging, parsing, NER, semantic roles, then coreference". Meaning that the representations in the earlier layers encode things like whether a token is a verb or noun, and later layers encode other, higher-level information. I've made an intro to these probing methods here: https://www.youtube.com/watch?v=HJn-OTNLnoE
A lot of applied work doesn't require interpretability and explainability at the moment, but I suspect the interest will continue to increase.
One of the most interesting presentations in the last session of the workshop is this talk by David Bau titled "Direct Model Editing and Mechanistic Interpretability". David and his team locate exact information in the model, and edit it. So for example they edit the location of the Eiffel Tower to be in Rome. So whenever the model generates anything involving location (e.g., the view from the top of the tower), it actually describes Rome
Talk: https://www.youtube.com/watch?v=I1ELSZNFeHc
Paper: https://rome.baulab.info/
Follow-up work: https://memit.baulab.info/
There is also work on "Probing" the representation vectors inside the model and investigating what information is encoded at the various layers. One early Transformer Explainability paper (BERT Rediscovers the Classical NLP Pipeline https://arxiv.org/abs/1905.05950) found that "the model represents the steps of the traditional NLP pipeline in an interpretable and localizable way: POS tagging, parsing, NER, semantic roles, then coreference". Meaning that the representations in the earlier layers encode things like whether a token is a verb or noun, and later layers encode other, higher-level information. I've made an intro to these probing methods here: https://www.youtube.com/watch?v=HJn-OTNLnoE
A lot of applied work doesn't require interpretability and explainability at the moment, but I suspect the interest will continue to increase.
Thanks, Jay!
I wasn't aware of that BERT explainability paper - will be reading it, and watching your video.
Are there any more recent Transformer Explainability papers that you would recommend - maybe ones that build on this and look at what's going on in later layers?
I wasn't aware of that BERT explainability paper - will be reading it, and watching your video.
Are there any more recent Transformer Explainability papers that you would recommend - maybe ones that build on this and look at what's going on in later layers?
Additional ones that come to mind now are:
Transformer Feed-Forward Layers Are Key-Value Memories https://arxiv.org/abs/2012.14913
The Dual Form of Neural Networks Revisited: Connecting Test Time Predictions to Training Patterns via Spotlights of Attention https://arxiv.org/abs/2202.05798
https://github.com/neelnanda-io/TransformerLens
Transformer Feed-Forward Layers Are Key-Value Memories https://arxiv.org/abs/2012.14913
The Dual Form of Neural Networks Revisited: Connecting Test Time Predictions to Training Patterns via Spotlights of Attention https://arxiv.org/abs/2202.05798
https://github.com/neelnanda-io/TransformerLens
Another piece of the puzzle seems to be transformer "induction heads" where attention heads in consecutive layers work together to provide a mechanism that is believed to be responsible for much of in-context learning. The idea is that earlier instances of a token pattern/sequence in the context are used to predict the continuation of a similar pattern later on.
In the most simple case this is a copying operation such that an early occurrence of AB predicts that a later A should be followed by B. In the more general case this becomes A'B' => AB which seems to be more of an analogy type relationship.
https://arxiv.org/abs/2209.11895
https://youtu.be/Vea4cfn6TOA
This is still only a low level mechanistic type of operation, but at least a glimpse into how transformers are operating at inference time.
In the most simple case this is a copying operation such that an early occurrence of AB predicts that a later A should be followed by B. In the more general case this becomes A'B' => AB which seems to be more of an analogy type relationship.
https://arxiv.org/abs/2209.11895
https://youtu.be/Vea4cfn6TOA
This is still only a low level mechanistic type of operation, but at least a glimpse into how transformers are operating at inference time.
That's great - thank you!
No one can tell you much about that. Interpretability is still very poor.
You don't know what they learn beforehand (else deep learning wouldn't be necessary) so you have to try and figure it out afterwards.
But artificial parameters aren't beholden to any sort of "explainabilty rule". No guarantee anything is wired in a way for humans to comprehend. And even if it was, you're looking at hundreds of billions of parameters potentially.
You don't know what they learn beforehand (else deep learning wouldn't be necessary) so you have to try and figure it out afterwards.
But artificial parameters aren't beholden to any sort of "explainabilty rule". No guarantee anything is wired in a way for humans to comprehend. And even if it was, you're looking at hundreds of billions of parameters potentially.
> not at the mechanistic level of the architecture, but in terms of what they learn (some type of world model ? details, not hand waving!)
https://imgs.xkcd.com/comics/tasks.png
https://imgs.xkcd.com/comics/tasks.png
Sure - but it's still the interesting part!
I'm sure some of key players know at least a little, but they don't seem inclined to share. In his Lex Fridman interview Sam Altam said something along the lines of "a LOT of knowledge went into designing GPT-4", and there's a time gap between GPT-3 (2020) and GPT-4 (2022) where it seems they spent a lot of time probably trying to understand it, among other things.
It seems the way values are looked up via query/key and added must constrain representations quite a bit, and comparing internal activations for closely related types of input might be one way to start to understand what's going on.
A high level understanding of what the model has learnt may be the last thing to fall, but understanding the internal representations would go a long way towards that.
I'm sure some of key players know at least a little, but they don't seem inclined to share. In his Lex Fridman interview Sam Altam said something along the lines of "a LOT of knowledge went into designing GPT-4", and there's a time gap between GPT-3 (2020) and GPT-4 (2022) where it seems they spent a lot of time probably trying to understand it, among other things.
It seems the way values are looked up via query/key and added must constrain representations quite a bit, and comparing internal activations for closely related types of input might be one way to start to understand what's going on.
A high level understanding of what the model has learnt may be the last thing to fall, but understanding the internal representations would go a long way towards that.
Are you saying no one really knows how these things work? I am very curious about if you can “peer into the weights”. I have seen simple examples of that with image recognition but only for early layers.
Hi Jay,
I really loved your [explainer on AI Art](https://www.youtube.com/watch?v=MXmacOUJUaw), and I've already added more of your videos and articles on my watch-later read-later lists! Can't wait to spend more time with them this weekend.
Thank you for creating such wonderful resources!
I really loved your [explainer on AI Art](https://www.youtube.com/watch?v=MXmacOUJUaw), and I've already added more of your videos and articles on my watch-later read-later lists! Can't wait to spend more time with them this weekend.
Thank you for creating such wonderful resources!
Love these.
I’ve also made some visual explanations for ml for Amazon, available at https://mlu-explain.github.io/
Big fan of your early work, Jay, a big inspiration for me!
I’ve also made some visual explanations for ml for Amazon, available at https://mlu-explain.github.io/
Big fan of your early work, Jay, a big inspiration for me!
That's beautiful! Hope you're getting to do more of these!
Thanks, am definitely trying to squeeze them out as I’m able!
You sir get an up vote for simply being Jay on HN. Thank you for all you do.
Nice work!
Minor nitpick: The intercom button obscures the topic expansion button for the final appendix in the nav menu. Maybe move intercom to the bottom right instead?
Minor nitpick: The intercom button obscures the topic expansion button for the final appendix in the nav menu. Maybe move intercom to the bottom right instead?
Looks great, thanks! It would be useful to add chapters indicators / links to jump directly to a specific news in the audio
Interesting, just yesterday I was googling something about transformers and had arrived on your page.
Looks great. Thank you.
Looks really helpful. Joined the community as well.
We've just published a lot of original, visual, and intuitive explanations of concepts to introduce people to large language models.
It's available for free with no sign-up needed and it includes text articles, some video explanations, and code examples/notebooks as well. And we're available to answer your questions in a dedicated Discord channel.
You can find it here: https://llm.university/
Having written https://jalammar.github.io/illustrated-transformer/, I've been thinking about these topics and how best to communicate them for half a decade. But this project is extra special to me because I got to collaborate on it with two of who I think of as some of the best ML educators out there. Luis Serrano of https://www.youtube.com/@SerranoAcademy and Meor Amer, author of "A Visual Introduction to Deep Learning" https://kdimensions.gumroad.com/l/visualdl
We're planning to roll out more content to it (let us know what concepts interest you). But as of now, it has the following structure (With some links for highlighted articles for you to audit):
---
Module 1: What are Large Language Models
- Text Embeddings (https://docs.cohere.com/docs/text-embeddings)
- Similarity between words and sentences (https://docs.cohere.com/docs/similarity-between-words-and-sentences)
- The attention mechanism
- Transformer models (https://docs.cohere.com/docs/transformer-models HN Discussion: https://news.ycombinator.com/item?id=35576918)
- Semantic search
---
Module 2: Text representation
- Classification models (https://docs.cohere.com/docs/classification-models)
- Classification Evaluation metrics (https://docs.cohere.com/docs/evaluation-metrics)
- Classification / Embedding API endpoints
- Semantic search
- Text clustering
- Topic modeling (goes over clustering Ask HN posts https://docs.cohere.com/docs/clustering-hacker-news-posts)
- Multilingual semantic search
- Multilingual sentiment analysis
---
Module 3: Text generation
- Prompt engineering (https://docs.cohere.com/docs/model-prompting)
- Use case ideation
- Chaining prompts
---
A lot of the content originates from common questions we get from users of the LLMs we serve at Cohere. So the focus is more on application of LLMs than theory or training LLMs.
Hope you enjoy it, open to all feedback and suggestions!