How LLMs Work
Learning and science
interactive explainer
SHOW HN2026
245points
55comments
#199of 4,366 in 2026
Top 5%of 2026 by points
How HN took it
RoastedHalf or more of the comments are critical
3 positive, 4 neutral or asking, 1 mixed, 22 critical, out of 30 comments about the project. Thread score -0.53.
Design: 0 for, 9 againstExecution: 1 for, 5 againstDocs: 1 for, 3 againstIdea: 0 for, 3 against
The thread, comment by comment
Each row is a top-level comment and its replies, in HN's order. Select a cell to open that comment on HN.
positive neutral or mixed critical about another commenter or something else the maker
What HN said
What people liked
Proper attribution to Karpathy's lecture
From 1 comment
Just want to give appreciation for proper attribution. I feel like still some people will say "Here's something I made" when the reality is, "Here's something I asked my AI to…
gslepak on HN
Well explained and useful
From 2 comments
I’ve had a look, and it’s very well explained! If you ever want to expand it, you could also add how embedded data is fed at the very final step for specific tasks, and how it…
learningToFly33 on HNThis is really well done. Curious how long it took to build something this interactive?
gn_central on HN
Objections
Content appears entirely AI-generated with little review
From 4 comments
> What does the input side of the neutral network look like? Is it enough bits to represent N tokens where N is the context size? Not quite. The raw text converted into IDs…
Udo on HNHave you reread what was produced by Claude Code before publishing ? This thing in one of the first paragraph jumps out: > you end up with about 44 terabytes — roughly what fits…
PetitPrince on HNAnother low effort, dark mode slopsite. You lost me at "44 terabytes" before I even got to the emdash in that sentence. @dang, when is the 'flag as slop' button coming?
arcza on HNyeah this seems entirely LLM-generated.
jblitzar on HN
Factual error about 44 TB fitting on a single hard drive
From 2 comments
Have you reread what was produced by Claude Code before publishing ? This thing in one of the first paragraph jumps out: > you end up with about 44 terabytes — roughly what fits…
PetitPrince on HNEven if those things are true, no reasonable person who knows what they're talking about, and who writes for an wider audience would say something like "you end up with about 44…
embedding-shape on HN
Generic AI-style CSS design
From 2 comments
I really dislike the default AI slop css - if you're going to do this - please have a design language and taste ideas beforehand. It can help so much in refining the look.…
endymion-light on HNI think the css here is what Claude spits out when you prompt it to make frontend
jblitzar on HN
Tokenization example is misleading
From 2 comments
The page does very poor job tokenizing phrase "Noinceolik fiyulnabmed fyvaproldge" into "Noinceolik fiyulnabm ed fyvaproldge", factoring only "ed" suffix. As if made up words…
thesz on HNI think that BPE visualization is slightly misleading, because it seems to imply that the "old" (smaller) tokens are thrown away and replaced with longer tokens, which is not the…
vova_hn2 on HN
Scroll jumping and layout overlap on iOS and narrow windows
From 2 comments
Page keeps annoyingly scroll-jumping a few pixels on iOS safari
lukeholder on HNLefthand labels (like Introduction) can overlap over main text content on the right in the central panel - may be able to trigger by reducing window width.
Barbing on HN
Requests
Explain embeddings and input handling more thoroughly
From 1 comment
I haven't found an explanation yet that answers a couple of seemingly basic questions about the LLMs: What does the input side of the neutral network look like? Is it enough bits…
gushogg-blake on HN
Cover how embeddings feed into final prediction tasks
From 1 comment
I’ve had a look, and it’s very well explained! If you ever want to expand it, you could also add how embedded data is fed at the very final step for specific tasks, and how it…
learningToFly33 on HN
Move tokenization to the end and start with byte-level model
From 1 comment
In these types of tutorial, can't we move tokenization to the end and just start with a model that inputs and outputs bytes (latin-1) instead? That would (1) take the reader to…
amelius on HN
Questions
Was this created with Claude Design?
From 1 comment
Was this created with claude design? it looks very similar to something it mocked for me last week
dylkil on HN
How long did it take to build?
From 1 comment
This is really well done. Curious how long it took to build something this interactive?
gn_central on HN
Alternatives named
Jay Alammar's The Illustrated GPT-2
From 1 comment
Highly recommend instead reading the human created "The Illustrated GPT-2" by Jay Alammar - https://jalammar.github.io/illustrated-gpt2/ And his similar work. He also has a free…
jasonjmcghee on HN
Karpathy's original lecture
From 1 comment
I agreed when I read your comment, but that turned out to be almost directly from the video https://youtu.be/7xTGNNLPyMI From around the 2:20 mark he says: “[…] actually ends up…
QuantumNomad_ on HN
tiktokenizer / gpt-tokenizer.dev
From 2 comments
The page does very poor job tokenizing phrase "Noinceolik fiyulnabmed fyvaproldge" into "Noinceolik fiyulnabm ed fyvaproldge", factoring only "ed" suffix. As if made up words…
thesz on HNI appreciate the feedback, I did notice that as well and I had this thought perhaps this is not worth fixing since I have a link to tiktokenizer. I decided to remove it and just…
ynarwal__ on HN