TokenTown: A visual way to understand how LLMs work

(laurentiugabriel.github.io)

56 points | by laurentiurad 2 hours ago

9 comments

  • foltik 34 minutes ago
    I read through the explanations and they’re all typical hand wavy LLM prose, full of jargon, and random tangential details. If you’re going to publicly present this as a learning tool, it would be great if you spent some actual human brain cycles writing and polishing it by hand, rather than just chucking the first thing that claude spits out over the fence. I can pretty much guess the prompt you used and probably get the same thing.
  • GCUMstlyHarmls 15 minutes ago
    "This is not a loop in code: it is a stack of distinct blocks"

    Proceeds to literally loop around 6 times.

  • Hugada 43 minutes ago
    I think you need to know a lot to understand this, so perhaps it should mention what you should already know in order for it to be comprehensible.
  • onlyrealcuzzo 35 minutes ago
    This would be really cool if it did 3Brown 1Blue's style of visualization - but interactively, and allowed you to explore at any stage.
  • markstos 21 minutes ago
    From this I learned LLMs are complex and compute-intensive.
  • lionel-messi 56 minutes ago
    this doesnt help at all
    • HarHarVeryFunny 14 minutes ago
      Not to mention that it's got all sort of bugs and traffic discontinuities. I just saw one of those "road train" things (a batch?!) on the central loop "crash" at the "Layer Counter Arch" while other cars just drove straight thru it.

      The "Output Plaza" doesn't seem very outputtish - cars/embeddings seem to loop around back into the input.

      Not sure why the Attention Plaza is occasionally shooting stuff down the storm drains/whatever in the center!

      The sampler as centrifuge is a strange visualization!

      So I guess the takeaway is that LLM are related to cars? Or maybe highways. Was this designed by Schmidhuber?

    • pluc 35 minutes ago
      Congratulations you're now a senior AI engineer
    • willchis 30 minutes ago
      Just softmax it over the KV cache dude!
  • hk__2 28 minutes ago
    Can we filter out random vibe-coded projects like this from the homepage?
    • simpledevraf 26 minutes ago
      filtering functionality would be great
  • simpledevraf 25 minutes ago
    interesting approach to describe this topic
  • laurentiurad 2 hours ago
    You can paste text and immediately see how it's split into tokens, along with token IDs, counts, and other details. It's useful for understanding why prompts cost what they do, why context limits behave the way they do, and why models sometimes split words in unexpected places.

    I made it because I kept explaining tokenization to friends and realized there wasn't a simple, interactive tool that focused on learning rather than documentation.

    I'd love feedback on:

    Features that would make it more useful UI improvements Bugs or edge cases you've found

    Hope it's useful!

    • hk__2 30 minutes ago
      > I made it

      *I asked a LLM to make it

      • JojiDoesCode 27 minutes ago
        Honestly this person's whole github pages screams tech grifter