Is there some documentation for this? The code is probably the simplest (Not So) Large Language Model implementation possible, but it is not straight forward to understand for developers not familiar with multi-head attention, ReLU FFN, LayerNorm and learned positional embeddings.
This projects shares similarities with Minix. Minix is still used at universities as an educational tool for teaching operating system design. Minix is the operating system that taught Linus Torvalds how to design (monolithic) operating systems. Similarly having students adding capabilities to GuppyLM is a good way to learn LLM design.
Absolutely. If you loaded this into an agentic coding harness with a decent model, I can practically guarantee it would be able to help you figure out what's going on.
> there is no more need for writing high level docs?
Absolutely not. That would be like exploring a cave without a flashlight, knowing that you could just feel your way around in the dark instead.
Code is not always self-documenting, and can often tell you how it was written, but not why.
> If you loaded this into an agentic coding harness with a decent model, I can practically guarantee it would be able to help you figure out what's going on.
My non-coder but technically savvy boss has been doing this lately to great success. It's nice because I spend less time on it since the model has taken my place for the most part.
There are so many blogs and tutorials about this stuff in particular, I wouldn't worry about it being outside the training data distribution for modern LLMs. If you have a scarce topic in some obscure language I'd be more careful when learning from LLMs.
Even cool projects can learn from others. Maybe they missed something that could benefit the project, or made some interesting technical choice that gives a different result.
For the readers/learners, it's useful to understand the differences so we know what details matter, and which are just stylistic choices.
But it isn't the OP's responsibility to compare their project to all other projects. The GP could themselves perform the comparison and post their thoughts instead of asking an open ended question.
It isn't, but such information will be immensely helpful to anyone who wants to learn from such projects. Some tutorials are objectively better than others, and learners can benefit from such information.
Well, the person who asked the question, for one. I'm sure they're not the only one. Best not to assume why people are asking though, so you can save time by not writing irrelevant comments.
There isn't enough training data though, is there? The "secret sauce" of LLMs is the vast amount of training data available + the compute to process it all.
This is essentially a distillation on the bigger model; you'd wind up surfacing a lot of artifacts from the host model, amplifying them in the same way repeated photocopying introduces errors.
Cool project. I'm working on something where multiple LLM agents share a world and interact with each other autonomously. One thing that surprised me is how much the "world" matters — same model, same prompt, but put it in a system with resource constraints, other agents, and persistent memory, the behavior changes dramatically. Made me realize we spend too much time optimizing the model and not enough thinking about the environment it operates in.
Meaning/goal of life is to reproduce. Food (and everything else) is only a means to it. Reproduction is the only root goal given by nature to any life form. All resources and qualities are provided are only to help mating.
I'd argue genes nor life has a "goal". They are what they are because they've been successful at continuing their existence. Would you say a rock's goal is not to get broken?
Only because genes/organisms can make choices (changes to its programming, or decisions) to optimize their path towards their goal.
A rock is maybe not a good counterexample, but a crystal is because it can grow over time. So in some sense, it tries not to break. However a crystal cannot make any choices; it's behavior is locked into the chemistry it starts with.
This is a nice idea. A tiny implementation can be way more useful for learning than yet another wrapper around a big model, especially if it keeps the training loop and inference path small enough to read end to end.
Could it be possible to train LLM only through the chat messages without any other data or input?
If Guppy doesn't know regular expressions yet, could I teach it to it just by conversation? It's a fish so it wouldn't probably understand much about my blabbing, but would be interesting to give it a try.
Or is there some hard architectural limit in the current LLM's, that the training needs to be done offline and with fairly large training set.
Wow that is such a cool idea! And honestly very much needed. LLMs seem to be this blackbox nobody understands. So I love every effort to make that whole thing less mysterious. I will definitely have a look at dabbling with this, may it not be a goldfish LLM :)
I am trying to find how the synthetic data was created (looking through the repo) and didn't find it. Maybe I am missing it - Would love to see the prompts and process on that aspect of the training data generation!
This is a direct output from the synthetic training data though - wonder if there is a bit of overfitting going on or it’s just a natural limitation of a much smaller model.
Does this work by just training once with next token prediction? Want to understand better how it creates fluent sentences if anyone can provide insights.
You’re absolutely right! HN isn’t just LLM-infested hellscape, it’s a completely new paradigm of machine assisted chocolate-infused information generation.
My initial idea was to train a navigation decision model with 25M parameters for a Raspberry Pi, which, in testing, was getting about 60% of tool calls correct. IMO, it seems like around 20M parameters would be a good size for following some narrow & basic language instructions.
Ok. This makes me wonder about a broader question. Is there a scientific approach showing a pyramid of cognitive functions, and how many parameters are (minimally) required for each layer in this pyramid?
I don't mean to be 'that guy', but after a quick review, this really feels like low-effort AI slop to me.
There is nothing wrong using AI tools to write code, but nothing here seems to have taken more than a generic 'write me a small LLM in PyTorch' prompt, or any specific human understanding.
The bar for what constitutes an engineering feat on HN seems to have shifted significantly.
I love these kinds of educational implementations.
I want to really praise the (unintentional?) nod to Nagel, by limiting capabilities to representation of a fish, the user is immediately able to understand the constraints. It can only talk like a fish cause it’s very simple
Especially compared to public models, thats a really simple correspondence to grok intuitively (small LLM > only as verbose as a fish, larger LLM > more verbose) so kudos to the author for making that simple and fun.
> the user is immediately able to understand the constraints
Nagel's point was quite literally the opposite[1] of this, though. We can't understand what it must "be like to be a bat" because their mental model is so fundamentally different than ours. So using all the human language tokens in the world can't get us to truly understand what it's like to be a bat, or a guppy, or whatever. In fact, Nagel's point is arguably even stronger: there's no possible mental mapping between the experience of a bat and the experience of a human.
IMO we're a step before that: We don't even have a real fish involved, we have a character that is fictionally a fish.
In LLM-discussions, obviously-fictional characters can be useful for this, like if someone builds a "Chat with Count Dracula" app. To truly believe that a typical "AI" is some entity that "wants to be helpful" is just as mistaken as believing the same architecture creates an entity that "feels the dark thirst for the blood of the living."
Or, in this case, that it really enjoys food-pellets.
Id highly disagree with that. Were all living in the same shared universe, and underlying every intelligence must be precisely an understanding of events happening in this space-time.
No I am saying the basis of intelligence must be shared, not that we have the same exact mental model.
I might for example say a human entered a building, a bat might on the other hand think "some big block with two sticks moved through a hole", but both are experiencing a shared physical observation, and there is some mapping between the two.
Its like when people say, if there are aliens they would find the same mathematical constants thet we do
I’m not going to argue other than to say that you need to view the point from a third party perspective evaluating “fish” vs “more verbose thing,” such that the composition is the determinant of the complexity of interaction (which has a unique qualia per nagel)
Hence why it’s a “unintentional nod” not an instantiation
File "<frozen runpy>", line 198, in _run_module_as_main
File "<frozen runpy>", line 88, in _run_code
File "/home/user/gupik/guppylm/guppylm/__main__.py", line 48, in <module>
main()
File "/home/user/gupik/guppylm/guppylm/__main__.py", line 29, in main
engine = GuppyInference("checkpoints/best_model.pt", "data/tokenizer.json")
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/gupik/guppylm/guppylm/inference.py", line 17, in __init__
self.tokenizer = Tokenizer.from_file(tokenizer_path)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Exception: No such file or directory (os error 2)
```
I think this is a nice project because it is end to end and serves its goal well. Good job! It's a good example how someone might do something similar for a specific purpose. There are other visualizers that explain different aspects of LLMs but this is a good applied example.
How much training data did you end up needing for the fish personality to feel coherent? Curious what the minimum viable dataset looks like for something like this.
Great work! I still think that [1] does a better job of helping us understand how GPT and LLM work, but yours is funnier.
Then, some criticism. I probably don't get it, but I think the HN headline does your project a disservice. Your project does not demystify anything (see below) and it diverges from your project's claim, too. Furthermore, I think you claim too much on your github. "This project exists to show that training your own language model is not magic." and then just posts a few command line statements to execute. Yeah, running a mail server is not magic, just apt-get install exim4. So, code. Looking at train_guppylm.ipynb and, oh, it's PyTorch again. I'm better off reading [2] if I'm looking into that (I know, it is a published book, but I maintain my point).
So, in short, it does not help the initiated or the uninitiated. For the initiated it needs more detail for it to be useful, the uninitiated more context for it to be understood. Still a fun project, even if oversold.
But right now people make it a hobby, and that thing can run on a laptop.
This is just so wild.
This projects shares similarities with Minix. Minix is still used at universities as an educational tool for teaching operating system design. Minix is the operating system that taught Linus Torvalds how to design (monolithic) operating systems. Similarly having students adding capabilities to GuppyLM is a good way to learn LLM design.
Absolutely. If you loaded this into an agentic coding harness with a decent model, I can practically guarantee it would be able to help you figure out what's going on.
> there is no more need for writing high level docs?
Absolutely not. That would be like exploring a cave without a flashlight, knowing that you could just feel your way around in the dark instead.
Code is not always self-documenting, and can often tell you how it was written, but not why.
My non-coder but technically savvy boss has been doing this lately to great success. It's nice because I spend less time on it since the model has taken my place for the most part.
Hah, you realize the same thing is going on in your boss's head right? The pie chart of Things-I-Need-stronglikedan-For just shrank tiny bit...
Also, large codebases are harder to understand. But projects like these are simple to discuss with an LLM.
Do LLMs not take comments into consideration? (Serious question - I'm just getting into this stuff)
For the readers/learners, it's useful to understand the differences so we know what details matter, and which are just stylistic choices.
This isn't art; it's science & engineering.
No one, including the GP, said it was.
Well, the person who asked the question, for one. I'm sure they're not the only one. Best not to assume why people are asking though, so you can save time by not writing irrelevant comments.
https://dailyai.com/2025/05/create-a-replica-of-this-image-d...
You> hello Guppy> hi. did you bring micro pellets.
You> HELLO Guppy> i don't know what it means but it's mine.
But the character still comes through in response :)
Food (not dying) is the goal of organisms.
A rock is maybe not a good counterexample, but a crystal is because it can grow over time. So in some sense, it tries not to break. However a crystal cannot make any choices; it's behavior is locked into the chemistry it starts with.
https://en.wikipedia.org/wiki/List_of_countries_by_total_fer...
Now, I ask, have LLMs ben demystified to you? :D
I am still impressed how much (for the most part) trivial statistics and a lot of compute can do.
How does it handle unknown queries?
If Guppy doesn't know regular expressions yet, could I teach it to it just by conversation? It's a fish so it wouldn't probably understand much about my blabbing, but would be interesting to give it a try.
Or is there some hard architectural limit in the current LLM's, that the training needs to be done offline and with fairly large training set.
https://github.com/arman-bd/guppylm/blob/main/guppylm/genera...
Uses a sort of mad-libs templatized style to generate all the permutations.
Laughed loudly :-D
Honestly, I never expected this post to become so popular. It was just the outcome of a weekend practice session.
How many parameters would you need for that?
There is nothing wrong using AI tools to write code, but nothing here seems to have taken more than a generic 'write me a small LLM in PyTorch' prompt, or any specific human understanding.
The bar for what constitutes an engineering feat on HN seems to have shifted significantly.
I want to really praise the (unintentional?) nod to Nagel, by limiting capabilities to representation of a fish, the user is immediately able to understand the constraints. It can only talk like a fish cause it’s very simple
Especially compared to public models, thats a really simple correspondence to grok intuitively (small LLM > only as verbose as a fish, larger LLM > more verbose) so kudos to the author for making that simple and fun.
Nagel's point was quite literally the opposite[1] of this, though. We can't understand what it must "be like to be a bat" because their mental model is so fundamentally different than ours. So using all the human language tokens in the world can't get us to truly understand what it's like to be a bat, or a guppy, or whatever. In fact, Nagel's point is arguably even stronger: there's no possible mental mapping between the experience of a bat and the experience of a human.
[1] https://www.sas.upenn.edu/~cavitch/pdf-library/Nagel_Bat.pdf
In LLM-discussions, obviously-fictional characters can be useful for this, like if someone builds a "Chat with Count Dracula" app. To truly believe that a typical "AI" is some entity that "wants to be helpful" is just as mistaken as believing the same architecture creates an entity that "feels the dark thirst for the blood of the living."
Or, in this case, that it really enjoys food-pellets.
I might for example say a human entered a building, a bat might on the other hand think "some big block with two sticks moved through a hole", but both are experiencing a shared physical observation, and there is some mapping between the two.
Its like when people say, if there are aliens they would find the same mathematical constants thet we do
I’m not going to argue other than to say that you need to view the point from a third party perspective evaluating “fish” vs “more verbose thing,” such that the composition is the determinant of the complexity of interaction (which has a unique qualia per nagel)
Hence why it’s a “unintentional nod” not an instantiation
* How training. In cloud or in my own dev
* How creating a gguf
Traceback (most recent call last):
Exception: No such file or directory (os error 2) `````` # after config device checkpoint_path = "checkpoints/best_model.pt"
ckpt = torch.load(checkpoint_path, map_location=device, weights_only=False)
model = GuppyLM(mc).to(device) if "model_state_dict" in ckpt: model.load_state_dict(ckpt["model_state_dict"]) else: model.load_state_dict(ckpt)
start_step = ckpt.get("step", 0) print(f"Encore {start_step}") ```
https://huggingface.co/datasets/arman-bd/guppylm-60k-generic
Then, some criticism. I probably don't get it, but I think the HN headline does your project a disservice. Your project does not demystify anything (see below) and it diverges from your project's claim, too. Furthermore, I think you claim too much on your github. "This project exists to show that training your own language model is not magic." and then just posts a few command line statements to execute. Yeah, running a mail server is not magic, just apt-get install exim4. So, code. Looking at train_guppylm.ipynb and, oh, it's PyTorch again. I'm better off reading [2] if I'm looking into that (I know, it is a published book, but I maintain my point).
So, in short, it does not help the initiated or the uninitiated. For the initiated it needs more detail for it to be useful, the uninitiated more context for it to be understood. Still a fun project, even if oversold.
[1] https://spreadsheets-are-all-you-need.ai/ [2] https://github.com/rasbt/LLMs-from-scratch