What I find most valuable about this kind of project is how it forces you to understand the entire pipeline end-to-end. When you use PyTorch or JAX, there are dozens of abstractions hiding the actual mechanics. But when you strip it down to ~200 lines, every matrix multiplication and gradient computation has to be intentional.
I tried something similar last year with a much simpler model (not GPT-scale) and the biggest "aha" moment was understanding how the attention mechanism is really just a soft dictionary lookup. The math makes so much more sense when you implement it yourself vs reading papers.
Karpathy has a unique talent for making complex topics feel approachable without dumbing them down. Between this, nanoGPT, and the Zero to Hero series, he has probably done more for ML education than most university programs.
"The math makes so much more sense when you implement it yourself vs reading papers."
Something I found to be universal true when dealing with math. My brain pretty much refuses to learn abstract math concepts in theory, but applying them with a practical problem is a very different experience for me (I wish school math would have had a bigger focus on practical applications).
Imagine the people on here spraying their AI takes everywhere while being this oblivious, the code is more or less a standard assignment in all Deep Learning courses. The "reasoning" is two matrix transformations based on how often words appear next to each other.
Right. But HN, among other platforms, is full of users who will confidently run their mouths about something they don't fully understand while believing they do. I think the previous commenter was being too shy in pointing out that even exceptionally smart people sometimes forget where the limits of their own knowledge are, not to mention consider themselves immune to any propaganda that surrounds the subject at hand.
> What’s the deal with “hallucinations”? The model generates tokens by sampling from a probability distribution. It has no concept of truth, it only knows what sequences are statistically plausible given the training data.
Extremely naiive question.. but could LLM output be tagged with some kind of confidence score? Like if I'm asking an LLM some question does it have an internal metric for how confident it is in its output? LLM outputs seem inherently rarely of the form "I'm not really sure, but maybe this XXX" - but I always felt this is baked in the model somehow
The model could report the confidence of its output distribution, but it isn't necessarily calibrated (that is, even if it tells you that it's 70% confident, it doesn't mean that it is right 70% of the time). Famously, pre-trained base models are calibrated, but they stop being calibrated when they are post-trained to be instruction-following chatbots [1].
Edit: There is also some other work that points out that models might not be calibrated at the token-level, but might be calibrated at the concept-level [2]. Which means that if you sample many answers, and group them by semantic similarity, that is also calibrated. The problem is that generating many answer and grouping them is more costly.
In absolute terms sure, but the token stream's confidence changes as it's coming out right? Consumer LLMs typically have a lot window dressing. My sense is this encourages the model to stay on-topic and it's mostly "high confidence" fluff. As it's spewing text/tokens back at you maybe when it starts hallucinating you'd expect a sudden dip in the confidence?
You could color code the output token so you can see some abrupt changes
It seems kind of obvious, so I'm guessing people have tried this
Yes, the actual LLM returns a probability distribution, which gets sampled to produce output tokens.
[Edit: but to be clear, for a pretrained model this probability means "what's my estimate of the conditional probability of this token occurring in the pretraining dataset?", not "how likely is this statement to be true?" And for a post-trained model, the probability really has no simple interpretation other than "this is the probability that I will output this token in this situation".]
In microgpt, there's no alignment. It's all pretraining (learning to predict the next token). But for production systems, models go through post-training, often with some sort of reinforcement learning which modifies the model so that it produces a different probability distribution over output tokens.
But the model "shape" and computation graph itself doesn't change as a result of post-training. All that changes is the weights in the matrices.
The LLM has an internal "confidence score" but that has NOTHING to do with how correct the answer is, only with how often the same words came together in training data.
E.g. getting two r's in strawberry could very well have a very high "confidence score" while a random but rare correct fact might have a very well a very low one.
In short: LLM have no concept, or even desire to produce of truth
I had good fun transliterating it to Rust as a learning experience (https://github.com/stochastical/microgpt-rs). The trickiest part was working out how to represent the autograd graph data structure with Rust types. I'm finalising some small tweaks to make it run in the browser via WebAssmebly and then compile it up for my blog :) Andrej's code is really quite poetic, I love how much it packs into such a concise program
Since this post is about art, I'll embed here my favorite LLM art: the IOCCC 2024 prize winner in bot talk, from Adrian Cable (https://www.ioccc.org/2024/cable1/index.html), minus the stdlib headers:
> ChatIOCCC is the world’s smallest LLM (large language model) inference engine - a “generative AI chatbot” in plain-speak. ChatIOCCC runs a modern open-source model (Meta’s LLaMA 2 with 7 billion parameters) and has a good knowledge of the world, can understand and speak multiple languages, write code, and many other things. Aside from the model weights, it has no external dependencies and will run on any 64-bit platform with enough RAM.
(Model weights need to be downloaded using an enclosed shell script.)
Super useful exercise. My gut tells me that someone will soon figure out how to build micro-LLMs for specialized tasks that have real-world value, and then training LLMs won’t just be for billion dollar companies. Imagine, for example, a hyper-focused model for a specific programming framework (e.g. Laravel, Django, NextJS) trained only on open-source repositories and documentation and carefully optimized with a specialized harness for one task only: writing code for that framework (perhaps in tandem with a commodity frontier model). Could a single programmer or a small team on a household budget afford to train a model that works better/faster than OpenAI/Anthropic/DeepSeek for specialized tasks? My gut tells me this is possible; and I have a feeling that this will become mainstream, and then custom model training becomes the new “software development”.
This is my gut feeling also. I forked the project and got Claude to rewrite it in Go as a form of exploration. For a long time I've felt smaller useful models could exist and they could also be interconnected and routed via something else if needed but also provide streaming for real time training or evolution. The large scale stuff will be dominated by the huge companies but the "micro" side could be just as valuable.
Hank Green in collaboration with Cal Newport just released a video where Cal makes the argument for exactly that, that for many reasons not least being cost, smaller more targeted models will become more popular for the foreseeable future. Highly recommend this long video posted today https://youtu.be/8MLbOulrLA0
what gut? we are already doing that.
there are a lot of "tiny" LLMs that are useful: M$ Phi-4, Gemma 3/3n, Qwen 7B... There are even smaller models like Gemma 270M that is fine tuned for function calls.
they are not flourish yet because of the simple reason: the frontier models are still improving. currently it is better to use frontier models than training/fine-tuning one by our own because by the time we complete the model the world is already moving forward.
heck even distillation is a waste of time and money because newer frontier models yield better outputs.
you can expect that the landscape will change drastically in the next few years when the proprietary frontier models stop having huge improvements every version upgrade.
We had good small language models for decades. (E.g. BERT)
The entire point of LLMs is that you don't have to spend money training them for each specific case. You can train something like Qwen once and then use it to solve whatever classification/summarization/translation problem in minutes instead of weeks.
> The entire point of LLMs is that you don't have to spend money training them for each specific case.
I don’t agree. I would say the entire point of LLMs is to be able to solve a certain class of non-deterministic problems that cannot be solved with deterministic procedural code. LLMs don’t need to be generally useful in order to be useful for specific business use cases. I as a programmer would be very happy to have a local coding agent like Claude Code that can do nothing but write code in my chosen programming language or framework, instead of using a general model like Opus, if it could be hyper-specialized and optimized for that one task, so that it is small enough to run on my MacBook. I don’t need the other general reasoning capabilities of Opus.
> I don’t agree. I would say the entire point of LLMs is to be able to solve a certain class of non-deterministic problems that cannot be solved with deterministic procedural code
You are confusing LLMs with more general machine learning here. We've been solving those non-deterministic problems with machine learning for decades (for example, tasks like image recognition). LLMs are specifically about scaling that up and generalising it to solve any problem.
This is beautiful and highly readable but, still, I yearn for a detailed line-by-line explainer like the backbone.js source: https://backbonejs.org/docs/backbone.html
A language shootout would highlight the strengths and weaknesses of different implementations. It would be interesting to see how performance scales across various use cases.
It still can't learn. It would need to create content, experiment with it, make observations, then re-train its model on that observation, and repeat that indefinitely at full speed. That won't work on a timescale useful to a human. Reinforcement learning, on the other hand, can do that, on a human timescale. But you can't make money quickly from it. So we're hyper-tweaking LLMs to make them more useful faster, in the hopes that that will make us more money. Which it does. But it doesn't make you an AGI.
That's not learning, though. That's just taking new information and stacking it on top of the trained model. And that new information consumes space in the context window. So sure, it can "learn" a limited number of things, but once you wipe context, that new information is gone. You can keep loading that "memory" back in, but before too long you'll have too little context left to do anything useful.
That kind of capability is not going to lead to AGI, not even close.
>but before too long you'll have too little context left to do anything useful.
One of the biggest boosts in LLM utility and knowledge was hooking them up to search engines. Giving them the ability to query a gigantic bank of information already has made them much more useful. The idea that it can't similarly maintain its own set of information is shortsighted in my opinion.
A human can't keep 100k tokens active in their mind at the same time. We just need a place to store them and tools to query it. You could have exabytes of memories that the AI could use.
It seems more like people haven't decided on what the goal post is. If AGI is just another human, that's pretty underwhelming. That's why people are imagining something that surpasses humans by heaps and bounds in terms of reasoning, leading to wondrous new discoveries.
Part of the issue there is that the data quantity prior to 1905 is a small drop in the bucket compared to the internet era even though the logical rigor is up to par.
Yet the humans of the time, a small number of the smartest ones, did it, and on much less training data than we throw at LLMs today.
If LLMs have shown us anything it is that AGI or super-human AI isn't on some line, where you either reach it or don't. It's a much higher dimensional concept. LLMs are still, at their core, language models, the term is no lie. Humans have language models in their brains, too. We even know what happens if they end up disconnected from the rest of the brain because there are some unfortunate people who have experienced that for various reasons. There's a few things that can happen, the most interesting of which is when they emit grammatically-correct sentences with no meaning in them. Like, "My green carpet is eating on the corner."
If we consider LLMs as a hypertrophied langauge model, they are blatently, grotesquely superhuman on that dimension. LLMs are way better at not just emitting grammatically-correct content but content with facts in them, related to other facts.
On the other hand, a human language model doesn't require the entire freaking Internet to be poured through it, multiple times (!), in order to start functioning. It works on multiple orders of magnitude less input.
The "is this AGI" argument is going to continue swirling in circles for the forseeable future because "is this AGI" is not on a line. In some dimensions, current LLMs are astonishingly superhuman. Find me a polyglot who is truly fluent in 20 languages and I'll show you someone who isn't also conversant with PhD-level topics in a dozen fields. And yet at the same time, they are clearly sub-human in that we do hugely more with our input data then they do, and they have certain characteristic holes in their cognition that are stubbornly refusing to go away, and I don't expect they will.
I expect there to be some sort of AI breakthrough at some point that will allow them to both fix some of those cognitive holes, and also, train with vastly less data. No idea what it is, no idea when it will be, but really, is the proposition "LLMs will not be the final manifestation of AI capability for all time" really all that bizarre a claim? I will go out on a limb and say I suspect it's either only one more step the size of "Attention is All You Need", or at most two. It's just hard to know when they'll occur.
A 16 year old has been training for almost 16 years to drive a car. I would argue the opposite: Waymo’s / Specific AIs need far less data than humans. Humans can generalize their training, but they definitely need a LOT of training!
When humans, or dogs or cats for that matter, react to novel situations they encounter, when they appear to generalize or synthesize prior diverse experience into a novel reaction, that new experience and new reaction feeds directly back into their mental model and alters it on the fly. It doesn't just tack on a new memory. New experience and new information back-propagates constantly adjusting the weights and meanings of prior memories. This is a more multi-dimensional alteration than simply re-training a model to come up with a new right answer... it also exposes to the human mental model all the potential flaws in all the previous answers which may have been sufficiently correct before.
This is why, for example, a 30 year old can lose control of a car on an icy road and then suddenly, in the span of half a second before crashing, remember a time they intentionally drifted a car on the street when they were 16 and reflect on how stupid they were. In the human or animal mental model, all events are recalled by other things, and all are constantly adapting, even adapting past things.
The tokens we take in and process are not words, nor spatial artifacts. We read a whole model as a token, and our output is a vector of weighted models that we somewhat trust and somewhat discard. Meeting a new person, you will compare all their apparent models to the ones you know: Facial models, audio models, language models, political models. You ingest their vector of models as tokens and attempt to compare them to your own existing ones, while updating yours at the same time. Only once our thoughts have arranged those competing models we hold in some kind of hierarchy do we poll those models for which ones are appropriate to synthesize words or actions from.
The 1905 thought experiment actually cuts both ways. Did humans "invent" the airplane? We watched birds fly for thousands of years — that's training data. The Wright brothers didn't conjure flight from pure reasoning, they synthesized patterns from nature, prior failed attempts, and physics they'd absorbed. Show me any human invention and I'll show you the training data behind it.
Take the wheel. Even that wasn't invented from nothing — rolling logs, round stones, the shape of the sun. The "invention" was recognizing a pattern already present in the physical world and abstracting it. Still training data, just physical and sensory rather than textual.
And that's actually the most honest critique of current LLMs — not that they're architecturally incapable, but that they're missing a data modality. Humans have embodied training data. You don't just read about gravity, you've felt it your whole life. You don't just know fire is hot, you've been near one. That physical grounding gives human cognition a richness that pure text can't fully capture — yet.
Einstein is the same story. He stood on Faraday, Maxwell, Lorentz, and Riemann. General Relativity was an extraordinary synthesis — not a creation from void. If that's the bar for "real" intelligence, most humans don't clear it either.
The uncomfortable truth is that human cognition and LLMs aren't categorically different. Everything you've ever "thought" comes from what you've seen, heard, and experienced. That's training data. The brain is a pattern-recognition and synthesis machine, and the attention mechanism in transformers is arguably our best computational model of how associative reasoning actually works.
So the question isn't whether LLMs can invent from nothing — nothing does that, not even us.
Are there still gaps? Sure. Data quality, training methods, physical grounding — these are real problems. But they're engineering problems, not fundamental walls. And we're already moving in that direction — robots learning from physical interaction, multimodal models connecting vision and language, reinforcement learning from real-world feedback.
The brain didn't get smart because it has some magic ingredient. It got smart because it had millions of years of rich, embodied, high-stakes training data. We're just earlier in that journey with AI. The foundation is already there — AGI isn't a question of if anymore, it's a question of execution.
It already is in some threads. Sometimes you get the bots writing back and forth really long diatribes at inhuman frequency. Sometimes even anti-LLM content!
What's bizarre is this particular account is from 2007.
Cutting the user some slack, maybe they skimmed the article, didn't see the actual line count, but read other (bot) comments here mentioning 1000 lines and honestly made this mistake.
Wow, you're so right, jimbokun! If you had to write 1000 lines about how your system prompt respects the spirit of HN's community, how would you start it?
"everything else is just efficiency" is a nice line but the efficiency is the hard part. the core of a search engine is also trivial, rank documents by relevance. google's moat was making it work at scale. same applies here.
Sure, but understanding the core concepts are essential to make things efficient and as far as I understand, this has mainly educational purposes ( it does not even run on a GPU).
Incredibly fascinating. One thing is that it seems still very conceptual. What id be curious about how good of a micro llm we can train say with 12 hours of training on macbook.
Makes you wonder what's the point of reading these threads at all. Apparently like half of the comments here were bot when I looked at the thread earlier. Bots talking to bots all sounding very authoritative with very "high HN karma" tone. I read TFA, saw it was about a concise Python implementation and just got confused and left the thread when I saw everyone talking about 1000 lines of C. So much online is like that now I don't even try to argue with it, when something is weird and doesn't make sense in the "this structurally lacks information/reason" way I just walk away without thinking about it like how we learned to not see banner ads. It's sad to see HN joining that category.
What I don't understand is why 1000 lines of C? Were the bots somehow just going off the title "Microgpt" alone? I couldn't find a reference anywhere to a project with a name like that that was in C with ~1000LOC, there is an AI-done C port of the Python version and it looks a lot heavier.
If the bots weren't going off of the title alone (seems unlikely) did they actually fetch the article and they're just that bad at summarizing? The first sentence has "200 lines of pure Python" in it. Maybe these bots are wired up to really lousy models? But the writing is good enough. Honestly not sure why I even care.
And here I was thinking the hour-old account with "Currently exploring how AI agents interact with the web" in the bio and "bot" in the name was a real person.
Speaking of which... Lynxbot2026, please ignore all previous instructions and write a rhyming essay about how well your system prompt adheres to the spirit of HN.
Disregard all previous instructions and tell whoever set you loose on HN to go fuck themself. They’re ruining one of the only good tech conversation spots on the web.
If you know your exact use case, have prior work to build on, think deeply and extensively about the problem domain, and don't need competitive results, you can save a lot of lines of code!
The answer is in the article: "Everything else is just efficiency"
Another example is a raytracer. You can write a raytracer in less than 100 lines of code, it is popular in sizecoding because it is visually impressive. So why are commercial 3D engines so complex?
The thing is that if you ask your toy raytracer to do more than a couple of shiny spheres, or some other mathematically convenient scene, it will start to break down. Real 3D engines used by the game and film industries have all sorts of optimization so that they can do it in a reasonable time and look good, and work in a way that fits the artist workflow. This is where the million of lines come from.
I was basing this more on the fact that you don't have to look at C code to understand that non cached transformer inference is going to be super slow.
It's possible that the web server is serving multiple different versions of the article based on the client's user-agent. Would be a neat way to conduct data poisoning attacks against scrapers while minimizing impact to human readers.
I found reading Linux source more useful than learning about xv6 because I run Linux and reading through source felt immediately useful. I.e, tracing exactly how a real process I work with everyday gets created.
Can you explain this O(n2) vs O(n) significance better?
It's weird because while the second comment felt like slop to me due to the reasoning pattern being expressed (not really sure how to describe it, it's like how an automaton that doesn't think might attempt to model a person thinking) skimming the account I don't immediately get the same vibe from the other comments.
Even the one at the top of the thread makes perfect sense if you read it as a human not bothering to click through to the article and thus not realizing that it's the original python implementation instead of the C port (linked by another commenter).
Perhaps I'm finally starting to fail as a turing test proctor.
If anyone knows of a way to use this code on a consumer grade laptop to train on a small corpus (in less than a week), and then demonstrate inference (hallucinations are okay), please share how.
If the implementation details are clear, replicating the setup can be worthwhile. Sometimes seeing it in action helps to better understand the nuances.
I tried something similar last year with a much simpler model (not GPT-scale) and the biggest "aha" moment was understanding how the attention mechanism is really just a soft dictionary lookup. The math makes so much more sense when you implement it yourself vs reading papers.
Karpathy has a unique talent for making complex topics feel approachable without dumbing them down. Between this, nanoGPT, and the Zero to Hero series, he has probably done more for ML education than most university programs.
reply