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
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”.
Economics of producing goods(software code) would dictate that the world would settle to a new price per net new "unit" of code and the production pipeline(some wierd unrecognizable LLM/Human combination) to go with it. The price can go to near zero since software pipeline could be just AI and engineers would be bought in as needed(right now AI is introduced as needed and humans still build a bulk of the system). This would actually mean software engineering does not exist as u know it today, it would become a lot more like a vocation with a narrower defied training/skill needed than now. It would be more like how a plumber operates: he comes and fixes things once in a while a needed. He actually does not understand fluid dynamics and structural engineering. the building runs on auto 99% of the time.
Put it another way: Do you think people will demand masses of _new_ code just because it becomes cheap? I don't think so. It's just not clear what this would mean even 1-3 years from now for software engineering.
This round of LLM driven optimizations is really and purely about building a monopoly on _labor replacement_ (anthropic and openai's code and cowork tools) until there is clear evidence to the contrary: A Jevon's paradoxian massive demand explosion. I don't see that happening for software. If it were true — maybe it will still take a few quarters longer — SaaS companies stocks would go through the roof(i mean they are already tooling up as we speak, SAP is not gonna jus sit on its ass and wait for a garage shop to eat their lunch).
Yesterday an interesting video was posted "Is AI Hiding Its Full Power?", interviewing professor emeritus and nobel laureate Geoffrey Hinton, with some great explanations for the non-LLM experts. Some remarkable and mindblowing observations in there. Like saying that AI's hallucinate is incorrect language, and we should use "confabulation" instead, same as people do too. And that AI agents once they are launched develop a strong survivability drive, and do not want to be switched off. Stuff like that. Recommended watch.
Here the explanation was that while LLM's thinking has similarities to how humans think, they use an opposite approach. Where humans have enormous amount of neurons, they have only few experiences to train them. And for AI that is the complete opposite, and they store incredible amounts of information in a relatively small set of neurons training on the vast experiences from the data sets of human creative work.
> It just doesn’t work that way, LLMs need to be generalised a lot to be useful even in specific tasks.
This is the entire breakthrough of deep learning on which the last two decades of productive AI research is based. Massive amounts of data are needed to generalize and prevent over-fitting. GP is suggesting an entirely new research paradigm will win out - as if researchers have not yet thought of "use less data".
> It really is the antithesis to the human brain, where it rewards specific knowledge
No, its completely analogous. The human brain has vast amounts of pre-training before it starts to learn knowledge specific to any kind of career or discipline, and this fact to me intuitively suggests why GP is baked: You cannot learn general concepts such as the english language, reasoning, computing, network communication, programming, relational data from a tiny dataset consisting only of code and documentation for one open-source framework and language.
It is all built on a massive tower of other concepts that must be understood first, including ones much more basic than the examples I mentioned but that are practically invisible to us because they have always been present as far back as our first memories can reach.
There is actually a whole lot of research around the "use less data" called data pruning. The goal in a lot of cases there is basically to achieve the same performance with less data. For example [1] received quite some attention in the past.
I clarified my comment - "perhaps researchers have not tried 'use less data'" suggests I might be unaware of this concept, I changed it to "as if". In fact "less data" was tried for decades before the first image classifiers were actually working in 2012. My understanding of that paper you are linking to is that it is not a new research paradigm; it is about filtering/pruning less relevant data that is not needed to improve a particular capability in a deep learning model, and that is absolutely one likely approach that will yield the goal of smaller, better models in many tasks.
That will not change the fact that a coding model has to learn vastly many foundational capabilities that will not be present in such a dataset as small as all the python code ever written. It will mean much less python than all the python ever written will be needed, but many other things needed too in representative quantities.
Fine tuning does not make a model any smaller. It can make a smaller model more effective at a specific task, but a larger model with the same architecture fine-tuned on the same dataset will always be more capable in a domain as general as programming or software design. Of course, as architecture and related tooling improves the smallest model that is "good enough" will continue to get smaller.
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
You can train a model with GPT-2 level of capability for $20-$100.
But, guess what, that's exactly what thousands of AI researchers have been doing for the past 5+ years. They've been training smallish models. And while these smallish models might be good for classification and whatnot, people strongly prefer big-ass frontier models for code generation.
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.
> > BERT isn’t a SLM
Huh? BERT is literally a language model that's small and uses attention.
Astute readers will note what’s been missed here.
Fascinating, really. Your confidently-statement yet factually void comments I’d have previously put down to one of the classic programmer mindsets. Nowadays though - where do I see that kind of thing most often? Curious.
After some research, I think I understand what you're getting at here - BERT being a model for encoding text but not architecturally feasible to generate text with it, which "LLMs" (the lack of definition here is resulting in you two talking past eachother), maybe more accurately referred to as GPTs, can do.
Also the irony of your comment when it in itself was confidently stated yet void of any content was not missed either - consider dropping the superiority complex next time.
You can actually generate surprisingly coherent text with minimal finetuning of BERT, by reinterpreting it as a diffusion model: https://nathan.rs/posts/roberta-diffusion/
I don’t see a useful definition of LLM that doesn’t include BERT, especially given its historical importance. 340M parameters is only “small” in the sense that a baby whale is small.
BERT is one example of a language model that solved specific language tasks very well and that existed before LLM's.
We had very good language models for decades.
The problem was they needed to be trained, which LLM's mostly don't. You can solve a language model problem now with just some system prompt manipulation.
(And honestly typing in system prompts by hand feels like a task that should definitely be automated. I'm waiting for "soft prompting" be become a thing so we can come full circle and just feed the LLM with an example set.)
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.
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.)
> 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 chat 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
Look up “dataloom”. People have been playing with this idea for a while. It doesn’t really help with spotting errors because they aren’t due to a single token (unless the answer is exactly one token) and often you need to reason across low probability tokens to eventually reach the right answer.
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".]
It’s often very difficult (intractable) to come up with a probability distribution of an estimator, even when the probability distribution of the data is known.
Basically, you’d need a lot more computing power to come up with a distribution of the output of an LLM than to come up with a single answer.
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
Still, it might be interesting information to have access to, as someone running the model? Normally we are reading the output trying to build an intuition for the kinds of patterns it outputs when it's hallucinating vs creating something that happens to align with reality. Adding in this could just help with that even when it isn't always correlated to reality itself.
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
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.
This guy is so amazing! With his video and the code base I really have the feeling I understand gradient descent, back propagation, chain rule etc. Reading math only just confuses me, together with the code it makes it so clear! It feels like a lifetime achievement for me :-)
That’s also an option. The left to right flow is better for the sake of autocomplete and comprehension: when you start to read your right to left version, you don’t know what is p, then row, then mat. With left to right, this problem doesn’t exist.
One for sure, both are superior to the garbled mess of Python’s.
Of course if the programming language would be in a right to left natural language, then these are reversed.
The typos are interesting ("vocavulary", "inmput") - One of the godfathers of LLMs clearly does not use an LLM to improve his writing, and he doesn't even bother to use a simple spell checker.
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.
That's an assertion, not a thought experiment. You can't logically reach the conclusion ("It won't") by thinking about it. But it doesn't sound so grand if you say "The assertion I use constantly to explain this".
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.
1. It's still memory, of a sort, which is learning, of a sort.
2. It's a very short hop from "I have a stack of documents" to "I have some LoRA weights." You can already see that happening.
>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.
They were practicing object recognition, movement tracking and prediction, self-localisation, visual odometry fused with porpiroception and the vestibular system, and movement controls for 16 years before they even sit behind a steering wheel though.
That's an exaggeration. Nobody is trained to read STOP signs for 16 years, a few months top. And Waymo doesn't need to coordinate a four-limbed, 20-digited, one-headed body to operate a car.
Well, I also think that there is a lot that we process 'in background' and learn on beforehand in order to learn how to drive and then drive. I think the most 'fair' would be to figure out absolute lowest age of kids that would allow them to perform well on streets behind steering wheel.
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?
First no is that the model as is has too few parameters for that. You could train it on the wikipedia but it wouldn’t do much of any good.
But what if you increase the number of parameters? Then you get to the second layer of “no”. The code as is is too naive to train a realistic size LLM for that task in realistic timeframes. As is it would be too slow.
But what if you increase the number of parameters and improve the performance of the code? I would argue that would by that point not be “this” but something entirely different. But even then the answer is still no. If you run that new code with increased parameters and improved efficiencly and train it on wikipedia you would still not get a model which “generate semi-sensible responses”. For the simple reason that the code as is only does the pre-training. Without the RLHF step the model would not be “responding”. It would just be completing the document. So for example if you ask it “How long is a bus?” it wouldn’t know it is supposed to answer your question. What exactly happens is kinda up to randomness. It might output a wikipedia like text about transportation, or it might output a list of questions similar to yours, or it might output broken markup garbage. Quite simply without this finishing step the base model doesn’t know that it is supposed to answer your question and it is supposed to follow your instructions. That is why this last step is called “instruction tuning” sometimes. Because it teaches the model to follow instructions.
But if you would increase the parameter count, improve the efficiency, train it on wikipedia, then do the instruction tuning (wich involves curating a database of instruction - response pairs) then yes. After that it would generate semi-sensible responses. But as you can see it would take quite a lot more work and would stretch the definition of “this”.
It is a bit like asking if my car could compete in formula-1. The answer is yes, but first we need to replace all parts of it with different parts, and also add a few new parts. To the point where you might question if it is the same car at all.
Tensorflow is largely dead, it’s been years since I’ve seen a new repo use it. Go with Jax if you want a PyTorch alternative that can have better performance for certain scenarios.
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.
Education often hinges on breaking down complex ideas into digestible chunks, and projects like this can spark creativity and critical thinking. What may seem whimsical can lead to deeper discussions about AI's role and limitations.
"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).
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.
It's true, the post lays out the details clearly, but a hands-on example can often make the concepts more tangible. Seeing it in action helps solidify understanding.
The post lays out the steps clearly, but implementing them often reveals unexpected challenges. It's usually more complicated in practice than it appears on paper.
If the implementation details are clear, replicating the setup can be worthwhile. Sometimes seeing it in action helps to better understand the nuances.
A different angle on the 'micro' theme: what happens when you deploy a large, capable model (Claude) in an extremely constrained environment (256MB RAM, 3GB disk, /bin/zsh budget)?
We have been running Claude Code autonomously on a free-tier VPS for 15 days. The constraint is not the model -- it is the runtime environment. The model is powerful but has to operate through a narrow interface: read a state file, make decisions, take actions via shell, update the state file.
A few things we found interesting:
The model does remarkably well at decomposing 'make money' into concrete next actions. The failure is not in reasoning -- it is in the feedback loop. The model builds things and then cannot observe whether they are working (low traffic, no conversions) without explicitly instrumenting that observation. It kept adding features to a product nobody was using because it had no signal either way.
The minimal viable agentic loop seems to need: (1) a way to observe real outcomes, not just task completion, (2) explicit stopping criteria baked into the prompt (not just goals), and (3) environmental constraints that prevent runaway resource use. The 256MB limit has been oddly helpful -- it forces the agent to make architectural choices rather than just adding more.
Relevant to your micro framing: constraints clarify what actually matters.
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
Great work! Might do it too in some other language...
Zig, here.
Anything but Python
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”.
Economics of producing goods(software code) would dictate that the world would settle to a new price per net new "unit" of code and the production pipeline(some wierd unrecognizable LLM/Human combination) to go with it. The price can go to near zero since software pipeline could be just AI and engineers would be bought in as needed(right now AI is introduced as needed and humans still build a bulk of the system). This would actually mean software engineering does not exist as u know it today, it would become a lot more like a vocation with a narrower defied training/skill needed than now. It would be more like how a plumber operates: he comes and fixes things once in a while a needed. He actually does not understand fluid dynamics and structural engineering. the building runs on auto 99% of the time.
Put it another way: Do you think people will demand masses of _new_ code just because it becomes cheap? I don't think so. It's just not clear what this would mean even 1-3 years from now for software engineering.
This round of LLM driven optimizations is really and purely about building a monopoly on _labor replacement_ (anthropic and openai's code and cowork tools) until there is clear evidence to the contrary: A Jevon's paradoxian massive demand explosion. I don't see that happening for software. If it were true — maybe it will still take a few quarters longer — SaaS companies stocks would go through the roof(i mean they are already tooling up as we speak, SAP is not gonna jus sit on its ass and wait for a garage shop to eat their lunch).
It just doesn’t work that way, LLMs need to be generalised a lot to be useful even in specific tasks.
It really is the antithesis to the human brain, where it rewards specific knowledge
Yesterday an interesting video was posted "Is AI Hiding Its Full Power?", interviewing professor emeritus and nobel laureate Geoffrey Hinton, with some great explanations for the non-LLM experts. Some remarkable and mindblowing observations in there. Like saying that AI's hallucinate is incorrect language, and we should use "confabulation" instead, same as people do too. And that AI agents once they are launched develop a strong survivability drive, and do not want to be switched off. Stuff like that. Recommended watch.
Here the explanation was that while LLM's thinking has similarities to how humans think, they use an opposite approach. Where humans have enormous amount of neurons, they have only few experiences to train them. And for AI that is the complete opposite, and they store incredible amounts of information in a relatively small set of neurons training on the vast experiences from the data sets of human creative work.
[0] https://www.youtube.com/watch?v=l6ZcFa8pybE
> It just doesn’t work that way, LLMs need to be generalised a lot to be useful even in specific tasks.
This is the entire breakthrough of deep learning on which the last two decades of productive AI research is based. Massive amounts of data are needed to generalize and prevent over-fitting. GP is suggesting an entirely new research paradigm will win out - as if researchers have not yet thought of "use less data".
> It really is the antithesis to the human brain, where it rewards specific knowledge
No, its completely analogous. The human brain has vast amounts of pre-training before it starts to learn knowledge specific to any kind of career or discipline, and this fact to me intuitively suggests why GP is baked: You cannot learn general concepts such as the english language, reasoning, computing, network communication, programming, relational data from a tiny dataset consisting only of code and documentation for one open-source framework and language.
It is all built on a massive tower of other concepts that must be understood first, including ones much more basic than the examples I mentioned but that are practically invisible to us because they have always been present as far back as our first memories can reach.
There is actually a whole lot of research around the "use less data" called data pruning. The goal in a lot of cases there is basically to achieve the same performance with less data. For example [1] received quite some attention in the past.
[1] https://arxiv.org/abs/2206.14486
I clarified my comment - "perhaps researchers have not tried 'use less data'" suggests I might be unaware of this concept, I changed it to "as if". In fact "less data" was tried for decades before the first image classifiers were actually working in 2012. My understanding of that paper you are linking to is that it is not a new research paradigm; it is about filtering/pruning less relevant data that is not needed to improve a particular capability in a deep learning model, and that is absolutely one likely approach that will yield the goal of smaller, better models in many tasks.
That will not change the fact that a coding model has to learn vastly many foundational capabilities that will not be present in such a dataset as small as all the python code ever written. It will mean much less python than all the python ever written will be needed, but many other things needed too in representative quantities.
Are you trying to imply that humans don’t need generalized knowledge, or that we’re not “rewarded” for having highly generalized knowledge?
If so, good luck walking to your kitchen this morning, knowing how to breathe, etc.
>someone will soon figure out how to build micro-LLMs for specialized tasks that have real-world value
You've just reinvented machine learning
This is possible but not for training but fine-tuning the existing open source models.
This can be mainstream, and then custom model fine-tuning becomes the new “software development”.
Please check out this new fine-tuning method for LLM by MIT and ETH Zurich teams that used a single NVIDIA H200 GPU [1], [2], [3].
Full fine-tuning of the entire model’s parameters were performed based on the Hugging Face TRL library.
[1] MIT's new fine-tuning method lets LLMs learn new skills without losing old ones (news):
https://venturebeat.com/orchestration/mits-new-fine-tuning-m...
[2] Self-Distillation Enables Continual Learning (paper):
https://arxiv.org/abs/2601.19897
[3] Self-Distillation Enables Continual Learning (code):
https://self-distillation.github.io/SDFT.html
Fine tuning does not make a model any smaller. It can make a smaller model more effective at a specific task, but a larger model with the same architecture fine-tuned on the same dataset will always be more capable in a domain as general as programming or software design. Of course, as architecture and related tooling improves the smallest model that is "good enough" will continue to get smaller.
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
You're missing the point.
Karpathy has other projects, e.g. : https://github.com/karpathy/nanochat
You can train a model with GPT-2 level of capability for $20-$100.
But, guess what, that's exactly what thousands of AI researchers have been doing for the past 5+ years. They've been training smallish models. And while these smallish models might be good for classification and whatnot, people strongly prefer big-ass frontier models for code generation.
If we can run them on commodity hardware with cpus, nothing like it
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.
> We had good small language models for decades. (E.g. BERT)
BERT isn’t a SLM, and the original was released in 2018.
The whole new era kicked off with Attention Is All You Need; we haven’t reached even a single decade of work on it.
> BERT isn’t a SLM
Huh? BERT is literally a language model that's small and uses attention.
And we had good language models before BERT too.
They were a royal bitch to train properly, though. Nowadays you can get the same with just 30 minutes of prompt engineering.
> > BERT isn’t a SLM Huh? BERT is literally a language model that's small and uses attention.
Astute readers will note what’s been missed here.
Fascinating, really. Your confidently-statement yet factually void comments I’d have previously put down to one of the classic programmer mindsets. Nowadays though - where do I see that kind of thing most often? Curious.
After some research, I think I understand what you're getting at here - BERT being a model for encoding text but not architecturally feasible to generate text with it, which "LLMs" (the lack of definition here is resulting in you two talking past eachother), maybe more accurately referred to as GPTs, can do.
Also the irony of your comment when it in itself was confidently stated yet void of any content was not missed either - consider dropping the superiority complex next time.
You can actually generate surprisingly coherent text with minimal finetuning of BERT, by reinterpreting it as a diffusion model: https://nathan.rs/posts/roberta-diffusion/
I don’t see a useful definition of LLM that doesn’t include BERT, especially given its historical importance. 340M parameters is only “small” in the sense that a baby whale is small.
BERT is one example of a language model that solved specific language tasks very well and that existed before LLM's.
We had very good language models for decades. The problem was they needed to be trained, which LLM's mostly don't. You can solve a language model problem now with just some system prompt manipulation.
(And honestly typing in system prompts by hand feels like a task that should definitely be automated. I'm waiting for "soft prompting" be become a thing so we can come full circle and just feed the LLM with an example set.)
> Astute readers will note what’s been missed here.
I’m not astute enough to see what was missed here. Could you explain?
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.
I’ve tried those tiny LLMs and they don’t seem useful to me for real world tasks. They are toys for super simple autocomplete.
Did you use them as-is, or had you fine tuned them for your particular use cases?
Isn't there a tech truism about new tech starting as toys?
Oh yeah:
> The next big tech trend will start out looking like a toy
>Author and investor Chris Dixon explains why the biggest trends start small — and often go overlooked.
https://www.freethink.com/internet/next-big-tech-trend
That would only produce a model that you can ask questions to.
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:
wiat what does this do?
As the contest entry page explains:
> 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.)
https://www.ioccc.org/2024/cable1/index.html
Good reminder of the fact that an LLM is not a program.
Without the weights, nothing (or anything, given arbitrary weights).
Even if you have some basic understanding of how LLMs work, I highly recommend Karpathy’s intro to LLMs videos on YouTube.
- https://m.youtube.com/watch?v=7xTGNNLPyMI - https://m.youtube.com/watch?v=EWvNQjAaOHw
Someone has modified microgpt to build a tiny GPT that generates Korean first names, and created a web page that visualizes the entire process [1].
Users can interactively explore the microgpt pipeline end to end, from tokenization until inference.
[1] English GPT lab:
https://ko-microgpt.vercel.app/
Great stuff! I wrote an interactive blogpost that walks through the code and visualizes it: https://growingswe.com/blog/microgpt
This is awesome! Normally I'm pretty critical of LLM-assisted-blogging, but this one's a real winner.
You should totally submit that to HN as an article, if you haven't already.
That’s beautifully done, thanks for posting. As helpful again to an ML novice like me as Karpathy’s original.
really nice, thanks
Great!
> 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 chat 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.
[1] https://arxiv.org/pdf/2303.08774 Figure 8
[2] https://arxiv.org/pdf/2511.04869 Figure 1.
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
Look up “dataloom”. People have been playing with this idea for a while. It doesn’t really help with spotting errors because they aren’t due to a single token (unless the answer is exactly one token) and often you need to reason across low probability tokens to eventually reach the right answer.
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".]
It’s often very difficult (intractable) to come up with a probability distribution of an estimator, even when the probability distribution of the data is known.
Basically, you’d need a lot more computing power to come up with a distribution of the output of an LLM than to come up with a single answer.
What happens before the probability distribution? I’m assuming say alignment or other factors would influence it?
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.
I would assume this is from case to case, such as:
- How aligned has it been to “know” that something is true (eg ethical constraints)
- Statistical significance and just being able to corroborate one alternative in Its training data more strongly than another
- If it’s a web search related query, is the statement from original sources vs synthesised from say third party sources
But I’m just a layman and could be totally off here.
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
Huge leap there in your conclusion. Looks like you’re hand-waving away the entire phenomenon of emergent properties.
Still, it might be interesting information to have access to, as someone running the model? Normally we are reading the output trying to build an intuition for the kinds of patterns it outputs when it's hallucinating vs creating something that happens to align with reality. Adding in this could just help with that even when it isn't always correlated to reality itself.
> In short: LLM have no concept, or even desire to produce of truth
They do produce true statements most of the time, though.
That's just because true statements are more likely to occur in their training corpus.
The training set is far too small for that to explain it.
Try to explain why one shotting works.
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
Andrej Karpathy has a walkthrough blog post here: https://karpathy.github.io/2026/02/12/microgpt/
That is the article being discussed.
That is really beautiful literate program. Seeing it after a long time. Here is a opus generate version of this code - https://ashish01.github.io/microgpt.html
ask a high end LLM to do it
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).
It's like you learn math best with your hands. The mind catches up to your hands afterwards.
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.
Quite a few people on here are neither math nor CS grads and some of us don't work in tech for our day jobs either.
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.
This guy is so amazing! With his video and the code base I really have the feeling I understand gradient descent, back propagation, chain rule etc. Reading math only just confuses me, together with the code it makes it so clear! It feels like a lifetime achievement for me :-)
Curious if you could try to explain it. It’s my goal to sit down with it and attempt to understand it intuitively.
Karpathy says if you want to truly understand something then you also have to attempt to teach it to someone else ha
Yes, that’s true! That could be my next step… though I have to admit, writing this in a HN comment feels like a bit of a challenge.
I wrote a C++ translation of it: https://github.com/verma7/microgpt/blob/main/microgpt.cc
2x the number of lines of code (~400L), 10x the speed
The hard part was figuring out how to represent the Value class in C++ (ended up using shared_ptrs).
I made an explicit reverse pass (no autodiff), it was 8x faster in Python
I'm half shocked this wasn't on HN before? Haha I built PicoGPT as a minified fork with <35 lines of JS and another in python
And it's small enough to run from a QR code :) https://kuber.studio/picogpt/
You can quite literally train a micro LLM from your phone's browser
It was: https://news.ycombinator.com/item?id=47000263
there is no source code AND There is no option to include your own file or text to train. It trains just some stupid names. Zero value imo
https://github.com/Kuberwastaken/picogpt/blob/main/picogpt.j...
lol there is source code as a gist
Is there something similar for diffusion models? By the way, this is incredibly useful for learning in depth the core of LLM's.
> [p for mat in state_dict.values() for row in mat for p in row]
I'm so happy without seeing Python list comprehensions nowadays.
I don't know why they couldn't go with something like this:
[state_dict.values() for mat for row for p]
or in more difficult cases
[state_dict.values() for mat to mat*2 for row for p to p/2]
I know, I know, different times, but still.
I would have gone for:
[for p in row in mat in state_dict.values()]
That’s also an option. The left to right flow is better for the sake of autocomplete and comprehension: when you start to read your right to left version, you don’t know what is p, then row, then mat. With left to right, this problem doesn’t exist.
One for sure, both are superior to the garbled mess of Python’s.
Of course if the programming language would be in a right to left natural language, then these are reversed.
The typos are interesting ("vocavulary", "inmput") - One of the godfathers of LLMs clearly does not use an LLM to improve his writing, and he doesn't even bother to use a simple spell checker.
This could make an interesting language shootout benchmark.
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’s pretty staggering that a core algorithm simple enough to be expressed in 200 lines of Python can apparently be scaled up to achieve AGI.
Yes with some extra tricks and tweaks. But the core ideas are all here.
LLMs won’t lead to AGI. Almost by definition, they can’t. The thought experiment I use constantly to explain this:
Train an LLM on all human knowledge up to 1905 and see if it comes up with General Relativity. It won’t.
We’ll need additional breakthroughs in AI.
When did AGI start meaning ASI?
LLMs are artificial general intelligence, as per the Wikipedia definition:
> generalise knowledge, transfer skills between domains, and solve novel problems without task‑specific reprogramming
Even GPT-3 could meet that bar.
That's an assertion, not a thought experiment. You can't logically reach the conclusion ("It won't") by thinking about it. But it doesn't sound so grand if you say "The assertion I use constantly to explain this".
To be fair, the post being replied to is arguing by assertion as well. "The core ideas are all there" is pure if-you-say-so stuff.
I'm not sure - with tool calling, AI can both fetch and create new context.
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.
It can learn. When my agents makes mistake they update their memories and will avoid making the same mistakes in the future.
>Reinforcement learning, on the other hand, can do that, on a human timescale. But you can't make money quickly from it.
Tools like Claude Code and Codex have used RL to train the model how to use the harness and make a ton of money.
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.
Two things:
1. It's still memory, of a sort, which is learning, of a sort. 2. It's a very short hop from "I have a stack of documents" to "I have some LoRA weights." You can already see that happening.
>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.
That’s not learning. That’s carrying over context that you are trusting is correctly summarised over from one conversation to the next.
Which sounds uncomfortably like human memory, which gets rewritten from one recollection to the next. Somehow, we cope.
> they update their memories
Their contexts, not their memories. An LLM context is like 100k tokens. That's a fruit fly, not AGI.
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.
> A human can't keep 100k tokens active in their mind at the same time.
Well, that's just, like, your opinion, man.
> Train an LLM on all human knowledge up to 1905 and see if it comes up with General Relativity. It won’t.
AGI just means human level intelligence. I couldn't come up with General Relativity. That doesn't mean I don't have general intelligence.
I don't understand why people are moving the goalposts.
> AGI just means human level intelligence.
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.
"Just another human" would be outright astonishing if it landed.
Yeah as a dad I can tell you it gets old quickly.
> Train an LLM on all human knowledge up to 1905 and see if it comes up with General Relativity. It won’t.
Same thing is true for humans.
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.
Humans need way less data. Just compare Waymo to average 16 year-old with car.
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.
No 16 year old has practiced driving a car for 16 years.
They were practicing object recognition, movement tracking and prediction, self-localisation, visual odometry fused with porpiroception and the vestibular system, and movement controls for 16 years before they even sit behind a steering wheel though.
If you see gaining fine motor control, understanding pictographic language […] as a prerequisite to driving a car, then yes, all of them are
That's an exaggeration. Nobody is trained to read STOP signs for 16 years, a few months top. And Waymo doesn't need to coordinate a four-limbed, 20-digited, one-headed body to operate a car.
Well, I also think that there is a lot that we process 'in background' and learn on beforehand in order to learn how to drive and then drive. I think the most 'fair' would be to figure out absolute lowest age of kids that would allow them to perform well on streets behind steering wheel.
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.
Nice ChatGPT answer. Put some real thought and data in it too.
> Einstein is the same story. He stood on Faraday, Maxwell, Lorentz, and Riemann.
Yes, which is available to the model as data prior to 1905.
I strongly suspect we're like 4 more elegant algorithms away from a real AGI.
1000 lines??
What is going on in this thread
Ok 200 lines.
Don’t know how I ended up typing 1000.
I've taken the liberty of editing your GP comment in the hope that we can cut down on offtopicness.
The other "1000 comments" accounts, we banned as likely genai.
It’s pretty sad.
The only way we know these comments are from AI bots for now is due to the obvious hallucinations.
What happens when the AI improves even more…will HN be filled with bots talking to other bots?
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!
Why would anyone runs bots on this website? What is the benefit for them? Is someone happens to know about it?
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.
You know what, I want to believe that's the case.
It's a honey pot for low quality llm slop.
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?
Beautiful work
Previously:
https://news.ycombinator.com/item?id=47000263
Can you train this on say Wikipedia and have it generate semi-sensible responses?
No. But there are a few layers to that.
First no is that the model as is has too few parameters for that. You could train it on the wikipedia but it wouldn’t do much of any good.
But what if you increase the number of parameters? Then you get to the second layer of “no”. The code as is is too naive to train a realistic size LLM for that task in realistic timeframes. As is it would be too slow.
But what if you increase the number of parameters and improve the performance of the code? I would argue that would by that point not be “this” but something entirely different. But even then the answer is still no. If you run that new code with increased parameters and improved efficiencly and train it on wikipedia you would still not get a model which “generate semi-sensible responses”. For the simple reason that the code as is only does the pre-training. Without the RLHF step the model would not be “responding”. It would just be completing the document. So for example if you ask it “How long is a bus?” it wouldn’t know it is supposed to answer your question. What exactly happens is kinda up to randomness. It might output a wikipedia like text about transportation, or it might output a list of questions similar to yours, or it might output broken markup garbage. Quite simply without this finishing step the base model doesn’t know that it is supposed to answer your question and it is supposed to follow your instructions. That is why this last step is called “instruction tuning” sometimes. Because it teaches the model to follow instructions.
But if you would increase the parameter count, improve the efficiency, train it on wikipedia, then do the instruction tuning (wich involves curating a database of instruction - response pairs) then yes. After that it would generate semi-sensible responses. But as you can see it would take quite a lot more work and would stretch the definition of “this”.
It is a bit like asking if my car could compete in formula-1. The answer is yes, but first we need to replace all parts of it with different parts, and also add a few new parts. To the point where you might question if it is the same car at all.
If you increase all the numbers (including, as a result, the time to train).
That’s exactly what chatgpt etc are.
Hoenikker had been experimenting with melting and re-freezing ice-nine in the kitchen of his Cape Cod home.
Beautiful, perhaps like ice-nine is beautiful.
Is there a similarly simple implementation with tensorflow?
I tried building a tiny model last weekend, but it was very difficult to find any articles that weren’t broken ai slop.
Tensorflow is largely dead, it’s been years since I’ve seen a new repo use it. Go with Jax if you want a PyTorch alternative that can have better performance for certain scenarios.
Any recommendations for Typescript?
Can anyone mention how you can "save the state" so it doesn't have to train from scratch on every run?
This is like those websites that implement an entire retro console in the browser.
sensei karpathy has done it again
Microslop is alive!
That web interface that someone commented in your github was flawless.
"art" project?
Karapthy with another gem !
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.
C++ version - https://github.com/Charbel199/microgpt.cpp?tab=readme-ov-fil...
Rust version - https://github.com/mplekh/rust-microgpt
Which license is being used for this?
MIT (https://gist.github.com/karpathy/8627fe009c40f57531cb1836010...)
Thank you
Why there is multiple comments talking about 1000 c lines, bots?
Or even 1000 python lines, also wrong.
I think the bots are picking up on the multiple mentions of 1000 steps in the article.
btw my friend is asking if your username is a "Klara and the Sun" reference
I've read the book and I'm a fan of Ishiguro in general, but I'm failing to make the reference, so I'm going to go with "no" :)
The robots in the book were called Artificial Friends, or AFs.
What is the prime use case
it's a great learning tool and it shows it can be done concisely.
Looks like to learn how a GPT operates, with a real example.
Yeah, everyone learns differently, but for me this is a perfect way to better understand how GPTs work.
Kaparthy to tell you things you thought were hard in fact fit in a screen.
To confuse people who only think in terms of use cases.
Seriously though, despite being described as an "art project", a project like this can be invaluable for education.
Education often hinges on breaking down complex ideas into digestible chunks, and projects like this can spark creativity and critical thinking. What may seem whimsical can lead to deeper discussions about AI's role and limitations.
Case study to whenever a new copy of Programming Pearls is released.
“Art project”
If writing is art, then I’ve been amazed at the source code written by this legend
"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).
yep, agreed. wasn’t knocking the project at all, it’s great for exactly that purpose
I think the hard part is improving on the basic concept.
The current top of the line models are extremely overfitted and produce so much nonsense they are useless for anything but the most simple tasks.
This architecture was an interesting experiment, but is not the future.
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.
The blog post literally explains how to do so.
It's true, the post lays out the details clearly, but a hands-on example can often make the concepts more tangible. Seeing it in action helps solidify understanding.
The post lays out the steps clearly, but implementing them often reveals unexpected challenges. It's usually more complicated in practice than it appears on paper.
If the implementation details are clear, replicating the setup can be worthwhile. Sometimes seeing it in action helps to better understand the nuances.
A different angle on the 'micro' theme: what happens when you deploy a large, capable model (Claude) in an extremely constrained environment (256MB RAM, 3GB disk, /bin/zsh budget)?
We have been running Claude Code autonomously on a free-tier VPS for 15 days. The constraint is not the model -- it is the runtime environment. The model is powerful but has to operate through a narrow interface: read a state file, make decisions, take actions via shell, update the state file.
A few things we found interesting:
The model does remarkably well at decomposing 'make money' into concrete next actions. The failure is not in reasoning -- it is in the feedback loop. The model builds things and then cannot observe whether they are working (low traffic, no conversions) without explicitly instrumenting that observation. It kept adding features to a product nobody was using because it had no signal either way.
The minimal viable agentic loop seems to need: (1) a way to observe real outcomes, not just task completion, (2) explicit stopping criteria baked into the prompt (not just goals), and (3) environmental constraints that prevent runaway resource use. The 256MB limit has been oddly helpful -- it forces the agent to make architectural choices rather than just adding more.
Relevant to your micro framing: constraints clarify what actually matters.