Back when ChatGPT came out, I was so shocked by how _good_ it was for an “AI” product that I simply had to know how it worked. Over the next month I ended up drawing out a block diagram on a whiteboard I have in my office, with the math involved next to each step in the blackboard. I’d puzzle about each step along the way, and the triumph of completing the drawing was also that of this sense of deep understanding. I kept that drawing up for many months after, and would gaze at it often during meetings and idle moments in wonder.
This is to say: the autoregressive decoder-only transformer llm architecture as pioneered by openai is wildly simple for how revolutionary its results are. I was reading about non-learned classical SLAM systems (uses video + handcrafted math to produce 3d mappings of physical spaces while also locating the camera in those spaces) at the time, and comparatively speaking I’d say the math is about as complicated as ONE of the components in those complex formulations. The only reason frontier LLMs need 6-figure computers to run is because the model designers made the middle bit in those models REALLY BIG, dimensionally speaking. They just took the steam engine, made a few gargantuan versions of it, and are selling them as the ultimate source of power.
This was openai’s entire breakthrough. Making this particular model architecture larger leads to emergent capabilities like being able to pick the best ending to a story/set of instructions or answer questions about broad factual knowledge. I’ve been meanwhile watching these AI companies attempt, successfully, to sell this capability as some sort of robot consciousness hand-crafted by supergeniuses. The fact that they are getting away with it is almost as shocking to me as the discovery itself.
"Attention is all you need" is actually a bad paper if you want to learn about autoregressive LLMs specifically, because it describes a more complicated encoder-decoder architecture while modern LLMs are decoder only. So it's an unnecessarily hard way to get into the subject. "Language Models are Unsupervised Multitask Learners" is probably what you are looking for (aka the GPT-2 paper). This was the first time LLMs really showed what is possible, i.e. they can learn to generalize very well from unstructured data. So no more human labelling necessary, which until then was the primary bottleneck in ML. The paper also lists several key ingredients beyond transformers that are mostly still in place today. This also highlights that there was more to it than just "scaling the transformer algorithm" like many people claim. Most developments since then were about improving training data, until "Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer" drastically changed the architecture landscape again. Later big developments like thinking/reasoning/chain of thought/inference time compute (whatever you want to call it nowadays) are actually all about training again. They work using the exact same architecture.
Yep. It's nearly identical to the neural nets we were using in the 90s. Back then even a supercomputer wasn't big enough or fast enough to do what we do today.
I have to wonder though. Is this all a human brain is? A similar thing to an LLM just scaled exponentially larger. I mean a brain is not just neurons with simple connections to each other. The neurons, axons, dendrites, <insert_unexplained_thing>, etc in a brain are all holding and processing information in different ways and doing it nearly 100% in parallel. That's a really big model.
The biological discoveries show how complex a biological brain actually is. Even the tiny brains in a bee or spider are able to solve puzzles and use tools. That's crazy.
No, it’s definitely not what a human brain is. That makes very little sense. The ways we interact with language (and thus conceptual memory) is completely and fundamentally different.
Indeed. It's pretty interesting to realize after implementing GPT-2 that the frontier models are scaled up versions of that, with various tweaks to improve performance, model-wise.
The secret sauce though is all the datasets, RL training, knowledge of what works from doing all kinds of ablation experiments, and a massive compute moat.
The secret sauce is also having the necessary 'creativity' to not get ceased and desisted into oblivion and jail from all the copyrighted material you trained your model on. Btw, not making a morla judgement, [0] shows Michael and Dalton from YC discussing why Ilya Sutskever had to leave Google to pursue what's now ChatGPT
How do we know that today's frontier models are merely scaled up versions of that? Genuine question, since the labs have narrowed what they share over the years to now almost nothing, in terms of how the model was trained and how it works under the hood.
- R1 https://arxiv.org/abs/2501.12948 (RL applied to ML models was well-known beforehand, but they show it in the open, at scale, on big models)
Then, there's the incentive analysis. If you can see that these models empirically get better with scale, why would you swap the main architecture? Those events will be pretty rare. I'm not saying there's noone cooking a new architecture, just that it is a pretty rare event. And it would have to come from some researchers that would be happy to not publish their findings, which is not really what a sizable portion of elite researchers (obviously not all) are incentivized to do.
Of course, it's a bit of a verbal compression to claim simply 'scaled up'. They are recognisable scaled up transformers, but most new models come with a few tricks, but we're at the point where those usually are not an architectural rewrite and added to solve an explicit problem, like hallucination, not for big new capability gains.
No they are clearly not just scaled up versions of gpt 2; there are different LLM architectures like mixture of experts etc that appeared relatively recently. I am not an expert though, far from it.
separately trained experts can surpass performance in their activated regime and DOES result in a smarter model, the Claude system cards talk about this and eg there is https://openreview.net/forum?id=iydmH9boLb to read...
For anyone who is curious about the first paragraph here, this is actually a great video overview of how LLM works and the tokenization part.
Tangentially related: This part always seemed fuzzy to me, especially when dealing with data scientists and how they talk about how 'ML' looks at problems. I had this issue when working at a SIEM vendor where they kept going on about use case development having to be designed a certain way to catch things. It was all very frustrating.
Architectures have evolved significantly since then. DeepSeek v4 =/= GPT-3. Even then, a great deal of complexity lies in everything surrounding the architectures e.g. how do you implement them performantly on modern accelerators, how do you distribute the model across a set of accelerators, how do you post-train, etc. And pre-training itself is a dark art. If you legitimately think that frontier labs are doing something equivalent to whatever you wrote on your whiteboard, you’re clueless.
We don't really know why language works with humans, either. If you raise a baby from birth, you kind of observe how it is learning language, but the process is also rather mysterious. My eldest son's first word was to actually imitate a cow mooing, and then after that to imitate a motor noise of a tractor or truck. And then after that a meow. (His first complete sentence was "King Graham fell"...)
My next child took a completely different path to language, including skipping all the non-verbal imitations.
And then at some point, you just suddenly can two-way communicate with them when you couldn't before, and then after that, they can engage in reasoning.
It’s interesting to me how similar attempting to understand LLMs is to neuroscience.
“When we turn this bit off, this other thing happens… if we change these weights the Eiffel Tower is now in Rome”
We’re basically just probing around and trying to reverse engineer an emergent system.
To your point, this system may be quite different from model to model (human to human) although some similarities likely occur.
The comment I was responding to tried to belittle the OP’s understanding of transformers, by mentioning that running an LLM at scale is much harder than the simple white board diagram.
My point was simply that we don’t know why they work, and all the extra optimizations isn’t the “thing” that makes it emergent.
Simply scaling the “GPT” is good enough to see it, so the OP’s awe should stand.
(On a side note, what other architectures can we scale to find similar emergent behavior?)
Human brain capabilities are truly amazing, imagine if people didn’t treat their children as if they are stupid and didn’t constantly lie to them, because kids are stupid right, they wouldn’t understand. What heights could be reached.
We don’t treat children like they’re stupid, we treat children like they’re children. A stupid adult is treated very differently than any child.
Adults are expected to have their world models approximately correct in terms of physical environment so they won’t accidentally kill themselves by falling off a cliff; then there are the social norms which adults are expected to conform to so everyone is kinda predictable to everyone else so adults don’t kill each other too often over food or mates. Understanding of neither is expected from children.
You may have been raised properly since you don’t get what I mean. I really envy kids with “Chinese parents” that had them learn math early on and not some bullshit like that if you put your tooth under your pillow, then a tooth fairy will come.
Waste 22 years of life without learning anything and then slave away at a 9-5 job you hate. Brilliant strategy. At least you had “fun”. Then blame billionaires or something.
It sucks that this article is clearly LLM edited, with common phrases like "same shape as", "the intuition: ", and the "tiny explainer" which clearly generalized from a prompt accidentally.
Good article, but when sharing it I will have to preface "yes it's slop, but it's a good explanation".
Absolutely embarrassing that the author didn't catch that these LLM-isms are a (and here I'll use one) bad signal.
We are living in a crazy science fiction world where on the top of the HN frontpage there is an article on how LLMs work which is likely itself LLM generated, and the only way to tell is its writing style rather than its factual accuracy.
Microgpt is really good, if you want to understand exactly what happens. I still thought that this article was a good, higher-level complement to that article though.
I learned TCP/IP by watching and reading raw packets over packet radio at 1200 baud.
I've noticed the same thing is possible if you watch the output of a slow LLM. Eventually you start to see the machinery. input tokens = output tokens, it's math. I can't exactly predict the tokens generated but I can see how they are formed. It's a lot like chess. You can't see every possible move but the mechanism is understandable.
normal people talk and write with some notion of meter, the cadence of communicating where pauses are inserted at places that naturally suit the speaker (and listener) to pause for thought. LLM's don't really do that, they just write a bunch of sentences.
> Researchers have found that some neurons inside the FFN are strongly associated with specific concepts or facts. One neuron might activate strongly on Eiffel-Tower-related text. Another on programming languages. Another on past-tense verbs.
People don't really write like this and they don't really talk like this (and no, people don't necessarily write exactly how they talk because they don't read exactly how they listen; the written word can be backtracked while the heard cannot, and speakers/writers know this, either consciously or unconsciously). A person would probably structure this more like:
> Researchers have found that some neurons inside the FFN are strongly associated with specific concepts or facts. For example, there could be one neuron that activates strongly on Eiffel-Tower-related text, another that activates strongly on programming languages, a third neuron activating on past-tense verbs, and so on.
Usually people wouldn't write "Another on programming languages." as a standalone sentence like that because the periods introduce an unnatural pause like they're giving a TED talk, unless of course they were punctuating that way for effect, but you'd essentially never communicate with that effect full time.
I don’t disagree with your conclusion that this is likely ai rewritten, but I do find it strange that you say “normal people don’t write like this” when it is mimicking how people write, and using patterns I have seen people write. I think models are at the point where style is not really reliable as an indicator anymore.
I'm sure there's plenty of writing in the above style to be found on the Internet, and hence having been trained on by the LLM. I'm also not a fan of this style, and in particular I'd say it's rarely or never found in scientific / technical writing meant to convey understanding rather than sell or hype. So here it's IMO more of a style mismatch.
people sure do write like that, in novels. nobody writes scientific articles like novels, because scientific articles don't need to maximally capture audience attention. the purpose of a scientific article is to convey information - this pursuit is not assisted by punchy prose.
considering they work with any architecture/configuration given enough compute, just more or less efficiently - then maybe it's fundamental, in the same sense as why electricity works...
Why does linear regression works? Why does computer works? Because it's about math and the encoding information. If we can encode words as numbers, then why can't we encode their order as a relation? It's just that neural networks are very apt at finding that relation even if it's noisy.
This is to say: the autoregressive decoder-only transformer llm architecture as pioneered by openai is wildly simple for how revolutionary its results are. I was reading about non-learned classical SLAM systems (uses video + handcrafted math to produce 3d mappings of physical spaces while also locating the camera in those spaces) at the time, and comparatively speaking I’d say the math is about as complicated as ONE of the components in those complex formulations. The only reason frontier LLMs need 6-figure computers to run is because the model designers made the middle bit in those models REALLY BIG, dimensionally speaking. They just took the steam engine, made a few gargantuan versions of it, and are selling them as the ultimate source of power.
This was openai’s entire breakthrough. Making this particular model architecture larger leads to emergent capabilities like being able to pick the best ending to a story/set of instructions or answer questions about broad factual knowledge. I’ve been meanwhile watching these AI companies attempt, successfully, to sell this capability as some sort of robot consciousness hand-crafted by supergeniuses. The fact that they are getting away with it is almost as shocking to me as the discovery itself.
[1] https://arxiv.org/abs/2207.09238
Language Models are Few-Shot Learners https://arxiv.org/abs/2005.14165
I also enjoyed the papers for DeepSeek and GLM for an overview of all the tricks you need to make these things work
DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models https://arxiv.org/abs/2512.02556
GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models https://arxiv.org/abs/2508.06471
I have to wonder though. Is this all a human brain is? A similar thing to an LLM just scaled exponentially larger. I mean a brain is not just neurons with simple connections to each other. The neurons, axons, dendrites, <insert_unexplained_thing>, etc in a brain are all holding and processing information in different ways and doing it nearly 100% in parallel. That's a really big model.
The biological discoveries show how complex a biological brain actually is. Even the tiny brains in a bee or spider are able to solve puzzles and use tools. That's crazy.
If we look beyond written languages which are late inventions of human civilization, oral languages are continuous and build with blocks not words.
Chomskyan school misled the entire field of linguistics for decades by ignoring spoken languages.
No, it's not. There are many animals that have extremely complex and even learned behaviour that have literally zero neurons.
Clearly "neurons" is an oversimplification just-so story, not a scientific theory.
The secret sauce though is all the datasets, RL training, knowledge of what works from doing all kinds of ablation experiments, and a massive compute moat.
[0] https://youtu.be/E8pvgN1j-Ck?t=748
- V3 https://arxiv.org/abs/2412.19437
- V2 https://arxiv.org/abs/2405.04434
- R1 https://arxiv.org/abs/2501.12948 (RL applied to ML models was well-known beforehand, but they show it in the open, at scale, on big models)
Then, there's the incentive analysis. If you can see that these models empirically get better with scale, why would you swap the main architecture? Those events will be pretty rare. I'm not saying there's noone cooking a new architecture, just that it is a pretty rare event. And it would have to come from some researchers that would be happy to not publish their findings, which is not really what a sizable portion of elite researchers (obviously not all) are incentivized to do.
Of course, it's a bit of a verbal compression to claim simply 'scaled up'. They are recognisable scaled up transformers, but most new models come with a few tricks, but we're at the point where those usually are not an architectural rewrite and added to solve an explicit problem, like hallucination, not for big new capability gains.
Tangentially related: This part always seemed fuzzy to me, especially when dealing with data scientists and how they talk about how 'ML' looks at problems. I had this issue when working at a SIEM vendor where they kept going on about use case development having to be designed a certain way to catch things. It was all very frustrating.
We still don’t really know why they work, we just know how to build them.
My next child took a completely different path to language, including skipping all the non-verbal imitations.
And then at some point, you just suddenly can two-way communicate with them when you couldn't before, and then after that, they can engage in reasoning.
It’s interesting to me how similar attempting to understand LLMs is to neuroscience.
“When we turn this bit off, this other thing happens… if we change these weights the Eiffel Tower is now in Rome”
We’re basically just probing around and trying to reverse engineer an emergent system.
To your point, this system may be quite different from model to model (human to human) although some similarities likely occur.
The comment I was responding to tried to belittle the OP’s understanding of transformers, by mentioning that running an LLM at scale is much harder than the simple white board diagram.
My point was simply that we don’t know why they work, and all the extra optimizations isn’t the “thing” that makes it emergent.
Simply scaling the “GPT” is good enough to see it, so the OP’s awe should stand.
(On a side note, what other architectures can we scale to find similar emergent behavior?)
Adults are expected to have their world models approximately correct in terms of physical environment so they won’t accidentally kill themselves by falling off a cliff; then there are the social norms which adults are expected to conform to so everyone is kinda predictable to everyone else so adults don’t kill each other too often over food or mates. Understanding of neither is expected from children.
The "bitter lesson" is that fake-it-till-you-make-it is a valid way of doing knowledge work.
(Or not make it, then people will just claim you're holding the LLM wrong and it's not the AI's fault.)
Good article, but when sharing it I will have to preface "yes it's slop, but it's a good explanation".
Absolutely embarrassing that the author didn't catch that these LLM-isms are a (and here I'll use one) bad signal.
https://github.com/space-bacon/SRT
I've noticed the same thing is possible if you watch the output of a slow LLM. Eventually you start to see the machinery. input tokens = output tokens, it's math. I can't exactly predict the tokens generated but I can see how they are formed. It's a lot like chess. You can't see every possible move but the mechanism is understandable.
I can only imagine what sort of visualizations are going on today inside of the AI labs.
https://archive.ph/aWtFG
it goes all over the place.
i'm not actually sure who your target audience is.
there's too many side tangents.
just like, structure it plz.
1. customer feels bad cuz they don't understand how llms work
2. provide high level abstracted explanation (don't dive into concepts yet)
3. provide breakdown guide of overall set of components.
4. walk through each component. don't side track. no need to explain, ROPE,GQA etc... it just distracts.
i.e. customers don't know how llms work, leading them to feel bad about their own intelligence.
at a high level llms take in words, do some math on them, and then produce words, one by one.
inside llms have these different components. we walk through them step by step.
1. tokenizer
2. embedding
3. attention
4. heads
5. ffn
6. sampling
## tokenizer
I imagine if resources were spent writing this text then one benefit of using it is not using more resources or the pollution caused from a chatbot.
> Researchers have found that some neurons inside the FFN are strongly associated with specific concepts or facts. One neuron might activate strongly on Eiffel-Tower-related text. Another on programming languages. Another on past-tense verbs.
People don't really write like this and they don't really talk like this (and no, people don't necessarily write exactly how they talk because they don't read exactly how they listen; the written word can be backtracked while the heard cannot, and speakers/writers know this, either consciously or unconsciously). A person would probably structure this more like:
> Researchers have found that some neurons inside the FFN are strongly associated with specific concepts or facts. For example, there could be one neuron that activates strongly on Eiffel-Tower-related text, another that activates strongly on programming languages, a third neuron activating on past-tense verbs, and so on.
Usually people wouldn't write "Another on programming languages." as a standalone sentence like that because the periods introduce an unnatural pause like they're giving a TED talk, unless of course they were punctuating that way for effect, but you'd essentially never communicate with that effect full time.
https://www.youtube.com/watch?v=5MdSE-N0bxs is remarkably prescient given that it was written before LLMs