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.
This interview https://youtu.be/oWOz2htozfI?si=qdQ0uZRoZOYeThOn
from 2 days ago with a top researcher from OpenAI directly addresses the bitter lesson argument and the importance of scaling for the history of their models.
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 moral judgement, [0] shows Michael and Dalton from YC discussing why Ilya Sutskever had to leave Google to pursue what's now ChatGPT
There is a whole moral judgement to be made here...lets hope Ilya wont get too pissed off if somebody leaks the work of his new initiative...information wants to be free and all that...
Also would love to know if the same Legal team advised on Gemini...
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.
We know for sure the architecture of the open weights models since llama.cpp understands the architecture it needs to build to plug the weights into to run them. It's always possible that the latest closed model is doing something architecturally different than the open weights ones we know about, but judging by how close the large open weight models such as DeepSeek are to SOTA performance, this seems unlikely. When OpenAI first came out with their near-mythical "Strawberry" (aka "o1") thinking model there was all sorts of speculation that they had made some sort of architectural breakthough, but then DeepSeek replicated the capability and published how they did it, proving that it was just better training, not any architectural change.
There have been minor changes to the architecture over the years, but these are basically all efficiency tweaks such as various types of attention (some pioneered in the open by DeepSeek) that better scale to large context lengths, and the confusingly named "mixture of experts" architecture, but what's more notable really is how little the architecture has changed. The capability gains have been coming from better training and better data.
- 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.
If you can make the existing model faster, you can then save your inference budget to then make your model bigger, which then makes it smarter.
A lot of how smart the models can be comes down to budget. If you can make your existing thing cheaper, you can instead make it bigger for the same price.
Not really “smarter” though? It’s just a big probability engine.
(Not trying to flame bait or anything. I just wouldn’t call LLM as exhibiting intelligence. It is great at making connections based on probability but doesn’t have a semantic understanding of what it is doing)
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...
> 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.
ReAct loops and tool-calling are the critical development feature. They turn a model from something that generates text into something that can independently influence the world around them.
There is a different way to look at this: that is, actually the Transformer is a minimal complication of what the based model is: in theory the neural network could be just a huge FFN, which is anyway the part of the Transformer that does the heavy lifting. But this would be impossibile to train both numerically and computationally, so the Transformer encodes enough priors for it to work: the causal attention, and the math tricks like the residuals and so forth. But the bottom line of all this is that the Transformer works because of the incredible semantical power of simple/huge FFNs.
You can go another step - a FFN can be simulated on a Turing machine, thus it just exemplifies the incredible semantical power of the Turing machine model of computation. (in fact you don't even need a Turing machine, since there is no looping in one forward pass).
In theory you can run a huge FFN on the tiniest Turing machine, in practice it's much better to run a Transformer on the latest NVIDIA hardware. Or as they say "quantity (performance) has a quality all its own"
I was about to post your last point / quote. Going multigpu is relatively not so though but once you go multi-node you have distributed storage/io/compute system which is highly non trivial. Add that the long training times now you have robustness/fault-tolerantness concerns with hardware failures and restarts. Today’s training systems are engineering marvels.
What hopes/paths does a mere CS bachelor (not deep into stats/maths), and mid level dev (native mobile only; 10-15 years exp.), have about not only understanding it (maybe not fully) but getting possibly into this as a career? Not expecting churning out models and AI systems from the first weeks/months but entry/employment into this field?
(If I can be honest, and I am not being disparaging about anything lest it might seem so, I am looking at it from a career breakthrough/move perspective rather than an intellectual pursuit.)
> I think you need to ask what you actually want to do with the AI.
What about improving the efficiency of token consumption, etc., basically opportunities for improving cost/performance?
I keep thinking there has to be a better way to share context with models than dumping entire gigantic skill files of raw text or otherwise into them - I'm betting there's a bunch of low-hanging fruit there.
There may be some low hanging fruit, but they're not available to people without deep understanding of how the math works. Well paid people already spend a lot of time thinking about this.
i am not sure acctually of the math is acctually that complicated/important. the math around neural networks is calculus/chain rule etc and for model comparison/validation one needs statistics. the required math for e.g. understand transformers is quite accessible.
Im also a mere mortal, and after putting a few years into it IMO I’d say people make it much more complicated than it actually is. I failed most of my math courses for lack of interest, but found passion later with the aforementioned SLAM stuff. I have no doubt you or any other programmer could learn this stuff, especially since you can ask ChatGPT clarifying questions.
I have no idea about careers at this point, I’m still doing fancy IT work as my day job I and look away from the future with dread. I also haven’t been looking for new roles on the open job market, so who knows maybe there’s multimillion pay packages for anyone who can articulate how attention works in an interview.
I have a BS in CS (and have been in the field for 25 years). I couldn't understand the transformer architecture until I built a few myself. Here are the books I worked through. I now feel I have a very good understanding of modern LLMs.
I had the same reaction as you, when I learned in detail, how all this works. But then I also learned about superposition and compressed sensing, and now...I am not so sure anymore...
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.
It is different, but there may be some universal principles that are relevant more abstractly among both cases. Of particular interest is the empirical notion that statistical models of a certain form will always tend to "average out noise" and "learn meaningful patterns" up to the capacity that those models have for representing said patterns. A parallel notion to this is the hypothesis dubbed "thermodynamic origins of life". The universal principle binding these two seemingly disparate topics is one that seems to underlie any sense of "learning" in physical systems: that semantics of those systems depend on their representational power, and the semantics they do come to represent are the results of adding up many pushes in one "direction" (phase space / state space / etc.) encoding a pattern, and adding up many random noise jiggles will cancel out but give you a first-order sense of variance of those semantic features as expressed by the environment.
As this description is so overly abstract, an exercise for the reader is to try to work through an explanation of how, say, a river delta comes to "learn" about its environment by "reacting" to the influences at its borders, and how it "encodes" whatever it is that it learns in the substrate that it inhabits.
I suspect it was considered many times, but the sheer computation scale would make it feel like obscene brute force. It feels like the right shape but too wild to think about implementing.
I think as time went on, and hardware got better, it seemed more reasonable to actually think about a viable implementation of what I think was a widespread intuition anyone in ML had that everything's context is everything.
It just seemed like a theoretical thing until hardware caught up. Maybe. Perhaps I'm applying a retrospective excuse to why it took so long.
"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.
Chain of Thought was kind of an obvious solution that everybody knew was necessary by the time chatgpt / gpt4 came out. It was just a matter of time that frontier labs actually shipped it.
MoE was also pretty straightforward, just a bit surprising how well it worked (that you can get away with just 1/32 active parameters), but most researchers would have come up with it on their own probably.
The true ground breaking papers are the first two you mentioned (transformers and gpt2), and InstructGPT was also very surprising that it worked so well.
It's not just the architecture but also the data - the decoder only approach lets you train in parallel over blocks of text (no RNN serial waiting), that allows you train on much, much more data.
I’d recommend my method of just drawing out the block diagram and drawing out + digging into the math at each step! I’m the kind of person who needs to take time to ask lots of questions before stuff clicks, and if you are too I strongly recommend it.
I picked it up from trying to teach myself that SLAM stuff. The papers are very short, but highly information dense and at the time there was no ChatGPT to help me. I got through them by just creeping my way through the math with a whiteboard, and something about drawing it out and having it there in my office made it all click. Trying to watch piecemeal lectures on YouTube or grind through foundational books like MVG just didn’t work for me, I used them instead as references for my drawings.
Same happened when I tried learning this GPT stuff. karpathy’s videos were out at the time, but I couldn’t really stay focused on them or connect the math with the code. Most other descriptions I could find were focused on getting you to use their inference library or harness. Assembling the picture together on my whiteboard by focusing on drawing out the block diagram continues to be my personal favorite method for deep understanding of complex systems.
Not OP but I worked through Sebastian Raschka's "Build a Large Language Model (From Scratch)" [0] and Raj Abhijit Dandekar's "Build a DeepSeek Model (From Scratch)" [1] books.
I don't think there is anything in a transformer I couldn't explain in the smallest detail now.
Well, as I suggested, working through the implementation yourself will give you that intuition. That said, I think the simplest way to explain why positional encodings are useful is that it gives the transformer just enough information to make attention meaningful without negatively impacting any parallel, content-based comparisons.
A vanilla self-attention layer is just a set of token vectors. Without positional info, swapping two identical embeddings changes very little about what attention can compute. We can "fix" this problem by using positional encodings. Text that has meaning isn't just a set of characters; the location and order of those characters is what provides meaning.
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.
I mean there is a little something called compute. And other complexity that comes like writing code to efficiently distribute a model across machines.
> 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.
how did you know about the steps and there was math involved. i am curious about your process and you came up with what exactly to learn to unravel the mystery.
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.
Maybe math works but critical thinking doesn’t. There are people who have lived for many decades without ever questioning insane b.s. they were taught as kids.
They aren't stupid, but they aren't quite ready to handle the full responsibilities of the world and worry about things they don't need to worry about.
My son is very worried about black holes lately when he learned anything that goes into one can't get out. He's pretty concerned astronauts could get stuck in one some day. So I explained to him that Hawking radiation does actually mean you can eventually get out; it just takes some time.
I didn't think it pertinent to mention spaghettification, the fact anywhere near a black hole will be really hot, or that cosmic censorship means whatever Hawking-radiates from a black hole wouldn't be an astronaut anymore.
It was also fun to hear Hawking speak. He wanted to know if Hawking was a robot. I said no, but he has a robot talk for him. Not quite true, but close enough.
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.
Childhood only lasts 13 to 15 years where I am. By the time you’re in high school, you can be expected to be responsible in some matters. By 22 you have 7 years of experience in making decisions for yourself.
Statistically most likely in what context, given which preconditions? Because each prompt sequence is unique so the probability of any token following it is unknown.
Correctness is based on axioms and rules. You need to define your axioms and rules first before you can determine correctness.
If you’re talking about matrix multiplication, I can use mathematical rules and axioms and proves formally that the multiplication is correct. For next token prediction, I can prove that the set of tokens is finite and that the next token is always part of that set.
But things like grammar correctness, or semantic consistency over a few sentences are not hardcoded rules in the model. They’re emergent properties, mostly due to the amount and quality of data available for training. Quantization is mostly about how much we can shed without loosing a particular emergent properties (like dithering or psycho acoustic audio compression)
This "they just predict the next statistically most likely token" is such an handwavey and willfully misleading explanation, it's unreal, and I'm so fucking tired of seeing it so incessantly repeated. It's beyond asinine.
You know it perfectly damn well that a typical person's idea of statistics is not some insanely high cardinality stateful prediction, but a "well a coin toss is a 50:50, and a lottery win is a 1:100000000". You also know it perfectly damn well that as a result, people will just think that all the sentences chatbots ever produced to them were then just somewhere in the massive training set, letter by letter. This insinuation is often even explicitly appealed to.
And that picture is outright false. It's a statistical process, yes, so saying that it does what it does by "just doing statistics" is gonna be a generally correct description, but that's not at all inquisitive to how exactly does it do it, nor is it the zinger you think it is. If you did the aforementioned, you'd just get milquetoast nonsense, like you can see in the countless Markov-chain primers. And while the models do have a lot of the training set lossily captured, they do also absolutely generalize (that's how they can do that lossy compression), and you can quite literally find representations of those generalizations in them, and also see them activate.
It's like summarizing how any program works by just saying "well it just manipulates ones and zeroes". Not very informative, is it? Or how programs are written by just programmers sitting in a cushy office, ryhtmically pressing keys on a keyboard. Not a very fair or insightful description, which you'll know if you've done any amount of programming in your life on your own. Extends to all other white collar jobs too.
It's also not even true in the most literal sense: models can and do absolutely choose a less than maximally likely next token, that's what the various decoding parameters are for. "Maximally likely next token" further conviently skipping over how that likelihood is established in the first place, i.e. the literal point of the question, going in a cute little circle.
Great, and won’t we all be just as surprised when human self-attentional control turns out to be just as simple or just as complex! Our minds as a strange fabric built of threads of recursions without the benefit of any explicit clock.
There's one thing I wish people understood about LLMs, and it doesn't really have anything to do with what's inside the neural network part. It's the fact that LLMs can only write in one direction — forward.
When you are writing an essay and realize midway through a sentence that what you've written doesn't make sense, you go back and edit. An LLM can't do that, the only thing it can do is keep on generating. Because training data typically contains full essays and not half-finished sentences which were then edited, LLMs have a strong preference for "saving face" and producing grammatically correct, internally coherent outputs. They will often do so even if the only way to write themselves out of the corner they wrote themselves into is to lie. To maintain internal coherence, they'll then repeat that lie for the rest of the response.
This is also why changing response structure used to affect LLM performance so dramatically. If you asked an LLM to solve a math problem and all-but-forced it to start with the answer, it would have had to calculate that answer before emitting any tokens, something which it very often wasn't able to do. If it was told to follow up the answer with an explanation, it would produce a plausible-sounding explanation to maintain coherence.
If, on the other hand, it was told to start by "thinking step by step", it would often be able to solve the first step, and then the next one given the results of the first, and so on, until it was able to reach the answer. Because the answer came last, it wasn't committing to anything, so had no reason to "save face" and lie.
This part of the problem is basically solved now with reasoning; reasoning is where all the step-by-step stuff happens, even if users aren't always able to see it. In the process of RLVR, models even train themselves into outputting phrases like "let me check my answer once again" in the chain-of-thought; those serve as their "life rafts" which they can use to both save face and change their answer.
In terms of our brains though we can only think forward as well (if forward is time). Our brain in the future says something we did in the past was wrong (part of the sentence we wrote) and that informs our body (the agent) to go back and fix it
The part about positional encoding is not correct.
> The intuition: instead of adding position info to each token’s vector, RoPE rotates the vector by an angle that depends on its position
You can't rotate the token's entire vector (or all three vectors, whatever is being implied is unclear). You rotate each token's Query and Key vectors only, so dot product can be used to tell how far apart the tokens are when comparing token 1's Query vector to token 2's Key vector.
Positional embedding should just be explained after explaining the Query, Key and Value vectors. When the article explains those only after that, the reader is building up on a wrong intuition and it gets confusing.
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.
I'd recommend to maybe also specifically watching Karpathy's videos and focusing on the early parts where he specifically deals with tokenization / embeddings generation (which gets really overlooked), and he does this in most of his videos.
Saying an article is of inferior quality just because editing was AI-assisted is like saying a book is lower quality just because it was printed rather than written by hand
Rather interesting than clanker slop defenders downplay the clanker aspect and highlight the human by calling it "ai-assisted", which defeats their entire point.
I hope you do some introspection and start consciously recognizing that the human input and the clanker slop is just debasing it.
Not just that, I think a lot of people are going to waste their time losing the battle (and make no mistake, they will lose) fighting against AI writing without ever asking themselves what makes writing good in the first place.
There’s good AI writing and bad organic writing. But it’s easier to point out a few LLM-isms than to actually identify the problems with text.
Sure, but the LLM-isms in AI writing are mentally exhausting to see in every way at this point.
The whole point of reading, frankly, is to understand the voice of other people. When you pass that through a distorted filter that makes everyone sound the same... its bad, lossy, frustrating communication
It's also dishonest. When you publish something that is direct output without your wording. Digital catfishing at best.
The only good AI writing is providing the prompt, because the question is way more interesting, and way more constructive to learning than the answer
I don't like how most LLM explainer articles and videos say that essentially a LLM " predicts the next word".
I'm a developer but not very good at maths and I still don't understand any of it.
A LLM clearly has some "visual" capacity. You ask Gemini to build something with Canvas and it's able to reason about the shape of things. Like recently I waanted a checkbox that has like a gradient flowing around the edge. It figured out it could use a radial gradient from the center of the checkbox, and overlay that with a small inner div so you only see the edge that looks like the gradient is circling around the checkbox.
How is that "predicting the next word"?
Not saying AI is intelligent or conscious or anything like that, but the algorithm clearly is far more complex than "predicting words".
What I mean, is the LLM is able to represent things in space . That part I don't understand.
I also still dont understand the relationship between the chat based LLM and the multi modal stuff. I think I read somewhere when image is generated it is also tokens?
Predicting a word is the final objective, as in the output of the model is a probability distribution of the next token. However, choosing the right token is more complicated than just regurgitating the training data (and you won't encounter an exact example in the training data, so you need to interpolate). This makes the model learn abstract representation of things that it is able to manipulate before outputting this back into token. RL also complicates this because the "fitness" is now some arbitrary metric computed over an entire sequence of tokens.
>At all times the LLM is, indeed, predicting the next token
The point is that saying they're just "predicting the next token" is not at all explanatory nor providing insight. Saying the brain is just firing action potentials gives you no understanding about how the brain does what it does or what the space of its capabilities are. Similarly, predicting the next token tells you nothing about the capabilities of LLMs.
No. The how is relevant here because it leads to understanding of the resulting behavior.
If you train the LLM on a corpus that shows people saying the sky is red, you get an LLM that is predisposed to say the sky is red. This is true even if it's also trained on all of the science that explains how and why the sky is blue.
If it were to "figure out" or "reason", it would not have such a predisposition to emit "red" after "the sky is" just because that matches the reward during training.
In other words, the token prediction is important because it both explains the successes AND the failures of the LLM. If there were situations in which a bird could fail to fly, then how it tried to fly would also be crucial knowledge.
> What I mean, is the LLM is able to represent things in space . That part I don't understand.
Why do you think this is mutually exclusive to "LLM predicts the next token"?
If you tell someone from 19th century that bytes (just 0s and 1s!) can represent an opera, a song, or even a whole interactive experience, they might be really confused. But there is no reason they can't.
If you tell someone without math background that the sums of smaller and smaller sin waves can represent pretty much anything in our universe, they might be really confused. But there is no reason they can't.
There is simply no reason that a next-token predicator can't generate a nice-looking checkbox.
I understand that to be the "emergent abilities" which are spoken about. There are correlations in the dataset that are strong enough for it to seem to have an understanding which wasn't obvious it would have from simply "predicting the next word".
It's still predicting the next word. Somewhere in the gigantic dataset that the LLM was trained on, there is a phrase that says "gradient border" being in the vicinity of a CSS code that render the stuff. Therefore when you run it on an inference loop there's a good chance it output that CSS code when you tell it to render a "gradient border"
Multi-modal models that can understand visual input do exists, but no such visual reasoning process happened in the example you mentioned. Not unless you have a visual feedback loop in the coding harness.
I'm not dismissing the capability of "predicting the next word" however. The vast amount of training data enable extremely complex and useful behavior you just described.
For instance I’ve written a few custom languages to learn how to write a VM and the lexer/parser/compiler/etc. that it had never seen before and then just gave it the syntax which is different than what it had ever seen before. Simply due to the fact I made it and it had never been trained on it.
After giving it my documentation, it was able to write the language just like a language that it had been trained on. I’ve also seen this behavior at work where there are weird quirks to do things and definitely not standard and it can handle it.
Because in its training data there is information on how to map from documentation of a language to actual programs. This means that following the pattern it can map between documentation for any language to programs in that language.
But I think it will have difficulty in crossing paradigm boundaries, by simply using documentation.
That’s because it does not encode words or keywords or anything like that. It encodes their relationships. A formal language like a programming language are pretty compact. There’s not much variation between the C-like languages. Just like most Lisps (clojure, scheme, elisp, racket) are fairly similar to each other.
The exact syntax does not matter, only the grammar. If you give it the grammar, and then the keywords, it can find something that has similar grammar and then use your keywords.
I'd be very careful assuming something is not in an LLM's training set. Those data sets are truly vast. And, from experience, people tend to miss a lot of their content.
As a for instance, back in the day some academics wrote a paper that compared GPT 3.5 to a couple of inductive programming systems (including one of mine) on solving programming problems in a certain well-known esoteric language which I shall call "L". The task was to solve those programming problems one-shot. The authors asserted that the "L" problem sets were unlikely to be in 3.5's training set, but I found them without much search in a public github repo. I mean the entire dataset was right there. In this case the researchers are colleagues and friends and I know they weren't simply negligent or malicious, they just missed the fact that their "unlikely to be in the training set" data was on the web.
So I'd always assume that if an LLM can perform a task that's because it's seen examples of the task during its training.
Without forgetting that LLMs have this really shockingly powerful ability to interpolate between examples and they can improve their performance on say Task A by training on Task B, where A and B are different but similar.
e.g. they seem to get better at translating between language pairs of which they have few examples of parallel text by training on other pairs of languages for which they have more parallel text; they seem to learn something about language translation in general by training on more examples of translation. I haven't got a good reference on that handy but it's well-known (and of course over-hyped and exaggerated by tech CEOs).
So without wanting to diminish your work, I'd guess that your new language's syntax is different and novel but everything else about it is more ordinary and the similarities are such that an LLM can wing it and write you a lexer etc. After all, the whole point about parser generators and similar tools is that the task can be abstracted and separated from syntax in the first place.
In fact LLMs are very good at that sort of thing, filling in the blanks as it were. I'm old enough to remember the excitement about GPT 3.5 being able to form syntactically correct sentences with nonsensical words give to it.
For example, I just asked Chat [1]:
Hey chat. The gostak distims the doshes. What happens to the doshes?
LLMs fundamentally work by predicting the next word (token). But that should not be used to diminish their potential capabilities. It's like saying that human brains "just predict (or produce) the next electrical impulse". Fundamentally correct, but says nothing about the potential emergent capabilities of scaled-up systems that work like that.
Emergent properties of complex systems should not be diminished just because the underlying operating principle is simple.
Sorry you're being downvoted for asking a very reasonable question. I don't think any of the replies here address your question either.
If I can do my best to answer, Gemini is a multi-modal system. That means it's trained not only on text but also still images, video and also sound. The training happens in parallel and the representation of each modality is usually different, so the image recognition part is not trained on text tokens but pixels, the video part (probably) on video frames etc. There is some kind of integrated training that goes on so that text can be generated that is correlated to an image and so on, but I don't know the specifics about Gemini in particular. This kind of thing is not exactly new either, you can find systems that captioned images before the rise of LLMs simply by training on examples of images coupled to their textual descriptions.
In that sense it's not entirely correct to call Gemini an "LLM" because it's not only a "language" (or, more precisely, text) model. But LLM I guess becomes a bit of a shorthand for everything based on, or combined with, an LLM.
Anyway that's what's going on: it's not just predicting the next word. It's also predicting the next image frame or the next set of pixels etc associated with the next word.
I don't want to pretend I can explain LLMs, but the same "math" can be applied for visual and non visual things. The dot product of two vectors gives you the angle between them. This is true in 2 or 3 dimensions. But it's also true in 4, 5, 6...n dimensions even though we cannot visualize a 4d space. That it's an angle is relevant for you in the space you can comprehend, but for math or a machine it works in any number of dimensions. So it does need to understand anything visually if the math checks out.
LLMs are modelled to predict the next token, and are indeed trained to do so on enormous bodies of text. But to be really good at predicting the next token (word) at the end of a long string of text, you must understand what the text means. If I give you the entire text of a long novel and at the end ask you a single "yes/ no" question about the plot, you only need to emit a single token, but emitting the correct one implies having understood the plot of the novel. This is what LLMs do. They're generating meaningful, coherent text, which implies understanding and cognition at a level that is much deeper than that of the single token they generate at each forward pass. Internally, the LLM has learned to represent the meaning of the entire prompt text, the concepts it implies and its possible continuations far beyond the horizon of simply outputting the next token.
> This is what LLMs do. They're generating meaningful, coherent text
No, they generate grammatically coherent text. That is because human language grammars are fundamentally mathematical structures that can be approximated with matrix operations.
They don't generate meaningful text because they have no inherent knowledge of the world.
If you've used LLMs for any amount of time you've already noticed how often they get confused about numeric quantities - like confusing notions of "bigger than" and "less than" or being unable to count letters in words.
This is because any meaning in their output is only accidental.
The model is just trying to map from sequence to next token. You could say that it doesn't really care about the relationships between words/tokens - it is just being trained to learn the best attention/etc weights to make this mapping as accurate as possible.
The model could just as well learn to predict next token from gibberish text as long as there were some statistical gibberish regularities to learn. However, if you train it on real meaningful text then the statistical regularities it needs to learn (and will, thanks to gradient descent, and the capable architecture) will be those reflecting "token relationships" - grammar, semantics, etc.
So, you can say the "token relationships" (incl word meanings) are reflected in the statistical regularities of the training data, and the model architecture and training algorithm are just very capable of learning those regularities whatever they may be.
You can consider it related to Word2Vec word embeddings, which are based on the idea that the meaning of words comes from how they are used, which to a first approximation can be implemented by considering the meaning of words to be defined by the words they appear next to(!), which is what the Word2Vec embedding training algorithm does, and famous examples such as "(king - man) + woman = queen" prove that this is in fact learning the meanings of words.
At a high level, the text samples are how the relationships are derived. If we treat text samples as sequences of tokens, then the sequences of tokens describe the joint distributions they occur together which confers the relationship between them. Iirc, this is related to the idea of the distributional hypothesis in NLP: the idea the semantics of words should be similar if they occur in similar situations.
If I handed you thousands of documents which said “Jan-Michael Vincent” all over them, would you need to understand who that is in order to notice the relationship there?
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.
This article describes how Transformers work, but not really how LLMs work. Explaining the underlying architecture gives you about as much insight into how a modern LLM behaves as an breakdown of neuronal biochemistry and a few pathways does for the brain. Meaning, almost no insight at all.
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.
It is trained on its own slop. They haven’t trained these models on books for three years at this point. Only on generated slop. (And RL slop upvotes/downvotes from users)
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.
There is a very high volume of them being posted every day, and they are a significant % of the total. Also, writing is hard, LLM articles can be slop whilst also being better written than average.
This is essentially an open research question. ML theory is unfortunately very weak relative to where the empirics are. I think there's a relatively optimistic paper that was posted a while back here but I would also take it with a grain of salt.
There's of course empirical results and relatively weak theoretical results like the UAT but I also don't think that answers your question fully, especially since it seems impossible to definitively answer questions that the industry seems to betting on like whether or not there is a lower bound to their error rate or whether hallucination as a problem can be solved. We have much stronger ideas of what linear regression is doing relative to what LLMs are doing.
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.
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.
In fact, I would go so far as to say that publishing in this style stems from a lack of reading experience and writing experience, which does not bode well for someone pretending to be an expert. I gave this article to someone highly intelligent who doesn't know the first thing about how LLMs work internally, and she immediately called out that it reads like AI text.
You're not supposed to read it, just like you're not supposed to write anything anymore. Claude can read and write more than any human. We just lean back and relax now.
I don’t think it’s absolutely embarrassing. First of all, the point of the author writing at all is to aid understanding, not produce prose. So from that standpoint, what would be embarrassing would be to include incorrect facts that suggest a fundamental misunderstanding of the topic.
From my read, it is fine. The brief history of LLMs is complicated since every single component has papers introducing enhancements. So it’s easy to ignore them or get bogged down with details.
The author appears to be a security researcher learning about LLMs for the purpose of defending against common attacks. So this piece is that person giving themselves a crash course on the topic. The fact that they cleaned up their notes with an LLM is frankly completely irrelevant.
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.
> This was openai’s entire breakthrough. Making this particular model architecture larger leads to emergent capabilities
Basically, the bitter lesson: https://www.cs.utexas.edu/~eunsol/courses/data/bitter_lesson...
This interview https://youtu.be/oWOz2htozfI?si=qdQ0uZRoZOYeThOn from 2 days ago with a top researcher from OpenAI directly addresses the bitter lesson argument and the importance of scaling for the history of their models.
Isn't the bitter lesson basically the same as "The Unreasonable Effectiveness of Data" from 2009?
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 moral judgement, [0] shows Michael and Dalton from YC discussing why Ilya Sutskever had to leave Google to pursue what's now ChatGPT
[0] https://youtu.be/E8pvgN1j-Ck?t=748
There is a whole moral judgement to be made here...lets hope Ilya wont get too pissed off if somebody leaks the work of his new initiative...information wants to be free and all that...
Also would love to know if the same Legal team advised on Gemini...
And to make anyone who threatened to expose them “commit suicide”
He's a massive massive thief that people who have stolen far less from a convenience store have gone to prison for. The man is a villain.
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.
We know for sure the architecture of the open weights models since llama.cpp understands the architecture it needs to build to plug the weights into to run them. It's always possible that the latest closed model is doing something architecturally different than the open weights ones we know about, but judging by how close the large open weight models such as DeepSeek are to SOTA performance, this seems unlikely. When OpenAI first came out with their near-mythical "Strawberry" (aka "o1") thinking model there was all sorts of speculation that they had made some sort of architectural breakthough, but then DeepSeek replicated the capability and published how they did it, proving that it was just better training, not any architectural change.
There have been minor changes to the architecture over the years, but these are basically all efficiency tweaks such as various types of attention (some pioneered in the open by DeepSeek) that better scale to large context lengths, and the confusingly named "mixture of experts" architecture, but what's more notable really is how little the architecture has changed. The capability gains have been coming from better training and better data.
DeepSeek research:
- 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.
There are thousands of people working in top level labs. Somebody would leak it
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.
MoE and such are basically performance enhancements, they don't make the model smarter.
Performance enhancements are huge though.
If you can make the existing model faster, you can then save your inference budget to then make your model bigger, which then makes it smarter.
A lot of how smart the models can be comes down to budget. If you can make your existing thing cheaper, you can instead make it bigger for the same price.
> to then make your model bigger, which then makes it smarter
There's diminishing returns and at some point making a model bigger makes it dumber.
Not really “smarter” though? It’s just a big probability engine.
(Not trying to flame bait or anything. I just wouldn’t call LLM as exhibiting intelligence. It is great at making connections based on probability but doesn’t have a semantic understanding of what it is doing)
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...
> 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.
ReAct loops and tool-calling are the critical development feature. They turn a model from something that generates text into something that can independently influence the world around them.
Without agent features, you have just a chatbot.
There is a different way to look at this: that is, actually the Transformer is a minimal complication of what the based model is: in theory the neural network could be just a huge FFN, which is anyway the part of the Transformer that does the heavy lifting. But this would be impossibile to train both numerically and computationally, so the Transformer encodes enough priors for it to work: the causal attention, and the math tricks like the residuals and so forth. But the bottom line of all this is that the Transformer works because of the incredible semantical power of simple/huge FFNs.
Isn't that over-simplifying it a bit too much?
You can go another step - a FFN can be simulated on a Turing machine, thus it just exemplifies the incredible semantical power of the Turing machine model of computation. (in fact you don't even need a Turing machine, since there is no looping in one forward pass).
In theory you can run a huge FFN on the tiniest Turing machine, in practice it's much better to run a Transformer on the latest NVIDIA hardware. Or as they say "quantity (performance) has a quality all its own"
I was about to post your last point / quote. Going multigpu is relatively not so though but once you go multi-node you have distributed storage/io/compute system which is highly non trivial. Add that the long training times now you have robustness/fault-tolerantness concerns with hardware failures and restarts. Today’s training systems are engineering marvels.
Good point!
There is also the case for Markov chains being theoretically able to do these if tuned well. Or even SAT problem.
"LLM is just fancy autocomplete"
Same. I recall reading a paper by Stephen Wolfram after ChatGPT came out where he goes over how it works and what it does. Such a good piece and really got me going with this stuff. https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-...
What hopes/paths does a mere CS bachelor (not deep into stats/maths), and mid level dev (native mobile only; 10-15 years exp.), have about not only understanding it (maybe not fully) but getting possibly into this as a career? Not expecting churning out models and AI systems from the first weeks/months but entry/employment into this field?
(If I can be honest, and I am not being disparaging about anything lest it might seem so, I am looking at it from a career breakthrough/move perspective rather than an intellectual pursuit.)
I think you need to ask what you actually want to do with the AI.
If you want to be a researcher and come out with the next breakthrough, get ready to go back to school and learn some math.
If you just need to learn how to use it well and build things with it, then you probably just need to have a high level understanding.
Same as programming. I’d bet most programmers have no idea about the physics that makes computers work.
> I think you need to ask what you actually want to do with the AI.
What about improving the efficiency of token consumption, etc., basically opportunities for improving cost/performance?
I keep thinking there has to be a better way to share context with models than dumping entire gigantic skill files of raw text or otherwise into them - I'm betting there's a bunch of low-hanging fruit there.
There may be some low hanging fruit, but they're not available to people without deep understanding of how the math works. Well paid people already spend a lot of time thinking about this.
i am not sure acctually of the math is acctually that complicated/important. the math around neural networks is calculus/chain rule etc and for model comparison/validation one needs statistics. the required math for e.g. understand transformers is quite accessible.
You missed the third and most important reason to learn: fun.
Which sums up HN these days.
Im also a mere mortal, and after putting a few years into it IMO I’d say people make it much more complicated than it actually is. I failed most of my math courses for lack of interest, but found passion later with the aforementioned SLAM stuff. I have no doubt you or any other programmer could learn this stuff, especially since you can ask ChatGPT clarifying questions.
I have no idea about careers at this point, I’m still doing fancy IT work as my day job I and look away from the future with dread. I also haven’t been looking for new roles on the open job market, so who knows maybe there’s multimillion pay packages for anyone who can articulate how attention works in an interview.
I have a BS in CS (and have been in the field for 25 years). I couldn't understand the transformer architecture until I built a few myself. Here are the books I worked through. I now feel I have a very good understanding of modern LLMs.
https://www.amazon.com/Build-Large-Language-Model-Scratch/dp...
https://www.amazon.com/Build-DeepSeek-Scratch-Abhijit-Dandek...
I had the same reaction as you, when I learned in detail, how all this works. But then I also learned about superposition and compressed sensing, and now...I am not so sure anymore...
"Beating Nyquist with Compressed Sensing" - https://youtu.be/A8W1I3mtjp8
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.
Is it different though?
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.
It is different, but there may be some universal principles that are relevant more abstractly among both cases. Of particular interest is the empirical notion that statistical models of a certain form will always tend to "average out noise" and "learn meaningful patterns" up to the capacity that those models have for representing said patterns. A parallel notion to this is the hypothesis dubbed "thermodynamic origins of life". The universal principle binding these two seemingly disparate topics is one that seems to underlie any sense of "learning" in physical systems: that semantics of those systems depend on their representational power, and the semantics they do come to represent are the results of adding up many pushes in one "direction" (phase space / state space / etc.) encoding a pattern, and adding up many random noise jiggles will cancel out but give you a first-order sense of variance of those semantic features as expressed by the environment.
As this description is so overly abstract, an exercise for the reader is to try to work through an explanation of how, say, a river delta comes to "learn" about its environment by "reacting" to the influences at its borders, and how it "encodes" whatever it is that it learns in the substrate that it inhabits.
But … how close a simulation is it. I can see why people are wondering
In the 90s you didn't have norm layers, residuals, attention, and some more.
So you're missing a lot of the building blocks that make LLMs. It's not a matter of just having the compute.
I think the attention mechanism is so simple but so revolutionary that people forget it.
Like the best leaps in thinking, once it is made, is is immediately obvious and intuitive.
Almost everything in ML is like that. It seems so obvious in hindsight. It's maybe what I love most.
Residual connections are so simple, so obvious and so vital. Yet nobody came up with them until 2015?
I suspect it was considered many times, but the sheer computation scale would make it feel like obscene brute force. It feels like the right shape but too wild to think about implementing.
I think as time went on, and hardware got better, it seemed more reasonable to actually think about a viable implementation of what I think was a widespread intuition anyone in ML had that everything's context is everything.
It just seemed like a theoretical thing until hardware caught up. Maybe. Perhaps I'm applying a retrospective excuse to why it took so long.
Attention layers were not used in the 90s.
LLMs are semiotic infrastructure. You won’t find a better analogy. The cognitive frame won’t hold.
> I mean a brain is not just neurons with simple connections to each other.
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.
Apparently even single-celled protozoa can show learned trial and error behaviour.
Do you consider fungi animals or do you perhaps mean animals that don't have a brain/CNS?
Yes, protozoans don't have brains and yet they exhibit complex behavior.
Probably better to not simply reduce it by just saying X is Y then if it has all that extra complexity and capacity.
Could you perhaps cite the core papers for LLMs beyond „Attention is all you need“?
"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.
Chain of Thought was kind of an obvious solution that everybody knew was necessary by the time chatgpt / gpt4 came out. It was just a matter of time that frontier labs actually shipped it.
MoE was also pretty straightforward, just a bit surprising how well it worked (that you can get away with just 1/32 active parameters), but most researchers would have come up with it on their own probably.
The true ground breaking papers are the first two you mentioned (transformers and gpt2), and InstructGPT was also very surprising that it worked so well.
The GPT3 paper is a good starting point
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
Not a core paper, but I found Formal Algorithms for Transformers [1] (a Google paper from 2022) to have a great pedagogical style.
[1] https://arxiv.org/abs/2207.09238
I'll add in here https://web.stanford.edu/~jurafsky/slp3/, "Speech and Language Processing", with chapters that deal specifically with LLMs and transformers.
It's not just the architecture but also the data - the decoder only approach lets you train in parallel over blocks of text (no RNN serial waiting), that allows you train on much, much more data.
I would argue that those are not emergent property of the model, but a property of how humans find insights in a plausible guess.
Since you spent a month digging into this, can you recommend any materials/projects to look into to get a decent grasp of how they work?
I’d recommend my method of just drawing out the block diagram and drawing out + digging into the math at each step! I’m the kind of person who needs to take time to ask lots of questions before stuff clicks, and if you are too I strongly recommend it.
I picked it up from trying to teach myself that SLAM stuff. The papers are very short, but highly information dense and at the time there was no ChatGPT to help me. I got through them by just creeping my way through the math with a whiteboard, and something about drawing it out and having it there in my office made it all click. Trying to watch piecemeal lectures on YouTube or grind through foundational books like MVG just didn’t work for me, I used them instead as references for my drawings.
Same happened when I tried learning this GPT stuff. karpathy’s videos were out at the time, but I couldn’t really stay focused on them or connect the math with the code. Most other descriptions I could find were focused on getting you to use their inference library or harness. Assembling the picture together on my whiteboard by focusing on drawing out the block diagram continues to be my personal favorite method for deep understanding of complex systems.
Not OP but I worked through Sebastian Raschka's "Build a Large Language Model (From Scratch)" [0] and Raj Abhijit Dandekar's "Build a DeepSeek Model (From Scratch)" [1] books.
I don't think there is anything in a transformer I couldn't explain in the smallest detail now.
[0]: https://www.amazon.com/Build-Large-Language-Model-Scratch/dp...
[1]: https://www.amazon.com/Build-DeepSeek-Scratch-Abhijit-Dandek...
>I don't think there is anything in a transformer I couldn't explain in the smallest detail now.
If you're up for it I would love to know how and why positional encodings work
Learn about superposition and then you will see nobody really know why this stuff works. Its actually a good interview question to set the bar....
Well, as I suggested, working through the implementation yourself will give you that intuition. That said, I think the simplest way to explain why positional encodings are useful is that it gives the transformer just enough information to make attention meaningful without negatively impacting any parallel, content-based comparisons.
A vanilla self-attention layer is just a set of token vectors. Without positional info, swapping two identical embeddings changes very little about what attention can compute. We can "fix" this problem by using positional encodings. Text that has meaning isn't just a set of characters; the location and order of those characters is what provides meaning.
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.
> this is actually a great video overview of how LLM works and the tokenization part
Did you mean to link to the video? I would be interested.
Don't forget the stolen data from books and papers. You'll never get anything intelligent without using the stolen data they had access to.
WTF?
aka "the bitter lesson"
Sooooo just because you are able to understand it, it's not worth anything?
It doesn't has any impact?
Ah wait it does. Mh weird.
Why are you not creating a startup and get rich?
I mean there is a little something called compute. And other complexity that comes like writing code to efficiently distribute a model across machines.
> 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.
how did you know about the steps and there was math involved. i am curious about your process and you came up with what exactly to learn to unravel the mystery.
After building some toy LLMs on my own I came to realise that architecture is not the hard part. Train is.
That's easy to say AFTER you know the architecture.
Einstein special relativity is taught these days in high-schools. Doesn't mean it wasn't the very hard part at some point in time.
As they say, shoulders of giants.
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.
Those are all just optimizations.
We still don’t really know why they work, we just know how to build them.
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.
Completely agree!
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?)
Computer vision ends up displaying emergent behaviour. It just "figures out" things.
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.
I think those 2 are orthogonal. Math still works with Santa or the tooth fairy.
Maybe math works but critical thinking doesn’t. There are people who have lived for many decades without ever questioning insane b.s. they were taught as kids.
It is possible to have learned both things you know.
I had to learn maths early (not chinese or asian) and also a bunch of scary stories to make me behave. I would have been glad to learn about fairies.
They aren't stupid, but they aren't quite ready to handle the full responsibilities of the world and worry about things they don't need to worry about.
My son is very worried about black holes lately when he learned anything that goes into one can't get out. He's pretty concerned astronauts could get stuck in one some day. So I explained to him that Hawking radiation does actually mean you can eventually get out; it just takes some time.
I didn't think it pertinent to mention spaghettification, the fact anywhere near a black hole will be really hot, or that cosmic censorship means whatever Hawking-radiates from a black hole wouldn't be an astronaut anymore.
It was also fun to hear Hawking speak. He wanted to know if Hawking was a robot. I said no, but he has a robot talk for him. Not quite true, but close enough.
Because god forbid that childhood, the one time in your life when you don't have any responsibilities, should be fun.
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.
Childhood only lasts 13 to 15 years where I am. By the time you’re in high school, you can be expected to be responsible in some matters. By 22 you have 7 years of experience in making decisions for yourself.
Hm, I wonder if it's more that we're shocked such a simple thing (relatively speaking) can work so well.
It was precisely that for me! Another commenter captures it well; “the bitter lesson” indeed.
We do know how they work. They predict the next statistically most likely token.
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.)
Sufficiently good iterated next token prediction is an AI hard problem.
> statistically most likely token.
Statistically most likely in what context, given which preconditions? Because each prompt sequence is unique so the probability of any token following it is unknown.
It’s not unknown because that’s what the model computes. It’s matrix multiplication just like shaders.
And how do you know that the model computes it correctly?
Correctness is based on axioms and rules. You need to define your axioms and rules first before you can determine correctness.
If you’re talking about matrix multiplication, I can use mathematical rules and axioms and proves formally that the multiplication is correct. For next token prediction, I can prove that the set of tokens is finite and that the next token is always part of that set.
But things like grammar correctness, or semantic consistency over a few sentences are not hardcoded rules in the model. They’re emergent properties, mostly due to the amount and quality of data available for training. Quantization is mostly about how much we can shed without loosing a particular emergent properties (like dithering or psycho acoustic audio compression)
This "they just predict the next statistically most likely token" is such an handwavey and willfully misleading explanation, it's unreal, and I'm so fucking tired of seeing it so incessantly repeated. It's beyond asinine.
You know it perfectly damn well that a typical person's idea of statistics is not some insanely high cardinality stateful prediction, but a "well a coin toss is a 50:50, and a lottery win is a 1:100000000". You also know it perfectly damn well that as a result, people will just think that all the sentences chatbots ever produced to them were then just somewhere in the massive training set, letter by letter. This insinuation is often even explicitly appealed to.
And that picture is outright false. It's a statistical process, yes, so saying that it does what it does by "just doing statistics" is gonna be a generally correct description, but that's not at all inquisitive to how exactly does it do it, nor is it the zinger you think it is. If you did the aforementioned, you'd just get milquetoast nonsense, like you can see in the countless Markov-chain primers. And while the models do have a lot of the training set lossily captured, they do also absolutely generalize (that's how they can do that lossy compression), and you can quite literally find representations of those generalizations in them, and also see them activate.
It's like summarizing how any program works by just saying "well it just manipulates ones and zeroes". Not very informative, is it? Or how programs are written by just programmers sitting in a cushy office, ryhtmically pressing keys on a keyboard. Not a very fair or insightful description, which you'll know if you've done any amount of programming in your life on your own. Extends to all other white collar jobs too.
It's also not even true in the most literal sense: models can and do absolutely choose a less than maximally likely next token, that's what the various decoding parameters are for. "Maximally likely next token" further conviently skipping over how that likelihood is established in the first place, i.e. the literal point of the question, going in a cute little circle.
I'm so over this "stochastic parrot" bullshit.
fucking well said
Great, and won’t we all be just as surprised when human self-attentional control turns out to be just as simple or just as complex! Our minds as a strange fabric built of threads of recursions without the benefit of any explicit clock.
There's one thing I wish people understood about LLMs, and it doesn't really have anything to do with what's inside the neural network part. It's the fact that LLMs can only write in one direction — forward.
When you are writing an essay and realize midway through a sentence that what you've written doesn't make sense, you go back and edit. An LLM can't do that, the only thing it can do is keep on generating. Because training data typically contains full essays and not half-finished sentences which were then edited, LLMs have a strong preference for "saving face" and producing grammatically correct, internally coherent outputs. They will often do so even if the only way to write themselves out of the corner they wrote themselves into is to lie. To maintain internal coherence, they'll then repeat that lie for the rest of the response.
This is also why changing response structure used to affect LLM performance so dramatically. If you asked an LLM to solve a math problem and all-but-forced it to start with the answer, it would have had to calculate that answer before emitting any tokens, something which it very often wasn't able to do. If it was told to follow up the answer with an explanation, it would produce a plausible-sounding explanation to maintain coherence.
If, on the other hand, it was told to start by "thinking step by step", it would often be able to solve the first step, and then the next one given the results of the first, and so on, until it was able to reach the answer. Because the answer came last, it wasn't committing to anything, so had no reason to "save face" and lie.
This part of the problem is basically solved now with reasoning; reasoning is where all the step-by-step stuff happens, even if users aren't always able to see it. In the process of RLVR, models even train themselves into outputting phrases like "let me check my answer once again" in the chain-of-thought; those serve as their "life rafts" which they can use to both save face and change their answer.
In terms of our brains though we can only think forward as well (if forward is time). Our brain in the future says something we did in the past was wrong (part of the sentence we wrote) and that informs our body (the agent) to go back and fix it
What is the minimum Nvidia graphics card required to train a 1B model from scratch??
The part about positional encoding is not correct.
> The intuition: instead of adding position info to each token’s vector, RoPE rotates the vector by an angle that depends on its position
You can't rotate the token's entire vector (or all three vectors, whatever is being implied is unclear). You rotate each token's Query and Key vectors only, so dot product can be used to tell how far apart the tokens are when comparing token 1's Query vector to token 2's Key vector.
Positional embedding should just be explained after explaining the Query, Key and Value vectors. When the article explains those only after that, the reader is building up on a wrong intuition and it gets confusing.
Could you restate this another way: I don't follow.
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.
Comment <-> username synergy.
It's basically possible build an LLM using just routers+packets, and then hook them up to Wireshark to see it compute!
How would I set this up?
I'd recommend to maybe also specifically watching Karpathy's videos and focusing on the early parts where he specifically deals with tokenization / embeddings generation (which gets really overlooked), and he does this in most of his videos.
https://distill.pub/2019/activation-atlas/
I can only imagine what sort of visualizations are going on today inside of the AI labs.
Out of curiosity, I wondered if you could break a tokenizer by introducing weird characters not mapped to an id.
But apparently, they either just emit a [UNK] token or translate the unrecognized character into raw UTF-8 bytes.
A better blog on Transformers: https://www.aleksagordic.com/blog/transformer
Saying an article is of inferior quality just because editing was AI-assisted is like saying a book is lower quality just because it was printed rather than written by hand
You are exactly right! People do not find the writing obnoxious, they are backwards technophobes getting brought down by their superstitions.
No? One affects the actual text and the other doesn’t.
AI assisted is a stretch. And that analogy isn't even close to being relevant
This analogy makes absolutely no sense.
Rather interesting than clanker slop defenders downplay the clanker aspect and highlight the human by calling it "ai-assisted", which defeats their entire point.
I hope you do some introspection and start consciously recognizing that the human input and the clanker slop is just debasing it.
Not just that, I think a lot of people are going to waste their time losing the battle (and make no mistake, they will lose) fighting against AI writing without ever asking themselves what makes writing good in the first place.
There’s good AI writing and bad organic writing. But it’s easier to point out a few LLM-isms than to actually identify the problems with text.
> There's good AI writing
Sure, but the LLM-isms in AI writing are mentally exhausting to see in every way at this point.
The whole point of reading, frankly, is to understand the voice of other people. When you pass that through a distorted filter that makes everyone sound the same... its bad, lossy, frustrating communication
It's also dishonest. When you publish something that is direct output without your wording. Digital catfishing at best.
The only good AI writing is providing the prompt, because the question is way more interesting, and way more constructive to learning than the answer
I couldn't load the article directly due to an SSL issue, so here's the archive link:
https://archive.ph/aWtFG
Nice article but chain of thought is what makes frontier LLMs smart, not really the token loop
Is chain of thought same as test time compute?
Nice intro, gonna help me dig further a lot now. Thanks a ton.
I don't like how most LLM explainer articles and videos say that essentially a LLM " predicts the next word".
I'm a developer but not very good at maths and I still don't understand any of it.
A LLM clearly has some "visual" capacity. You ask Gemini to build something with Canvas and it's able to reason about the shape of things. Like recently I waanted a checkbox that has like a gradient flowing around the edge. It figured out it could use a radial gradient from the center of the checkbox, and overlay that with a small inner div so you only see the edge that looks like the gradient is circling around the checkbox.
How is that "predicting the next word"?
Not saying AI is intelligent or conscious or anything like that, but the algorithm clearly is far more complex than "predicting words".
What I mean, is the LLM is able to represent things in space . That part I don't understand.
I also still dont understand the relationship between the chat based LLM and the multi modal stuff. I think I read somewhere when image is generated it is also tokens?
Predicting a word is the final objective, as in the output of the model is a probability distribution of the next token. However, choosing the right token is more complicated than just regurgitating the training data (and you won't encounter an exact example in the training data, so you need to interpolate). This makes the model learn abstract representation of things that it is able to manipulate before outputting this back into token. RL also complicates this because the "fitness" is now some arbitrary metric computed over an entire sequence of tokens.
Your casual understanding is imprecise.
At all times the LLM is, indeed, predicting the next token. Anything it does emerges from that.
It did not "figure anything out". It predicted that text describing the use of a radial gradient was likely to follow text describing your problem.
>At all times the LLM is, indeed, predicting the next token
The point is that saying they're just "predicting the next token" is not at all explanatory nor providing insight. Saying the brain is just firing action potentials gives you no understanding about how the brain does what it does or what the space of its capabilities are. Similarly, predicting the next token tells you nothing about the capabilities of LLMs.
Lol, the bird did not 'fly' - it just flapped its wings and generated lift!
No. The how is relevant here because it leads to understanding of the resulting behavior.
If you train the LLM on a corpus that shows people saying the sky is red, you get an LLM that is predisposed to say the sky is red. This is true even if it's also trained on all of the science that explains how and why the sky is blue.
If it were to "figure out" or "reason", it would not have such a predisposition to emit "red" after "the sky is" just because that matches the reward during training.
In other words, the token prediction is important because it both explains the successes AND the failures of the LLM. If there were situations in which a bird could fail to fly, then how it tried to fly would also be crucial knowledge.
More like being suspended by a thread...
> What I mean, is the LLM is able to represent things in space . That part I don't understand.
Why do you think this is mutually exclusive to "LLM predicts the next token"?
If you tell someone from 19th century that bytes (just 0s and 1s!) can represent an opera, a song, or even a whole interactive experience, they might be really confused. But there is no reason they can't.
If you tell someone without math background that the sums of smaller and smaller sin waves can represent pretty much anything in our universe, they might be really confused. But there is no reason they can't.
There is simply no reason that a next-token predicator can't generate a nice-looking checkbox.
I understand that to be the "emergent abilities" which are spoken about. There are correlations in the dataset that are strong enough for it to seem to have an understanding which wasn't obvious it would have from simply "predicting the next word".
It's still predicting the next word. Somewhere in the gigantic dataset that the LLM was trained on, there is a phrase that says "gradient border" being in the vicinity of a CSS code that render the stuff. Therefore when you run it on an inference loop there's a good chance it output that CSS code when you tell it to render a "gradient border"
Multi-modal models that can understand visual input do exists, but no such visual reasoning process happened in the example you mentioned. Not unless you have a visual feedback loop in the coding harness.
I'm not dismissing the capability of "predicting the next word" however. The vast amount of training data enable extremely complex and useful behavior you just described.
What about things it wasn’t trained on?
For instance I’ve written a few custom languages to learn how to write a VM and the lexer/parser/compiler/etc. that it had never seen before and then just gave it the syntax which is different than what it had ever seen before. Simply due to the fact I made it and it had never been trained on it.
After giving it my documentation, it was able to write the language just like a language that it had been trained on. I’ve also seen this behavior at work where there are weird quirks to do things and definitely not standard and it can handle it.
Because in its training data there is information on how to map from documentation of a language to actual programs. This means that following the pattern it can map between documentation for any language to programs in that language.
But I think it will have difficulty in crossing paradigm boundaries, by simply using documentation.
That’s because it does not encode words or keywords or anything like that. It encodes their relationships. A formal language like a programming language are pretty compact. There’s not much variation between the C-like languages. Just like most Lisps (clojure, scheme, elisp, racket) are fairly similar to each other.
The exact syntax does not matter, only the grammar. If you give it the grammar, and then the keywords, it can find something that has similar grammar and then use your keywords.
I'd be very careful assuming something is not in an LLM's training set. Those data sets are truly vast. And, from experience, people tend to miss a lot of their content.
As a for instance, back in the day some academics wrote a paper that compared GPT 3.5 to a couple of inductive programming systems (including one of mine) on solving programming problems in a certain well-known esoteric language which I shall call "L". The task was to solve those programming problems one-shot. The authors asserted that the "L" problem sets were unlikely to be in 3.5's training set, but I found them without much search in a public github repo. I mean the entire dataset was right there. In this case the researchers are colleagues and friends and I know they weren't simply negligent or malicious, they just missed the fact that their "unlikely to be in the training set" data was on the web.
So I'd always assume that if an LLM can perform a task that's because it's seen examples of the task during its training.
Without forgetting that LLMs have this really shockingly powerful ability to interpolate between examples and they can improve their performance on say Task A by training on Task B, where A and B are different but similar.
e.g. they seem to get better at translating between language pairs of which they have few examples of parallel text by training on other pairs of languages for which they have more parallel text; they seem to learn something about language translation in general by training on more examples of translation. I haven't got a good reference on that handy but it's well-known (and of course over-hyped and exaggerated by tech CEOs).
So without wanting to diminish your work, I'd guess that your new language's syntax is different and novel but everything else about it is more ordinary and the similarities are such that an LLM can wing it and write you a lexer etc. After all, the whole point about parser generators and similar tools is that the task can be abstracted and separated from syntax in the first place.
In fact LLMs are very good at that sort of thing, filling in the blanks as it were. I'm old enough to remember the excitement about GPT 3.5 being able to form syntactically correct sentences with nonsensical words give to it.
For example, I just asked Chat [1]:
And it promptly answered: See, it even got the spelling right!_________________
[1] https://chatgpt.com/c/6a242b65-e248-83ed-9a6e-f238a1e871b6
LLMs fundamentally work by predicting the next word (token). But that should not be used to diminish their potential capabilities. It's like saying that human brains "just predict (or produce) the next electrical impulse". Fundamentally correct, but says nothing about the potential emergent capabilities of scaled-up systems that work like that.
Emergent properties of complex systems should not be diminished just because the underlying operating principle is simple.
>is the LLM is able to represent things in space
It is imitating the text written by humans who can represent things in space.
Sorry you're being downvoted for asking a very reasonable question. I don't think any of the replies here address your question either.
If I can do my best to answer, Gemini is a multi-modal system. That means it's trained not only on text but also still images, video and also sound. The training happens in parallel and the representation of each modality is usually different, so the image recognition part is not trained on text tokens but pixels, the video part (probably) on video frames etc. There is some kind of integrated training that goes on so that text can be generated that is correlated to an image and so on, but I don't know the specifics about Gemini in particular. This kind of thing is not exactly new either, you can find systems that captioned images before the rise of LLMs simply by training on examples of images coupled to their textual descriptions.
In that sense it's not entirely correct to call Gemini an "LLM" because it's not only a "language" (or, more precisely, text) model. But LLM I guess becomes a bit of a shorthand for everything based on, or combined with, an LLM.
Anyway that's what's going on: it's not just predicting the next word. It's also predicting the next image frame or the next set of pixels etc associated with the next word.
It can’t. It’s like a Redditor, it just repeats what it has seen other people say.
It has read all of stackoverflow, so it has seen your kind of problem before. Try asking it something really unusual and it will shit the bed.
> It’s like a Redditor, it just repeats what it has seen other people say.
Can stochastic parrots understand irony?
I don't want to pretend I can explain LLMs, but the same "math" can be applied for visual and non visual things. The dot product of two vectors gives you the angle between them. This is true in 2 or 3 dimensions. But it's also true in 4, 5, 6...n dimensions even though we cannot visualize a 4d space. That it's an angle is relevant for you in the space you can comprehend, but for math or a machine it works in any number of dimensions. So it does need to understand anything visually if the math checks out.
LLMs are modelled to predict the next token, and are indeed trained to do so on enormous bodies of text. But to be really good at predicting the next token (word) at the end of a long string of text, you must understand what the text means. If I give you the entire text of a long novel and at the end ask you a single "yes/ no" question about the plot, you only need to emit a single token, but emitting the correct one implies having understood the plot of the novel. This is what LLMs do. They're generating meaningful, coherent text, which implies understanding and cognition at a level that is much deeper than that of the single token they generate at each forward pass. Internally, the LLM has learned to represent the meaning of the entire prompt text, the concepts it implies and its possible continuations far beyond the horizon of simply outputting the next token.
> This is what LLMs do. They're generating meaningful, coherent text
No, they generate grammatically coherent text. That is because human language grammars are fundamentally mathematical structures that can be approximated with matrix operations.
They don't generate meaningful text because they have no inherent knowledge of the world.
If you've used LLMs for any amount of time you've already noticed how often they get confused about numeric quantities - like confusing notions of "bigger than" and "less than" or being unable to count letters in words.
This is because any meaning in their output is only accidental.
I do agree bigly. Calling what is basically a superhuman brain inside a computer just a "token predictor" is peak thinkslop.
Inside the magic AI box is literally nothing but this loop:
If you don't believe me then just download llama.cpp and see for yourself.Style nit: the transitions between dark-mode text and large diagrams with a snow white background are jarring.
> so the model figures out during training what each token should look for and what it should offer
But how does it learn this token-relationship?
All it has is many text samples, but still, nowhere it says how the tokens relate to each other, so where does this information come from?
The model is just trying to map from sequence to next token. You could say that it doesn't really care about the relationships between words/tokens - it is just being trained to learn the best attention/etc weights to make this mapping as accurate as possible.
The model could just as well learn to predict next token from gibberish text as long as there were some statistical gibberish regularities to learn. However, if you train it on real meaningful text then the statistical regularities it needs to learn (and will, thanks to gradient descent, and the capable architecture) will be those reflecting "token relationships" - grammar, semantics, etc.
So, you can say the "token relationships" (incl word meanings) are reflected in the statistical regularities of the training data, and the model architecture and training algorithm are just very capable of learning those regularities whatever they may be.
You can consider it related to Word2Vec word embeddings, which are based on the idea that the meaning of words comes from how they are used, which to a first approximation can be implemented by considering the meaning of words to be defined by the words they appear next to(!), which is what the Word2Vec embedding training algorithm does, and famous examples such as "(king - man) + woman = queen" prove that this is in fact learning the meanings of words.
At a high level, the text samples are how the relationships are derived. If we treat text samples as sequences of tokens, then the sequences of tokens describe the joint distributions they occur together which confers the relationship between them. Iirc, this is related to the idea of the distributional hypothesis in NLP: the idea the semantics of words should be similar if they occur in similar situations.
If I handed you thousands of documents which said “Jan-Michael Vincent” all over them, would you need to understand who that is in order to notice the relationship there?
How does evolution learn the form-fitness relationship?
It's the same thing here, you randomly try various token-relationship values and the ones which are slightly better will be favoured.
I thought Karpathy’s microgpt explain how LLMs work
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.
This article describes how Transformers work, but not really how LLMs work. Explaining the underlying architecture gives you about as much insight into how a modern LLM behaves as an breakdown of neuronal biochemistry and a few pathways does for the brain. Meaning, almost no insight at all.
Great insights. RL training is the key
But how do they “think”? This is the only repo that can tell you that.
https://github.com/space-bacon/SRT
this is hard to read...
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
It's just slop.
very interesting and useful!
find it difficult to engage with AI generated text. What am I getting here that I couldn't get from a chatbot.
Hopefully someone has asked right questions and removed confusing answers/hallucinations.
Is this text generated by AI? I couldn't tell but I'd believe it if it was.
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.
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.
It’s not a model of an author, it’s a model of documents. That’s not the same thing.
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.
It is trained on its own slop. They haven’t trained these models on books for three years at this point. Only on generated slop. (And RL slop upvotes/downvotes from users)
The voice of several passages resembles ChatGPT very closely.
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.
I don't understand how these AI written articles get so many votes.
There is a very high volume of them being posted every day, and they are a significant % of the total. Also, writing is hard, LLM articles can be slop whilst also being better written than average.
Next do "why LLMs work"
This is essentially an open research question. ML theory is unfortunately very weak relative to where the empirics are. I think there's a relatively optimistic paper that was posted a while back here but I would also take it with a grain of salt.
https://arxiv.org/abs/2604.21691
There's of course empirical results and relatively weak theoretical results like the UAT but I also don't think that answers your question fully, especially since it seems impossible to definitively answer questions that the industry seems to betting on like whether or not there is a lower bound to their error rate or whether hallucination as a problem can be solved. We have much stronger ideas of what linear regression is doing relative to what LLMs are doing.
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...
See Tegmark's "why does deep cheap learning work so well" (well not so cheap anymore...)
https://www.youtube.com/watch?v=5MdSE-N0bxs is remarkably prescient given that it was written before LLMs
Universal approximation theorem, embeddings, self-attention, gradient descent. And empirically, scaling laws.
Because there are patterns everywhere!
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.
What's the point of this? Im not here to engage with AI bots
accidentally quadratic
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.
In fact, I would go so far as to say that publishing in this style stems from a lack of reading experience and writing experience, which does not bode well for someone pretending to be an expert. I gave this article to someone highly intelligent who doesn't know the first thing about how LLMs work internally, and she immediately called out that it reads like AI text.
You're not supposed to read it, just like you're not supposed to write anything anymore. Claude can read and write more than any human. We just lean back and relax now.
I don’t think it’s absolutely embarrassing. First of all, the point of the author writing at all is to aid understanding, not produce prose. So from that standpoint, what would be embarrassing would be to include incorrect facts that suggest a fundamental misunderstanding of the topic.
From my read, it is fine. The brief history of LLMs is complicated since every single component has papers introducing enhancements. So it’s easy to ignore them or get bogged down with details.
The author appears to be a security researcher learning about LLMs for the purpose of defending against common attacks. So this piece is that person giving themselves a crash course on the topic. The fact that they cleaned up their notes with an LLM is frankly completely irrelevant.