[malwrar]

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.

[ekunazanu]

> 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...

[williamstein]

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.

[jochembrouwer]

So the take-away here is that we (as humans) try to model these AIs like humans, but eventually these AIs get better. Which to me seems like a logical conclusion if they can do "things" (like "learning" or pattern matching) much faster than we can (the compute). Then language in LLMs is a bottleneck, the AI is constrained by the language, and thus if we want to scale further we could let AI create its own language (we would then have to translate whatever it creates back to a language we understand). It is the same for instance if we check the language of Inuit (people who live in the north and make temorary shelters like igloos in the snow) they have multiple words/verbs to describe the snow, while in English we only have one (?): snow. In English we don't need more words (we can explain snow state using multiple words) but for the Inuit language it makes sense to create these new terms (would also make it easier and faster to communicate). So in some sense, all languages are then "newspeak" to whatever a general language is what researches or AI might come up with. If this sounds dumb let me know, but if you know some research in this general language direction (I'd assume general AI research) would love to see it!

[Vachyas]

I think one hypothesis along these lines is that, if allowed, due to the limitations of human language you described, LLMs will gravitate towards "inventing" their own language (which, due to training pressures, may even resemble english from the outside, but contain deeper, "true", meaning within), but that we should do our best to prevent this even if it bottlenecks reasoning capabilities since it would cut off our ability to read its "true" thoughts and detect misalignment

See: https://openai.com/index/chain-of-thought-monitoring/

Quote below:

  Chain-of-thought (CoT) reasoning models “think” in natural language understandable by humans. Monitoring their “thinking” has allowed us to detect misbehavior such as subverting tests in coding tasks, deceiving users, or giving up when a problem is too hard.

  We believe that CoT monitoring may be one of few tools we will have to oversee superhuman models of the future.

  We have further found that directly optimizing the CoT to adhere to specific criteria (e.g. to not think about reward hacking) may boost performance in the short run; however, it does not eliminate all misbehavior and can cause a model to hide its intent. We hope future research will find ways to directly optimize CoTs without this drawback, but until then

  We recommend against applying strong optimization pressure directly to the CoTs of frontier reasoning models, leaving CoTs unrestricted for monitoring.

  We understand that leaving CoTs unrestricted may make them unfit to be shown to end-users, as they might violate some misuse policies. Still, if one wanted to show policy-compliant CoTs directly to users while avoiding putting strong supervision on them, one could use a separate model, such as a CoT summarizer or sanitizer, to accomplish that.

[cavemandaveman]

My understanding is that the Inuit snow claim is a bit of a myth. Beyond English words like slush, sleet, powder, hardpack, flurry, blizzard, etc you can also say "fluffy snow", "wet snow", etc. The Inuit language is just smooshing the adjective so you get something like "wetsnow" as one word.

[xnx]

Isn't the bitter lesson basically the same as "The Unreasonable Effectiveness of Data" from 2009?

[swyx]

not exactly, bitter lesson is one meta-level up from "scale eats everything". this is a common misunderstanding of bitter lesson that rich sutton has been fighting ever since the thing was written. in rich's own words[1], the modern summary is

> Don’t be distracted by human knowledge, as AI has been historically.

> Instead focus on methods for creating knowledge that scale with computation, like search and learning.

so the lesson is choose methods that scale with computation, not just that blindly scaling up anything (data, params, people, whatever) works, it is choosing the right x axis and the right scaling laws consistently wins out in the long run despite short term wins from other methods.

1: https://x.com/RichardSSutton/status/2056419165502935198

[jfim]

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.

[gobdovan]

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

[root-parent]

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...

[miltonlost]

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.

[someguyiguess]

And to make anyone who threatened to expose them “commit suicide”

[achrono]

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.

[HarHarVeryFunny]

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.

[gobdovan]

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.

[swyx]

> 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

c.f. hardware lotter https://arxiv.org/abs/2009.06489

[matusp]

There are thousands of people working in top level labs. Somebody would leak it

[ai_slop_hater]

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.

[otabdeveloper4]

MoE and such are basically performance enhancements, they don't make the model smarter.

[yababa_y]

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...

[jmalicki]

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.

[TheHalfDeafChef]

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)

[otabdeveloper4]

> 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.

[fizx]

Performance enhancements are what allow you to train a bigger model.

[locknitpicker]

> 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.

[galaxyLogic]

The big breakthrough is we can interact with the agents using natural language - because of the LLM.

It is the combination of LLM and agent-harnesses that make it look really smart. Agent-harness is a programmatic device that lets us tap into the vast knowledge in the LLM.

It is probabaly true that many TV-commentators fail to appreciate this fact and therefore think LLMs are super-intelligent. No, it is the combination of LLM and the programmatic agent-haness that is the breakthrough.

An interesting thought is that the LLM could in theory code the agent-harrness, start it running every time we interact with it. Currently the agent-harrness I think is pretty static I think. In theory it could be dynamically created for every task. Would that make it better don't know.

[locknitpicker]

> The big breakthrough is we can interact with the agents using natural language - because of the LLM.

Without ReAct and tool calling, all you have is a chatbot. That's useful, but it's just a chatbot.

ReAct loops and tool calling is what unblocks high value usecases. It enables systems to actually address free-form problem statements, gather data that is not a part of their training set, inspect the current state of services,and trigger actions in external systems. This goes well beyond mere chatbots.

> It is the combination of LLM and agent-harnesses that make it look really smart.

It's really not about "smart". It's about autonomous systems, and being able to consume and analyze new data, and trigger actions in external systems.

[Chu4eeno]

It's not very novel, though, it's a fairly obvious step once you get something that can operate iteratively and largely independent, there were a ton of people trying to get LLMs to loop on their own even before deepseek r1.

And I remember talking about goal directed behavior (which what people are calling "agents" now don't seem to properly have) and autonomous operation decades ago in the intelligent agent course at uni, including react loops.

So no, the huge step with LLMs really was just that attention mechanism from that translation paper everyone forgot until Google brought its marketing to it, everything else is either just optimization/scaling, more money or old ideas suddenly relevant.

[locknitpicker]

> It's not very novel, though (...)

I completely disagree. The rollout of agentic tools, and even support for agent mode in IDEs, is the whole value proposition of AI code assistant services.

Otherwise you'd just have a glorified search engine in a chat window.

> (...) it's a fairly obvious step once you get something that can operate iteratively and largely independent,

There's some confusion in your reply. ReAct loops is exactly what this "operate iteratively and largely independently" represents.

[forestsitter]

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-...

[antirez]

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.

[dist-epoch]

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"

[musebox35]

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.

[zbendefy]

Good point!

There is also the case for Markov chains being theoretically able to do these if tuned well. Or even SAT problem.

[CGMthrowaway]

"LLM is just fancy autocomplete"

[galaxyLogic]

LLM is an Oracle

[Chu4eeno]

Well, it seems obvious (in hindsight) that attention is vital to actually do anything useful with all those matrices getting smashed together in increasingly unhinged ways in more and more layers.

Which is incidentally more or less the only thing I remember about Global Workspace Theory (attention facilitating consciousness in a way iirc).

[10GBps]

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.

[ctolsen]

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.

[rfv6723]

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.

[uoaei]

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.

[0xbadcafebee]

Chomsky did the opposite of what you're saying. He didn't ignore spoken language. He said that human vocalization is independent of language, and that the way our brains can manipulate and use sound (a cognitive capability, not specifically an aural one) is the fundamental differentiator that allows us to make compound ideas, and our specific use of language is a byproduct.

Example: a programming language's capability to produce complex software does not come from some inherent quality of language. It comes from binary. 0's and 1's, representing basic logic, and that being built on top of with an abstract "tool" called a language. If the binary logic didn't work, the language wouldn't do anything.

A dolphin can make sounds, and technically has a language, but they can't manipulate or recursively compound concepts (as far as we can tell) in order to create modified ideas. If they could, they probably would have come up with vastly more advanced fishing methods than the (admittedly novel) ones they have now.

[zaphirplane]

But … how close a simulation is it. I can see why people are wondering

[otabdeveloper4]

> 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.

[adammarples]

Apparently even single-celled protozoa can show learned trial and error behaviour.

[formerly_proven]

Do you consider fungi animals or do you perhaps mean animals that don't have a brain/CNS?

[otabdeveloper4]

Yes, protozoans don't have brains and yet they exhibit complex behavior.

[redox99]

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.

[sirsinsalot]

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.

[bonoboTP]

Yes, but it wasn't invented from nothing in 2017. Soft attention existed in other applications like information retrieval, Nonlocal networks had similar ideas as well. But it wasn't seen or used as a fundamental building block. But it wasn't something out of the blue either.

[redox99]

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?

[sirsinsalot]

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.

[redox99]

People definitely wanted to train deep networks before, but didn't know how. They evdn tried things like training layers independently.

I don't think it was intuitive to anyone back then, the vanishing gradient problem was a big deal since the dawn of NNs. I'm not sure what you mean by sheer computation, residuals allow you to have deep networks instead of shallow and wide ones. You can have equivalent parameter count.

[bonoboTP]

Attention layers were not used in the 90s.

[foxes]

Probably better to not simply reduce it by just saying X is Y then if it has all that extra complexity and capacity.

[spacebacon]

LLMs are semiotic infrastructure. You won’t find a better analogy. The cognitive frame won’t hold.

[crossroadsguy]

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.)

[2muchcoffeeman]

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.

[bluerooibos]

> 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.

[coliveira]

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.

[la_fayette]

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.

[sirsinsalot]

You missed the third and most important reason to learn: fun.

Which sums up HN these days.

[malwrar]

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.

[LatencyKills]

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...

[tinktank]

Has it given you enough of an understanding that you can pick up and follow research papers or did you have to do more to achieve that?

[LatencyKills]

I went this route because I had difficulty visualizing the content of the Attention Is All You Need paper. After going through both books, I can now understand every part of that paper.

I'm currently working on a robotics project that uses Nvidia's GR00T N1 model, and I was able to understand the research paper. [0]

[0]: https://arxiv.org/abs/2503.14734

[tinktank]

Thank you for the information.

[wuschel]

Could you perhaps cite the core papers for LLMs beyond „Attention is all you need“?

[sigmoid10]

"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.

[redox99]

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.

[sigmoid10]

Reasoning is a little bit more than just "baked in" chain of thought prompting. The important takeaway here was that it is not realized at the architecture level of the neural network. And you could say that all these things regarding LLMs were pretty straightforward. But only in hindsight, otherwise there wouldn't have been so much time and effort spent on intermediaries. Breakthroughs mean people simply didn't know stuff before, even if it seems easy with the benefit of hindsight.

[blackbear_]

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

[sharma-arjun]

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

[barrenko]

I'll add in here https://web.stanford.edu/~jurafsky/slp3/, "Speech and Language Processing", with chapters that deal specifically with LLMs and transformers.

[root-parent]

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

[dominotw]

> 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.

[GardenLetter27]

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.

[darksim905]

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.

[cloche]

> 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.

[darksim905]

Oops! Here you go - https://www.youtube.com/watch?v=7xTGNNLPyMI

[sesm]

I would argue that those are not emergent property of the model, but a property of how humans find insights in a plausible guess.

[coliveira]

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.

[giardini]

WTF?

[bluerooibos]

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?

[LatencyKills]

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...

[hackinthebochs]

>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

[LatencyKills]

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.

[root-parent]

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....

[malwrar]

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.

[Gmolomo]

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?

[sarjann]

I mean there is a little something called compute. And other complexity that comes like writing code to efficiently distribute a model across machines.

[golergka]

After building some toy LLMs on my own I came to realise that architecture is not the hard part. Train is.

[dist-epoch]

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.

[pkoird]

aka "the bitter lesson"

[faurroar]

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.

[jumploops]

Those are all just optimizations.

We still don’t really know why they work, we just know how to build them.

[trollbridge]

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.

[jumploops]

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?)

[galaxyLogic]

Isn't the LLM simply predicting what should be the next sentences after user's input, using its algorithm and data it has exatrcted from existing texts on the internet. The algorithm that does that could have many different designs, some better some worse for the purpose of predicting what output makes most sense next?

So what is it that we don't understand about why theyr work? The algorithm? We have the code. Why the specific algorithm makes such good predictions? I see it as a generalization of trying to predict who wins Kentucky Derby.

[trollbridge]

Computer vision ends up displaying emergent behaviour. It just "figures out" things.

[ai_slop_hater]

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.

[baq]

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.

[Izkata]

Another example, my parents taught me to read at about 4 years old. When I started kindergarten (the year before 1st grade in the US), the teachers and principal didn't believe I could read and I had to prove it by reading a book to them I'd never seen before.

I think they're right that kids (at least in the US) are generally treated as less capable than they are, and it ends up slightly delaying their development.

[ai_slop_hater]

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.

[trollbridge]

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.

[pmg101]

Because god forbid that childhood, the one time in your life when you don't have any responsibilities, should be fun.

[ai_slop_hater]

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.

[slopinthebag]

Hm, I wonder if it's more that we're shocked such a simple thing (relatively speaking) can work so well.

[malwrar]

It was precisely that for me! Another commenter captures it well; “the bitter lesson” indeed.

[otabdeveloper4]

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.)

[throw310822]

> 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.

[skydhash]

It’s not unknown because that’s what the model computes. It’s matrix multiplication just like shaders.

[throw310822]

And how do you know that the model computes it correctly?

[perching_aix]

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.

[stevenhuang]

I don't even try anymore. The people who still parrot the stochastic parrot bit this late in the game will simply never understand it.

[otabdeveloper4]

LLMs predict next token one at a time. (Stochastically.) Literally. It's what they do. That's how they literally work.

If you don't believe me, download llama.cpp and see for yourself.

P.S. I write inference backends in C++ every day. The gall of people like you who figured out how to prompt Claude and think they're hot shit now is simply unbelievable.

[klempner]

Sufficiently good iterated next token prediction is an AI hard problem.

[firemelt]

fucking well said

[robwwilliams]

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.