How LLMs work • Hacker News Board

Re: How LLMs work

#211

Post by Izkata » Sat, Jun 06, 2026, 7:19 PM UTC

mattnewton wrote: post #28

zemo wrote: post #21

Earlier quoted context omitted.

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-Towe…

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.

A lot of the common patterns people ping as AI (like "it's not X, it's Y"*) are marketing-speak, of which there's a lot of on the internet. It's applying existing patterns in unusual locations, ignoring the original context.

The one they're pointing out (the short punchy sentences) also apply to things like politicians and news articles. Blog posts are a bit odd.

* And here I mean those literal exact words. People are also extrapolating to similar patterns that use different or more words than "it's not" and "it's", but those flow better and aren't what I'm referring to here.

Re: How LLMs work

#212

Post by layla5alive » Sat, Jun 06, 2026, 7:34 PM UTC

antran22 wrote: post #91

AltruisticGapHN wrote: post #88

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…

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…

I do not think you are correct.

Re: How LLMs work

#213

Post by layla5alive » Sat, Jun 06, 2026, 7:39 PM UTC

YeGoblynQueenne wrote: post #164

360MustangScope wrote: post #109

Earlier quoted context omitted.

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 la…

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…

I could answer the same query the same way as a child.

Re: How LLMs work

#214

Post by layla5alive » Sat, Jun 06, 2026, 7:42 PM UTC

Marha01 wrote: post #122

AltruisticGapHN wrote: post #88

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…

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 beca…

So much this - so many people seem to miss the forest from the trees that emergent properties are not bound to the complexity of the underlying mechanics.

All of life arises (maybe) from very simple subatomic particles, and at each stage you can repeat this refrain, complexity increasing as you stack.

Re: How LLMs work

#215

Post by layla5alive » Sat, Jun 06, 2026, 7:55 PM UTC

raincole wrote: post #176

AltruisticGapHN wrote: post #88

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…

> 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 backgroun…

You're talking about simple compression and encoding mechanisms and by implication you're drawing an analogy to an LLM encoding/compressing the information..

And sure, it does, but the person you're replying to was trying to understand why it also seems to reason about the query to give an answer consistent with it, despite not being trained on that query or answer. Your answer seems to imply that its just another slick complex encoding.

But the emergent property of trillions of digital neurons predicting the next token is that in the process of being trained to do so, they can also learn to reason.

At some scale, it is efficient to encode cognition which is capable of mimicing the cognition which generated the input tokens.

Re: How LLMs work

#216

Post by galaxyLogic » Sat, Jun 06, 2026, 7:59 PM UTC

locknitpicker wrote: post #181

jfim wrote: post #17

Earlier quoted context omitted.

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 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.

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.

Re: How LLMs work

#217

Post by galaxyLogic » Sat, Jun 06, 2026, 8:03 PM UTC

CGMthrowaway wrote: post #180

zbendefy wrote: post #129

Earlier quoted context omitted.

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"

LLM is an Oracle

Re: How LLMs work

#218

Post by Gareth321 » Sat, Jun 06, 2026, 8:12 PM UTC

spaceisballer wrote: post #207

Solid read especially for someone not in this field. While everything I’ve learned about LLMs has been pretty interesting, all I can say for sure is that I’m more and more skeptical about wide scale adoption. Consumers are being pushed almost to the level of coercion to utilize LLMs. Especially in the case of the government, who in most cases will get a free pass for a year to help build that addiction before the rea…

I’m kind of amazed when I read comments like this, but I have to remind myself that I work in an industry which use these tools at the cutting edge and see what they can really do. In the space of 18 months I changed from skeptic to the belief that our world is going to RAPIDLY change, and soon.

I sense that statistics and benchmarks and research and statements from the world’s greatest academics won’t sway you, so maybe I’ll give you a personal anecdote. I have suffered from a condition my whole life called bile acid malabsorption. It caused chronic diarrhoea, pain, arthritis, dehydration, insomnia, and more. I spent decades searching for an answer. Dozens of different tests. Eventually doctors just said I was depressed and prescribed me antidepressants. They didn’t help. On the bad days I considered ending my life.

In desperation I turned to ChatGPT. Over months I described my symptoms, triggers, diet, timing, etc. We “sparred” with each other over assumptions and ideas. I gave it all my medical history. All the tests. Eventually it concluded that BAM was likely (plus another few options). So I pushed my doctor for a specialist referral. The specialist agreed to a scan based on the symptoms. It was confirmed. I’ve been taking some cheap medication each day now and it has changed my life.

I know others for whom ChatGPT has changed their lives in similar ways. Research shows LLMs are better than doctors already in many cases at diagnosis. They are improving at an exponential rate.

Re: How LLMs work

#219

Post by hiworld6543 » Sat, Jun 06, 2026, 8:14 PM UTC

vocram wrote: post #68

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

Saying the article was AI assisted is also wrong. Current consumer models can read a vehicle user manual in moments and indicate probable writing errors. This article had a few stylistic errors (or choices) that were irrelevant, except to prove the author is human. If he/she had used AI even in an assistive manner, knowing the nitpicking behavior of the intended audience, it would have had none.

Also, the author’s other public writings have similar errors/choices in style. When “consumer AI” writes or rewrites, it’s impossible for it to mimic one’s writing style so similarly. Literally impossible, because it can’t disregard everything else that it “knows” (voluminous training, guardrails, interface design, social boundaries) for it to 1. Disregard all that training after processing the user’s prompt 2. output a complete article in the user’s style 3. Turn back on its knowledge 4. then continue to function. That’s just not how consumer products work.

Re: How LLMs work

#220

Post by Borealid » Sat, Jun 06, 2026, 8:14 PM UTC

layla5alive wrote: post #203

Borealid wrote: post #148

Earlier quoted context omitted.

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"…

You can also teach humans science and math and then they can be trained by a cult to not use any of that reasoning when emitting canned responses that they were rewarded by the cult for internalizing during their training. "Fake News!" You're caught up on the mechanics of token processing (floating point matrix ALU math) and ignoring the context that p(next token) as a function being "computed" is doing so over a tri…

I completely agree that humans sometimes are not applying reasoning to things.

I'm not trying to argue a model cannot "reason" or have "cognition", whatever those things are. I'm only saying that it's absolutely the case that whatever those things are, they come from its mechanism of predicting one token at a time ad infinitum, and that throwing away a deep understanding in favor of a shallow one is foolish. Just because it might seem to be "reasoning" does not mean it IS doing so, and certainly giving the appears of reasoning does not mean it is NOT a token predictor.

If I knew deeply how the human brain works I would use that understanding instead of saying things like "this person reasons" or "this person thinks".

In summary, I'm not "caught up in" anything - I'm just trying to point out that the original poster here is incorrect in saying that clearly LLMs aren't working through token prediction. They are, and all their behavior is 100% explained by token prediction. That's more than enough for interesting behavior!