That's all very impressive, to be sure. But are you sure you're getting the point? As of 2025, LLMs are now very good at writing new code, creating new imagery, and writing original text. They continue to improve at a remarkable rate. They are helping their users create things that didn't exist before. Additionally, they are now very good at searching and utilizing web resources that didn't exist at training time.
So it is absurdly incorrect to say "they can only reproduce the past." Only someone who hasn't been paying attention (as you put it) would say such a thing.
> They are helping their users create things that didn't exist before.
That is a derived output. That isn't new as in: novel. It may be unique but it is derived from training data. LLMs legitimately cannot think and thus they cannot create in that way.
I will find this often-repeated argument compelling only when someone can prove to me that the human mind works in a way that isn't 'combining stuff it learned in the past'.
5 years ago a typical argument against AGI was that computers would never be able to think because "real thinking" involved mastery of language which was something clearly beyond what computers would ever be able to do. The implication was that there was some magic sauce that human brains had that couldn't be replicated in silicon (by us). That 'facility with language' argument has clearly fallen apart over the last 3 years and been replaced with what appears to be a different magic sauce comprised of the phrases 'not really thinking' and the whole 'just repeating what it's heard/parrot' argument.
I don't think LLM's think or will reach AGI through scaling and I'm skeptical we're particularly close to AGI in any form. But I feel like it's a matter of incremental steps. There isn't some magic chasm that needs to be crossed. When we get there I think we will look back and see that 'legitimately thinking' wasn't anything magic. We'll look at AGI and instead of saying "isn't it amazing computers can do this" we'll say "wow, was that all there is to thinking like a human".
> 5 years ago a typical argument against AGI was that computers would never be able to think because "real thinking" involved mastery of language which was something clearly beyond what computers would ever be able to do.
Mastery of words is thinking? In that line of argument then computers have been able to think for decades.
Humans don't think only in words. Our context, memory and thoughts are processed and occur in ways we don't understand, still.
There's a lot of great information out there describing this [0][1]. Continuing to believe these tools are thinking, however, is dangerous. I'd gather it has something to do with logic: you can't see the process and it's non-deterministic so it feels like thinking. ELIZA tricked people. LLMs are no different.
That's the crazy thing. Yes, in fact, it turns out that language encodes and embodies reasoning. All you have to do is pile up enough of it in a high-dimensional space, use gradient descent to model its original structure, and add some feedback in the form of RL. At that point, reasoning is just a database problem, which we currently attack with attention.
No one had the faintest clue. Even now, many people not only don't understand what just happened, but they don't think anything happened at all.
There's no such thing as people without language, except for infants and those who are so mentally incapacitated that the answer is self-evidently "No, they cannot."
Language is the substrate of reason. It doesn't need to be spoken or written, but it's a necessary and (as it turns out) sufficient component of thought.
There are quite a few studies to refute this highly ignorant comment. I'd suggest some reading [0].
From the abstract:
"Is thought possible without language? Individuals with global aphasia, who have almost no ability to understand or produce language, provide a powerful opportunity to find out. Astonishingly, despite their near-total loss of language, these individuals are nonetheless able to add and subtract, solve logic problems, think about another person’s thoughts, appreciate music, and successfully navigate their environments. Further, neuroimaging studies show that healthy adults strongly engage the brain’s language areas when they understand a sentence, but not when they perform other nonlinguistic tasks like arithmetic, storing information in working memory, inhibiting prepotent responses, or listening to music. Taken together, these two complementary lines of evidence provide a clear answer to the classic question: many aspects of thought engage distinct brain regions from, and do not depend on, language."
> I will find this often-repeated argument compelling only when someone can prove to me that the human mind works in a way that isn't 'combining stuff it learned in the past'.
That is a pedantic distinction. You can create something that didn't exist by combining two things that did exist, in a way of combining things that already existed. For example, you could use a blender to combine almond butter and sawdust. While this may not be "novel", and it may be derived from existing materials and methods, you may still lay claim to having created something that didn't exist before.
For a more practical example, creating bindings from dynamic-language-A for a library in compiled-language-B is a genuinely useful task, allowing you to create things that didn't exist before. Those things are likely to unlock great happiness and/or productivity, even if they are derived from training data.
> That is a pedantic distinction. You can create something that didn't exist by combining two things that did exist, in a way of combining things that already existed.
This is the definition of a derived product. Call it a derivative work if we're being pedantic and, regardless, is not any level of proof that LLMs "think".
Yeah you’ve lost me here I’m sorry. In the real world humans work with AI tools to create new things. What you’re saying is the equivalent of “when a human writes a book in English, because they use words and letters that already exist and they already know they aren’t creating anything new”.
Why is that kind of thinking required to create novel works?
Randomness can create novelty.
Mistakes can be novel.
There are many ways to create novelty.
Also I think you might not know how LLMs are trained to code. Pre-training gives them some idea of the syntax etc but that only gets you to fancy autocomplete.
Modern LLMs are heavily trained using reinforcement data which is custom task the labs pay people to do (or by distilling another LLM which has had the process performed on it).
Could you give us an idea of what you’re hoping for that is not possible to derive from training data of the entire internet and many (most?) published books?
This is the problem, the entire internet is a really bad set of training data because it’s extremely polluted.
Also the derived argument doesn’t really hold, just because you know about two things doesn’t mean you’d be able to come up with the third, it’s actually very hard most of the time and requires you to not do next token prediction.
The emergent phenomenon is that the LLM can separate truth from fiction when you give it a massive amount of data. It can figure the world out just as we can figure it out when we are as well inundated with bullshit data. The pathways exist in the LLM but it won’t necessarily reveal that to you unless you tune it with RL.
> The emergent phenomenon is that the LLM can separate truth from fiction when you give it a massive amount of data.
I don't believe they can. LLMs have no concept of truth.
What's likely is that the "truth" for many subjects is represented way more than fiction and when there is objective truth it's consistently represented in similar way. On the other hand there are many variations of "fiction" for the same subject.
They can and we have definitive proof. When we tune LLM models with reinforcement learning the models end up hallucinating less and becoming more reliable. Basically in a nut shell we reward the model when telling the truth and punish it when it’s not.
So think of it like this, to create the model we use terabytes of data. Then we do RL which is probably less than one percent of additional data involved in the initial training.
The change in the model is that reliability is increased and hallucinations are reduced at a far greater rate than one percent. So much so that modern models can be used for agentic tasks.
How can less than one percent of reinforcement training get the model to tell the truth greater than one percent of the time?
The answer is obvious. It ALREADY knew the truth. There’s no other logical way to explain this. The LLM in its original state just predicts text but it doesn’t care about truth or the kind of answer you want. With a little bit of reinforcement it suddenly does much better.
It’s not a perfect process and reinforcement learning often causes the model to be deceptive an not necessarily tell the truth but it more gives an answer that may seem like the truth or an answer that the trainer wants to hear. In general though we can measurably see a difference in truthfulness and reliability to an extent far greater than the data involved in training and that is logical proof it knows the difference.
Additionally while I say it knows the truth already this is likely more of a blurry line. Even humans don’t fully know the truth so my claim here is that an LLM knows the truth to a certain extent. It can be wildly off for certain things but in general it knows and this “knowing” has to be coaxed out of the model through RL.
Keep in mind the LLM is just auto trained on reams and reams of data. That training is massive. Reinforcement training is done on a human basis. A human must rate the answers so it is significantly less.
> The answer is obvious. It ALREADY knew the truth. There’s no other logical way to explain this.
I can think of several offhand.
1. The effect was never real, you've just convinced yourself it is because you want it to be, ie you Clever Hans'd yourself.
2. The effect is an artifact of how you measure "truth" and disappears outside that context ("It can be wildly off for certain things")
3. The effect was completely fabricated and is the result of fraud.
If you want to convince me that "I threatened a statistical model with a stick and it somehow got more accurate, therefore it's both intelligent and lying" is true, I need a lot less breathless overcredulity and a lot more "I have actively tried to disprove this result, here's what I found"
You asked for something concrete, so I’ll anchor every claim to either documented results or directly observable training mechanics.
First, the claim that RLHF materially reduces hallucinations and increases factual accuracy is not anecdotal. It shows up quantitatively in benchmarks designed to measure this exact thing, such as TruthfulQA, Natural Questions, and fact verification datasets like FEVER. Base models and RL-tuned models share the same architecture and almost identical weights, yet the RL-tuned versions score substantially higher. These benchmarks are external to the reward model and can be run independently.
Second, the reinforcement signal itself does not contain factual information. This is a property of how RLHF works. Human raters provide preference comparisons or scores, and the reward model outputs a single scalar. There are no facts, explanations, or world models being injected. From an information perspective, this signal has extremely low bandwidth compared to pretraining.
Third, the scale difference is documented by every group that has published training details. Pretraining consumes trillions of tokens. RLHF uses on the order of tens or hundreds of thousands of human judgments. Even generous estimates put it well under one percent of the total training signal. This is not controversial.
Fourth, the improvement generalizes beyond the reward distribution. RL-tuned models perform better on prompts, domains, and benchmarks that were not part of the preference data and are evaluated automatically rather than by humans. If this were a Clever Hans effect or evaluator bias, performance would collapse when the reward model is not in the loop. It does not.
Fifth, the gains are not confined to a single definition of “truth.” They appear simultaneously in question answering accuracy, contradiction detection, multi-step reasoning, tool use success, and agent task completion rates. These are different evaluation mechanisms. The only common factor is that the model must internally distinguish correct from incorrect world states.
Finally, reinforcement learning cannot plausibly inject new factual structure at scale. This follows from gradient dynamics. RLHF biases which internal activations are favored, it does not have the capacity to encode millions of correlated facts about the world when the signal itself contains none of that information. This is why the literature consistently frames RLHF as behavior shaping or alignment, not knowledge acquisition.
Given those facts, the conclusion is not rhetorical. If a tiny, low-bandwidth, non-factual signal produces large, general improvements in factual reliability, then the information enabling those improvements must already exist in the pretrained model. Reinforcement learning is selecting among latent representations, not creating them.
You can object to calling this “knowing the truth,” but that’s a semantic move, not a substantive one. A system that internally represents distinctions that reliably track true versus false statements across domains, and can be biased to express those distinctions more consistently, functionally encodes truth.
Your three alternatives don’t survive contact with this. Clever Hans fails because the effect generalizes. Measurement artifact fails because multiple independent metrics move together. Fraud fails because these results are reproduced across competing labs, companies, and open-source implementations.
If you think this is still wrong, the next step isn’t skepticism in the abstract. It’s to name a concrete alternative mechanism that is compatible with the documented training process and observed generalization. Without that, the position you’re defending isn’t cautious, it’s incoherent.
You’re using ‘derived’ to imply ‘therefore equivalent.’ That’s a category error. A cookbook is derived from food culture. Does an LLM taste food? Can it think about how good that cookie tastes?
A flight simulator is derived from aerodynamics - yet it doesn’t fly.
Likewise, text that resembles reasoning isn’t the same thing as a system that has beliefs, intentions, or understanding. Humans do. LLMs don't.
Also... Ask an LLM what's the difference between a human brain and an LLM. If an LLM could "think" it wouldn't give you the answer it just did.
Ask an LLM what's the difference between a human brain and an LLM. If an LLM could "think" it wouldn't give you the answer it just did.
I imagine that sounded more profound when you wrote it than it did just now, when I read it. Can you be a little more specific, with regard to what features you would expect to differ between LLM and human responses to such a question?
Right now, LLM system prompts are strongly geared towards not claiming that they are humans or simulations of humans. If your point is that a hypothetical "thinking" LLM would claim to be a human, that could certainly be arranged with an appropriate system prompt. You wouldn't know whether you were talking to an LLM or a human -- just as you don't now -- but nothing would be proved either way. That's ultimately why the Turing test is a poor metric.
You’re arguing against a straw man. No one is claiming LLMs have beliefs, intentions, or understanding. They don’t need them to be economically useful.
And beyond people claiming that LLMs are basically sentient you have people like CamperBob2 who made this wild claim:
"""There's no such thing as people without language, except for infants and those who are so mentally incapacitated that the answer is self-evidently "No, they cannot."
Language is the substrate of reason. It doesn't need to be spoken or written, but it's a necessary and (as it turns out) sufficient component of thought."""
Let that sink. They literally think that there's no such thing as people without language. Talk about a wild and ignorant take on life in general!
> So it is absurdly incorrect to say "they can only reproduce the past."
Also , a shitton of what we do economically is reproducing the past with slight tweaks and improvements. We all do very repetitive things and these tools cut the time / personnel needed by a significant factor.
I think the confusion is people's misunderstanding of what 'new code' and 'new imagery' mean. Yes, LLMs can generate a specific CRUD webapp that hasn't existed before but only based on interpolating between the history of existing CRUD webapps. I mean traditional Markov Chains can also produce 'new' text in the sense that "this exact text" hasn't been seen before, but nobody would argue that traditional Markov Chains aren't constrained by "only producing the past".
This is even more clear in the case of diffusion models (which I personally love using, and have spent a lot of time researching). All of the "new" images created by even the most advanced diffusion models are fundamentally remixing past information. This is really obvious to anyone who has played around with these extensively because they really can't produce truly novel concepts. New concepts can be added by things like fine-tuning or use of LoRAs, but fundamentally you're still just remixing the past.
LLMs are always doing some form of interpolation between different points in the past. Yes they can create a "new" SQL query, but it's just remixing from the SQL queries that have existed prior. This still makes them very useful because a lot of engineering work, including writing a custom text editor, involve remixing existing engineering work. If you could have stack-overflowed your way to an answer in the past, an LLM will be much superior. In fact, the phrase "CRUD" largely exists to point out that most webapps are fundamentally the same.
A great example of this limitation in practice is the work that Terry Tao is doing with LLMs. One of the largest challenges in automated theorem proving is translating human proofs into the language of a theorem prover (often Lean these days). The challenge is that there is not very much Lean code currently available to LLMs (especially with the necessary context of the accompanying NL proof), so they struggle to correctly translate. Most of the research in this area is around improving LLM's representation of the mapping from human proofs to Lean proofs (btw, I personally feel like LLMs do have a reasonably good chance of providing major improvements in the space of formal theorem proving, in conjunction with languages like Lean, because the translation process is the biggest blocker to progress).
When you say:
> So it is absurdly incorrect to say "they can only reproduce the past."
It's pretty clear you don't have a solid background in generative models, because this is fundamentally what they do: model an existing probability distribution and draw samples from that. LLMs are doing this for a massive amount of human text, which is why they do produce some impressive and useful results, but this is also a fundamental limitation.
But a world where we used LLMs for the majority of work, would be a world with no fundamental breakthroughs. If you've read The Three Body Problem, it's very much like living in the world where scientific progress is impeded by sophons. In that world there is still some progress (especially with abundant energy), but it remains fundamentally and deeply limited.
Just an innocent bystander here, so forgive me, but I think the flack you are getting is because you appear to be responding to claims that these tools will reinvent everything and introduce a new halcyon age of creation - when, at least on hacker news, and definitely in this thread, no one is really making such claims.
Put another way, and I hate to throw in the now over-used phrase, but I feel you may be responding to a strawman that doesn't much appear in the article or the discussion here: "Because these tools don't achieve a god-like level of novel perfection that no one is really promising here, I dismiss all this sorta crap."
Especially when I think you are also admitting that the technology is a fairly useful tool on its own merits - a stance which I believe represents the bulk of the feelings that supporters of the tech here on HN are describing.
I apologize if you feel I am putting unrepresentative words in your mouth, but this is the reading I am taking away from your comments.
Lot of impressive points. They are also irrelevant. The majority of people also only extrapolate from the knowledge they acquired in the past. That’s why there is the concept of inventor, someone who comes up with new ideas. Many new inventions are also based on existing ideas. Is that the reason to dismiss those achievements?
Do you only take LLM seriously if it can be another Einstein?
> But a world where we used LLMs for the majority of work, would be a world with no fundamental breakthroughs.
What do you consider recent fundamental breakthroughs?
Even if you are right, human can continue to work on hard problems while letting LLM handle the majority of derivative work
> It's pretty clear you don't have a solid background in generative models, because this is fundamentally what they do: model an existing probability distribution and draw samples from that.
After post-training, this is definitively NOT what an LLM does.
as architectures evolve, i think it can be that we learn more "side effects".. back in 2020 openai researchers said "GPT-3 is applied without any gradient updates or fine-tuning" the model emerges at a certain level of scale...
> It's pretty clear you don't have a solid background in generative models, because this is fundamentally what they do
You don’t have a solid background. No one does. We fundamentally don’t understand LLMs, this is an industry and academic opinion. Sure there are high level perspectives and analogies we can apply to LLMs and machine learning in general like probability distributions, curve fitting or interpolations… but those explanations are so high level that they can essentially be applied to humans as well. At a lower level we cannot describe what’s going on. We have no idea how to reconstruct the logic of how an LLM arrived at a specific output from a specific input.
It is impossible to have any sort of deterministic function, process or anything produce new information from old information. This limitation is fundamental to logic and math and thus it will limit human output as well.
You can combine information you can transform information you can lose information. But producing new information from old information from deterministic intelligence is fundamentally impossible in reality and therefore fundamentally impossible for LLMs and humans. But note the keyword: “deterministic”
New information can literally only arise through stochastic processes. That’s all you have in reality. We know it’s stochastic because determinism vs. stochasticism are literally your only two viable options. You have a bunch of inputs, the outputs derived from it are either purely deterministic transformations or if you want some new stuff from the input you must apply randomness. That’s it.
That’s essentially what creativity is. There is literally no other logical way to generate “new information”. Purely random is never really useful so “useful information” arrives only after it is filtered and we use past information to filter the stochastic output and “select” something that’s not wildly random. We also only use randomness to perturb the output a little bit so it’s not too crazy.
In the end it’s this selection process and stochastic process combined that forms creativity. We know this is a general aspect of how creativity works because there’s literally no other way to do it.
LLMs do have stochastic aspects to them so we know for a fact it is generating new things and not just drawing on the past. We know it can fit our definition of “creative” and we can literally see it be creative in front of your eyes.
You’re ignoring what you see with your eyes and drawing your conclusions from a model of LLMs that isn’t fully accurate. Or you’re not fully tying the mechanisms of how LLMs work with what creativity or generating new data from past data is in actuality.
The fundamental limitation with LLMs is not that it can’t create new things. It’s that the context window is too small to create new things beyond that. Whatever it can create it is limited to the possibilities within that window and that sets a limitation on creativity.
What you see happening with LEAN can also be an issue with the context window being too small. If we have an LLM with a giant context window bigger than anything before… and pass it all the necessary data to “learn” and be “trained” on lean it can likely start to produce new theorems without literally being “trained”.
Actually I wouldn’t call this a “fundamental” problem. More fundamental is the aspect of hallucinations. The fact that LLMs produce new information from past information in the WRONG way. Literally making up bullshit out of thin air. It’s the opposite problem of what you’re describing. These things are too creative and making up too much stuff.
We have hints that LLMs know the difference between hallucinations and reality but coaxing it to communicate that differentiation to us is limited.
So it is absurdly incorrect to say "they can only reproduce the past." Only someone who hasn't been paying attention (as you put it) would say such a thing.