This is definitely fascinating - being able to do AI brain surgery, and selectively tuning its knowledge and priors, you'd be able to create awesome and terrifying simulations.
LLMs are just seemingly intelligent autocomplete engines, and until they figure a way to stop the hallucinations, they aren't great either.
Every piece of code a developer churns out using LLMs will be built from previous code that other developers have written (including both strengths and weaknesses, btw). Every paragraph you ask it to write in a summary? Same. Every single other problem? Same. Ask it to generate a summary of a document? Don't trust it here either. [Note, expect cyber-attacks later on regarding this scenario, it is beginning to happen -- documents made intentionally obtuse to fool an LLM into hallucinating about the document, which leads to someone signing a contract, conning the person out of millions].
If you ask an LLM to solve something no human has, you'll get a fabrication, which has fooled quite a few folks and caused them to jeopardize their career (lawyers, etc) which is why I am posting this.
Sure, LLMs do not think like humans and they may not have human-level creativity. Sometimes they hallucinate. But they can absolutely solve new problems that aren’t in their training set, e.g. some rather difficult problems on the last Mathematical Olympiad. They don’t just regurgitate remixes of their training data. If you don’t believe this, you really need to spend more time with the latest SotA models like Opus 4.5 or Gemini 3.
Nontrivial emergent behavior is a thing. It will only get more impressive. That doesn’t make LLMs like humans (and we shouldn’t anthropomorphize them) but they are not “autocomplete on steroids” anymore either.
This is just an appeal to complexity, not a rebuttal to the critique of likening an LLM to a human brain.
> they are not “autocomplete on steroids” anymore either.
Yes, they are. The steroids are just even more powerful. By refining training data quality, increasing parameter size, and increasing context length we can squeeze more utility out of LLMs than ever before, but ultimately, Opus 4.5 is the same thing as GPT2, it's only that coherence lasts a few pages rather than a few sentences.
And I know not everyone thinks in a literal stream of words all the time (I do) but I would argue that those people's brains are just using a different "token"
However, what it is doing is layered autocomplete on itself. I.e. one part is trying to predict what the other part will be producing and training itself on this kind of prediction.
What emerges from this layered level of autocompletes is what we call thought.
Other posters already noted other reasons for it, but I will note that you are saying 'similar to autocomplete, but obviously' suggesting you recognize the shape and immediately dismissing it as not the same, because the shape you know in humans is much more evolved and co do more things. Ngl man, as arguments go, it sounds to me like supercharged autocomplete that was allowed to develop over a number of years.
Or in other words, this thread sure attracts a lot of armchair experts.
... but we also need to be careful with that assertion, because humans do not understand cognition, psychology, or biology very well.
Biology is the furthest developed, but it turns out to be like physics -- superficially and usefully modelable, but fundamental mysteries remain. We have no idea how complete our models are, but they work pretty well in our standard context.
If computer engineering is downstream from physics, and cognition is downstream from biology ... well, I just don't know how certain we can be about much of anything.
> this thread sure attracts a lot of armchair experts.
"So we beat on, boats against the current, borne back ceaselessly into our priors..."
Think about an average dinner party conversation. Person A talks, person B thinks about something to say that fits, person C gets an association from what A and B said and speaks...
And what are people most interested in talking about? Things they read or watched during the week perhaps?
Conversations would not have had to be like this. Imagine a species from another planet who had a "conversation" where each party simply communicated what it most needed to say/was most benefitial to say and said it. And where the chance of bringing up a topic had no correlation at all with what previous person said (why should it?) or with what was in the newspapers that week. And who had no "interest" in the association game.
Humans saying they are not driven by associations is to me a bit like fish saying they are not noticing the water. At least MY thought processes works like that.
Prior to LLMs, there was never any suggestion that thoughts work like autocomplete, but now people are working backwards from that conclusion based on metaphorical parallels.
Roots of predictive coding theory extend back to 1860s.
Natalia Bekhtereva was writing about compact concept representations in the brain akin to tokens.
Yes, you can draw interesting parallels between anything when you're motivated to do so. My point is that this isn't parsimonious reasoning, it's working backwards from a conclusion and searching for every opportunity to fit the available evidence into a narrative that supports it.
> Roots of predictive coding theory extend back to 1860s.
This is just another example of metaphorical parallels overstating meaningful connections. Just because next-token-prediction and predictive coding have the word "predict" in common doesn't mean the two are at all related in any practical sense.
Predictive coding theory was formalized back around 2010 and traces it roots up to theories by Helmholtz from 1860.
Predictive coding theory postulates that our brains are just very strong prediction machines, with multiple layers of predictive machinery, each predicting the next.
Fascinating framing. What would you consider evidence here?
That is to say, they are equally likely if you don't do next token prediction at all and instead do text diffusion or something. Architecture has nothing to do with it. They arise because they are early partial solutions to the reconstruction task on 'all the text ever made'. Reconstruction task doesn't care much about truthiness until way late in the loss curve (where we probably will never reach), so hallucinations are almost as good for a very long time.
RL as is typical in post-training _does not share those early solutions_, and so does not share the fundamental problems. RL (in this context) has its own share of problems which are different, such as reward hacks like: reliance on meta signaling (# Why X is the correct solution, the honest answer ...), lying (commenting out tests), manipulation (You're absolutely right!), etc. Anything to make the human press the upvote button or make the test suite pass at any cost or whatever.
With that said, RL post-trained models _inherit_ the problems of non-optimal large corpora reconstruction solutions, but they don't introduce more or make them worse in a directed manner or anything like that. There's no reason to think them inevitable, and in principle you can cut away the garbage with the right RL target.
Thinking about architecture at all (autoregressive CE, RL, transformers, etc) is the wrong level of abstraction for understanding model behavior: instead, think about loss surfaces (large corpora reconstruction, human agreement, test suites passing, etc) and what solutions exist early and late in training for them.
Probably you believe that humans have something called intelligence, but the pressure that produced it - the likelihood of specific genetic material to replicate - it is much more tangential to intelligence than next-token-prediction.
I doubt many alien civilizations would look at us and say "not intelligent - they're just genetic information replication on steroids".
Second: modern models also under go a ton of post-training now. RLHF, mechanized fine-tuning on specific use cases, etc etc. It's just not correct that token-prediction loss function is "the whole thing".
Invoking terms like "selection mechanism" is begging the question because it implicitly likens next-token-prediction training to natural selection, but in reality the two are so fundamentally different that the analogy only has metaphorical meaning. Even at a conceptual level, gradient descent gradually honing in on a known target is comically trivial compared to the blind filter of natural selection sorting out the chaos of chemical biology. It's like comparing legos to DNA.
> Second: modern models also under go a ton of post-training now. RLHF, mechanized fine-tuning on specific use cases, etc etc. It's just not correct that token-prediction loss function is "the whole thing".
RL is still token prediction, it's just a technique for adjusting the weights to align with predictions that you can't model a loss function for in per-training. When RL rewards good output, it's increasing the statistical strength of the model for an arbitrary purpose, but ultimately what is achieved is still a brute force quadratic lookup for every token in the context.
Second, to autocomplete the name of the killer in a detective book outside of the training set requires following and at least some understanding of the plot.
This tells me that you haven't really used Opus 4.5 at all.
I wasn’t arguing that LLMs are like a human brain. Of course they aren’t. I said twice in my original post that they aren’t like humans. But “like a human brain” and “autocomplete on steroids” aren’t the only two choices here.
As for appealing to complexity, well, let’s call it more like an appeal to humility in the face of complexity. My basic claim is this:
1) It is a trap to reason from model architecture alone to make claims about what LLMs can and can’t do.
2) The specific version of this in GP that I was objecting to was: LLMs are just transformers that do next token prediction, therefore they cannot solve novel problems and just regurgitate their training data. This is provably true or false, if we agree on a reasonable definition of novel problems.
The reason I believe this is that back in 2023 I (like many of us) used LLM architecture to argue that LLMs had all sorts of limitations around the kind of code they could write, the tasks they could do, the math problems they could solve. At the end of 2025, SotA LLMs have refuted most of these claims by being able to do the tasks I thought they’d never be able to do. That was a big surprise to a lot us in the industry. It still surprises me every day. The facts changed, and I changed my opinion.
So I would ask you: what kind of task do you think LLMs aren’t capable of doing, reasoning from their architecture?
I was also going to mention RL, as I think that is the key differentiator that makes the “knowledge” in the SotA LLMs right now qualitatively different from GPT2. But other posters already made that point.
This topic arouses strong reactions. I already had one poster (since apparently downvoted into oblivion) accuse me of “magical thinking” and “LLM-induced-psychosis”! And I thought I was just making the rather uncontroversial point that things may be more complicated than we all thought in 2023. For what it’s worth, I do believe LLMs probably have limitations (like they’re not going to lead to AGI and are never going to do mathematics like Terence Tao) and I also think we’re in a huge bubble and a lot of people are going to lose their shirts. But I think we all owe it to ourselves to take LLMs seriously as well. Saying “Opus 4.5 is the same thing as GPT2” isn’t really a pathway to do that, it’s just a convenient way to avoid grappling with the hard questions.
No it isn't.
> ...fool you into thinking you understand what is going on in that trillion parameter neural network.
It's just matrix multiplication and logistic regression, nothing more.
The sequence of matrix multiplications are the high level constraint on the space of programs discoverable. But the specific parameters discovered are what determines the specifics of information flow through the network and hence what program is defined. The complexity of the trained network is emergent, meaning the internal complexity far surpasses that of the course-grained description of the high level matmul sequences. LLMs are not just matmuls and logits.
Yes, so is logistic regression.
And if you want to get pedantic and technical, you didn't even get the reductionism right! Modern LLMs don't use the logistic regression sigmoid function for network activation nonlinearity anymore, they use things like ReLU or GELU. You're about 15 years behind.
Reductionism is counterproductive in biology ("human brains are voltage spikes across membranes, nothing more") and it's counterproductive here as well. LLMs have nontrivial emergent behavior. The interesting questions are all around what that behavior is and how it arises in the network during training, and if you refuse to engage beyond bare reductionism you won't even be able to ask those questions, let alone answer them.
You still need to hand hold it all the way as it is only capable of regurgitating the tiny amount of code patterns it saw in the public. As opposed to say a Python project.
But regardless, I don’t think anyone is claiming that LLMs can magically do things that aren’t in their training data or context window. Obviously not: they can’t learn on the job and the permanent knowledge they have is frozen in during training.
For someone speaking as you knew everything, you appear to know very little. Every LLM completion is a "hallucination", some of them just happen to be factually correct.
Most modern post training setups encourage this.
It isn't 2023 anymore.
> What did I have for breakfast this morning?
> I don’t know what you had for breakfast this morning…
LLMs are like a topographic map of language.
If you have 2 known mountains (domains of knowledge) you can likely predict there is a valley between them, even if you haven’t been there.
I think LLMs can approximate language topography based on known surrounding features so to speak, and that can produce novel information that would be similar to insight or innovation.
I’ve seen this in our lab, or at least, I think I have.
Curious how you see it.
BINGO!
(I just won a stuffed animal prize with my AI Skeptic Thought-Terminating Cliché BINGO Card!)
Sorry. Carry on.
Well, no, they are training set statistical predictors, not individual training sample predictors (autocomplete).
The best mental model of what they are doing might be that you are talking to a football stadium full of people, where everyone in the stadium gets to vote on the next word of the response being generated. You are not getting an "autocomplete" answer from any one coherent source, but instead a strange composite response where each word is the result of different people trying to steer the response in different directions.
An LLM will naturally generate responses that were not in the training set, even if ultimately limited by what was in the training set. The best way to think of this is perhaps that they are limited to the "generative closure" (cf mathematical set closure) of the training data - they can generate "novel" (to the training set) combinations of words and partial samples in the training data, by combining statistical patterns from different sources that never occurred together in the training data.
Transformers allow for the mapping of a complex manifold representation of causal phenomena present in the data they're trained on. When they're trained on a vast corpus of human generated text, they model a lot of the underlying phenomena that resulted in that text.
In some cases, shortcuts and hacks and entirely inhuman features and functions are learned. In other cases, the functions and features are learned to an astonishingly superhuman level. There's a depth of recursion and complexity to some things that escape the capability of modern architectures to model, and there are subtle things that don't get picked up on. LLMs do not have a coherent self, or subjective central perspective, even within constraints of context modifications for run-time constructs. They're fundamentally many-minded, or no-minded, depending on the way they're used, and without that subjective anchor, they lack the principle by which to effectively model a self over many of the long horizon and complex features that human brains basically live in.
Confabulation isn't unique to LLMs. Everything you're saying about how LLMs operate can be said about human brains, too. Our intelligence and capabilities don't emerge from nothing, and human cognition isn't magical. And what humans do can also be considered "intelligent autocomplete" at a functional level.
What cortical columns do is next-activation predictions at an optimally sparse, embarrassingly parallel scale - it's not tokens being predicted but "what does the brain think is the next neuron/column that will fire", and where it's successful, synapses are reinforced, and where it fails, signals are suppressed.
Neocortical processing does the task of learning, modeling, and predicting across a wide multimodal, arbitrary depth, long horizon domain that allow us to learn words and writing and language and coding and rationalism and everything it is that we do. We're profoundly more data efficient learners, and massively parallel, amazingly sparse processing allows us to pick up on subtle nuance and amazing wide and deep contextual cues in ways that LLMs are structurally incapable of, for now.
You use the word hallucinations as a pejorative, but everything you do, your every memory, experience, thought, plan, all of your existence is a hallucination. You are, at a deep and fundamental level, a construct built by your brain, from the processing of millions of electrochemical signals, bundled together, parsed, compressed, interpreted, and finally joined together in the wonderfully diverse and rich and deep fabric of your subjective experience.
LLMs don't have that, or at best, only have disparate flashes of incoherent subjective experience, because nothing is persisted or temporally coherent at the levels that matter. That could very well be a very important mechanism and crucial to overcoming many of the flaws in current models.
That said, you don't want to get rid of hallucinations. You want the hallucinations to be valid. You want them to correspond to reality as closely as possible, coupled tightly to correctly modeled features of things that are real.
LLMs have created, at superhuman speeds, vast troves of things that humans have not. They've even done things that most humans could not. I don't think they've done things that any human could not, yet, but the jagged frontier of capabilities is pushing many domains very close to the degree of competence at which they'll be superhuman in quality, outperforming any possible human for certain tasks.
There are architecture issues that don't look like they can be resolved with scaling alone. That doesn't mean shortcuts, hacks, and useful capabilities won't produce good results in the meantime, and if they can get us to the point of useful, replicable, and automated AI research and recursive self improvement, then we don't necessarily need to change course. LLMs will eventually be used to find the next big breakthrough architecture, and we can enjoy these wonderful, downright magical tools in the meantime.
And of course, human experts in the loop are a must, and everything must be held to a high standard of evidence and review. The more important the problem being worked on, like a law case, the more scrutiny and human intervention will be required. Judges, lawyers, and politicians are all using AI for things that they probably shouldn't, but that's a human failure mode. It doesn't imply that the tools aren't useful, nor that they can't be used skillfully.
Source needed RE brain.
Define innovate, in a way that a LLM can't and we definitively can prove a human can.