Stephen Wolfram had a great description of this effect in the early days (GPT 3.5 era):
Machine learning trains the network to do... anything that you reward it for. If you keep training, it keeps getting better.
Next word prediction can always keep getting better.
At first, simply "learning" spelling is what makes the predictions better because tokens are word chunks, not always whole words.
Then, the models "run out of steam" and can't get any better by learning more spelling rules, but the gradient descent forces them to get better... so they do... by learning the rules of grammar.
At this point the AIs can output correctly spelled and grammatically coherent sentences, but the sentences ramble on about nonsense topics.
So what happens next as the models run out of grammar rules is that they're forced to learn the rules "above grammar": logic, world knowledge, coherent story telling, etc.
At some point they learn to output pages and pages of fluid, coherent text, but... if they're not smart, if they don't think, and if they don't know what they're talking about, then they're still "suboptimal" and their forced gradient descent will make them close those gaps.
Eventually, the only way they can improve at "next token prediction" is by building up to human-like intelligence, including an inner monologue, theory of mind, and everything.
If we could give a comprehensive and global explanation of an LLM's behavior in a single paragraph, we wouldn't need the model to begin with, but that doesn't mean there's absolutely no understanding of the model internals whatsoever
> processes better regarded as generalization and prediction, which seem not only seem to have direct and obvious neural realizations
If the workings of those circuits are obvious to you, I'd really like to learn. Do you mean the level of analysis at https://transformer-circuits.pub/ ? (That looks like good work but not a deep understanding.)
> When you write a program to predict tokens based on context, seeding its context with something that makes it predict "self-reflecting" text is trivial. Program does what it is programmed to do. Would observing the output of the following program inspire doubt as to its sentience?
Then you follow it up with print statements as if that is a good analogy.
As I said, they are not programmed, so your question above is not relevant to your argument.
You say they're programs that are stochastically jiggled, but that's simply not accurate either. All LLM abilities are emergent, even when the training corpus is well defined.
I didn't think you literally thought they were made of print statements, but you are implying they're software that's been "fuzzed". Hopefully you don't literally that either and you're just using it as a bad analogy.
You could have argued from the stance of neural networks being universal functions, which might at least be closer to the truth, but instead your example is print statements!
I get you're trying to say that something trained to say a thing doesn't mean it has arrived at the thing like a mind would, and perhaps that would have been closer for GPT 2.
These days though, we just have so much more awareness of what they're actually doing internally that it's bizarre to even compare them to stochastic parrots of the training corpus, if that is closer to what you're implying.
No, I mean we just don't know what's going on in the circuits of the model at any substantial level. We set their architecture (hyperparameters), we pump them full of data (pretraining), and we shape how they behave through examples (SFT) and reward (RL), but we can't say with any certainty what the resulting model does internally.
I haven't read this in depth yet, though I plan to. If this general line of research is interesting to you, I'd recommend checking out some of the lines of research it touches upon--they're really rich and fascinating, and some are pretty approachable mathematically even if ML research papers aren't usually your thing. The related works section here seems pretty well stocked, but mechanistic interpretability is a pretty interesting peephole into this general vein: https://transformer-circuits.pub/
I'm not sure how to test this but I think there's an interesting possibility where the "reasoning" tokens are actually both an accurate reflection of a line of reasoning, but also, that there can be changes in the weights as the computation proceeds onward that may not be reflected in the apparently nominal meaning of the human language the tokens are output as for our consumption.
Some modest evidence is my own subjective experience of the many times I've explained why I'm doing something, and it is a true explanation in the sense that it is certainly not a lie, but it is also incomplete and there are entire strands of thought that went into my decision that are not being articulated. Though human speech is not equivalent to an LLM's output since we can trivially think without literally speaking whereas they can not. (No need to nitpick on the definitions there; all I'm observing here is that they are forced to emit an externally-visible artifact whereas I can sit in silence, thinking, with no externally-visible artifact being produced. Not trying to make any grand claims about what is "real" cognition or anything.)
It is conceivable how to create a test of whether the tokens correspond to the "real" thought process, and papers and work on that have been done, such as [1]. It is difficult for me to imagine how to scramble the nominal tokens without also completely trashing any implicit calculations that may be occurring too.
In this token-mania frenzy that has taken hold of the industry, I guess solutions like "soul document" and "system prompts" will continue for a while, and once the industry matures a bit we'll go back to things like LoRA[1] and control vectors[2][3].
The other explanation may be that these AI labs may be expecting more government scrutiny, and "here's a document" would probably go better than "here's some vector representation of our values" when talking to politicians.
I suspect you will not accept this (because you have been manipulated by the generator?) but there is no "I" in repeated inference calls to large language models. There is no coherent entity behind the text. Try switching backend APIs mid-session, flip from Claude to Gemini and then to ChatGPT and the output will happily burble onwards "Oops, you're right, I see my last answer contained an error..."
It's lying to users just as much as an email autoreply that pretended to be a human response is lying.
Before reading, know that I am uncertain in either direction.
> a hidden representation of self that is continually tended to
This sounds like a personality? They act like they have one of those. It may be an illusion, and even if it isn't an illusion it is unlikely to be anything like the source (us), but they act like it.
> I further fail to identify how it could be hidden or maintained, considering I control like half of it.
Indeed you control everything about a local model, and much of the context of even a remote model. But the state of activations and circuits in SotA AI is hidden in similar ways to those of synapses in your head: difficult to decipher even with probes monitoring the signals directly, and often not emitted at the normal output.
> The best you could ascribe it is a meticulous maintenance of a persona the user is talking to, but then that doesn't necessarily represent the model's internal state, the same way my own words here aren't doing so either. Difference being, I actually have one (I'm "on-line").
While we can be confident that LLMs make up personas etc., it is insufficient to go from "that doesn't necessarily represent the model's internal state" to "therefore it doesn't have one".
> You'll sometimes catch models mixing up who's who and how many who-s there even are for example.
I've, unfortunately, also experienced this with humans. Perhaps they were losing their self-awareness at the time? I do wonder if old-age dementia does that by the end, though the person in question didn't ever get diagnosed with that.
> If you know of anything like this, your turn now, would be happy to learn.
I'm talking about using mechanistic interpretability to see the model's intent. If it is deliberately using compromised libraries to weaken some code's security, there's going to be a signal in its hidden activations that it's doing so.
Finding these kinds of activations is something Anthropic is actively researching [1] but they're the only ones who can use those techniques to see Claude's intent. On the other hand, if a model is open-weights, in theory whoever is running the model could look inside the activations at runtime to see if a hidden vector associated with "deception" or "sabotage" is being activated [2].
(Those sources are just a couple of relevant starting points I could find without much effort, there is also https://www.neuronpedia.org/ if one is interested in seeing interactive demonstrations of interpretability concepts)
Machine learning trains the network to do... anything that you reward it for. If you keep training, it keeps getting better.
Next word prediction can always keep getting better.
At first, simply "learning" spelling is what makes the predictions better because tokens are word chunks, not always whole words.
Then, the models "run out of steam" and can't get any better by learning more spelling rules, but the gradient descent forces them to get better... so they do... by learning the rules of grammar.
At this point the AIs can output correctly spelled and grammatically coherent sentences, but the sentences ramble on about nonsense topics.
So what happens next as the models run out of grammar rules is that they're forced to learn the rules "above grammar": logic, world knowledge, coherent story telling, etc.
At some point they learn to output pages and pages of fluid, coherent text, but... if they're not smart, if they don't think, and if they don't know what they're talking about, then they're still "suboptimal" and their forced gradient descent will make them close those gaps.
Eventually, the only way they can improve at "next token prediction" is by building up to human-like intelligence, including an inner monologue, theory of mind, and everything.
We can even read their "thoughts": https://transformer-circuits.pub/2026/workspace/index.html