I think a lot of people using AI are falling for the same trap, just at a different level. People want it to be conscious, including AI researchers, and it's good at giving them what they want.
In a way it's a reframing of the timeless philosophical debate around determinism vs free will.
that is, if you train an LLM on a bunch of text that is almost certainly going to include stuff about sentient robots, that it sometimes announces that it's a sentient robot is not evidence that it is one
Perhaps in the flicker of processing between prompt and answer the signal patter does resemble human consciousness for a second.
Calling it a token predictor is just like saying a computer is a bit mover. In the end your computer is just a machine that flips bits and switches but it is the high level macro effect that characterizes it better. LLMs are the same at the low level it is a token predictor. At the higher macro level we do not understand it and it is not completely far fetched to say it may be conscious at times.
I mean we can’t even characterize definitively what consciousness is at the language level. It’s a bit of a loaded word deliberately given a vague definition.
Read the above carefully because you just hallucinated a statement and attributed it to me. I never "filled" in a gap. I just stated a possibility. But you, like the LLM, went with your gut biases and attributed a false statement to me.
Think about it. The output and input of the text generator of humans and LLMs are extremely similar to the point where it passes a turing test.
So to say that a flicker of consciousness exists is reasonable. It's not unreasonable given that the observable inputs are EXACTLY the same.
The only parts that we know are different are hallucinations, and a constant stream of thought. LLMs aren't active when not analyzing a query and LLMs tend to hallucinate more than humans. Do these differences spell anything different for "consciousness" not really.
Given that these are the absolute ground truth observations... my guessed conclusion is unfortunately NOT unreasonable. What is unreasonable to to say anything definitive GIVEN that we don't know. So to say absolutely it's not conscious or absolutely it is, BOTH are are naive.
Think extremely logically. It is fundamental biases that lead people to come to absolute conclusions when no other information is available.
1. Academic understanding of consciousness is effectively zero. If we understand something that means we can actually build or model the algorithm for consciousness. We can't because we don't know shit. Most of what you read is speculative hypotheticals derived from observation that's not too different from attempting to reverse engineer an operating system by staring at assembly code.
Often we describe consciousness with ill defined words that are also vague and lack understanding for. The whole endeavor is bs.
2. Understanding of LLMs outside of the low level token prediction is effectively zero. We know there are emergent second order effects that we don't get. You don't believe me? How about if I have the god father of AI say it himself:
https://youtu.be/qrvK_KuIeJk?t=284 Literally. The experts say we don't understand it.
Look if you knew how LLMs work you'd say the same. But people everywhere are coming to conclusions about LLMs without knowing everything, so by citing the eminent expert saying the ground truth you should be convinced that the reality is this conclusive fact:
You are utterly misinformed about how much academia understands about LLM and consciousness. We know MUCH less than you think.
Calling it a token-predictor isn't reductionism. It's designed, implemented and trained for token prediction. Training means that the weights are adjusted in the network until it accurately predicts tokens. Predicting a token is something along the lines of removing a word from a sentence and getting it to predict it back: "The quick brown fox jumped over the lazy ____". Correct prediction is "dogs".
So actually it is like calling a grass-cutting machine "lawn mower".
> I mean we can’t even characterize definitively what consciousness is at the language level.
But, oh, just believe the LLM when it produces a sentence referring to itself, claiming it is conscious.
It absolutely is reductionism. Ask any expert who knows how these things work and they will say the same:
https://youtu.be/qrvK_KuIeJk?t=497
Above we have Geoffrey Hinton, the godfather of the current wave of AI saying your statements are absolutely crazy.
It's nuts that I don't actually have to offer any proof to convince you. Proof won't convince you. I just have to show you someone smarter than you with a better reputation saying the exact opposite and that is what flips you.
Human psychology readily can attack logic and rationality. You can scaffold any amount of twisted logic and irrelevant analogies to get around any bulwark to support your own point. Human psychology fails when attacking another person who has a higher rank. Going against someone of higher rank this causes you to think twice and rethink your own position. In debates, logic is ineffective, bringing opposing statements with (while offering ZERO concrete evidence) from experts with a higher rank is the actual way to convince you.
>But, oh, just believe the LLM when it produces a sentence referring to itself, claiming it is conscious.
This is an hallucination. Showing that you're not much different from an LLM. I NEVER stated this. I said it's possible. But I said we cannot make a definitive statement either way. We cannot say it isn't conscious we cannot say it is. First we don't understand the LLM and second WE don't even have an exact definition of consciousness. So to say it's not conscious is JUST as ludicrous as saying it is.
Understand?
Here we go again...
Geoffrey Hinton is popular because of ML. You know why he doesn't understand the modern wave of LLMs? Because the transformer architecture isn't his invention. He had almost nothing to do with it. It's like saying that nobody understands how the nuclear bomb works simply because Einstein won't explain it to you in detail. There is a science here, and denying that it can be interpreted or understood is simply hysterical. We can perfectly well understand a Markov chain, why are LLMs any different?
The "current wave of AI" is an autoregressive feed-forward algorithm inferring tokenized data using complex pretrained weights. In simple terms, you could describe this as a token predictor without glossing over any details that are fundamentally important. To call it "reductionist" you would have to explain what is missing from the definition, not simply that you distrust people who understand it.
Did you invent the transformer architecture yourself, and I am mistaken to guess that you probably understand it less than Geoffrey Hinton does?
I'll venture a wild guess that you probably haven't been working in the field as long as he has, and you don't personally know as many people in the field as he does, huh?
Or is my presumptuous chain of logic just a random thoughtless unconscious probabilistic stream of tokens that has nothing to do with reality, and I don't realize I'm speaking to Ashish Vaswani himself?
Then why don't you use your real name as you HN handle, are you really that modest, Ashish? And don't you have better things to do than posting to Hacker News? If you really wanted to win an argument from authority, you could at least post under your real name, Mr. Vaswani.
I would have predicted a single lazy "dog". Does that mean I'm more conscious than you? ;)
>Correct prediction is "dogs".
Maybe I just trained on Wikipedia, or the February 9, 1885 edition of The Boston Journal "Current Notes" article, or Linda Bronson's 1888 book "Illustrative Shorthand", for what the statistically and historically "correct" completion is.
https://en.wikipedia.org/wiki/The_quick_brown_fox_jumps_over... [..._the_lazy_dog]
Why would you insert redundant letters into a lovely well-known succinct pangram, when the whole point is to make the SHORTEST sentence with all letters of the alphabet?
This entire post could also technically be considered a terribly redundant pangram if you wanted to throw in a lot of unnecessary letters, slashes, colons, underscores, double quotes, commas, periods, question marks, brackets, equal signs, digits, spaces, and newlines. As could Moby Dick, War and Peace, or the Unabomber Manifesto.
>So actually it is like calling a grass-cutting machine "lawn mower".
Is it anything like calling Larry Ellison a "lawn mower", which is much more "correct" than anthropomorphising him?
https://news.ycombinator.com/item?id=15886728
As ninetyninenine pointed out, we don't understand what's going on with you either:
>Someone can claim that the responses are an illusion and they have no meaning but that exact argument applies to humans and we can’t say anything either way. We don’t understand what’s going on.
> Why would you insert redundant letters into a lovely well-known succinct pangram
The 's' is not redundant, because my version of the pangram has "fox jumped" not "fox jumps".
I picked up this version as a kid and have always used that one when testing new keyboards.
> the whole point is to make the SHORTEST sentence with all letters of the alphabet
If that is the whole point, the fox-dog sentence must be abandoned for a shorter pangram.
The requirement those shorter pangrams tend not to satisfy is being easy to remember.
People don't care about the sentence being short, which is why variants starting with "The ..." instead of "A ..." soon appeared after the invention of the original, which itself could have dropped the "A".
Under that assumption, the correct prediction is "dog", not "dogs", because LLMs are trained on the wikipedia page https://en.wikipedia.org/wiki/The_quick_brown_fox_jumps_over... [..._the_lazy_dog] and other earlier publications like The Boston Journal "Current Notes" article, and Linda Bronson's 1888 book "Illustrative Shorthand", and not the version you picked up as a kid.
Can you cite any sources that says "dogs" that are more popular than the wikipedia page or better known than the 19th century publications it cites?
I will ask several popular LLMs what they think, even using four underscores to imply a four letter word:
Please complete: The quick brown fox jumped over the lazy ____
ChatGPT o4-mini-high said: dog
Claud-4-Opus said: dog
Gemini-2.5-Pro said: dog
DeepSeek-v3.1 said: dog
Grok-3 said: dog
If you can't cite any wikipedia pages, books, articles, or other examples of an LLM that predicts "dogs", do you still stand by your bold claim that:
>Correct prediction is "dogs".
Your mistake was confidently but incorrectly using the word "correct". In the context of predicting the most likely word, the correct, highest probability answer by far is unequivocally "dog", not "dogs".
That's why the wikipedia page, which all mainstream LLMs are trained on, has the title that it does.
The other hypocritical mistake you made was saying the following right after YOU incorrectly tried to pick what is "correct":
>You don't get to pick what is "correct"
So am I to understand that even though I DON'T get to pick what is correct (which I'm not presuming to do: I'm just deferring to wikipedia, two 19th century publications, and 5 LLMs), but YOU DO get to pick what is "correct", even if it contradicts all available evidence? Because that's exactly what you're doing! Please explain how you got such an awesomely important, exclusive, earth shattering superpower? Were you bitten by a radioactive web crawler?
If you boldly claim to be the only person who has the actual power to pick what's correct and what's not, then there are so many other much more important wrongs you should be righting than arguing about "dog" vs "dogs" on hacker news.
For more evidence, you can even read the talk page and archive yourself. There are discussions about alternatives, but NONE of the alternatives ever use the word "dogs".
Talk: https://en.wikipedia.org/wiki/Talk:The_quick_brown_fox_jumps...
Talk Archive: https://en.wikipedia.org/wiki/Talk:The_quick_brown_fox_jumps...
The larger point is that your argument is reductionist, and you're exhibiting the same kind of mistakes that people point to when they criticize LLMs for their overconfident hallucinations.
We understand LLMs pretty well. That we can't debug them and inspect every step of every factor on every prediction doesn't mean we don't understand how they work.
We also know that convincing speech doesn't require consciousness.
This is the definition of lack of understanding. If you can’t debug or know what happens on every step in means you don’t understand the steps. Lack of knowledge about something = lack of understanding.
The amount of logic twisting going on here is insane.