Hacker News Reader

Comment by root_axis | Hacker News Reader

I won't speak for what's "all the way down", but on whatever level humans perform reflective reasoning, it's self-evident that we don't simply "predict the next word" in the way a LLM does.

Whether or not a LLM is capable of "creating anything new" is a different argument altogether. "New" is in eye of the beholder and it is often said that nothing is new under the sun, but we already know that unthinking processes within nature are capable of producing beauty far and beyond anything that even the greatest human artist could ever hope to accomplish. That a large LLM is unthinking doesn't preclude it from producing art.

View on HN

Please provide evidence that intelligent speech isn’t simply predicting the next word, as it’s obviously not self-evident to me.

I'm not sure it counts as "evidence" because I can't show it to you, but based on personal introspection, whenever I write something it starts with a barely-linguistic thought I want to communicate that gradually gets expanded into as much prose as is necessary to get across the concept whilst establishing context and defending from the most probable detractors.

I literally never just start writing and then just keep adding words.

It’s not evidence, sure, but what you describe, the “feel of it” is most of what we have to work with at the moment, unfortunately.

But it clearly is not self-evident that intelligence is or isn’t prediction.

Nothing is self-evident in a field we know so very little about. That’s my main objection, we should be humble and curious, not dogmatic.

It's pretty much self-evident when you worked with 3-6 that this isn't how it work. The idea come before the vocalization.

Or even when you talk in a foreign language you're semi-competent with.

I’ve spent many years with infants and learned 3 languages aside from my mother tongue. I don’t find it self-evident at all.

You’re gonna need to elaborate why you think so if you’re willing to convince someone.

That the idea form before it's vocalized?

A three year old reaching toward any toy, then saying 'mine!'.

Stumbling on a specific word, stopping your sentence, then proceed to replace it with a combination of words.

Or any time you paraphrase basically.

Not true in my experience. To me, words come first, the idea is a consequence I realize after a few of them have been stringed together.

Knowing where the words come from, how “the brain” knows the next one before my conscious self does has always been a mystery to me. I’ve heard Sam Harris describing a similar feeling in his podcast.

You never paraphrased something despite knowing the exact word?

[edit] Or explained high level CS concepts to an undergrad at your school?

Sure, but how's that indicative that prediction isn't taking place?

I have the idea before i have thought of the word i would use to express it. Sometimes, i don't even have word to express it, "you know what i mean?".

And like i said. Children will often act, then talk. Or seek their words for long period of time, and when you propose a word to help them vocalize, they can accept it or reject it.

I have another example. Do you ever have slip of the tongue? You're talking about "yoga", tell yourself while speaking "i really shouldn't say 'yoda' here", and say the word "yoda" anyway? This happened to me a lot when i was younger (my speed of thought is lower those days, i don't think i slipped this way in two or three years). I still think a bit faster than i speak tbh.

All of that happens to me as well. I just don’t see how it couldn’t be explained by prediction from probability distribution.

The fact that you can’t find a word could be that you’re stuck in a Markov chain wall or loop:

https://news.ycombinator.com/item?id=34688673

The thing is, we just don’t know how any of that work. Our speculation is as good as any.

I've always hated the idea promoted by some AI researchers that "intelligence is just prediction." Specifically in the context of your challenge, since a brain must decide the next word before uttering it, is it not vacuously true that any conceivable method of producing speech is "predicting" it? You're asking for evidence that it isn't; I'm asking for evidence that question has any meaning.

Sure, any production in that case can be classified as prediction. What I’m saying is that such distinction, if at all possible, is far from self-evident.

Anytime you say “uh, what’s that word again?”

So you’re thinking of a concept, you can picture it and it’s impact on your environment, but can’t quite remember the word for it.

Or, you reached a point in the Markov chain where none of the next guesses rank high enough if the whole paragraph is taken into account. It often feels like that, in fact, you can’t figure it out until you take another “route”.

Of course, my anecdote is just as valid as yours. My point is we have no freaking idea what’s going on and shouldn’t jump to conclusions.

The content of intelligent speech is not driven by statistical relationships between the spoken lexemes, this is self-evident.

For example, let's say you're planning to approach your boss about receiving a raise, the contents of your speech will be driven by a personal theory of mind regarding how your boss perceives you, your own memory of who you are and how you like to be perceived, the cadence and tenor of speech you imagine your boss is receptive to, information contextually relevant to the purpose of the speech like your proven history of success within the organization, as well as a hundred other variables driven by physiological factors, culture, social habituation and more.

"Predict the next word" just doesn't describe intelligent speech in any meaningful way.

So, you thrown in theory of mind, physiological factors and culture without defining what that means or showing that they don't all derive from the same underlying statistical relationships in order to make your argument more robust? That's not how it works.

Defining intelligence is fundamental philosophical problem, a discussion that's thousands of years old in the making. To call something in this field self-evident is either ignorance or self denial.

Someone deaf and blind still form thoughts but don’t output them in words. Words follow cognition, it’s sometimes obvious when you have a clear concept in mind but forget the word.

Edit: also when programming, I form mental structures and code them without verbalizing them, same with maths or when drawing or making music. A thought is maybe a prediction, but language seems not to be the abstraction it’s operated on.

Of course they do. They still have language. How else would they interact with the world?

You're asking for the evidence of a non existing factor in a complex process not fully understood... I'd say it's on you to provide it.

Why is predicting the next word proof of intelligent speech?

That’s not what I’m asking.

The grandparent said it’s self-evident that humans do not simply predict the next word. I asked for evidence of that self-evidence.

Reply on news.ycombinator.com