Hacker News Reader

Comment by kla-s | Hacker News Reader

Genuine question, how does multi step reasoning work for you then? Like eg if you have some math problem that's trivially to solve individually but needs multiple steps, lets say 16 * 3 + 5? How does 16 * 3 = 48 land in some 'register' of your brain (short term memory), so that you can then add 5 to get to 53? Maybe 16 * 3 + 5 is to easy for you and you'll just 'see' it but the question still stands, just choose a more complex problem.

Isn't the same meta process at play when thinking about more fuzzy topics?

View on HN

Not that poster, but for me it's directly manipulating numbers (for example, "16×3 + 5" turns into "10×3 + 6×3 + 5" into "30 + 18 + 5" into "30 + 10 + 8 + 5" into "40 + 13" into "53"). There's no language involved, though in some cases I might use some spatial reasoning by doing something like associating given chunks of an equation with different fingers.

I don't really follow. Say, during the in-between step "10×3 + 6×3 + 5", how do you store and cognize the individual numeric and operator elements?

Surely, even if the arithmetics can be simplified and "lookup-table'd", you are still aware of the numbers in Arabic form or whatever equivalent you're using, right? Or do you somehow have 53 individual blobs swirling inside your consciousness?

Language (human) and language (theoretical computer science) are different concepts.

Not poster you replied to but it sounds like we might have a similar internal model.

I store numbers as pictures of numbers, or a geometric representation depending on how big or precise the number is.

Are you saying when you think of the concept of 'twelve plus twelve' you have the equivalent of someone in your head saying 'hmm, well twelve is 2 more than ten, so if I add up ten and ten and two and two I get twenty four?'

That's wild if so.

For your reference, I would follow the procedure above approximately, but visually with numbers that just do the thing that feels right. I think under the hood we're probably doing the same thing, just with a different interface layer

I can just speak for me, obviously, but yes, that is what's happening. But it's not someone, it is more like me explaining / telling it to myself. Depending on the complexity this can be more or less verbal - the more complex, the less verbal I would say.

Not sure if it helps you make sense of it, but at least myself I only verbalize more complicated things. 10x3 for example I just "see" as 30 without needing to "think" or "speak" about it.

If you see a cat walking along the road, do you have to think to yourself "oh, that's a cat" or do you just know that it's a cat without verbalizing anything? It has its own abstract concept, right? Same thing with sufficiently simple numeric transformations.

> you are still aware of the numbers in Arabic form or whatever equivalent you're using, right?

Not sure if you're in a Fahrenheit or Celsius sort of place but if someone says that it's 70 degrees out do you really think in terms of numbers? Or do you just "know" what 70 degrees is without thinking about it?

In short, I just do. It's not a linguistic representation of the number 10, it's the concept of the number 10, the same way as (to extend on the mention of another comment) I don't need the word "cat" to know how to interact with a pet cat.

It's also very probable that the verbalization the majority does internally is just that - a verbalization of the actual underlying thought process. That is what much of current cognitive linguistics points to as far as I have understood.

(Also a reason why I'm very sceptical that the current LLM approach will eventually lead to AGI, BTW)

I think you're probably right that the verbalization is the 'interface layer' but why does that mean LLMs can't approach AGI? They also only use words as an 'interface' layer. Underlying weights are vectors in an abstract space.

With humans, it has been shown that reasoning processes for different aspects of human behavior are completely distinct from verbalization, they work fully autonomous, and the (inner) verbalization comes afterwards.

For LLMs, the tokens (i.e. words) are what the weights are based on, as there isn't other input into them.

I believe I have an internal narrator but I’m not certain exactly what others mean by that so I don’t know for sure.

However the way I think about math is different than the way I plan my day or other things. In my case, it is very much like I have registers that would hold the result of 16 x 3 in it so I can add the 5 to it later. I have a certain number of registers and with effort like repeating what Ive already solved I could temporarily create more.

It also feels somewhat physical, as if the register is an actual box or has a “location” or like I’ve put the answer down on the desk like a part of something I’m building. Perhaps not coincidentally I am one of the many people who have a “calendar shape” for the months.

I speak out loud or write on paper, or just do it a tiny bit slower and sometimes have to redo steps when I forget a result.

Reply on news.ycombinator.com