I’ve proposed this argument before, but I want to know what you all think - and if you can poke some holes in it.
Often times, I’ve heard from others that AI isn’t capable of original thought. So, I propose the argument: is a model able to interpolate? In other words, given a set of data, can they fill in the missing data in-between?
Humans are obviously capable of this, and I would argue that AI models, too, are capable of this. Then, comes the more difficult question:
If a model can effectively interpolate, can we have models with effective extrapolation?
If so, what differentiates organic thought? Humans can extrapolate; this is most of what we do with “new” thought, or when we have a revelation. Because of this, as AI becomes sufficiently advanced, what will truly differentiate organic extrapolation from artificial extrapolation? Will it matter, assuming AI is able to make strides in thought that humans still struggle to?
But isn’t the answer here simply that AI can’t extrapolate?
As far as I know, All AI models today only interpolate. They can only model a distribution. They can’t extrapolate.
Even the very advanced ones: we know GPTs and Generative Models in general aren’t extrapolating. They’re simply very good at large-scale interpolation.
I don’t really believe we can begin to discuss whether we can differentiate between organic and artificial thought if we haven’t taken a single convincing step towards artificial extrapolation, even with all the advancements in AI we’ve achieved today.
It seems to be a major fundamental problem. Interpolation is relatively easy, as you know the starting and ending points, and you just need to fill in the middle. There are many methods for that process.
Extrapolation, on the other hand, is computationally hard; you do not know where the trend is going, and it cannot be predicted reliably from the previous data. Extrapolation will always provide a massive amount of inaccuracy. Case in point: Solar power.
Every time so far, extrapolating the potential growth of solar power has been too low as per the economist screenshot here(article paywalled).
The problem with extrapolation is the fact that reliable extrapolation doesn’t exist. Trends can change unexpectedly. AI, as we know it, is a tool designed to mimic human behaviour from data patterns that already exist. This makes interpolation easy for AI, but extrapolation, a trend we cannot do, is also something AI cannot do.
Ah, but I would be inclined to argue that a lot of large language models do primarily extrapolation. Given an input, the model has no defined endpoint, so it extrapolates word-by-word for the entire response. It’s not filling in blanks, it’s responding until it decides it’s said enough. Or, do you think that these models are (in some way) provided an endpoint?
The token-by-token generation of LLMs is extrapolation–it does not have an established endpoint–but I think hameddaoud is talking more about the kinds of concepts the AI is capable of representing. For example, ChatGPT’s image generation used to be unable to create full glasses of wine (until the people at openAI updated it for that specific situation.) This is good evidence that AI can only interpolate. It is a very simple extrapolation to draw a full glass of wine if you know how to draw a half-full glass and an empty glass. If one of the largest AIs is incapable of extrapolating something that simple, can any AI do it?
Hmm, okay, I see what you’re saying. So then, let’s say that we expand the model to have a better concept of “thought” - let’s say it first visualizes a glass of wine in 3D, and uses data about the 3D shape to generate a 2D image. Could we engineer a model to, instead of just guessing based on pixel data, actually use physical data to have a more thorough concept of what it’s representing? In that case, would it be the more-formal definition of extrapolation?
That’s good. It also gives a better framework to talk about extrapolation.
I think the big issue with the glass of wine is the difference between categories and spectrums. As humans, we think ‘fullness’ is a spectrum, where one side is a completely empty container, and the other side is a completely full container. We understand ‘half full’ and ‘quarter full’ as lying on the same continuous line. But for an AI, ‘emptiness’ and ‘fullness’ are two categories which an object can be in. Since people never fill a glass of wine to the brim, we say a glass is full when it really isn’t, and the AI picks up on that. It puts an actually half-full glass of wine in the ‘fullness’ category, and an empty glass into the ‘emptiness’ category. Since it only understands these categories, it doesn’t understand something being fuller than full, and so it can’t fill the glass to the brim.
So is that to say that AI is learning based off of cultural and societal norms instead of actual physical reality? I imagine that’s what we’d want, more or less, as the alternative is a very literal interpretation of prompts and the like. What do you think would be better?
I think it’s good if it can recognize the context and decide whether to use a more direct meaning rather than filtering it through cultural meaning. But we’re getting off-topic here. The main point I’m trying to make is AI struggles to extrapolate because it defines concepts in terms of categories, so a fuller glass of wine wouldn’t correspond to a glass filled to the brim, but something nonsensical. (like maybe the wine in the glass is just redder)
