[fluoridation]

>A bird doesn't learn gravity or aerodynamics, it has no 'sense of physics'.

That's not what I said. What I said was that it's physics that provides the ground truth.

>you could, at least in theory, falsify the entire experience of the bird

It wouldn't be a bird anymore, but a dysfunctional cyborg with false perceptions.

>There is nothing in a bird's brain that directly percieves reality.

Yes, of course there is. Animal sensory organs do not produce false information, nor do they provide the brain an interpretation of what they perceive. The brain may fail to distinguish hallucinations from reality, but those are processes internal to the brain. What it gets from the body is raw physical measurements, and what it sends out is raw motor commands.

A multimodal model doesn't have the same direct access to reality, it just has collections of words and images. It has no capacity to determine the reality of a photograph of a sunset or a CGI render of a dragon. The word "real" is itself meaningless to it; they're both real in that they appear in its training corpus. It lacks the capacity to investigate these stimuli in any way, and can just learn to associate different stimuli in arbitrary ways that appear to make sense to us, but nothing else.

[throwaway7356]

> nor do they provide the brain an interpretation of what they perceive

> What it gets from the body is raw physical measurements

No, sensory organs like eyes do a lot of processing ("interpreation"). They certainly don't send "raw physical measurements" to the brain.

[supern0va]

And even then, it's just sampling. Much of what we "see" is a prediction, and there are plenty of optical illusions out there premised on that (plus VR techniques like foveated rendering that take advantage).

[ToValueFunfetti]

>It wouldn't be a bird anymore, but a dysfunctional cyborg with false perceptions.

The criticism in the paper is of the architecture of LLMs, isn't it? The paper contends

"""Text generated by an LM is not grounded in communicative intent, any model of the world, or any model of the reader’s state of mind. [...] an LM is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning: a stochastic parrot"""

They're saying that the model cannot learn anything about reality irrespective of training data. Your point is an interesting one, but I think it's distinct. To your point though, this is unfalsifiable from the perspective of the "bird". I can't prove that I'm not a dysfunctional cyborg with false perceptions, which makes me wonder if that's a meaningful distinction.

>Animal sensory organs do not produce false information,

I happen to be in possession of some of these and I think this needs a "usually, under ordinary conditions". (Nitpicky and not critical to my point, but I liked the beginning of this sentence too much to edit it out)

>nor do they provide the brain an interpretation of what they perceive

Whereas this I'd argue is not true at all. My cones interact with wave-particle photons at particular wavelengths. I can't even conceptualize wave-particle duality (though some humans can), but "red" and "blue" are the bread and butter of my visual consciousness. These correspond to firings of my sensory neurons much more than they correspond to anything in reality. If that's not interpretation, what is it/what is interpretation?

>The word "real" is itself meaningless to it; they're both real in that they appear in its training corpus.

I expect we agree that I can show a multimodal model a real and a CGI picture and it can tell me which is which. I can take a CGI dragon to GPT 5 and say "look what I found in my backyard" and it will say "Yeah right". Are you saying this is something only possible thanks to RLHF or other modern techniques? That may be the case, unsure how to test that without access to pretraining-only models. Or would you say my experiment is faulty here and doesn't get to your underlying claim?

On the flipside, I could show some meh drawings of fairies to Arthur Conan Doyle, and he'd say "Whoa, this changes everything". I consider him to be one of the great rational minds of history, but he was unable to pass your test here. (In fairness, he was in his 60s and his senses may have dulled, though his belief in spiritualism at large dates to his prime).

Appreciate the conversation!

[fluoridation]

>I can't prove that I'm not a dysfunctional cyborg with false perceptions

It doesn't matter if you are one. If you were an AI researcher and encountered a model that saw things for what they really are you would deem it to be malfunctioning and discard it. Any functional model will always be at least one degree further removed from a raw sensory experience that agrees with yours, than you. Its perception will be invariably filtered through the lens of labeled human output.

>I happen to be in possession of some of these and I think this needs a "usually, under ordinary conditions".

I won't dwell much on what you mean, since you said it's unimportant, but it takes a lot for sensory organs to malfunction, and even when they do, they produce corrupted, not false, information; black blotches, not pink elephants. I assume you were thinking of alcohol or something; drugs that affect perception affect the brain, not the other organs.

An interesting edge case is stuff like entoptic phenomena, but those aren't false perceptions; they're, if you will, hypertrue perceptions (something so true, we would rather not see it).

>but "red" and "blue" are the bread and butter of my visual consciousness. These correspond to firings of my sensory neurons much more than they correspond to anything in reality. If that's not interpretation, what is it/what is interpretation?

"Red" and "blue" are your brain's interpretation of the signals it receives from the eye. Notice how "red" and "blue" are fully abstract words, decoupled from anything physical, whereas if I were to refer to the actual encoding of visual information that passes through the optic nerve, I'd have no choice but to reference physical processes (probably talk about voltage and action potentials; I honestly don't know how the optic nerve works). That's because a retinal cell is a simple transducer. The eye doesn't interpret, it merely converts and encodes. Interpretation is a higher level operation.

You cannot compare the simplicity of the mechanism of a whole eye to the indescribable complexity that is between the sentence "roses are red and violets are blue" written on a book, and the raw perception of red roses and blue violets. But a model will only ever be exposed to that distant interpretation, not to roseness or redness.

>I expect we agree that I can show a multimodal model a real and a CGI picture and it can tell me which is which.

Of course, but obviously that's not something it knows inherently. There's nothing about the image intrinsically that says it's fake; someone has to label it such that the model can associate it with unreality (according to our own parameters). That's not how an animal works. An animal assumes what its senses perceive is real and can distinguish its own thoughts from its sensory input.

>On the flipside, I could show some meh drawings of fairies to Arthur Conan Doyle, and he'd say "Whoa, this changes everything". I consider him to be one of the great rational minds of history, but he was unable to pass your test here.

That's a slight equivocation. He would not have mistaken the drawings of fairies for raw perceptions of fairies, he would have simply been swayed by the rhetorical strength of the testimony implicit in the drawing (and perhaps an explicit one that accompanied it). If you want to make a true parallel to a multimodal model you'd have to compare a CGI fairy and a photograph of someone holding a drawing of a fairy. The CGI is as raw to the model as the signals passing through your optic nerves right now, but the drawing is one level of abstraction further away.