Are you being sarcastic? There's no way you believe people couldn't wash dishes or clean floors because of all the emails and birthday greetings they had to write, or that people have almost completely lost the ability to consider making art.
>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.
>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.
>If multimodal models were still stochastic parrots by the original argument, humans would have to be as well; we don't have any way to ground anything beneath sense data
Animals don't passively learn from their perceptions, don't have a separation between training and inference, and don't have a prompt-response execution model. Besides its fundamental biology, the grounding an animal brain has is that when it outputs a motor signal it receives some feedback as to the effects of a signal of that strength. A bird learns to fly because the grounding truth of aerodynamics and gravity consistently respond in a specific way to the flapping of its wings. It doesn't learn by passively replaying thousands of hours of somatosensory recordings of flights.
A multimodal model doesn't have the capacity to do much with a prompt. It has no head to turn to look at an image from a slightly different angle to attempt to gleam more information, doesn't have the capacity to interact with the real thing the image represents in any way, and even if it requests another angle and is given it, it lacks the capacity to learn that new information permanently. A multimodal model knows about images of pipes and facts about pipes, but doesn't know pipes; it doesn't have literally first-hand experience with them.
>evolution can't have programmed some innate grounding into us because it didn't either.
What do you mean? Of course genetics programs ground truths. For example, "forward is the way your face points when your neck is relaxed" and "if you can feel it, then it's part of your body".
To add to dwa3592's comment, a sentence is not a self-contained idea. The sentence doesn't include what any of the words in it mean, nor what "this sentence" refers to. The exact same sentence can mean different things depending on the text that surrounds it.
Translation is good enough, but quite often no translation exists. MTL is at this point good enough for simple texts and for everyday living like restaurant menus and shopping, but any even mild complexity in a text, like long-running context (necessary for consistent translation of terminology), puns, purposeful ambiguity, etc., MTL is completely unable to cope with. It just doesn't work and it will likely never work.
But it still counters the claim that multilingualism is a proxy for high sociability. I speak two languages (and I'm working on a third) and most days I don't speak to anyone.
Steelmanning, I think the argument is that a vindictive but fair AI would not torture someone who did not cooperate with it because they didn't believe the threat was real, because they were merely mistaken, not malicious. It's similar to how a just god would not damn sincere unbelievers, it would only damn true believers who nevertheless refuse to worship it.
I only counted the cost of the satellite itself. I assumed you can get it into orbit for free, so if it costs a cent more than that, it's a further argument against.
>The max steering angles of the phased arrays are much higher
You can't steer the antenna back and forth for every exchange between station and customer. What the steering may get you is increasing the coverage of an area currently underserved by the constellation, and maybe a slight increase in diameter of ground covered due to the geometry, at the cost of lower signal strength.
>And for the last few years, lasercoms can route traffic inside the constellation so a given sat doesn't need to be within sight of a ground station.
Did they finally implement satellite-to-satellite links? Fine, if that actually works, they can indeed extend the range much further. I don't know if I believe it, though.
I wonder about the economics, though. Intuitively it doesn't seem like it can be more efficient to launch constellations of satellites than run kilometers of cables, even if you have to run 20 km for each customer. That's, what, $10k a pop? So around two orders of magnitude cheaper than a satellite? Something isn't adding up for me.
That's not semantic nitpicking, that's just taking words at face value. I wasn't aware we could retroactively change their meaning when it became convenient. Can I try? What I meant by my original comment was you made a very eloquently put point that made a masterful use of irony in its rhetoric.
I don't understand. Starlink satellites are just routers to ground stations. Why are there no wired connections available? Do the connections not reach the famous last mile?