Gábor

I don't know (of course) what new advances will come but I strongly suspect that "mere philosophers" will no longer be able to tackle it instead of people with much deeper mathematical and comp science background. I clearly think this is the end of the line for any argument like the CRA. That's why I considered the approach presented in the paper of Roddus futile: it's nothing more that the age old "let's list what previous philosophers have said and add one more" approach.

To be clearer: I don't say CRA has already failed. I posit it seems already pretty obvious that it will and it will not be a final, universal argument that we can keep refer to and consider valid. And I also posit that CRA was about the last philosophical argument in this area that had the chance to be as simple as a few sentences written so that everybody can easily understand.

We don't yet know whether such an artificial intelliegence is achieveable in the future (although I don't think too many of us doubt that it will eventually happen) but if it does, CRA will obviously not apply to it as it's formulated now.

If we suppose there will be some artificial intelligence (let's say an android like Data in Star Trek TNG or anything similar) then it will be immediately obvious that whatever argument CRA has about the actor closed inside the Chinese Room not understanding Chinese is false; such an artificial intelligence will be able to acquire the actual understanding of Chinese just as well as it can acquire the actual understanding of any other subject.

CRA claims to be an universal argument and that's why I say it's starting to fail. ChatGPT is not intelligent, as we have already established multiple times, and therefore, it's clearly still subject to CRA. However, CRA cannot say anything about possible and probable future advances that will not fit its rulebook approach any more.

I don't think randomness is in any way important to the discusssion at hand. It was just a side remark to mention that being deterministic is not necessarily a rigid requirement any more, depending on what you want to accomplish.

Contemporary conputers are no longer necessarily monotasking, not since practically all our mainstream processors (desktop and mobile) are multi-core and since we have parallel processors (CPU and GPU, more recently NPU) in our machines.

"Two centuries ago" was a very general reference to the idea of "thinking machines", not anything specific to today's computers.

And this accumulated knowledge base will grow and new computing techniques will come, who know what quantum computers will add to this area (and no longer deterministic, by the way, by virtue of the underlying physical processes).

Of course, this is still a metaphor, ChatGPT accumulated a lot but not the "whole knowledge of humanity". Still, you get the metaphor. The LLM has not only access to some arbitrary rules but to a vast repository of knowledge. Much larger than anybody ever dreamt of half a century ago and much larger than could be formulated into predetermined programs by programmers.

And the question of metaphor is not an irrelevant one. The CRA says that a human without understanding but following the rulebook can appear as someone with understanding to the outside world. But this fails to graps the idea that, instead of a mere rulebook, the computer program we try to compare to is no longer a rulebook but the accumulated knowledge of humanity.

@haxor789 The "set of pipes" never was a good metaphor, a computer doesn't work that way. I can understand why a thinker two centuries ago might have used this metaphor but it's fundamentally flawed.

But that's not the real point. The real point is that we're going beyond fixed programs with fixed (albeit very large and complex) statements, conditions and rules, written by programmers line by line, and followed by the computer.

ChatGPT does use randomness while generating the answer, this is well documented in many descriptions.

Yes, I know well what the CRA says. I simply say it's no longer relevant. The rule book was a metaphor back then. It no longer is an applicable metaphor. Or, at least, we're on the verge of not being able to use this metaphor any more.

This is exactly where CRA starts to fall apart. There's no "rulebook" that the actor inside the CR can follow to make the outside world believe. Not even in theory any more.

@haxor789 Well, a computer program doesn't need to be deterministic, even on the very same input. And no, it isn't necessarily true that you can guess from the source code what its output will be. ChatGPT itself is a prime example, I seriously doubt it would be possible (not humanly, of course, because that's obviously impossible but even theoretically) to determine by its "source code" what its output will be.

Turing: But then he also said the idea of a thinking a machine is largely meaningless. Ditto. We respect Turning for everything he did but why is it important what he thought about thinking machines back when computers were hardly more than the maximum amount of relays that could be stuffed into a room? You need to consider that computers have grown not only quantitatively but qualitatively since then. And will continue to do so.

This is a moving target, and a very rapidly moving target at that. There are areas where old adages can still be valid but this one is definitely not among those. You really should try to get away from the approach of "let's see what earlier philosophers said, let's quote them and let's say something so that we will be one more philosopher in the long list that the future will quote". This simply doesn't work here any more.

I know you have some philosophical background and I know philosophy still works largely along these lines but seriously, why is it important what somebody said 70 years ago? But seriously? Yes, the claimed so, so what? We've come an enormously long way since then. Newel and Simon simply didn't have clue back then. Searle didn't have a clue back when he formulated CRA, either.

But even if you do, there's hardly any need for a "scientific paper". Just a newspaper-level description of what ChatGPT actually does is sufficient to show that it's nothing near intelligent, just the next step in our searching machines. It's impressive at that and even if it gets the facts wrong in many cases, it can wrap them into nice sentences and that's an achievement.

@Roddus I really don't understand why you consider it important to find an justification to the conclusion that ChatGPT isn't intelligent. It isn't. It was never claimed to be. The I in the AI is just a marketing moniker. You -- presumably -- don't author a paper to prove that cars aren't intelligent because it's common knowledge that they are not and there's no opposite claim that merits a rebuttal.

All these conversions have one very important thing in common: they are intentionally created unambiguous, so whatever two representations they convert between, they are all designed to retain the same text.

Which is, actually, the exact same reason why you could read my sentences in spite of the fact that there's probably an ocean between the two of us. Because my computer receives the text from me and yes, after many various conversions, it reaches you and you see it after even more various conversions.

The text "cat" is not these three characters you see on your monitor now. It is any unambiguous representation of the symbol <C> followed by any unambiguous representation of the symbol <A> followed by any unambiguous representation of the symbol <T>.

Trying to put it even more simply: text is content, not external representation.

You already received heaps of examples like Braille for blind people, Morse code, we could even add arbitrary new ones like assigning a specific flower to each and every letter and planting a garden with rows of flowers representing the same text, all that is text and all these representations denote and contain the same text. As long as a computer is able to use any of these representations, it operates on the text itself.

As long as you stick to shapes, you simply won't understand. Text is not shapes. No matter what it was during a few centuries when printing (or writing) was the only means of storing text, text is a sequence of symbols. Its actual representation is absolutely, completely and utterly irrelevant.

@WeatherVane Yes, I know, I read your initial comments. Just that you asked later whether the paper had any definition of "text" and I replied that it did. But I don't think the real misunderstanding has anything to do with the communication between server and browsers. OP simply starts with the patently false claim that "text is only that which we can read with our eyes" and reaches the outlandish "conclusion" that computers or the internet know no text at all.

@WeatherVane Page 17 of the linked PDF: "Text is written or printed language such as we find in textbooks. It comprises instances of shapes." This is the basic problem of this "reasoning": the initial claim is completely bogus.

@WeatherVane Actually, it does and a fundamentally and completely flawed one: it only recognizes text as printed text.

As to the ANN, of preferably neural networks, it seems quite certain that this is the path forward. But, even if we're on Philosophy.SE, I actually think that this question is well above the paygrade of any philosopher, Searle or other. The basic papers that ChatGPT fundamentally relies on are a very tough nut to crack even for well versed mathematicians. It seems increasingly unlikely that by sitting back and thinking up a nice little CRA with a few sentences is something that will ever just scratch the outermost surface of what comes.

In other words, you won't be able to present a paper (even if it was based on valid premises, unlike yours at the moment) that gives a completely new spin on how to go forward towards real artificial intelligence. This research and this endeavor now, irrevocably, requires such a deep mathematical background and apparatus that I seriously doubt any contemporary philosopher comes even remotely close to it. Relevant works today need to contain a myriad of cryptic Greek letters of formulae rather than a myriad of quotes of what other philosophers said a decade or so ago. A whole new ball game.

@Roddus you're coming closer but no cigar. :-) You still claim the "usual sense of text". No. You still get the usual sense of text wrong. Consult the dictionary to see the meanings (plural!) of text, there will be about half a dozen. When we speak about computers, internet, digital formats, everybody uses a different meaning than you do, and that's so obvious that it shouldn't need specification. And this leads to the situation that any of your reasoning based on another meaning simply has no bearing on the issue at hand. It's simply a non sequitur.

So, no, thanks, I won't go out and listen to this and that. Irrelevant. What's relevant is what was said in all these answers and comments. ChatGPT doesn't understand because it wasn't made to understand and because we are not yet able to make it understand. And not because of some laughably false definition of text. ChatGTP (and computers) operate on text, just they do something different than understanding.

I'm very well aware that most in philosophy is based on authority of somebody having said something and all quote that but this isn't important. That fact that Searle is an experienced philosopher is completely irrelevant. Even the least experienced non-philosopher can prevail over the most experienced philosopher if he happens to say the truth. Argue with content and not quotes.

@Roddus The time will come when, not ChatGPT but some later system will understand, this already seems inevitable. Not everybody agrees, of course, but I guess the majority already thinks so. However, the fact the ChatGPT is not yet there doesn't make your paper and the premise inside any more valid. It's fundamentally flawed, whether you accept it or not. You simply decided to create a very restrictive interpretation of "text" for yourself that nobody accepts as valid -- you can see it from the answers, there's not a single one that would give the slightest credence to it.

@JD As a side note, while we cannot yet prove CRA's invalidity, I don't think there's much hope in anybody that Searle will prevail. It was formulated in an era when programming was really only the computer following specific rules written by specific programmers. While ChatGPT and the likes are only the very beginning of this journey, what we already can tell for sure is that computers are gaining ways to operate that are beyond what the CRA "rulebook" analogy suggests. While CRA was valid yesterday and still is today, it's a terribly weak argument for a yet unforeseeable future.

@Roddus By the way, I saw it now in your paper that "enthusiasts say ChatGPT is learning the meanings of the text (shapes)". No matter if shapes or not, nobody says that, at least not anybody with a modicum of knowledge of what ChatGPT is. There's absolutely and positively no question of "learning the meaning". It's a statistical system that tries to determine what word should come next after the current one. There's no meaning and no understanding of any kind involved, and it was never claimed there was.

You confuse a single (albeit widespread) representation of text (written or printed text) with the second dictionary meaning of text, ie. content. Computers, if so instructed, interpret text (=content) and the representation they use to store that text is absolutely irrelevant, it's just a question of convention and practicality. The whole basis of your reasoning is fundamentally and irreparably flawed and doesn't hold an ounce of water. ChatGPT is fed and works with text all right, even if this doesn't make it intelligent.

@Roddus No. Shape is but a happenstance. Blind people see no shapes, still can read and understand text. I could write this paragraph in an alphabet you never met, using shapes you never saw and have no idea of what they represent, still, it would remain exactly the same text. I could convert it to Morse codes that have nothing to do with the shapes of the letters of the Latin alphabet, and it would still remain the same text. And if I don't write it to in dots and dashes but whistle it to you, it would still be the same text.

Forget shapes completely. There are just one representation of the very many that we use to create a physical representation of text. Shapes do not text make. You claim that the only place you can find text in a computer are the markings on its capacitors. Then I'll claim there's no text there at all, just protons and electrons inside a pigment that smeared on another surface, also made of electrons and protons, wrapped around some other material, also made of electrons and protons. Is it nonsense? Yes, it is. Just as nonsense as your original claim.

@KenShirriff -- the world is so full of urban legends that this is hardly a surprise. But this story is so outlandish that you don't need formal proof to be sure that it's bogus. It's written all over it: in uppercase, shouting letters. :-))

Teletype conventions aside, this is the base, really no need to try to make an ideology of the contrary. Teletype had obvious mechanical reasons to settle on a lettershape that had no size variations, just like the first typewriters invented (both Hansen and Sholes) but legibility of longer passages of text was certainly not among those.

Well, come on, clearly, the first and most basic tenet of typography is that uppercase is less readable because it's less distinctive (yes, mostly due to ascenders and descenders but not only). We have 500 years behind us with practically all printed material, the bulk of the text I mean, always set in lowercase just because of that.

For your purposes, the version of the PDF is of no significance. Later versions learned new tricks that mostly deal with images, transparencies, cool effects. For plain text or even plain images, a PDF from the time before Columbus discovered the Americas is still just fine.

What you actually lose when converting to curves is the so-called hinting information, fonts have more than just the actual outline of the letters, extra instructions that help the rendering. But this was much more crucial for 300 and 600 dpi laser printers than the 2400 dpi or higher that's used today when creating the printing plates (or, 10-15-20 years ago, when we were creating films first, and then the films were used to create the plates, it's a single step now).

The printer certainly isn't using Corel Draw just for the printing, CDR is a graphics package (very much like Adobe Illustrator or Inkscape) but yes, you can produce PDF directly from it, and you could always "convert to curves" directly from CDR, probably that's the reason for the reference.

And if you follow the objects, you'll see that it's really simple, in the end. The Pages contains the Kids, a list of the actual pages, each kid contains the size boxes plus the actual page contents, and the contents all describe the actual text and images.