I fully agree as it highlights the difference between the algorithm and the model (transformer).
This is true and not true. It is true in the sense that certainly with neural networks in general and extreme large ones like that used for ChatGPT we cannot say “this probability comes as a cause of X”. I.e., the resulting neural network is a black box. However, we do know how it works, and while outside the ability for most humans (due to time mainly), analyzing the training data you can give you some clues. Additionally, running a network backwards (a so-called dreaming neural network) can also gives some sense of why the neural networks exists the way it does. But again, you are correct in that neural networks are mainly considered black boxes outside of trivial examples (e.g, a two number adder).
If you mention “a blue dog jumping in pink grass” to any human, they’d be able to draw a picture of it. Maybe not an artistically-great picture, but most humans would be able to produce something that was identifiably representative of that phrase. And I would maintain that the ability to draw that picture is an inherent part of what it means for a human to understand that phrase. Even a human whose art abilities are so poor that they can’t draw that picture, would still recognize it if they saw it. But ChatGPT cannot draw or recognize that picture. It’s simply completely incapable of it.
Now, there are a number of other AIs out there right now that can draw a reasonable picture of “a blue dog jumping in pink grass”, so that ability clearly isn’t beyond the capability of AIs in general. And it may well be that, in the very near future, someone will create an AI that combines the capabilities of ChatGPT and Dall-E (in fact, I’d be surprised if this isn’t done, within a single-digit number of years). When that happens, that new AI will be closer to working like a human brain does. And closer yet, when they make an AI that can process sound as well as text and images, and then other senses.
The question was rhetorical, intended as a claim that there is no difference. Your response doesn’t seem to contradict that in any meaningful way. I mean, if I can have a back-and-forth conversation with ChatGPT about recipes, getting useful information in the process, surely that is “interpreting” my input, and doing so successfully and in a way that is genuinely useful! Parsing natural language was once considered a difficult task, and extracting semantic meaning in the intended context was judged a daunting task that was well in the distant future.
There were many famous examples of common sayings and expressions that were used to illustrate the difficulties of machine translation. One, for instance, was “the spirit is willing but the flesh is weak”. Absent any contextual understanding, this was translated into Russian as meaning roughly “the liquor is still good but the meat has gone bad”.
This is what I got out of ChatGPT:
Please parse the sentence “the spirit is willing but the flesh is weak”.
Sure! Here is the grammatical parsing of the sentence “the spirit is willing but the flesh is weak”:
“the” is a definite article, indicating that the nouns it precedes refer to specific things or groups of things.
“spirit” is a singular noun, the subject of the sentence, referring to a person’s non-physical essence or character, often associated with motivation or intention.
“is” is a linking verb, connecting the subject “spirit” to its predicate, “willing”.
“willing” is an adjective that describes the subject “spirit”, indicating that it is eager or ready to do something.
“but” is a coordinating conjunction, joining two clauses that contrast with each other.
“the” is a definite article, referring to the second subject of the sentence, “flesh”.
“flesh” is a singular noun, the object of the sentence, referring to the physical body of a person or animal.
“is” is a linking verb, connecting the object “flesh” to its predicate, “weak”.
“weak” is an adjective that describes the object “flesh”, indicating that it is lacking in strength or ability.
Overall, the sentence means that while a person’s motivation or intention may be strong, their physical body may not have the strength or ability to carry out the intended action.
It’s not actually “interpreting”, you say? It’s analyzing my request, analyzing the syntax and semantics of the example sentence with the semantics derived from context, and returning an accurate interpretation of a sentence that once posed such challenges to machine translation. Based in the principle of empirical evaluation of performance, it’s “interpreting” in any and every meaningful sense of the word.
(And I can discuss soup recipes with it, too. It seems like a good cook.)
This is a misunderstanding. Turing explicitly bracketed ‘internal’ considerations, allowing him to focus on behavior; his point was precisely to replace the question of whether machines think (which he thought ill-defined) with one that can be answered in a purely functional manner. To then claim that hence, all questions regarding machine cognition are just functional is both question-begging and antithetical to Turing’s argument. And that understanding reduces to the behavioral level, nobody has seriously held since the collapse of behaviorism in the '60s.
My argument is that there is nothing to deny, because there is no basis on which understanding can be claimed for ChatGPT. It explicitly works in such a way as to not pay any heed whatsoever to the semantic properties of tokens. What any given sequence of words mean is of no consequence to it; indeed, even if they mean anything doesn’t matter. What matters is that there is a sufficiently large corpus where words are arranged showing regular patterns. Then, it can replicate these patterns—that’s all. What it’s talking about simply doesn’t enter into the process at any point. This isn’t something we need evidence to decide, it’s just how the thing works.
You can’t decipher a language if all you have access to is texts written in that language. The relationships between the words don’t suffice to triangulate their meaning. But that’s all that ChatGPT has access to.
I haven’t denied that. But that alone doesn’t make him or his views immune from criticism, much less right.
But that’s the difference: ChatGPT doesn’t form concepts. The next word is fully and solely a function of the preceding words, without even any conceptual room for the consideration of what those words mean. ChatGPT does not know what it’s talking about when it’s talking about pink grass and blue dogs, because its architecture doesn’t care about what the words mean. There’s nothing in there that’s even sensitive to word meanings—they simply play no role for it.
Again, consider a version of ChatGPT trained on Voynichese texts. It will be able to produce Voynichese outputs, convincingly. If anybody ever manages to translate Voynichese, they’ll even be meaningful. Does that mean that it’s learned Voynichese, just from the text? Of course not, because that’s impossible. What it means is that the ability to produce intelligible text is not predicated on having a productive understanding of the text—which is a far more remarkable insight than ChatGPT understanding the text would be. (After all, we have perfectly good understanders already!)
I think it’s here that the compression-view becomes useful. Suppose you measure, say, the rate of decay of some radioactive isotope. You get out a sequence of data points, activity over time. You feed those data to a computer, which then computes a best-fit function. Using this function, if predicts the data for future points in time. You go out and measure the activity at that point, and find its predictions confirmed (within experimental tolerances).
Has the computer understood radioactive decay? Does it know anything at all about activity, time, half-life and so on? No: in fact, those considerations never entered into its prediction. To the computer, these were just numbers with a certain regular pattern. Tokens entirely divorced from what they represent. That the computer could make a successful prediction does nothing to imply otherwise. It could just as well have been, say, the rate of discharge of a capacitor, or anything else.
ChatGPT works exactly like that: using patterns in a large dataset to extrapolate further points of data. What the data represent, what the tokens mean, is of no consequence to the whole process at any point. Give it artificial, meaningless data without any mapping to reality: it will just as happily reproduce the patterns in that data.
We found out many posts ago that chatGPT can “speak” different languages: German, Spanish, Hebrew, Japanese… I wonder whether someone with the ability to judge the quality of the output has asked it to compose music using the usual pentagram and notes notation. I can’t read music myself, so I could not judge the result. Just asking.
Reminds me of a former professor of mine who sometimes said “How am I supposed to know what I mean before I have said it?” But I think he said it tongue in cheek.
I"m nore sure what you mean by “pentagram” in this sense, but there are plenty of videos of people using ChatGPT for music composition. I have tried it in various ways, from creating guitar and bass tablature to giving me chord progressions and continuing chord progressions and the such, and it is – at the moment – not good at the task. There are music AIs being developed, too. I know Google has one – I forget its name at the moment – and that one produces actual audio. From what little I’ve seen of it, it’s definitely better than ChatGPT, but still has a ways to go. Give it a year or two.
Oh, sorry, I see now that you call it staff in English. I assumed that with the five lines it would be called pentagram in all western languages, my bad.
ChatGBT balked at first but warmed to the idea after a couple of exchanges:
ChatGBT response
Indeed, in a high-stakes context like naval operations, the primary concern is often the timely and accurate analysis of data, rather than the cost of developing a custom model or algorithm. In such a scenario, a custom model based on the Transformer architecture could potentially offer advantages over more traditional methods of data analysis.
For example, the Transformer’s ability to model complex relationships between input data could potentially help identify patterns in sensor data that might be difficult or impossible to detect using more traditional data analysis methods. Additionally, the ability of the Transformer to generate predictions based on input data could potentially be useful in helping commanders make decisions quickly and accurately.
Of course, the development of a custom model for warship damage analysis would likely require significant resources and expertise, both in terms of data collection and processing, as well as in model development and testing. Nonetheless, for high-stakes applications like naval operations, the potential benefits of such a model might justify the investment.
I was wondering a few minutes after I posted that whether that is what you meant, when I realized the staff has five lines. I don’t know if the word is ever used to refer to the musical staff in English, but my idea of a pentagram is a type of five-pointed-star symbol.
Anyway, it won’t do a true staff, that I can tell, but there may be some sort of prompting that will get it there, if there is some sort of standardized ASCII representation like there is for guitar. It will spit out some tablature for you, though. It’s – uh – given me some odd results. I will ask it for something like a single note bassline and it’ll give me bass tabs with four-note chords on it.
Like I just asked it for a walking bass over C - Am - F - G, which each chord lasting a half note, and each bass note an eighth. It did give me singe notes this time (yay!), but it confined itself completely to the D string, and gave me this tablature on that string:
Alright, as promised, here’s what I got out of trying to get the AI to write in various ancient languages.
First, I asked ChatGPT to write the Pledge of Allegience in Ancient Egyptian Hieroglyphics
I then copied the hieroglyphics back in and asked what they meant. Response:
I looked into it, and as best as I can tell ChatGPT is a very weird mix of correct and incorrect here. This is an inscription for a pharaoh, but not Akhenaten, but instead Neferneferuaten who was in fact a FEMALE pharaoh - so references to Akhenaten and what ‘he’ did are incorrect. However, both names reference the god Aten?
So when I asked it to deliver the Pledge in Ancient Hieroglyphics, it instead found some ancient text (the inscription of the name of Neferneferuaten) and tried to pass it off as the pledge. When challenged to explain the text, it revealed that it wasn’t the pledge, but still lied about what the actual text is.
This brings me back to that AI Alignment problem. The example given was asking ChatGPT to write various types of poems, where it would reply with a generic rhyming poem; this is because most people training ChatGPT didn’t know enough about poems to penalize it for delivering the wrong type of poem, so it was trained to give a response that’s ‘close enough’ for whoever is rating its answers. I doubt that ChatGPT’s training focused on transliterating text into ancient hieroglyphics.
I then challenged the response ChatGPT gave:
Response:
I ask ChatGPT to try again:
I ask it to identify the hieroglyphics it provided. Response:
Well, it either figured out how to give me what I want, or it figured out how to lie about what it has provided. Let’s find out which…
So! Now we are getting somewhere! It’s showing its work! Let’s check on that.
𓉐 = “I” (as in “I pledge”)
But when I look up 𓉐, it actually stands for a house, and relates to the sound ‘pr’ (pronounced ‘per’) meaning house. It is related to the English B and the Hebrew Bet, and many other letters like that. But in no way does it correspond to “I” as in “I Pledge”.
𓃀 represents a foot, and corresponds to the sound ‘b’. 𓇋 represents a reed, and makes the sound ‘j’. Not “a”, as stated.
We can keep going, but you get the point. ChatGPT will happily deliver a breakdown of the work it did that has absolutely no correspondence with reality.
I think this goes to back up what @Half_Man_Half_Wit is saying, even though I was initially disagreeing. ChatGPT will deliver a response that has all the structure and appearance of a perfectly reasonable, well thought out response; but in actuality, it is meaningless.
Not at all. My point is precisely to replace the question of whether ChatGPT possesses “understanding” (which I think ill-defined) with one that can be answered in a purely functional manner. I make the same category of claim as Turing, and I make no other.
“Behaviourism” seems to be generally used in the context of psychology or in philosophy in reference to human psychology. The funny thing is, the aforementioned “collapse of behaviorism in the '60s” (or thereabouts) coincided with the founding and rise of the new interdisciplinary field of cognitive science, and of luminaries like Jerry Fodor and his colleagues and many others during that period. None of which has any relevance to my functional approach to assessing the performance of ChatGPT.
