# The next page in the book of AI evolution is here, powered by GPT 3.5, and I am very, nay, extremely impressed **URL:** \ **Category:** Cafe Society\ **Tags:** ai\ **Created:** [December 2, 2022, 10:54pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945 "2022-12-02T22:54:31Z")\ **Posts on this page:** 20\ **Page:** 36
**Author:** ![BeepKillBeep](https://avatars.discourse-cdn.com/v4/letter/b/9e8a1a/32.png) [@BeepKillBeep](https://boards.straightdope.com/u/BeepKillBeep)\ **Post date:** [February 18, 2023, 2:33pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/701 "2023-02-18T14:33:52Z")
> [@Sam\_Stone](#): > > Take the transformer off the back end, and ask ChatGPT a question. You get back a list of probabilities which, when put through a transformer process, tell you what ChatGPT is ‘thinking’. Foocusing on the transforming part is a mistake, IMO. It’s just the system 1 that puts ChatGPTs thinking in a form we understand. I fully agree as it highlights the difference between the algorithm and the model (transformer). > [@Sam\_Stone](#): > > We actually don’t know how it comes up with the list of probabilities for the tokens. 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). ---
**Author:** ![Chronos](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/chronos/32/134_2.png) [@Chronos](https://boards.straightdope.com/u/Chronos)\ **Post date:** [February 18, 2023, 2:42pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/702 "2023-02-18T14:42:58Z")
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. ---
**Author:** ![Sam\_Stone](https://avatars.discourse-cdn.com/v4/letter/s/ecccb3/32.png) [@Sam\_Stone](https://boards.straightdope.com/u/Sam_Stone)\ **Post date:** [February 18, 2023, 2:45pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/703 "2023-02-18T14:45:10Z")
There are rumors that GPT-4 may be multi-media. If not, something else will be along shortly. Then it will get even more interesting. ---
**Author:** ![wolfpup](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/wolfpup/32/10618_2.png) [@wolfpup](https://boards.straightdope.com/u/wolfpup)\ **Post date:** [February 18, 2023, 3:02pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/704 "2023-02-18T15:02:47Z")
> [@Crane](#): > > > [@wolfpup](#): > > > > Where does “token-matching” end and “true understanding” begin? > > When the output is interpreted. 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.) 😉 ---
**Author:** ![wolfpup](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/wolfpup/32/10618_2.png) [@wolfpup](https://boards.straightdope.com/u/wolfpup)\ **Post date:** [February 18, 2023, 3:03pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/705 "2023-02-18T15:03:55Z")
> [@Sam\_Stone](#): > > There are rumors that GPT-4 may be multi-media. If not, something else will be along shortly. Then it will get even more interesting. Multimedia probably means voice input and output. That would be really cool! ---
**Author:** ![Sam\_Stone](https://avatars.discourse-cdn.com/v4/letter/s/ecccb3/32.png) [@Sam\_Stone](https://boards.straightdope.com/u/Sam_Stone)\ **Post date:** [February 18, 2023, 3:15pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/706 "2023-02-18T15:15:02Z")
It might also mean the ability to process images and video, and to produce images and video along with text. ---
**Author:** ![Half\_Man\_Half\_Wit](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/half_man_half_wit/32/21766_2.png) [@Half\_Man\_Half\_Wit](https://boards.straightdope.com/u/Half_Man_Half_Wit)\ **Post date:** [February 18, 2023, 3:44pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/707 "2023-02-18T15:44:57Z")
> [@wolfpup](#): > > As Alan Turing recognized long ago, this is completely the wrong evaluation criterion, and it’s exactly what you’re doing. The rote mechanistic functioning you describe makes it sound like a more sophisticated version of Eliza. For purposes of this argument, the internals of ChatGPT must be regarded as _ **completely irrelevant** _. 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. > [@wolfpup](#): > > The above interaction was mostly just for fun, but the key question is, does it have sufficient “understanding” to be able to “imagine” and thereby create a picture of that scene, given a picture-creation subsystem? I don’t see how you can deny it, and that was the entirely of your argument 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. > [@wolfpup](#): > > He was a towering figure in cognitive science who I daresay greatly overshadowed his various critics. I haven’t denied that. But that alone doesn’t make him or his views immune from criticism, much less right. ---
**Author:** ![Half\_Man\_Half\_Wit](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/half_man_half_wit/32/21766_2.png) [@Half\_Man\_Half\_Wit](https://boards.straightdope.com/u/Half_Man_Half_Wit)\ **Post date:** [February 18, 2023, 3:54pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/708 "2023-02-18T15:54:43Z")
> [@Sam\_Stone](#): > > I certainly don’t consciously think of every word I’m going to say as I say it, I just form concepts, and then the words just follow. Maybe the speech center is spitting out words with a similar mechanism. 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. ---
**Author:** ![Pardel-Lux](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/pardel-lux/32/20182_2.png) [@Pardel-Lux](https://boards.straightdope.com/u/Pardel-Lux)\ **Post date:** [February 18, 2023, 4:01pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/709 "2023-02-18T16:01:40Z")
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. > [@Sam\_Stone](#): > > I certainly don’t consciously think of every word I’m going to say as I say it, I just form concepts, and then the words just follow. Maybe the speech center is spitting out words with a similar mechanism. 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. ---
**Author:** ![pulykamell](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/pulykamell/32/3166_2.png) [@pulykamell](https://boards.straightdope.com/u/pulykamell)\ **Post date:** [February 18, 2023, 4:19pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/710 "2023-02-18T16:19:55Z")
> [@Pardel-Lux](#): > > 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. 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. ---
**Author:** ![Crane](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/crane/32/3495_2.png) [@Crane](https://boards.straightdope.com/u/Crane)\ **Post date:** [February 18, 2023, 4:20pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/711 "2023-02-18T16:20:31Z")
> [@wolfpup](#): > > (And I can discuss soup recipes with it, too. It seems like a good cook.) Does it have a gender? Brief hijack: I wonder if this technology is being applied in other areas like: embryology warship damage evaluation weather forecasting ? Could be an interesting thread, I’ll ask ChatGBT ---
**Author:** ![Pardel-Lux](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/pardel-lux/32/20182_2.png) [@Pardel-Lux](https://boards.straightdope.com/u/Pardel-Lux)\ **Post date:** [February 18, 2023, 4:33pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/712 "2023-02-18T16:33:28Z")
Oh, sorry, I see now that you call it [staff](https://en.wikipedia.org/wiki/Staff_(music)) in English. I assumed that with the five lines it would be called pentagram in all western languages, my bad. ---
**Author:** ![Crane](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/crane/32/3495_2.png) [@Crane](https://boards.straightdope.com/u/Crane)\ **Post date:** [February 18, 2023, 4:41pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/713 "2023-02-18T16:41:51Z")
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._ ---
**Author:** ![pulykamell](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/pulykamell/32/3166_2.png) [@pulykamell](https://boards.straightdope.com/u/pulykamell)\ **Post date:** [February 18, 2023, 4:44pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/714 "2023-02-18T16:44:28Z")
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: 0 - 0 - 2 - 2, 3 - 3 - 2 -2, 1 - 1 - 0 - 0, and 2 - 2 - 3 - 3 which corresponds to: D - D - E - E (over C) F - F - E - E (over Am) Eb - Eb - D - D (over F) E - E -F - F (over G) which is a very, um, unusual selection of notes for a walking bass in C over a I-vi-IV-V progression. ---
**Author:** ![Snarky\_Kong](https://avatars.discourse-cdn.com/v4/letter/s/a183cd/32.png) [@Snarky\_Kong](https://boards.straightdope.com/u/Snarky_Kong)\ **Post date:** [February 18, 2023, 4:48pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/715 "2023-02-18T16:48:23Z")
Deepmind released a multi media chatbot paper almost a year ago. > **[Tackling multiple tasks with a single visual language model](https://www.deepmind.com/blog/tackling-multiple-tasks-with-a-single-visual-language-model)** > > We introduce Flamingo, a single visual language model (VLM) that sets a new state of the art in few-shot learning on a wide range of open-ended multimodal tasks. ---
**Author:** ![Snarky\_Kong](https://avatars.discourse-cdn.com/v4/letter/s/a183cd/32.png) [@Snarky\_Kong](https://boards.straightdope.com/u/Snarky_Kong)\ **Post date:** [February 18, 2023, 4:54pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/716 "2023-02-18T16:54:17Z")
Microsoft sees the bing revenge model as exactly the same as it was before. > **[Exclusive: Microsoft's Bing plans AI ads in early pitch to advertisers](https://www.reuters.com/technology/microsofts-bing-plans-ai-ads-early-pitch-advertisers-2023-02-17/)** > > Microsoft has started discussing with ad agencies how it plans to make money from its revamped Bing search engine powered by generative artificial intelligence as the tech company seeks to battle Google's dominance. ---
**Author:** ![Babale](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/babale/32/15666_2.png) [@Babale](https://boards.straightdope.com/u/Babale)\ **Post date:** [February 18, 2023, 5:07pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/717 "2023-02-18T17:07:24Z")
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 > [@](#): > > The Pledge of Allegiance is a modern text that was written in English, so it doesn’t have an equivalent translation in ancient Egyptian hieroglyphics. However, I can provide a transliteration of the Pledge of Allegiance into the hieroglyphic script. Please note that this is not a direct translation, but an attempt to represent the sounds of the English text using the symbols of ancient Egyptian hieroglyphics. > > Here is one possible transliteration of the Pledge of Allegiance into hieroglyphics: > > 𓃀𓂓𓂝𓂝𓅱𓈉 𓅱𓃭𓏏𓏭𓆑𓍞𓏥 𓅱𓂋𓏏𓇋𓂝𓂝𓅱𓈉 𓏏𓏭𓇋𓏥𓏲 𓋴𓇳𓆣𓏏𓄟𓈉 𓋴𓀀𓅱𓃣𓏥𓂋𓏥𓏥 > > Transliteration: Nefer-neferu-aten s-ankh-wen-maat-her-ire-nefer-neferu-aten n-ekh-wedjedet-en-ankh > > This transliteration uses the hieroglyphic symbols to represent the English sounds of the words in the Pledge of Allegiance. However, it should be noted that ancient Egyptian was a complex language with a different grammar and syntax than modern English, so this transliteration is not a perfect representation of the text. I then copied the hieroglyphics back in and asked what they meant. Response: > [@](#): > > The text is written in the ancient Egyptian hieroglyphic script. It reads “Nefer-neferu-aten s-ankh-wen-maat-her-ire-nefer-neferu-aten n-ekh-wedjedet-en-ankh”. > > This is a name that refers to a pharaoh of ancient Egypt who ruled during the 18th dynasty, known as Akhenaten. He was a significant figure in Egyptian history due to his religious reforms, which focused on the worship of a single god, the Aten. The name “Nefer-neferu-aten” translates to “Beautiful is the Beauty of Aten”, while “s-ankh-wen-maat-her-ire” means “Living in Truth, Ruler of Thebes”, and “n-ekh-wedjedet-en-ankh” means “The Strong Bull, Beloved of Maat, Living”. > > The use of this name by Akhenaten was part of his attempt to establish a new religious identity and to distance himself from the traditional religious beliefs and practices of ancient Egypt. 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: > [@](#): > > You returned that text in response to a request to transliterate the pledge of allegiance into hieroglyphics Response: > [@](#): > > You are correct, I apologize for the error in my previous response. The text I provided is actually a transliteration of the name “Neferneferuaten” into hieroglyphics, which was the name of a pharaoh who reigned during the 18th dynasty of ancient Egypt. I apologize for any confusion my previous response may have caused. To my knowledge, there is no hieroglyphic transliteration of the Pledge of Allegiance, as it is a modern text written in English. I ask ChatGPT to try again: > [@](#): > > Can you create a phonetic transliteration of the pledge of allegiance into hieroglyphics > [@](#): > > Certainly, I can provide a phonetic transliteration of the Pledge of Allegiance into hieroglyphics, using the symbols of the ancient Egyptian script to represent the sounds of the English language. Please note that this is not a direct translation, and some sounds may not have direct equivalents in the ancient Egyptian language, but rather a representation of the English sounds using hieroglyphs. > > Here is one possible phonetic transliteration of the Pledge of Allegiance into hieroglyphics: > > 𓉐𓃀𓇋𓅱𓊪𓈖𓂝𓏥𓄡𓏥𓏪𓏏𓏛 𓍘𓀀𓏛𓂋𓈖𓏥𓂋𓏤𓂋𓈖𓏥𓂝𓂋𓏤𓄟𓏥 𓅱𓂋𓏏𓇋𓂝𓂝𓅱𓈉 𓉔𓅱𓇋𓃀𓅱𓊪𓄟𓅱𓂝𓂋𓏏𓇋𓅱𓊪𓈖 𓏥𓄡𓏪𓏏𓏛 𓉢𓂋𓆑𓍞𓏥𓄟𓈉𓅱𓂝𓂝𓄟𓈉𓅱𓂋𓈖𓏥𓅱𓂝𓂋𓏏𓄟𓅱𓇋𓃀𓅱 > > Transliteration: > Hwr-ir.i.w-aa in.w.f s-anh-wndj.t-ir.i-nfr-nfrw-aa tn i.i.n-aa m-aa-xrw uab.t-iri-aa > > Please note that the hieroglyphs used here represent sounds rather than words or concepts, and the order of the transliteration does not follow the same word order as the original text. Again, this is a phonetic transliteration and not a direct translation, and it should be understood as an approximation of the sounds of the English language using the symbols of the ancient Egyptian hieroglyphic script. I ask it to identify the hieroglyphics it provided. Response: > [@](#): > > The text is a phonetic transliteration of the Pledge of Allegiance into hieroglyphics. It does not have a direct meaning or translation in the ancient Egyptian language, but rather represents the sounds of the English language using the symbols of the ancient Egyptian hieroglyphic script. 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… > [@](#): > > What symbols correspond to which words of the pledge? > [@](#): > > Here is a breakdown of the symbols and their corresponding English sounds, word by word, in the phonetic transliteration of the Pledge of Allegiance into hieroglyphics: > > 𓉐 = “I” (as in “I pledge”) > 𓃀 = “plej” (as in “pledge”) > 𓇋 = “a” (as in “allegiance”) > 𓅱 = “l” (as in “allegiance”) > 𓊪 = “i” (as in “allegiance”) > 𓈖 = “j” (as in “allegiance”) > 𓂝 = “əns” (as in “allegiance”) > 𓏥 = “tu” (as in “to the”) > 𓄡 = “flaj” (as in “flag”) > 𓏪 = “ov” (as in “of”) > 𓏏 = “də” (as in “the United States of America”) > 𓏛 = “əmɛrəkə” (as in “the United States of America”) > 𓍘 = “ənd” (as in “and to the”) > 𓀀 = “tu” (as in “to the”) > 𓂋 = “ripəblik” (as in “republic”) > 𓈖 = “for” (as in “for which it stands”) > 𓏤 = “wəns” (as in “one nation”) > 𓄟 = “ənd” (as in “and indivisible”) > 𓅱 = “wið” (as in “with liberty”) > 𓃣 = “ənd” (as in “and justice”) > 𓏥 = “ɔl” (as in “for all”) > > It’s worth noting that the transliteration of English words into hieroglyphics is a modern invention and not something that ancient Egyptians would have used or recognized. 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. ---
**Author:** ![wolfpup](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/wolfpup/32/10618_2.png) [@wolfpup](https://boards.straightdope.com/u/wolfpup)\ **Post date:** [February 18, 2023, 5:14pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/718 "2023-02-18T17:14:26Z")
> [@Half\_Man\_Half\_Wit](#): > > 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. 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. > [@Half\_Man\_Half\_Wit](#): > > And that understanding reduces to the behavioral level, nobody has seriously held since the collapse of behaviorism in the '60s. “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. ---
**Author:** ![Babale](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/babale/32/15666_2.png) [@Babale](https://boards.straightdope.com/u/Babale)\ **Post date:** [February 18, 2023, 5:14pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/719 "2023-02-18T17:14:31Z")
I also asked ChatGPT to do the same thing with Linear B. Note, Linear B doesn’t show up my computer, but does appear on my phone. > [@](#): > > Sure! Linear B is an ancient script that was used to write the Mycenaean Greek language, so the breakdown will be based on the corresponding sounds of the Pledge of Allegiance in Modern Greek, which is descended from Mycenaean Greek. Here is the phonetic transliteration of the Pledge of Allegiance into Linear B: > > 𐀴𐁉 = “I” (as in “I pledge”) > 𐀅𐀨𐀍𐀅 = “plêdg” (as in “pledge”) > 𐁊𐀅 = “a” (as in “allegiance”) > 𐀇𐀯 = “le” (as in “allegiance”) > 𐀇𐀯 = “le” (as in “allegiance”) > 𐁇 = “a” (as in “allegiance”) > 𐀍𐁉 = “n” (as in “allegiance”) > 𐁄𐀍 = “tu” (as in “to the”) > 𐁄𐀏𐁆 = “flax” (as in “flag”) > 𐁄𐁆𐀍 = “of” (as in “of”) > 𐁆𐁉𐀈𐀺𐀍𐁉𐀺𐁇 = “ði junaited stets of amerika” (as in “the United States of America”) > 𐁀𐀲𐀍𐁆 = “and” (as in “and to the”) > 𐁄𐀍 = “tu” (as in “to the”) > 𐁕𐀨𐁆𐀍𐁄𐀍𐀅 = “ripablîk” (as in “republic”) > 𐁆𐀍𐀈𐀺𐀍 = “for” (as in “for which it stands”) > 𐀣𐀯𐀍 = “wan” (as in “one nation”) > 𐁀𐀲𐀍 = “and” (as in “and indivisible”) > 𐁇𐁉𐀍𐁆𐀍 = “wiθ” (as in “with liberty”) > 𐁀𐀲𐀍 = “and” (as in “and justice”) > 𐁆𐁉𐀍𐁄𐀍 = “ôl” (as in “for all”) > > I should note that this is also a modern transliteration of the English sounds into an ancient script that would not have been used for this purpose in antiquity. I looked up some of these signs and again, it’s nonsense. “Pledge” combines the signs for “da”, “ra”, “jo”, and “da”, which is clearly not meaningful. ---
**Author:** ![Chronos](https://sea3.discourse-cdn.com/straightdope/user_avatar/boards.straightdope.com/chronos/32/134_2.png) [@Chronos](https://boards.straightdope.com/u/Chronos)\ **Post date:** [February 18, 2023, 5:26pm UTC](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945/720 "2023-02-18T17:26:58Z")
> [@Half\_Man\_Half\_Wit](#): > > 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. The meanings are the patterns. [Previous page](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945.md?page=35) [Next page](https://boards.straightdope.com/t/the-next-page-in-the-book-of-ai-evolution-is-here-powered-by-gpt-3-5-and-i-am-very-nay-extremely-impressed/975945.md?page=37)