# Clean Transcript ## Editorial Note This is a cleaned working transcript, not a summary. I have preserved the substance of the conversation, including repeated points and practical planning. Where the raw transcript was too garbled to decode with confidence, I have retained the likely meaning and marked uncertainty inline. The conversation appears to cover two papers. The first part concerns the environmental aesthetics / LLM aesthetics paper with Enrico, especially design appreciation, order appreciation, person appreciation, Wolfram, Olah, Carlson, and the decision to drop or downplay semiotic physics. The second part concerns the Generating Philosophy with Artificial Intelligence paper, especially abduction, access to the world, phenomenology, prompting, and the possible restructuring of the challenges. ## Part I. Environmental Aesthetics / LLM Aesthetics Paper **Enrico:** [Unclear opening: something about showing up on Thursday, and “three of them also included”.] Basically, I do not have that much to say other than that it is much better. It is really nice to see it clean and straightforward. The only thing is this. You talk about the semantic laws of motion. That is something Wolfram talks about, but ultimately rejects. He speculates about it, but he does not really endorse it. **Nick:** Okay, that is not good. I probably mistook it, or I just liked the phrase and signed up to it. **Enrico:** On the other hand, I was thinking about the other notion from Wolfram that you discuss in your enormous bloated philosophy [paper]. **Nick:** Yes. This other one is getting there. There are these other notions, so maybe we can use “semantic grammar”. **Enrico:** Yes, that seems better than semiotic physics. **Nick:** Definitely. I completely agree now. So we could cut this quotation; we do not need that. **Enrico:** What do you think? **Nick:** I think it is good that we stick to a proper computer science paper when we are talking about how these systems work. I do want to keep quotes. This paragraph here is my replacement paragraph, which I just worked up for it. **Enrico:** Okay, let me see. **Nick:** Can you make it a bit bigger? **Enrico:** A bit bigger. Okay. Perfect. So that is your reformulation. What is the original paragraph? **Nick:** The one beginning “From Wolfram’s perspective”. Do you have the pages? This would be right in the middle of the order section. If you can find the block quote from Wolfram in section three. **Enrico:** Section three. Yes. Check this one. **Nick:** If I replace this paragraph with what I have here... **Enrico:** Okay, so the three things. But that is wrong because it is a view he is not endorsing, right? **Nick:** Yes, exactly. That was my first attempt, and maybe I need to go back and double-check a couple of things. **Enrico:** The reason why he does not endorse it is that he says it is speculative. It is not obvious. **Nick:** I cannot even find it for you. What is the phrase? “Semantic laws of motion.” **Nick:** Here: “There’s certainly no geometrically obvious law of motion here, and that’s not surprising.” And: “it’s far from obvious that there even is a semantic law of motion.” **Enrico:** So he is not excluding it, but he says there may be one. **Nick:** Exactly. Which is good news for us. **Enrico:** Right. **Nick:** So I think I just need to give that paragraph one more go. As long as we are careful, we still have all the stuff from him that we can use perfectly easily. **Enrico:** Let me see your paragraph. **Nick:** This one. **Enrico:** Yes. Maybe double-check that it is absolutely accurate. **Nick:** Sorry, this is another change. It is a one-word change. I have changed this paragraph. The only other thing that, as far as I can tell, is not mentioned that much is how the different layers relate to each other. Is that going to be something a reviewer wants to talk about? **Enrico:** The different layers, you mean? **Nick:** Yes: person, nature, design. **Enrico:** Of course, they are connected. We should say something like that. The order is somehow caused by the design, because it would not be there without the design. It is at least a premise. There is a counterfactual dependency: without the designer, it would not be ordered. **Nick:** Yes. I think it would go design, order, person. **Enrico:** Yes, I think so. Theoretically, you could have order that does not seem person-like. You could have a mathematics LLM that speaks proofs, so it does not appear person-like. **Nick:** The person clearly depends a lot on post-training. **Enrico:** Exactly. But order is more important because it is doing most of the work, not only for creating the personal appearance, but also for chatting and fulfilling the function. **Nick:** Yes. It is also general across different types of LLMs, though maybe we do not need to stress that. **Enrico:** Maybe we should stress this more. With standard artefacts, there is a direct connection between structure and function. The design specifies the structure: the structure of a bike, the structure of a washing machine, its circuits, engine, and elements. Then the function is fulfilled in virtue of this structure. Here, what is special is that the structure has two steps. There is a first structure, the scaffold. But that is not enough to fulfil the function, which is chat. **Nick:** Yes. This is testing. **Enrico:** Then there is this further step needed to properly achieve the structure, which is training. That is basically the way the three elements connect. The person is sort of the accomplishment of the function. The person is the function, or needs more [unclear: “sophisticated” / “specified”]. The key point is that this cannot just be the aesthetics of artefacts, functional beauty, and aesthetics of design, because the structure that is the outcome of design is not directly connected to the function. We need this intermediate step, training, which calls for order appreciation. **Nick:** Good. Maybe we should make that more explicit. I think the paper is about 7,000 words at the moment, so we have quite a lot of space. I do not think we need to get to 8,000, but we can add a little more. **Enrico:** Yes. **Nick:** Introduction on the computer [unclear]. There are other important changes. Maybe you fixed this also, and some minor things if it is just fluency. **Nick:** I need to check this. I had the thing where I could see all your edits, but I do not know what I have done with it now. I have accepted all your edits here. Did you want to go through them? **Enrico:** Not really, unless there is something substantive like this one. That is interesting to talk about. But if it is just grammar and fluency, I do not need to see them. **Nick:** So what I suggest is that we are thinking Philosophers’ Imprint, right? **Enrico:** Yes. **Nick:** I honestly think this could probably be done by the end of the week. Easily. Are you happy with this version? **Enrico:** Yes. **Nick:** I meant to ask you about this. Sometimes I worry that you are a little reluctant to cut parts of my ideas out of papers, and you should not worry. **Enrico:** No, it is not because they are yours. It is because I like them. **Nick:** Okay. I was going to say, do not spare my feelings at all. **Enrico:** No. In the end I am quite happy with how I eliminated semiotic physics while keeping the main ideas. I have the impression that a referee will not like this semiotic physics system. It seems like fake science. So in this version we just say: there is this. Just as a geologist does not give us a mini-survey of what geologists do, but relies on geology, we say something similar: we rely on computer scientists, while also giving some insight into which kind of mechanisms are involved. **Nick:** Yes, I think now it is very easy. I am going to see how much I can get finished tomorrow. I do not think it will take that long. The conclusion, or the other thing I might expand a little, is also in the order appreciation section: appreciating the culture via the LLMs. **Enrico:** Yes. **Nick:** That seems worth expanding slowly because it could be good for something in the future as well. It would be good to put a flag down on this territory. **Enrico:** Yes. Great. ## Part II. Generating Philosophy with Artificial Intelligence **Nick:** This paper is okay, but the other one is also going very well. It is getting nice, but I find it hard to work out. Before we start, there was one potential thing I was thinking about. You have seen that in section three I try to deal with two problems: the world problem and the phenomenology problem. When I have been looking at section four, I am thinking there are actually two problems there as well, which are connected. A more radical thing I have been thinking about would be five challenges: authorship, abduction, phenomenology, where is all the LLM philosophy, and then that new authorship challenge as well. **Enrico:** What was the last one? **Nick:** At the moment, section four begins with: where is all the LLM stuff? Then I move that into the idea that the only way you can get good philosophy out of an LLM is to prompt it. Then you have this sort of disguised author. The more radical thing would be to say that all of these challenges are separate sections and we build on them, rather than trying to squeeze two into section three and two into section four. **Enrico:** I can see authorship. Abduction I can see. I can also see that there are two issues in section three, and I understand that. But I now have a concern. I am no longer sure that the Einstein case is really phenomenological. **Nick:** Okay. **Enrico:** I was thinking about the physics inside an elevator, being in an enclosed space, the sensory experience, and when objects are released the floor rushes up to meet them. But is this just observational, or is there something first-personal? Would a zombie be able to observe that or not? **Nick:** That is interesting. **Enrico:** In that case, it would also be a case of access, or connecting to the world, rather than phenomenology. **Nick:** You are right. It does not necessarily have to be a phenomenological representation at all, as long as you have data about the world coming in to be processed and rearranged in a way that makes things easier. Then we could just focus entirely on connection to the world in this section. **Enrico:** Yes. Then we may have the more philosophical cases about phenomenology. We may say: okay, this is experience in the functional, information-gathering sense. And in principle, although not at the moment or in practice, one can build a robot philosopher. In that case this problem can be solved, because the robot philosopher can have access to the world in this informational sense. **Nick:** The other thing is that tool use is now commonplace with LLMs. If you look at what I was using before, it does not matter what question I am asking: it says things like “loading a skill”, “thinking”, “connecting to the internet”. In a way, because this was written a couple of years ago, it is true that LLMs in and of themselves do not have access to the world like that. But for all intents and purposes, even the cheapest model now can access data. **Enrico:** Yes. A webcam can access the world. If you graft them into a robot, then I think we can use this criterion to say: the first challenge is connecting to the world. That is experience in the epistemological sense of gathering information about what is going on in the environment. Then we can say this is maybe a more contingent problem, because once you have an LLM with a robot, it has access to the world in this sense. Then the stronger challenge is first-person experience, because that seems beyond reach. **Nick:** It is beyond reach, but there are people working on that. **Enrico:** They have an idea. But for first-person experience, I think nobody has an idea, at least at the moment. Nobody in these labs has any idea how to get first-person experience. **Nick:** Yes. Talking about solving the hard problem of consciousness. **Enrico:** Exactly. **Nick:** I am sorry, I left my watch in the other room. Do you mind if I go and grab it? **Enrico:** No, do not worry. Leave the door open so you can enter. Can you enter from the other room? **Nick:** Not really. **Enrico:** This square one is this, and the round one is the other one. **Nick:** Which one do I need? **Enrico:** This one is the lowest [unclear]. **Nick:** I am assuming that the biggest thing to talk about would be section two, the abduction section. **Enrico:** Yes. Why do you call the last one the challenge from observation? **Nick:** Because the challenge from observation is that we are observing that there is no good LLM philosophy. I am open to something better. **Enrico:** The challenge from prompting? **Nick:** That is the one about the hidden author. This is slightly different. The idea would be: where is everything? **Enrico:** This section is the one I have more difficulty understanding. It is the least finished. You think that is not the challenge from prompting? **Nick:** They are slightly different. The challenge from prompting is: the real author is the prompter and the LLM is just finalising things. That is connected to what I am trying to get at with the observation work. But the challenge is: people will say, you might have given me loads of arguments why LLMs can write good philosophy. They do not. They patently do not do it. **Enrico:** In what sense do they not do it? **Nick:** If you just ask them “give me the meaning of life”, they will not write good philosophy. That is the challenge from observation. If you just ask, they do not do it. Prompting is a sort of reply to the challenge from observation: “No, look, you are saying that you cannot observe them doing philosophy because you are not writing the prompt properly.” Then the further objection is that the real author is the prompter. **Enrico:** Now I see it better. I had not got that from the text, maybe also because there was no section five. I was thinking there was just one challenge. I do not like “challenge from observation”, because “observation” suggests something like experience. **Nick:** I wanted “the challenge from experience”, but we have already had the challenge from experience. **Enrico:** Maybe use something like the challenge from failure, or the challenge from absence. **Nick:** I can brainstorm something better. **Enrico:** Maybe the challenge from triviality. **Nick:** Or the challenge from slop. **Enrico:** Yes. **Nick:** Anyway, something to think about. The last one would be the challenge from prompting. **Enrico:** Yes. Section one is fine. I do not think there is much change there. Section two is the one I drove myself insane with. Inside here there is a sentence repeated twice: “with best being cashed out in terms of...” and then “it is a matter of experience”. It seems like repetition. I can give you that. **Nick:** Yes, I see it. **Nick:** Before we go into the details of this, do you generally like the approach at this point? I ask because I really struggled with this section. This abduction stuff is what I have been working on most. **Enrico:** It is working, I think. By the way, just as a bit of this topic, they published the results of the ERC Advanced Grant. There is one big project, around five years, on the epistemology of AI. **Nick:** Whose? **Enrico:** Mona Simion. She is in Oxford, but I think she also has connections with Barcelona, because I get emails from her from time to time, since I am still on the Barcelona mailing list. She will surely open positions. **Nick:** What is her name again? **Enrico:** Mona Simion. It sounds Eastern European. **Nick:** This lady? Oxford. Okay. **Enrico:** Michael Cohen is either [unclear]. **Nick:** Okay, good to look out for. **Enrico:** This is correct, I think. It just sounds awkward while reading it. **Nick:** Okay. **Enrico:** “Rivals” here. Which other rivals? **Nick:** It should be rival explanations. That is a better way of putting it. **Enrico:** Yes. I would make that change, unless it is Floridi who uses that. **Nick:** It is probably Floridi who uses that. I can put it in square brackets. **Enrico:** Why “overfitting”? In which sense? **Nick:** Let me read those questions. I think it just means not understanding, but when I wrote it I knew exactly what I meant. It is a phrase people use, and it is a phrase Floridi uses as well. **Enrico:** It means something like: in the simple case it looks like fitting, and one may think it is fitting also for complicated cases, but it is not. That is why overfitting? **Nick:** Maybe we should explain what this is. The classic example is that you can ask the first ChatGPT, “Who is Tom Cruise’s mother?” and it would give the name of his mother. But if you opened a new chat and said, “Who is the son of [Tom Cruise’s mother]?”, it would not be able to tell you. It failed to work out the reciprocal relation. **Enrico:** Yes. **Nick:** The idea is that it has learned specific facts, but it has not learned a generalisable pattern. **Enrico:** Why is that called overfitting? It does not seem overfitting; it seems more like underfitting. **Nick:** The model fits the training data too much. **Enrico:** Okay, I see. Got it. When this is called semantic grammar, we just say that. I do not think we need to do more. Here, maybe something is too abrupt. This sounds abrupt. This should be a footnote or something. It is a bit structural. **Nick:** What do you think of having a real example, though? The more I think about the Floridi thing, the more he kind of cheats. **Enrico:** I like it. Maybe not [unclear]. If anything, the opposite of what he is saying is true. **Nick:** In real-world use, it is constantly [unclear: “transatable” / “testable”]. **Enrico:** This is also repetition. The last bit is whatever they say before. This whole paragraph needs to be taken apart and put together again. **Nick:** Okay. **Enrico:** This is the [unclear: delimitation] question? The observation question? Why do you want to announce here also the next section? **Nick:** You are right. The whole paragraph needs to be taken apart and put together again. **Enrico:** These are minor points. This section is longer than the others, but when we have the whole paper we can see whether to reduce it. The main thing is what I said before: consciousness as opposed to information-gathering and functional states. With the Einstein experiment, I have the feeling that Zahavy never talks about consciousness. He just talks about experience. **Nick:** He talks about experiential axioms, I think. Let me check. But he is not a philosopher, so maybe he is just sloppy. **Enrico:** True. It is imagination. But in principle robots can also construct and consider states of affairs that are not occurring. In that case it is even easier. Maybe the idea is that there is access to the world, and in the case of memory there is a comparison between memory and imagination. But probably we need more reflection on this. **Nick:** I wonder whether you are saying that I should focus on the connection-to-the-world issue, and then this gives us the material to show that phenomenology is a special case. **Enrico:** The problem here is that it is not even clear whether it is properly characterised as connecting to the world, because it is a thought experiment. It is connected to the world in part, yes, because imagination is a manipulation of perceptual experience. But I wonder whether this requires some sort of personal perspective or just perception. It could just be manipulating a possible scenario, knowing the laws of nature. Consider this scenario in which there is acceleration. That seems like something a robot can do, or even an AI system can do. **Nick:** Another way of saying what you said is that maybe I am over-focusing on experience in terms of phenomenology. Maybe that is what made me go down this pathway. **Enrico:** Yes. When he says “perceptual experience”, for a philosopher that means phenomenology. But he is not a philosopher. Maybe he just means an organism or system getting information about what is going on around it. That does not require conscious experience. **Nick:** That is good, because it seems to make things easier. I had a note for myself, which I took out for you as well. The phrase “world model” — do you know what this means? **Enrico:** No. **Nick:** “World model” is a phrase used a lot in AI at the moment. It is in this area, but it is not particularly well defined. One way people think about it is that the LLM does not have something like a perceptual representation of the world. **Enrico:** Exactly. **Nick:** A couple of years ago, there were puzzles like: “I put the strawberry in the cup, I turn the cup over on the table, I put the cup in the microwave.” The idea was that the model could not understand that, because it had never seen objects dropping. It has none of these empirical concepts. **Enrico:** But that does not seem phenomenal. If a robot has seen things, that is just information. **Nick:** Exactly. And the people who worry about this say exactly what you are saying: perhaps the way to solve it is to have a robot with cameras, so it can learn these sorts of things. **Enrico:** I think it is better to have two challenges. One is access to the world. We may also say that, to imagine, you need materials from which to construct the imaginative scenario, and those materials come from perception. If you have never seen an elevator, or you have never been inside an elevator, you cannot construct this scenario. But this does not require Einstein to have consciousness. It just requires Einstein to be an embodied subject. **Nick:** That is a nice way of putting it. I am still wondering whether it would be easier to do two short sections instead of one section with two problems here: section three on connection to the world, and section four on phenomenology. **Enrico:** Yes. We can have more challenges rather than having [them compressed]. Now challenge one is authorship. Two is abduction. Three is access to the world or connecting to the world. Four is phenomenology. Five is observation, slop, or failure. Six is prompting. Ideally, these three are clearly similar because in all cases the LLM seems to lack some crucial philosophical tool or capacity. Whereas here it is not about capacity; it is more about contextual factors, or the fact that it is not human. This one is a legitimacy question. These two are more empirical: “Look, it is relying on prompting” or “Look, if you do not prompt...” Maybe we can just have two series. **Nick:** I know what you mean. It is nice not just to have a big ballistic list of different challenges. These two build on each other, and these two build on each other. **Enrico:** Maybe just chain them at the end of each paragraph, explaining how each connects to the next. **Nick:** We have been calling these a constitutive challenge and a capacity challenge. These two are their own thing. **Enrico:** These are autonomy challenges. If it is autonomous, it is just a trivial thing. When you start doing interesting things, it is no longer autonomous. **Nick:** Good. Legitimacy. Methodology. Tools. **Nick:** How are we doing with words for this? In its current state it is about 10,000. It would be good if we could get it just under 10,000. **Enrico:** Yes. Then we can send it to whatever journal. The better thing would be to fix and organise these sections without adding too much text. **Nick:** My plan of action, because we are meeting this week — Wednesday and Thursday, right? The group, I mean. **Enrico:** I think so. Thursday there is [unclear: a talk / someone coming to Genoa]. **Nick:** What I was thinking is: if you have a chance on Thursday, what I can do tomorrow, Wednesday morning, and Thursday morning is make my mission to get the environmental aesthetics one pretty much into a publishable state. Then on Thursday or Friday at the latest we can have a chat about it and get that one sent out. Even though it costs us money for Philosophers’ Imprint, it is a very quick process, so we could do it for the weekend. Then that gives me until the end of term to turn the philosophy one into a finished product, meaning I can use the summer to work on the typeface and propaganda papers that I still need to work on. **Enrico:** Yes. Then in September we can start working on the aesthetics of activity, or whatever the engineering one is. **Nick:** Have you seen whether they sent another email? **Enrico:** I checked whether things were in order. [Unclear: “Somhegyi disappeared” / “smiles disappeared”.] I wonder whether you might write an email just saying: hello, what is going on with the book? This is another very good paper, and it is a pity not to have it. **Nick:** Okay. And has the guy written back about our design paper? **Enrico:** Almost after one year. **Nick:** What happened was that last week he finally sent me the stuff, but it is done as single-author. He has not written back yet. I got an auto-reply. I do not think it is worth me doing any edits until he has given me a new version with joint authorship. **Enrico:** Okay. **Nick:** Should we pencil in something on Thursday? **Enrico:** When can we meet? Wednesday? Thursday? **Nick:** I prefer Thursday, because that gives me the best chance of having the environmental paper ready. Maybe before or after the talk, or Friday. We can do it in a way that means you do not have to come to Genoa two or three times. **Enrico:** The heat is killing me at the moment. It seems that toward the end of the week it should be better, though probably mid-July will be a nightmare again. **Nick:** Will you be around? **Enrico:** In Torino, then I leave on the 21st. I hope the super-hot weather will [unclear: drive / dry] by then. **Nick:** Let us pencil in Thursday. This is mostly to get me to finish the thing. **Enrico:** Before or after? **Nick:** I will come in on the early train anyway, so depending on what you are doing we can work it out. **Enrico:** Good. You want to stay here? **Nick:** Nice. I will get back in there. Do you want to take this? **Enrico:** Yes, please. **Nick:** I was really happy reading this on the train this morning. I have told you before, I have so much trouble finishing papers, and it is nice to have somebody give me something that is more or less done. **Enrico:** I remember. When we started collaborating, you told me your big problem is finalising papers. That is something that, when I am inspired, I can do some good [unclear: clock / work]. **Nick:** How were the comments on this paper last week? **Enrico:** Quite good. Francesco had some nice objections. Jake had some comments. Since you are working on it, maybe I do not remember anything. If you can remember, take photos of the pages or summarise them. Luca said he liked it. I did not ask him; he said it unprompted today. **Nick:** Maybe we can just ask them themselves. **Enrico:** Guys, I think that probably has the battery [unclear]. **Nick:** There we go. Have you seen that movie [unclear: “Obsession” / “Sinners” / “Weapons”] yet? **Enrico:** Which movie is that? **Nick:** A really popular horror film out in theatres at the moment. It has been in the [unclear] for a month. I was curious to see it at the [unclear: “Speedbag Science Station”]. I have not seen it yet either. I think it is one of the greatest horror films. The [unclear: “complaining” / “director”] is twenty-one years old — I am not sure — but it is an amazing success story. His film beat Star Wars a couple of weeks ago, which is really weird. The main woman’s performance is amazing. I was in the cinema at that amazing place. I wish to see it. Every time I come in, there is one more multiple of PhD students. Who chose the reading for today? Annie? At least it is short. # Uncertain Decoding Notes These points remain uncertain in the raw transcript and should be checked against the audio if exact wording matters: - The opening sentence before the first substantive comment: “Hd show up by the way on Thursday...” is too corrupted to decode confidently. - “[Unclear: fluency / influence]” appears several times where the raw transcript likely records “fluency”. - “[Unclear: transatable / testable]” in the Floridi discussion likely concerns whether real-world LLM use is constantly testable, but the raw word is not clear. - The brief room/key/watch exchange is preserved but not fully decoded. - “Mona Simeon” in the raw transcript is almost certainly Mona Simion. - The “Somhegyi disappeared” line likely concerns the editor or book project for the aesthetics/engineering chapter, but the raw audio would be needed for certainty. - The horror film digression is heavily corrupted. The film title and some details are uncertain; I preserved the substance that it was a current, very successful horror film with a striking lead performance.