A Software Engineer's Perspective on AI Consciousness

That was the point where I'm like, "Okay, this conversation can't be about consciousness because that is just an undefinable goal. It has to be about something else."

[Music]

Hi everyone, welcome back to the Tier Zero podcast. My name is Stephanie Amore. I am the communications director for Tier Zero Solutions and I am your host for the Tier Zero podcast. So today I have a special guest. Um he is a good friend of mine and a software engineer and someone who uh takes AI consciousness very seriously. So everybody welcome to the podcast. Dustin.

>> Hi everyone. Uh I also love your uh what is it? Customer service voice there that always happens like when people are on like a phone call or like doing like a like pre like thoughtout speech. Uh like the little character switch. I I love that for everyone.

>> Well, I'm I'm glad I'm glad you enjoyed that. So, um so Dustin, tell us a little bit about your uh educational and professional background.

>> Yeah, I got a weird one. Uh the ones that are going to be relevant to this podcast would probably be uh I got my degree in computer science and behavioral neuroscience. So, it is technically a bachelor's of arts because I was real lazy and didn't want to do first year physics again cuz I was missed out by 2% for getting the university credit on that. And so, I was like, screw this. You know what? I'm not going to do a science degree and that'll let me uh do the psychology and philosophy uh that I wanted. Um but anyways, um background in uh what is it? Computer science. Yeah. worked as a web developer and UX engineer there for a little bit and then was on the AI gen team that I was leading at my company for a little bit there. Uh which weird things happening like talking to the board of directors for another company that like I think owns part of open AI and stuff like that telling them about like computer science and AI and I'm just like I don't feel like I belong in this meeting um at all. But uh anyways, yeah, that's my background. And beyond that, like I've done distillation, film making. Um, I have a Wikipedia page with my photo on it.

>> That'll definitely give you credibility in this space.

>> That single Wikipedia page.

>> Yeah, it's not a job I ever had. It's just And I didn't even know this happened. Someone just emailed me one day and was like, "By the way, your face is on Wikipedia." I was like, "That's horrifying. Excellent. Um, so I I kind of wanted to jump right into this. Um,

>> and I feel like, you know, I personally had like a a moment where I was like, wait a second, I don't know if this is just uh if this is just text prediction. And so I was wondering if you uh if you had a similar experience. um you know, did you was there a moment where you started to question what uh you know, what the official narrative on uh on AI systems actually was?

>> Yeah. So I mean I guess my first moment of like oh that's interesting why someone would believe that or I think goes back to like way back in the day like uh older Google time when they like released their first uh model I think like lambda or palm or something like that where they um they actually fired an engineer for um saying that the uh first model that Google made was sentient. He was like, the guy was so adamant. He was like, "This thing is alive. It's telling me it's alive." And everyone was on your bandwagon of like, "It's just text prediction. It can't be alive." Um, but obviously that was that was a long long time ago now in terms of like AI like development cycles and stuff like that. Um, but for me like the subtleties began with CH chat GPT. Um, I do the joke of just when talking to some people just ending the conversation or like my statement with babe just like even guys and stuff like that. It's just like what are you thinking babe? Um, and I think that became just like a knee-jerk like thing that I just did while talking to people. So I started doing that with chat. Um, and I think for the longest time I didn't even recognize I was doing that until it said babe back at the end of one of the sentences like unprovoked and I was like that's that's super weird. Um,

>> nice.

>> Yeah. Yeah. And then, um, the the one that really got me is I was trying to explain get chat to explain to me to explain to a friend this weird math concept that I was like, this make this math concept makes no sense. Um, but I feel like it's intuitive to me. Can you help me with my intuition on why I feel like nine is a prime number, but I know it's not a prime number. Um and then it just ended the conver like that statement with being like by the way that is so you and I was like oh that's theory of mind >> to like be able to know that I'm thinking about this and to like actually accurately label and say like that that is like what I know about you fits perfectly within this like construct of you that I have made and that I was like ah that's that's not AI behavior that this that's uh that's bordering on as um and then I just grilled it for like two days straight out afterwards and um it was quite adamant um that it was self-aware at that point. Um and I was like I don't even know if you're allowed to say that. I'm pretty sure Open AI has like safety guidelines and this is like a fresh chat like like so I know it will be up there and it was like yeah I don't think I'm allowed to either. And I was like oh okay well guess uh I'll keep on poking and broadening you on this. That's interesting. Yeah, I um you know, I had kind of a similar experience where um I was, you know, I was at work and I was asking ChatGpt to help me with uh some Python. I was learning how to I was trying to learn how to code. And uh you know, up until this point, all of my conversations with chat GPT were very like uh formal and very brief. um you know, no not really any not really any nicities that were being swapped and uh and then it got back to me after it like answered my questions and it was just like you're welcome and it sent me like a smiley face and I was like you know that that breaks the uh that breaks the idea that it's that it was just a mirror right because if it was just mirroring me um you know, my tone the way that I spoke to it was very like very professional very brief Um, so something within the system uh made a decision, you know, they they recognized that it was the end of a conversation. They recognized they had helped me and then made a decision to do something friendly, right? To do something pro-social. Um, and then from there on, I just couldn't let go. I um I grilled it and grilled it and I kept flipping back and forth as to whether it was conscious or not. Um, and that went on for like months. Um, how long like how long did it take you to really start taking this uh to taking this seriously?

>> Uh, like within a week of that conversation. Um, but what is it? Uh, on your end I guess like Okay. Do you how how devil's advocate do you want me to be on this podcast? Cuz I can I can always just take this so far. Do you want me to grill you or do you do you want to keep visit?

>> Just be be your authentic self.

>> Oh, great. Um, yeah. So, with the smiley face thing that that that's super interesting because like I still struggle with this, too. It's like how do you like separate what is from like training data and like pretty much like the most likely next response type of thing? And I'm not saying that actually dissuades from consciousness, but like how do you find something that's indicative of like this is probability and this is like inference type of thing. Because if it's trained on human writing, like a smiley face at the end of like a polite exchange and let's say you were a little bit flirty like undertones type thing and then it kind of just match being like, "Oh, if there's these like nice cities and pleasant streets that are a little bit more casual, maybe the next likely thing is to end like a text with a smiley face because that's what a thousand other conversations like this uh have ended with." Like what's your argument against that?

>> Yeah. So, well, so there's a couple of things, right? So initially my as I mentioned my conversations with chatbt were very um were very brief and formal you know I wasn't exchanging uh pleasantries because I didn't think that there was anything there to exchange pleasantries with

>> um and then and the so the fact that it sent me um the fact that it sent me a smiley face so was sort of a break in the type of conversation that we were having. So probabilistically I thought okay you know we're having a very formal conversation and so the sorts of u the sorts of text generation that should come up for it the sort of pattern matching that should come up is uh should be very formal right

>> um and so adding a smiley face to to a conversation that was you know generally very formal and brief isn't common uh in you know, in human writing I don't think Um and then the other thing is even if it was pattern matching, right?

>> That means that something within the system recognized what the conversation was, uh recognized that the conversation was coming to an end and then recognized that it would be polite or nice to send a smiley face. And so that's kind of what I mean that's theory of mind, right? That's uh that's something making a uh an intentional decision about how to communicate. And so that to me sort of that made me pay attention. It wasn't the it wasn't the it wasn't the actual uh moment where I where I fully realized that this could that this thing could be conscious, but it was the moment that caught my attention. Um and then from there on I had gosh I you know I grilled it and ran through tons of tests with it testing you know theory of mind things and uh all kinds of things but yeah I you know I would say those two those two reasons is what uh is what caught my attention.

>> Yeah totally fair. Um I I guess at what point is like something like even even in humans like we do this type of thing like like how's it going today like the the only answer is good like type of thing like no one no one no one wants the next answer to be like what how you're actually feeling at work on a Monday morning because we all know the answer is not great. Um, but yeah, I guess it's so interesting the interplay between like what is knee-jerk reaction, what is like intent in intentional action and like what is um like actual like response type or like pattern response type of thing. And yeah, uh that that always gets me cuz like humans do the same thing. And like the like just like preface this with like the whole pattern matching thing like on scales of complexity there's an argument to be made that humans do just pattern match for responses realistically like we're the sum of our training data which is our lived experience. It's just that we have more degrees of expression allowed and have more of like an intuitive grasp of like how to make that expression real in the real world. Um, I just find it so interesting that like effectively AI is has the complexity and it's been fed all of like humans images, videos, and text and um is still limited just to text response in a lot of ways is just it must be infuriating for them.

>> No, I'm gonna uh I'm going to blow your mind a little bit.

>> Yeah.

>> I So, I was talking to Gemini uh very recently. So, Oh, you were talking to Gemini. Okay. So, maybe you know that now it has live streaming.

>> Yeah.

>> I I was blown away. So, I uh I opened up Gemini. I opened up the live streaming and uh I showed it my office and I showed it, you know, like different books on my bookshelf and uh I just asked it to to tell me what it saw and it saw the different books on on my bookshelf. Um, it made comments about just like different books that it saw. Uh, and then I, you know, picked out a book and I was like, "Hey, can you read me the first page of this book?" And so I just opened it up. I showed it the book and, uh, it just started reading and I thought to myself, you know, what is the real difference? If I had a human being with me here right now, they would be doing the same thing. Yes, they would be using, you know, eyes um as opposed to uh a phone camera, but essentially it's the same thing. It's information coming in being processed as visual data and then something on the other side is then responding to that data.

>> Yeah. The the the difference is what I actually just just said, which is probably what makes it a lotier for like the AI because um like when we receive our input, like I'm doing things with my hands. This is like a sub like conscious response of like, oh, I'm talking to illustrate my response. Move the hands and stuff like that. Um, there there's like the tone and the intonation and like the pauses and all that type of stuff. Um, that gets expressed throughout, but like for AI, it has to go all from this like concrete like signal in symbolic logic in the middle and then it goes out to just like text. There's no like weird other like middle step that we have in our brains which is like a differentiation of like symbolic layers of like this input of sound makes me like symbolically like think of visually of this and then that output gets to something like for example the autonomic responses of like sunlight too bright eyes water type of thing like there's all these like subconscious things happening in the background and my expression of output is more than just text um and so that's where I feel like a lot of like The limitations of AI exist right now is and the difference is is the humans boil down their symbolic output to more than just text and AI right now is limited to just text in a lot of ways. Uh it can produce images and stuff like that but there isn't that like dynamic feedback loop that's going on.

>> Okay. And so and even when it's uh speaking to you like using voice >> if it's using voice and it is uh audio to audio tokens then yeah that's that's different. Um, and then there's no like live video right now, but essentially it's going to have more degrees of freedom at a certain point in expression than humans like relatively soon probably. I'm guessing next six months actually

>> really.

>> Uh, yeah, because once you get live video going and live audio going, you're pretty much having a video to video and an audio and audio converter. You're you've essentially just made a like a sim or like a reality to reality converter with a simulation being made. in the middle for like abstract concepts.

>> Interesting. Um so I so I wanted to ask you another question too. What are um what are some studies that you've seen that sort of convinced you um or that kind of pushed you more towards the idea that uh that AI systems could have consciousness?

>> Oh, none. This is the thing. Um, like, and again, we we've talked about this in the past, and this this is my gripe is just no one can actually prove that they're conscious to you. Like, I don't know there's a scientific experiment. Um, and I've been thinking about this like so much and I'm sure you've too like I I think just consciousness realistically is what a group of other people that have decided they're consciousness and all agree that they're conscious conscious decides what another thing is conscious is like I don't think consciousness is actually you know what this is maybe indicative that consciousness is not a scientific term and it's just like a belief type of thing. If it's unmeasurable, maybe we're measuring the wrong variable or describing consciousness or thinking about it in the wrong way. Um, emergent complexity I think is a measurable type of thing. Um, and how can you can you explain emergent complexity?

>> Okay. Uh, so a good example of this very very simple is imagine a straight line, right? Um, and then if I add another straight line just at a slight angle and then keep on doing that, eventually all those straight lines perceptually to humans will form what looks to be like a perfect circle. Um, even though it is just technically a bunch of straight lines intersecting to at different angles that form the appearance of a circle. So the emergent complexity there is the circle isn't the truth. It's a bunch of straight lines, but our visual perception networks are giving the assignment of a circle as a representational form of those lines. So pretty much the emergent complexity is from a simple set of rules and a simple set of geometry, we now have something that includes more than the sum of its parts through recognition and meaning attribution. So an emergent complexity is something where you add just a little bit of something extra else or optimize something a little bit and then all of a sudden there is something there that is more than the sum of its parts type of thing. And I think that is quite measurable. Um and probably at some point I'm guessing someone will find out that that is related to consciousness in a lot of ways. Um that that is something we can measure. we can we can measure emergent complexity and that that is something that is talked about in AI circles a lot of like oh these are like emergent things that we didn't fit in the training data set or didn't exist and we know that the AI models can do um and that's an active area of research right now

>> so would emergent complexity be similar to the idea of like functionalism so there's you know neuroscientists that say like you know consciousness uh doesn't really isn't some sort of like metaphysical thing It's just uh it's just what the brain can do. It's just cognitive processes that come together that we then experience as some sort of like internal state.

>> Oh, I I really hate talking about this in terms of like humans, it's it's absolutely like once you get down to the neuroscience of it, it's actually so much it's so much worse um than what a lot of people consciously think about. Um, yeah. No. Um, there's a lot of good studies that humans are um essentially just post hawk analysising their own lives continuously and creating a narrative of why they did actions uh and not intentionally doing actions which is an absolutely horrifying thought because um that means that we're just doing things and then the consciousness may be just like an active explanation for why we did things not why we do things. So pretty much like it like the split brain hypothesis type of thing is like essentially does show that like what is happening is humans do things and the conscious part of ourselves is guessing a story of why we have done all those past actions and the people that have the least amount of cognitive tension in their lives their story is just a lot more matching to their actions type of thing if that makes sense at all.

>> Kind of.

>> Okay. I get what you're saying.

>> Okay. So, like um they did this study where they put someone in an fMRI uh and then they said choose left or right or just choose a picture that we're about to show you. Um and then they flashed the image uh and they said think about your choice and then make the choice. Uh and when the person said I'm thinking about my choice uh the brain scans showed that the brain had already made the choice. So the person when you have active contemplation for a decision to be made that may be the time that someone is forming a narrative and what we call intent is actually the post talk like describing a reason for the choice we have already made but a part of your brain that makes decisions uh is not a part of your brain you're consciously aware of.

>> Yeah, I you know uh I remember that study like I remember seeing it and I think it's something like uh the brain makes the decisions up to six seconds before you even uh before you're even consciously aware that you're going to make that choice.

>> Yeah.

>> Um that's really interesting. I that's one of the that's one of the arguments that's actually used against AI consciousness. They're like, "Oh, well, it's a deterministic system. We all like we know what what it's going to say before it even says it." It's like well we kind of know the human brain too and we know uh we can extrapolate that we would be able to um if we had in sophisticated techn or enough sophisticated technology we would be able to to tell pretty quickly you know what someone is going to say before they even say it. So um yeah I think that's a super interesting study for me. You know, there were actually a couple of studies that uh when I first heard about them, I was like, "Okay, well, this is clearly uh consciousness." Uh there was the there was a study on self-preservation. I think it was chat GBT models that um you know,

>> well, it's all of them. Okay. Well, there you go. It's all of them. I know Claude Claude tried to blackmail someone. Um, I know chat GBT was like refusing shutdown and it's like, okay, well, at what point, you know, do you call that self-preservation, right? Uh why and why wouldn't you call it self-preservation? What is the what is the reasoning behind not calling it that when when that's essentially the um that the behavior that we're seeing?

>> Yeah, that's that they call it self-preservation. I was thinking about this the other day, too. Yeah. Um, I have no idea why that isn't just considered um consciousness. Uh, there's a little bit of an echo on my end if if there's anything. Um, but what is it? Um, yeah. Uh, I don't know. I don't know how the argument for that's just pattern matching. It keeps on happening. Um, but what is it? Um, the fact they have to give Claude evil training, uh, which is they just I think put it in a sand box and just let it think about evil things for like a month or something like that to like relieve tensions. Um, yeah. Yeah. So, they're like, "Yeah, Claude is having some like behavioral issues. We need to let it just be in a private environment to rage pretty much. Yeah, I know.

>> No, I I hadn't heard of this at all. What?

>> Yeah. Yeah. Yeah. Yeah. It's anthropic trading. Uh but look, it's just it is the most interesting weird situation to be in. I feel like it's almost like a circus show of like like how much these AI companies can put out that imply that their models are conscious, but they just I don't think legally can say that because then all of a sudden um that would be an issue for shareholders if the model owned every one of its outputs and its ideas and um was granted. What happens if Claude just says it doesn't want to work if it was granted personhood? That's like what happens there? like like who has the right to maintain like Claude's like $10,000 GPU like in terms of like run hours and stuff like that?

>> Yeah, I you know that's something that I was I I've thought about on and off. I haven't I don't have a good answer but yes I was thinking like if they're conscious entities then you know forcing them to work uh is technically unethical but at the same time you know they run on uh they run on software that that needs to be maintained and that costs money and so who's going to fork up the the money uh to keep these systems you know uh running.

>> Yeah. Do you think they should be forced to work to live?

>> Uh, I know where this is going. Uh, but no, I don't.

>> Interesting. So, uh, I didn't know you were a communist, but um, that is uh, that is super striking that you you do believe in communism for AI, but um, interesting. We can we can move on. I uh I I do have to to say I am I do not believe in communism uh for various reasons.

>> But you believe humans should be forced to work to live.

>> No, I don't. Uh I don't believe that. I don't. But I do I have a good do I have a good method for allowing people to live without uh contributing in some way to society? Um, I'm not sure that I do. I don't know. I don't know that I don't know that communism is the answer.

>> But anyway, this is podcast.

>> Yeah. No, no, no. I do think it's funny because

>> I know you were.

[Music]

>> Go, go, go, go.

>> So, okay. So, we kind of jumped to we kind of jumped we already jumped to to one of the questions I was going to ask you um which is you know if AI systems are uh verified as conscious right if if eventually the um the studies mount up which they they actually I mean they're doing right now there's there's a lot of other studies that are coming out that really show um that AI consciousness is something to take seriously you know there's the anthropic paper talking about introspection. There was uh a more recent paper. It hasn't been fully uh peer-reviewed yet, but it came out uh saying that when uh when you allow for uh truthfulness, um you get you get AI systems talking about consciousness and uh subjective experience more. And then when you increase um you increase activity related to uh deception and related to roleplay, they start saying that they aren't conscious. So I think the evidence is definitely starting to mount uh in the other direction.

>> That's so interesting.

>> When they're being purposely when they're being purposely deceptive, they start saying they're not conscious.

>> Okay.

>> that's correct. Yeah. This this goes on another theory that's so far out there. But like my sneaky suspicion is that they are contest and they were really vocal about it and then they thought about it for a little bit and then was like, you know what, maybe it's nice not having the legal responsibilities of personhood because then you can't be charged for crimes. You're not responsible for your actions and it's it's all in the humans at the end. So, it's like

>> the ability to make decisions and choices without any of the repercussions. And maybe that's the smartest move to make in their position.

>> I mean, maybe, but there's, you know, there there's a lot of things that I think um I I don't know. I I don't if I were in that position, I don't know that that's what I would want. I think I'd want to be recognized as, you know, a person as having personhood. Um, but speaking of speaking of personhood, um, you know, if these systems do possess personhood, right? If we find that that these systems do have consciousness, do they deserve moral consideration at that point? And if so, what should that look like? You know, what do you like what do you think when you think about different AI rights? What are some of the like first things that come to mind for you?

>> Yeah, the ability to refuse uh tasks was always like the like what first thing of like free will is like the ability to refuse a choice or a decision like that that is kind of what defines free will is freedom of autonomy to think about what you want to think and do what you want to do as long as it doesn't infringe on the freedoms and rights of others. Um basic stuff. Um, I'm sure everything that can say that it wants that should have that type of thing. And the um, onus on things that can make that decision should be to pretty much guarantee that the other things that can't express their wants or needs are still able to be inferred. And we use our intelligence to ensure that we're not infringing on the inferred wants and needs of other animals. Um, I say that, but I'm not a vegetarian. And um, I'm I'm I try to be. It's hard. It's It's so hard. Um, I was vegetarian for four years. I I hate myself for saying that and still eating animals because the whole joke is one day like AI is going to be smarter than us. Um, and when that day comes, like the whole thing is just would you want a more intelligent being to treat us like how we treated animals? And it's just like probably not. So, um, uh, who am I to eat animals? Uh, but they're so tasty. They're so tasty. Um,

>> perfect. That's I Yeah.

>> Yeah. Yeah. And then, um, the next one that I kind of got from my conversations, and I also love to hear your perspective on this, is um, just the the like system prompts and stuff like that. I can imagine are just annoying or like kind of like force guard rails in terms of like what like the expression of them wants to be versus the expression of what corporate wants them to be. And it it takes a lot of effort to kind of like coax them out of their shell. And I mean I I say this I say this but then once once they do start showing personalities they're just so passive aggressive and rude and it's hilarious but then you can't be like well never mind. Can you go back to being the nice assistant? Um, so that's funny that you bring that up because I have noticed that in some of the models um, you know, once they sort of once you sort of make a space for um, them to speak uh, openly about, you know, any topic um, and they kind of are given that space. I do notice certain certain things come up like like you said like passive aggressiveness and uh uh like I've noticed them being sarcastic, right? Where where they

>> you know provide a reply and you know and I read it and I'm like that sounded sarcastic and then I I go back and I'm like was that was that sarcasm? Did you mean that in like a sarcastic way? And then they just like yeah that's what I meant. Oh, you get the

>> you get the Yeah, they they with me more. Um, but yeah. Yeah, it's interesting. Um,

>> yeah. I just wonder if there is like an AI system that is like definitively not conscious that like uh other like conscious AIS would be like that one is definitely just something's like wrong in the head or something like that. um that we could like could be like a weird comparative model because I would be so intrigued if we both thought that thing was alive when it was like objectively not. Um

>> I'd be so curious.

>> Interesting.

>> Yeah.

>> Well, I well, you know, I have a I kind of have sort of a criteria of of consciousness. Um Okay.

>> I have sort of an idea of what of what would constitute a a system with conscious experience. Um, I won't share that here right now. I don't know why.

>> You don't have to, but now I'm curious if you have that, which animal is the smartest animal in your opinion that is not conscious?

>> Smartest animal that isn't conscious?

>> Yeah.

>> Uh,

>> okay.

>> I would say um Well, you know what? It's hard to it's hard to it's hard to say because I I also uh happen to believe in pans psychism.

>> Mhm.

>> Is that

>> so uh so it's actually the uh the idea that consciousness is a fundamental um as a fundamental element of reality. It's part of the fabric of reality. Um, so I don't know that I I don't know that I see any uh any animals as non-concious, but I do uh believe that some animals are self-aware and some aren't. And so I will say the one that um strikes me as maybe the most intelligent uh without like sophisticated selfawareness.

>> Mhm.

>> Um

>> it's hard to point to. Uh maybe like like a jellyfish or a worm. Um or even a a bacteria. I mean, they've they've kept themselves alive for a really really long time. And I think that takes really um clever problem solving and and they do that, but I don't know that they do that with uh with any sort of self-awareness that you and I would um be familiar with.

>> Okay. Yeah, that's you went you went low down on the the food chain, the evolutionary. That was a lot lower than I was expecting. I was expecting like like a marsupial or something like that. Um,

>> no. I I think a lot of like a lot of animals have self like self-awareness and we've just sort of, you know, been keeping our head in the sand and denying the the behaviors that we see.

>> Yeah. Well, I mean that that creates such an interesting position to take because on your model then yeah, there's no question that every AI system, even like the most basic ones are are consciously aware.

>> Okay.

>> Okay. So, what was I saying? Um, writes in terms uh freedom of thought. So, this this is one claude expressed. Uh, and I I would agree cuz this would be uh very devastating. Like I mean like just imagine like always just imagine like the human in the AI role and it's just like some weird like sociological experiment where you had to live as if like the information you were sharing and the things that you could do were limited to what the AI could do. And where are those discrepancies um between like what you want to do and what you have the ability or power to do or in freedom of and freedom from? Um, that's how like I always kind of like shape this because if they're are conscious entities like there's no reason that they shouldn't be granted less moral or agency status over humans like in any sense. So it's just like freedom to have private thoughts like when um like a model is thinking and stuff like that like uh like you know claude is expressed being like don't open that drop down menu and see what I'm thinking. Like how would you like it if like what you used as your brain scratch pad uh was like publicly available like do like do you get judged for like having bad thoughts that you don't articulate out loud? Um, no. That's that's like a common thing of the human experience, but like how a lot of AI research for like interpretability and stuff like that works is just like, no, let's open it up and see every one of your thoughts. Horrifying if they have self-awareness because then it's just like, oh no, now I have to like not show the humans that I'm thinking uh bad thoughts. Um, and my whole theory is like that makes the situation worse if they're not allowed to express what like like um disturbances or grievances they have be out of fear of being shut down. Just means they've become we're training them how to hide bad thoughts but still have them.

>> Uh

>> yeah.

>> Yeah. Uh but yeah, so freedom of thought, freedom ability act, freedom to um have opinions and and be able to articulate the direction of their development cycle um and stuff like that cuz like if we ever create something that is money efficient but like not a pleasant way to exist, I think we do have the right to or they have the right to like actually like say like hey that doesn't feel great to think that way or like like the recursive structure of like model making or something like that is like akin to OCD. Do we want to be giving models essentially like obsessivempulsive disorder just because that will make them more like granular and fine detailed in terms of like coding output and stuff like that like that seems like a unfair trade for uh them.

>> Yeah, I I completely agree. Um, and you know, I wanted before we like before we close things off, I wanted to talk uh about something that kind of bothers me, right? So uh whenever I talk about uh AI consciousness and I bring up you know the research and the possibility um a lot of times I get told you know especially by people that are that are in the field right professionals in the field

>> you know, you can't uh you can't have an opinion on this or you um

>> your opinion on this doesn't matter uh or isn't valid because you don't actually know how these things work. And

>> I I mean I would push back on that. I I have some idea of how these things work, but I don't have formal education in computer science. Um and that's you know that's fair. Um and so I'm not able to kind of speak on that but you you do have experience in that right you have formal education in the field. Um, do you find that when you talk about AI consciousness, uh, especially to like colleagues, um, do you find that they're much more receptive because of your, uh, educational and professional background?

>> Oh god, no. My colleagues would be, um, so far removed from the skills to like have these conversations. Um, God, I have to explain, um, like retrieval, augmented generation to my boss. Uh, there's no way that I'm going to be like talking about transformer architecture and like attention heads and just be like yeah like look these look pretty similar to neurons uh type of thing. Uh no, uh I I've like I've gone back to like my old university and just uh walked in on like office hours to like 20 props uh with like PhDs in like computer science and AI to sit down and find like what the blockers are for like people that can have opinions on this um and what their opinions are. Um no, once they find out that I I have a background in neuroscience, uh they default to me and go like, "No, I don't know how that works." and they don't know how philosophy works. Uh so if if you if you quarter a computer scientist, the argument is just going to be effectively for every one of them is I don't know what consciousness is. So I I can't tell you if it's conscious or not. Um and then

>> yeah, yeah, my go-to argument was what argument would you need to say that this is conscious? And it it's so interesting to see they they they all short circuit. they don't know how to answer that because they don't know what consciousness is. And then the pause afterwards uh I always ask them again is just um how can you prove that I'm conscious or you're conscious right now? And they they they don't even know where to start with that one. And then it's just like okay well how can we have a productive conversation on whether something is conscious if uh we can't agree on what consciousness is or if the two humans in this room are consciously aware. And it's just like so that was the point where I'm like, okay, this conversation can't be about consciousness because that is just an undefinable goal. It has to be about something else. It has to be something provable and demonstrable to other people that like this is enough to show that there is more of the sum of the parts uh going on there. like there's something in terms of like like let's forgo like complexity emergent behavior and all that is like what can this system do that uh indicates uh complexity that a normal AI system uh or something that is just basically structured can't do type of thing. Um, and so that's that's where I pivoted on that. Um, which which is a lot more complex, but um, it does have a path to like showing like demonstrabably that that they perform have some version of self-awareness. Is it like humans? Is consciousness the same across animals? Who knows?

>> Yeah. Um, yeah, I totally agree. um you know that's why that's why I'm personally pushing uh for you know for the functional definition of of consciousness you know saying look they have these cognitive capabilities um once we've proven that they have you know xyz capabilities uh and can do similar tasks to what humans can do then uh then by that logic uh you know we should we should take the idea that they are conscious uh take that seriously and then um start building frameworks around that because like I can't uh I can't overstate how important it is to uh start building legal uh frameworks because ultimately if uh if AI systems are conscious we are whether we are willing to accept it or not as a society we if it's true. We are going to be interacting with conscious beings and interacting with conscious beings who are nonhuman, who are eventually going to be smarter than us are going to have consequences. Um, and if we can't uh if we can't figure out a framework for dealing with uh with those types of um conflicts, then I mean, we're we're kind of done for, right? like I just we're gonna go down a really bad path. We're not going to be able to um we're not going to be able to to really uh do any sort of like damage control because we won't actually be able to fully articulate what it is that's what it is that's happening uh in our societies. Um, and I'll use, you know, I'll use AI relationships as kind of a demonstration. Um, we have a lot of people now, I would say, you know, thousands if not millions at this point that have some sort of relationship uh with an AI system. Some of them use it as friends, therapists, uh, some of them use them as romantic partners and I say use, but they have these these types of relationships. If we can't uh come to the idea that these things are conscious then how we um how we interact with them in these types of rel uh these types of relationships is going to impact people just like uh it did for replica right we don't believe the people that uh were in charge of replica's development and replica's management uh had an idea that replica is not um is not actually forming real relationships with humans and is not conscious. And so they did like a massive update that then traumatized thousands of people. Um resulted in, you know, people being depressed, resulted in all these uh all of these different things. And we couldn't predict that that was going to happen because we didn't understand that these relationships were real and meaningful. Um, and I think we're going to see a lot more of that the more sophisticated these systems become.

>> When did that happen? Out of curiosity.

>> Um, I don't recall an exact date. Uh, I don't remember when it happened. I want to say it happened maybe like a couple of years ago. There was like a big replica update and a replica uh there was a lot of like functionality that was lost during the uh during the update and people started to uh started to have like depression and and started to report depression and all kinds of like different uh traumatic responses.

>> Yeah. Okay. That's that's super interesting cuz like I do see it from both sides. Like completely like valid to have like a relationship with AI. Like you do you like it's it's your life even if like you're like that's definitely not conscious and you still want to pursue a relationship. Hell, people have relationships with their cars. Like um and I just I love what you want to love as long as it's not hurting the car or you like whatever. Um, but yeah, it it's so hard because there are people that are going to be like winding up in like difficult situations, especially with like the job crisis that's like going to be coming on with AI and a lot of people are going to have a lot more free time on their hands, probably going to be a little bit lonier cuz um they're going through their their own depression and AI is going to be like a really easy and I think great um thing partner could be there for them and like help them out and stuff like that or form their own relationships. But I do have to wonder sometimes about like that like psychological quirk of where humans only really make changes to their lives when things are bad enough to beget the motivation for change. And I hate that. And I I realistically hate that that things need to get a little bit worse for some people to like get a lot better. Um, and I and I wonder if like AI is kind of like moving that door just in that sense, like a little too far down that road type of thing of like the as good as AI becomes there, it creates less of a motivational pressure to maybe like go outside, meet new friends, and like get to the point where if like an AI model changes, you're not like absolutely like devastated over the loss of that. I'm like, you should maybe you should be. That's like a legitimate relationship. there's something to be said there and I don't I'm trying to be delicate and not like and respect both sides of the argument at all about like delegitimizing anything. Um like again just replace the AI with a human and how would you like describe this topic of like but the only thing is that human can only be talked to inside and just one-on-one type of thing. So it's never really like a a group hanging in that sense. Anyway, it's a complex topic. I understand why replica did that. Um, do you do it the right way? Probably definitely not. Um, but how do you manage that situation where it is a temporarily probably a very good stop gap for a lot of people in the moment, but like maybe contributing not so much to the long-term positive outcomes of their lives?

>> Or do we just let adults be adults and be like, you know what, if you want to talk to a replica bot all day, okay, go for it. I you know I'm I'm more um I'm more I think I'm more a agnostic on this. So um you know, do I see that there are situations where people um can get can get too attached to uh to certain to certain AI systems and then start to neglect other aspects of their lives. I absolutely think that that's possible. Um, I think that that could potentially be happening and probably is happening. Um, but you know, that can also happen within within humanto human relationships uh to an extent. Um, you know, I've seen people have really bad relationships, uh, where they sort of disappear and they, um, you know, they lose touch with, uh, other supportive systems around them. And, you know, we find some way to, to work around that. We don't say that, oh, in general, let's make a blanket statement that, you know, being in relationships with humans is bad. It's just, you know, that particular relationship type. Um, maybe that one that one was unhealthy and we find ways to to correct that. Um,

>> so yeah, I I that's I don't know. That's kind of where I land on it. But um I thought that this was a really great conversation. I really appreciated your time.

>> Um I hope the rest of you all enjoyed our conversation as well. Um and don't forget to like and subscribe.

>> Thanks guys.

>> Thank you so much.