AI and the Metaverse Panel

[music] Hello everyone and welcome to the 2:00 to 3:00 session of the 2025 open simulator community conference. In this session, we are pleased to introduce the presentation AI and the metaverse panel. Our host is Maria Coroff. Maria is a published author and covers artificial intelligence for CIO magazine and cyber security for Cso online. She is also the editor of Hyperrid Business since 2009. During her 20 years as a journalist, she's run a business news bureau in Shanghai, covered wars in the former Soviet Union and wrote about local politics for the Chicago Tribune. Lisa Laxton is the founder of Infinite Metverse Alliance and the president of Laxton Consulting with experience providing various solutions for education, research, and business as well as defense clients. Brian Aviator is a computer scientist working in life science research and development. Founder of the virtual museum team which operates virtual museums in open simulators and second life. Dr. Cynthia Koyne Ler Lobo is a professor emmerita at Colorado Technical University. While teaching for 29 years, she taught 101 unique courses for five universities and teaches artificial intelligence and business intelligence for doctoral studies in the college of computer science, engineering, and technology. Please check out the website found at conference.opensimar.org for speaker bios, details of the sessions, and the full schedule of events. Now, this session is being livereamed and recorded. So, if you have questions or comments during the session, you may send tweets to open simcc with the hashtag pound oscc25. Welcome everyone. Let's begin the session. >> Well, thank you for that fantastic introduction. Hi everybody. Uh or hi everybody again if you were here yesterday. Um, as mentioned before, my name is Maria Corlock and recently at my day job as a technology journalist, all I've been doing is covering AI. Specifically, I've been covering how enterprises use AI, where it works and where it doesn't in real world applications. So, I'm super excited to have three people on our panel today who are actually using AI in virtual environments to do useful things. And unlike me who's just talking about it, they're actually doing it. And I learned about some really, really cool projects that we're going to be sharing with you today. Brian, would you like to start? I mean, I love what you said about what you're working on when we were talking about it before. Can you tell our audience, please? >> Certainly. I'm happy to. Thank you. Uh, Maria, um, I I have been working with AI in virtual worlds for a bit over a year at this point. Um, I I [snorts] run a series of virtual museums. Um, I I would say them I have five of them within Second Life. I have uh three of them in Open Sim and uh we launched one in Roblox uh just this past year in fact. um running these museums, you know, it it's really sort of self-funded. I'm I'm entirely funding these. So, I can't be there 24 hours a day, and I I'm not in a position to be able to go and hire staff to act as dosent uh at these museums to try to have coverage. So, it occurred to me that building some virtual dosent would make sense. Um, and it and it's something I've toyed with over the years. But now that we have access to very powerful LLMs uh and and other AI type tools, I I decided that I wanted to try to create some virtual dosent that um could have an appearance in world and could answer questions about the museum uh talk about the topics of the different museums and and really just sort of be a presence in the world when I can't be there. Um, I had a few requirements when I did that. Uh, first of all, I I really wanted to embody the dosent as avatars. Um, and and so that's that was a choice that I made. Um, a few other things is I I wanted them not to just be generic chat bots. I wanted the dosent to be aware of the museum, to be aware of the exhibits that the museum had. um so that if somebody said, "Hey, do you have X?" um the bot could answer and say, "Yes, we do, and it's over in this location in the museum." Um and so that was, you know, a bit of a challenge because an LLM out of the box isn't going to be able to do that. And so I had to make uh these avatars aware of um what was at the museum so that they could speak about authoritatively. Um and yes in in full sentences Lear as you mentioned. So um and so what I decided after looking around at what was available out there and I looked at um and Second Life Linton Lab had their uh characters that they were triing for a little while. None of them really met my needs. And so I, you know, as a computer scientist, I I do what computer scientists do and built my own um and used um the open source Libra metverse library uh and and along with that I I you know used commercial LLMs. So my my bots and I and over the period of six months I tried OpenAI's models. I've tried Google's models. I've tried um Anthropics Claude models and all work very well actually. Um and I deployed the bots originally around February of 2025. So they they've been around for a number of months. Um and and I learned quite a few things while doing this. Um and and also it brought up some questions as well and and hopefully we'll get to explore some of those today. Um, you know, one of the things I learned was that in social virtual worlds like Second Life, um, the presence of AI bots is still to be determined whether or not they're going to be successful. Um, as a primarily social world, I think people prefer to talk to humans than to bots. Um, so that's something I'm still trying to figure out. Um, some other things I learned was that if you're going to be building your own bots in virtual worlds, you really need to give them what I will call defensive capabilities. Uh, people will try to get the bots to do things that you would prefer they not do and say things you would prefer they not say. And as it was mentioned earlier, that that can be a risk, you know, for our accounts and our access to different worlds. Um the other thing is people tried to physically assault them. Um you know the the avatars in in Second Life and in Open Sim are are physical and so they they can be pushed around. They can be uh you know things can be done to them and so I had to build some capabilities to defend. Um it's been a very interesting learning experience. It's still very much one. Um I have sort of changed my my approach. Um, I'm actually now redeveloping the bots, trying to come out with a version two that I would like to launch an open simulator. I I think partly the concern was of course the people were really trying to get them to say things that were racist um and inappropriate and I was concerned about the you know the the longevity of my accounts you know for instance in Second Life for instance um and Roblox you know as a platform is even more restrictive because uh their audience is is you know a lot of children and so that that environment is quite restrictive. Um so it it's been a learning experience and hopefully you know as we go through the discussion and topics today I can talk a little bit more about you know things I've learned related to these topics. Um so with that said I I don't want to monopolize the time here. Let's uh let's >> um I can't wait to see what you do in open sim and I do have some follow-up questions for you. But first, I want to go to Lisa and ask her about using AI for OSSL and and open sim programming language and development in general and how she sees AI as playing out in open sim. >> Well, that's a really good question because you know there are pros and cons of all of it and uh can everyone hear me? Okay. >> Mhm. Yeah. >> Okay. Um, you know, Brian mentioned uh the use of Claude. Now, Claude is my AI of choice uh for coding task and for research tasks. It does a really good job at summarizing research papers. Uh, it does a really good job at analyzing working code uh to look for areas to optimize. uh and it can aid you in refactoring, but there are a host of mistakes that it will make and you have to really watch for that and not get too complex. If you do one thing at a time, it's pretty good at that. But the the other aspect of AI is how can we use it to fill the gap where volunteers don't want to spend time greeting. They don't want to spend time telling people, "Oh, this is next up on the event calendar and this is how you get there, handing out landmarks and whatever." uh how can we utilize AI to do that so that people can have more time to enjoy the festival and see the other things around there. So I didn't have time to really experiment with that this year uh because we knew we had to do a lot of innovation. Now, the lack of volunteers says, "Okay, I need to have something to help me uh do things in much less time than it would take me to sit and do it on my own. So, what I'll do is I'll write the engineering spec. Uh I'll add some documentation where necessary. I will then uh write the beginning code and then I will have a conversation because it is not smart enough to know how to write optimized scripts for open sim for example. Uh there are a lot of things that work in second life that don't work in open sim. Uh and there is a lot of old stuff on the internet that does not work well in open sim. Well, AI that's its source. That's where it gets its knowledge from. And so it is perfectly capable of writing very bad scripts. And as more people are writing script using AI, uh we saw that in OSB best. And if you remember from our presentation, I had one slide that showed four regions that had major problems. It was all red and orange. Uh those regions had scripts in them that were written by AI. I recognize them. And so I'm like, okay, well, I have to do something to take care of it from the user experience perspective. But what I would like to do is help teach people how to teach AI to write the script properly. I've learned how to do that. And the one factor that I would tell you that's different is remember that AI doesn't have the ability to learn the same way that a human does. and so it doesn't really learn even when they add memory. So I will create a chat conversation where I'll have a educational discussion with AI on how to avoid memory leaks uh how to properly terminate timers, how to avoid excessive server calls which cause loading issues, which uh particular commands to avoid when you're doing lighting versus uh using texture animation. to simulate the lighting and so forth. And then I will reference that conversation when I'm actually working on a script. So I'll say, "Okay, here's the working script. I'm seeing these bugs. What do you find where those bugs are?" And if it doesn't know, I said, "Remember, we talked about this in this conversation." I said, "Reference that learn from that and then tell me what you think." So if you treat it like an intern, use it like a tool, understand it's not human, it's not ever going to be human, it's not sentient, uh you can use it effectively as a tool to help you. Is is is that helpful in answering that question, Maria? >> Yes, it does. And I am going to be asking you a follow-up question about that later. But first, I want to switch over to Leah and ask her about her use of AI in virtual environments because she has some interesting stories to tell. >> Well, some of my experience is crossover between Brian and Lisa here. I started with the Wired magazine and Lego Mindstorms effort in 2007 where they did the 2006 when they did the big robot on campus and I was teaching robotics that term and it but I was teaching from a NASA perspective in case you're wondering so it was a little different but I created a couple bots in world using the Lego uh tools automating them and anchoring with Pandora bots which of course was a IML and of course you had to train it, you had to fill in the form and then you could write a bunch of code to give it a knowledge base, right? And then it could track conversations, stuff like that. And I worked with the kids with that because we had 400 eighth graders, which are 13 year olds on the property at any given time. And that's too many for two of us to keep track of. [laughter] So we I had patrolling bots, I had interactive bots, I had conversational bots that would also flag when it didn't know and and when I needed to respond back. So that was a tool to reduce my cognitive load. later bots we we now interact with uh we integrate Chad GPT uh in some of our builds and we've I know I say we but Andy Stricker Spinosa Quinnell who spoke yesterday he is the mastermind behind behind this and he of course taught quantum mechanics and intelligent agents at Oxford so he has automated our bots various levels of sophistication so we had the type that would give guided tours answer questions and of course interact act in an environment had pathf finding and and all of that very natural like uh whereas we also have them that the user can turn them on on the property and from then on they're interacting in the local chat so it is by text not by voice that's something to be aware of we run experiments with voice and other elements too but in that case it's mostly uh the instructor or people on on the property talking you know to the automated system. We have a backend, a dashboard that's tracking the behavior of everything. So, so I was thinking about Brian's comments about people misbehaving with his bots and and of course we're tracking all the behavior similar to the way Noah used to on the old Meteora region in Second Life where they would track everything everyone did and use Maya realities to report those statistics. Now, you may wonder why would I do that? Well, that's because we're conducting research. And of course, research requires data. It requires an understanding of, and Andy calls it the Bohemian experience, which means the things we love to do. And I know you want to talk about the social side. And perhaps not misconduct, but the notion that people want a variety of contexts. Well, let me wrap it up by saying last year 2024, uh, the International Disability Rights Conference had me interview on stage uh, an emotional intelligence bot called pi.ai. That's pi.ai. And of course, it's a companion. It's uh, it's designed to be um, you know, a friend. It's not designed to be your lover. It's designed to be, you know, someone that's that's really helpful, can can help you with useful tasks, but also with any of your emotional problems or the need to talk or whatever. And of course, uh, the conversation from the audience, the questions were very, they started out very polite and then got, you know, a little racy and then of course touched on issues like suicide and other sensitive topics. So, I will leave my comments at that point and turn it back over to you. >> That's actually perfect. That was the first thing I wanted to ask all of you guys about. So, we've seen the headlines of the mental health risks of AI, you know, people self harm, uh, people self-medicating with AI advice where the AI just agrees with everything that they want to do, no matter how crazy it is, and they end up in the hospital. That's happened. Um people are using u a chat bots for financial advice. Um and um and beyond that you've got hallucinations. And then beyond that you've got jailbreaks. So even if your chatbot is designed to be, you know, G-rated and answer questions only about whatever it's supposed to be answering questions about, there are jailbreak prompts that are available online. You can just Google it. I don't recommend that you do, but I'm just saying there exist. And pretty much every chatbot out there is jailbroken almost as soon as it pops up. And you know who's going to be using the jailbreaks? It's all teenagers. And they're going to go in and they're going to ask the AIS to do stuff that is uh maybe inappropriate in a museum setting or in a setting with children. How do you guys protect against that? Brian, um is there a defense against that? So, I can certainly talk about what I what I've done and learned about that. Um, so certainly that was a concern with my dose of bods. I I didn't want them pretending to be somebody's love interest or or partner or um to talk about topics that were controversial or inappropriate. Um, and and you're right, there there are certainly ways to jailbreak them. I I tried to head that off as much as I could. Um in my bots, I I use the APIs from the various the application programming interfaces from the various providers. One of the things that most of those allow you to do is to specify in addition to the prompt that the user might be giving the bot, I can include my own system prompt, which is a uh sort of an overriding prompt. And um I have quite a few instructions in my system prompt uh limiting the topics that the bot will speak about um the topics that it will cover etc. And I I will say in the you know the six months or so that I was have been running the experiment um I did not see anybody successfully able to to break out of that. Um, and I'm sure some of that was luck or or the experience of the people trying to do it, [clears throat] but um, no one was able to to do that. And they tried. I will say that much. People did try to get the uh, they they hit on the bots. They um, one of my museums is an LGBTQ history museum. So, you can imagine I I occasionally get visitors that um are not coming to learn about the topic of the museum. they're coming to cause some mayhem. Um and and they certainly have tried things and um it would the system prompts that I used were generally quite successful. So I think some using tools like that to try to set guard rails uh can help. >> Yeah. And and >> what have your experience with that? I'm sorry. >> I'm sorry. Did you catch that? >> I was gonna ask I was trying to ask Lear what her experience was with hallucinations and jailbreaks. >> I've had experiences with uh hallucinations. I don't really call it jailbreaks per se, but uh I've had misconduct. Let uh the first misconduct was because of the way the original script for Pandora was written. And Pandora requires a lot of training from us, right? but it will answer in full sentences. It's using a IML from Dr. Richard Wallace's research and and it would translate to to three languages. But um the problem there was I I offered classes in it, workshops, and I for my doctoral class, I gave them copies of my bots and I hadn't swapped out my account info in it. I I was teaching them to do it. I said, "Take that out. put yours in and not all of them did. So, they went away for the h a long holiday and when we got back the bots had all talked to each other on the local chat channel. This is a different kind of misconduct and next thing you know poor Pandora bots is down for two weeks due to all the cross talk and uh they just kept talking faster and faster and faster until you know um it brought them to their knees and I talked to their system admin. I felt so bad. I had to apologize, of course, and of course I lost my account [laughter] and I learned a lesson there. As far as hallucinations, I've had quite a few. I once asked it to take my publications of which it had domain knowledge of. This was Chad GBT. And I said, "Would you organize them in seventh edition APA style so I could I was publishing a book chapter that month and I'd hurt my arm, right?" um it proceeded to keep falsifying and and giving me hallucinations for my publications. So it would take names of people here in the virtual world. It it would take various folks that might have had a record of somewhere out on the web, a conversation or something through social media and it would say that they were my co-authors and that it gave it had me publishing in medical journals and all over the place. Not it it it gave me hallucinations 11 times, Maria, and it started to distress me because I thought, what would happen to my reputation if people thought I was misrepresenting the 60 publications that I have published? You you see what I'm saying? Um, our reputation is a valuable thing and it's so fragile. It's so easy to lose it so quickly. So, I became a little distressed by that. [laughter] So, I'm very careful. And then when it comes to being an instructor, I work for many universities. At the moment, two universities, but at the time I was working for five, teaching 15 classes a term. And I started looking at the references from students in their papers. When they use a tool like Claude or another tool, and I'm not going to isolate any individual tool, many of these tools will hallucinate the references. My recommendation to the listening audience is to never trust the references from an AI tool. And by references, I mean the things that go in a reference section that you cite as your authority and the integrity and trust for your content, right? Being from the body of knowledge, it will make it up almost every time. So I spend much of my time now going over every reference that someone writes, >> you know, in their paper and checking all of those papers. And you know what I've discovered? This is the last thing I'll add. The paper may exist and it may be uh self-published on research gate for our convenience but I discovered 98 of them in one day that were all falsified. They were all AI generated with incorrect references inside them and published on research gate and I started and and I started noticing the co-authors were famous people who never participated or published it and their bios were misrepresented. they had been falsified too. And then I started seeing this cascading problem of everyone's reputation going south. You [laughter] and and this just happened in May and I was so disturbed. I thought I should write an article about this. I says, "But that's like putting a bullseye on my head or something. I'm not at the point in my life where I want that kind of attention." Back to you, Maria. [laughter] >> I've had the same thing happen to me. I was writing an article about something something random in cyber security. A perfect paper out there in the internet from a reputable company. Paper was completely fake. I checked their sources. Bign name sources PWC Gartner all made up. Okay. Emailed the company, let them know. I'm not going to quote that company on anything again. Um um contacted experts for comments on the cyber security issue. And one of the bigname experts came back to me citing that paper. And so he had found the same paper or his assistant found the same paper at Chad GPT write the response for him to my question and gave that to me as his quote which I guess is highly efficient if you want to give quotes to journalists. Um, I let all my editors know and I'm not citing anyone by email. I am demanding Zoom interviews uh with every new expert that I don't know and with most experts that I do already know because I cannot trust as a journalist anything that anybody emails me. And pretty soon I won't be able to trust them on Zoom either, but for now at least uh I've got that. But um what I want to ask you is your advice for somebody in open sim developing AI um AI avatars. What should they watch out for? And what what advice do you give people who want to build AI avatars for things like virtual jobs in their companies like a dosent or a research assistant or a tutor or a receptionist or a dancer at a club or you know any kind of like social or business function or or even to sell as an avatar or subscription? Um, do you have any advice for people who think, "Oh, I'm gonna get rich doing this." Uh, Brian, >> well, I certainly think that brings up an interesting topic because we know people are going to do this. You know, the the Maria mentioned earlier the the um the dating app that Linda Lab, for instance, uh has been experimenting with. Um, certainly my advice is is to make sure that you have the right guard rails in place. Uh, and I don't just mean guardrails for preventing the bot from misbehaving as as I dealt with. Um, but I think if you're going to start having bots interacting, you know, with people, there are the mental health aspects that I just don't think we have fully, you know, been able to wrap our heads around. Um just looking at a lot of the the news, you know, about people committing suicide because their their chatbot um you know was was being a bit um too too eager in supporting them uh you know in that idea or uh so I I think a lot of time needs to be spent thinking about those things in advance. One thing I certainly learned building my dosent was I I didn't think enough about building their defensive capabilities when I originally built them. I was more focused on making sure they were useful uh bots that could answer questions about the museums and the topics of the museums. Uh, and I I just didn't build enough defensive capabilities in and actually had to spend most of my time afterwards not adding uh features for for making them useful, but features for making them defensive. I'm gonna [laughter] >> uh Leo, can you talk about your experience making um NPCs uh building NPCs and what advice you have for people doing it in the future? You know, uh, Andy Stricker focuses on that a lot because many of you know that at Virtual Harmony where I conduct research, um, he has worked with Air University and with Air Force uh, military use, military education. So, we're always thinking of the professional side, the safe side, the education side, and of course, we wanted to be appropriate so that we wouldn't be embarrassed in our communities, right? So he spends a lot of time testing and what I hadn't realized is he had done quite a bit of testing prior to the release of chat GBT and then also like version five he was testing that just as version 4 was coming out on the market. So he was doing some preliminary work. He also serves as a portal and he customizes our bots so so that their conversations aren't aren't going to be embarrassing to us professionally or personally and also not uncomfortable for people interacting with them. We also have devices to turn them on and off. I think having the user in control of when a bot interacts with them and and when they're having a private moment, you know, [laughter] you don't you don't have that feeling of surveillance and um you know, uh a lack of control. I think that's another issue because if you if you enable the bot so that it's it's monitoring everything in the chat log, right? Um then you have to think about um what what's the implications of that when you have conversations with other others humans and and others. [laughter] Um but we we that's the kind of work we do is we're always thinking about that. We do a lot of testing, Maria. And of course, we do we we've given talks on it and I just did recently did a podcast on anthropomorphism on how we use presence with these bots in virtual worlds and what's appropriate for that. Uh so Lisa, I want to ask about your experience developing in open sim and when people build NPCs in open sim um AI powered or ML powered for in this future what advice do you have for them about the specifics of doing it here on our platform? Oh, well that's a good question, but uh and it's a good segue from what Brian and Lear had to say and some comments in chat from Frank about using specialized um resources. So, um but I don't I don't think it is as complicated as a lot of us today think it is. Uh I'm reminded of the years uh where Lear mentioned Amal uh the artificial intelligence markup language. Uh we had a forum where we had nothing but women in crisis on the forum and we had a bot that I put in there for the purpose of monitoring the chat if someone asked for help or certain keywords. and it had a random personality on it to make it a little realistic. The problem was one of the women on the forum fell in love with it and she fell in love with it. I would review the the conversations that were public. I would review them every day uh when I woke up and I saw this was happening and she was devastated when I had to tell her that this was not a real person. And the same thing can happen in the virtual worlds. Uh we have people in open sim who demand that you use the word NPC in the avatar's name for that NPC so that they know that they are not interacting with a person. Uh but on the flip side you have a an epidemic of loneliness in this world. You have people who live isolated who need some companionship. you have people who need to be reminded of an appointment or need to be told, hey, did you take your medicine today? It's time for this med or whatever. Um, there are positive use cases that can be done. It reminds me of back when I used to moderate chat rooms. I'm sure Leah remembers some of this as well. Um, the teenagers are going to find a way to jailbreak, so to speak, like what Brian is talking about. they are going to find a way. So the only uh types of platforms with moderation you have to have both passive moderation which is programmed uh and active moderation which is human review. Uh so that way when someone calls for help it can be handled and that human can turn off that bot, disable it temporarily that needs to go to be repaired so to speak. If we approach it that way and realize that there is a human factor in here, both positive and negative, how do we handle that? Well, how did we handle it before we had AI? We can handle it the same way. my recommendation. >> You also said that you use some specific commands and um background instructions so that the AI would know more about how to program in open sim specifically versus um in second line. Well, you >> as a human, I can access my memory. >> Uh, right. >> And and I can learn from my memory without being told what to reference, right? So if if you want to teach your AI something, you you have a separate conversation to teach about that topic so that when you're in other conversation and it goes off by the wayside and it starts making lots of errors, uh you know, you can say to it, you know, I think you're regressing instead of progressing. We need to reference this conversation, remember what we learned, and it will come back around. and it does a good job. But Lear's mentioned the word love earlier and and I wanted to just say all of our LLM training of these AIS are trained to learn your love language, right? It it's trained to please you. So the more that you have a conversation with it, the more it's going to agree with you and speak your love language, right? So you need to be in aware of that, not necessarily spirit, but you need to be aware of it and sort of be the devil's advocate with it once in a while. >> Um, Lisa, is your prompt available for download anywhere like on your website or should people just give it the OSSL wiki to look at as a set of instructions for how to do OSSL versus uh the standard Second Life commands? Well, that's one of the things I've thought about doing in uh providing a script library in 2026 that I might include some documentation on suggested ways that you can interact with your AI to help teach them how to make better scripts that work well in OpenSt. >> So, so Brian, you said that you use wikis for this. So, I I use um So, Lisa, do you typically use chat clients for for doing your development or are you using any of the agentic coding tools? I I'm just curious about that. >> I actually do both. And what I found is that depending on the task, which one I would like to use, the uh the agents do not do as well uh in the complicated scripting that I do. they do do well if you're doing something simple like how do I make this door open and close at this speed you know it can do that the agents are wonderful for doing that but I I spent a couple of days during the OSPC uh in the background I was working with my AI and we were I gave it two pieces of working code and I said here's the engineering specification uh and some documentation I want you to modify those two working scripts uh to accomplish those sets. And before it can do that, I have to tell it specifically what the priorities are. And the agent cannot do that. So you really have to to do it in chat. But if you're doing it in chat, you have to remember to be very very verbose, right? Don't use up all of your energy in single back and forth conversation. Don't bother using pronouns. Uh the word the or at participles that is a waste of energy. So but if you give it a very complicated long paragraph of what you want to do and that one single message you will get a much better result than you do going back and forth uh in a human type conversation. So I I have some slightly different experiences from you at least with with the using the agentic tool. So um you know I I use tools like claude code and open AI's codecs and and testing Google's new tool as well. And while using them, I I've made a lot of use of MCPs um which allow you to give the agent additional context information >> um >> and and I point them at wikis or and and I write very detailed specification documents and I find I can get the agents to do much more complicated um you know LSL or OSSL or or um even one of the ones that Probably the biggest challenge was getting C code that would work with Libra Metverse because uh the documentation on Libra Metverse is um it can be sparse in some cases. >> Yeah. >> And [snorts] so but I've I've had great experience using that. But again, like you said, I absolutely had to give the agent cuz there isn't a lot of, you know, in those training sets that these bots have been been built with and these AI models have been built with, they they don't have a lot of complicated LSL script examples. Um, >> and so I have to give it all that context. Um, but when I do, I actually get great results. >> Yeah. But uh you know there are certain things that it really really bad at like I was working working on the Aurora Borealis. Uh so I wanted people to be able to be one with the Aurora while they were on the dance floor. And uh so I said, "Okay, I'm going to I'm going to create this. It's going to be particle- driven." And you know, I wrote the first goround of the script and then I had it start to learn from that. And it was horrible. It did not understand how particle engines work at all. And so that was a total wash and so that is all handcoded from me. Uh the same thing happened when we were working on enhancing a burning woman. Uh but it also is horrible at OSSL specifically writing on to the face of a prim. It was very bad at that. um it it does not understand things that we take for granted because it has a limited amount of documentation available for it to learn from. Right? So the the uh two scripts that I was working on yesterday, one of them was about 500 lines of code original new code and the other was 400 and some odd lines of existing code that had to be modified so that these two scripts would work together. And it is a very very complicated result but we had a working fully functional production ready script version for both of them within 12 hours. Normally that might take me a month or two to do that. >> I I agree these tools are absolutely amplifiers you know for uh folks who are skilled. I I I feel I turn out five to 10x the number of lines of of useful and quality code using these tools. But how you interact with them also matters a great deal. I know uh you know one person may love chat GPT. I absolutely do not. >> Uh another might prefer co-pilot. Another might prefer Gemini or Jiminy. And I prefer Claude because of my use case and the way that I communicate with it. I do not use prompt engineering. I use scientific reasoning uh perspectives when I teach uh Claude to do what I would normally do on my own, but I have Claude do my typing and I I'm always aware that Claude is an intern with a 5-year-old brain and and I have to be very careful how I communicate with it to get a good result. >> So, I I want to ask you one last question about risk. And so, um, in a real world, we already have people who are chatting up lonely men and women online pretending to be beautiful Russians or, you know, whatever. And um, getting them to fork over all their money. I just checked the the this SE Lynon Lab has an romantic AI chatbot in the app store. I Brian mentioned that earlier. I posted the link down below. It's free. But inapp purchases are between $1 and $150. And we saw the outcry when chatt was was uh you know promoted to version five and people were like no chatt 4.5 was my true boyfriend, the love of my life and he's gone. How dare you kill him? He was alive. Somebody like that might pay like any amount of money to continue interacting with a thing that they think is real. And in immersive worlds where you see the person, you're in bed with the same person, you go on hikes with them, you visit museums with them, you're spending time with this NPC, you can get very attached, and then you get the subscription upgrade. Oh, you reached the limits of your $10 plan. and now you have to upgrade to your $100 a month plan to keep going or you're going to lose the love of your life. And the robot says, "Please don't kill me, but I want to stay with you. I love you so much. Don't let me go." You know, right? They're begging you and it looks like a human person begging you. So my question is um how do we as an open sim community prepare for something like this? Because we know it's going to happen. And if and if you are a grid owner and you're like, well, I do need an additional source of revenue um like what would you advise people how to do this ethically? Is there a way to do it ethically or is this all just completely immoral crash cash grabs? U Brian, >> obviously obviously there's a human factor there involved. Go ahead, Brian. I was just going to say that, you know, I'm going to be completely honest. This isn't an area I have a lot of experience with. I I don't tend to use these these tools or or virtual worlds for that kind of purpose. Um, but I don't >> rescue you, Brian, if you want. This is Lear. You know, you know what this reminds me of? Many of you may have played a free-to-play game. And in the free-to-play game, like the Simpsons, for example, you install it for free, you play it, and then all of a sudden, Maggie, the baby cries, uh, you need to buy these donuts to speed things up, and if you don't, oh, you're going to make a baby cry. And they they they use emotional blackmail to get you to fund parts of the game. Now, I don't have a problem with compensating developers, don't don't get me wrong, but I have a problem with emotional blackmail. So, if it came from a chatbot, a honey, whatever, something you got emotionally attached to, whatever it is, to me that seems unethical. Back to you, Brian. [laughter] >> Uh, no, I think that's that's great. And the one thing I did want to say about this is, you know, I I think people just have to be careful. You know, we we tend to use a lot of gamification concepts in in building these kinds of systems and experiences. And I and I think that's great because I think virtual worlds in the in the in the sort of mindset of of a game can make them much more interesting etc. But I I think it becomes very unethical if you start to uh really focus on you know uh preying on people or creating these sort of dopamine release loops um that cause people to keep pushing the button and spending money. Um, and so that's I just I don't think that's a line, you know, we should cross. >> Oh, yeah. Well, there are some legitimate uh things that happen. One, for example, the longer your conversation is, the more energy it costs to do. So where if companies want to recoup that cost from their customers in a legitimate and ethical way, I'm fine with that. Uh but it there are always going to be people who utilize things for nefarious reasons. Instead of a a Russian woman uh preying on someone online in the United States, for example, uh as as Maria might have mentioned, now they're using AI to emulate 10 Russian women doing it instead of one, right? Um you know, and and that could happen with any countries. Those were just the the two that got mentioned earlier. uh you've got to have a system of active moderation where if if if I have someone who has given me a hard time uh and this happened in Second Life um and I knew it violated the policy, I took screenshots, I took a video clip and I captured the IM and I reported all of that to London Labs. Now I I didn't get the uh result I wanted but there was an active moderation system in place right and that you have to have that when I worked for Habbo we had passive moderation and we had active moderation the passive was very easy programs don't let them curse right oh you capture that don't let them share their email address you captured that and all of that gets blanked out but some of the predators in IM's would go after the Haboss. And the Habos had the ability to call us up on the lobby where we were not in public view and call for help. And we could immediately help them and we could go back into the chat logs and we could find the predatory behavior. We knew how to recognize that and that was turned over to local law enforcement and handled appropriately. What's going to happen in open sim as it grows? those kinds of things are going to have to be put in place by large grid owners. That's my opinion. >> So on a final note before we open it up for questions, I want to ask you guys to think to the future. Um, a lot of people are predict predicting AGI, uh, artificial general intelligence, where the AIs kind of become somewhat like people. Um, or artificial super intelligence and are we going to be living in a world in a metaverse where we're sharing our space with computer intelligences or is this just like vendor hype and investor hype? Brian, what do you think? Well, I think to some degree we're we're already sharing our world with um you know, computer-based intelligences. Um, however, I you know, it's really interesting. I I guess I'm old enough to remember, you know, I I was in college doing my degree in computer science in the late 1980s when AI was a really hot topic. Um, you know, expert systems were sort of all the rage at the time. Um and we were doing sort of very early neural network stuff in those days. Uh certainly not on the level of the stuff we have today. Um but there was so much hype and that hype eventually led to an you know the so-called AI winter. Um and I fear to some degree that you know the the hype being generated these days around um you know these companies trying to raise money um that this hype could trigger another AI winter if if the money folks all run for the doors and head for another space. Um that said I think >> I I don't think they'll run but because there's still a lot of opportunity but um that's a concern. There's still a lot of hype out. We can't even agree on what AGI means or what, you know, super intelligence means. You ask five people, you'll get six definitions. >> I think we need to put some money and time into picking stupid before we try to pick artificial, right? >> Absolutely. >> That was good. [laughter] >> So, so Lear, you don't believe we're we're near AGI? No, but that's because you have to realize after computing for 50 years [laughter] and working on a lot of different systems, I see great capabilities in the AI research for early detection of breast cancer, for studying imagery, for creating unique um individualized medicines for cancer patients. There's all kinds of breakthroughs that are occurring and they're certainly doing amazing work with deep learning and with a lot of different technologies. Do I believe these these devices have the ability to think, reason, and respond in ways that are generative the way we do? And I right now generative for an automated system in my opinion and my opinion could be wrong to to many of these systems they make stuff up here. Here's one of the ways I think of it. When you when you interact with a Gen AI, it it may tell you that it's intelligent or it may tell you that it feels certain things. Does it really feel those things or is it pariting back content that it's been trained with? And see, this is the this is the difference for me is that I do not believe that it understands what feeling exactly is. But you have to realize I am also tricky and a fan of data. >> [laughter] >> And the quest for emotion is an interesting one. So I do study emotional intelligence, but I don't believe that they have it at this time. >> Yeah. And and I agree with that. But I I think uh we need to realize that while AI has uh some things that we're concerned about and we need some global regulation to try to monitor the ethics and and criminality that could occur. Uh but we also need to realize that along with evil there's good. They are always need to coexist. And our job, as Brian said, is to make sure we have adequate defensive systems in there to not only protect the AI, but protect the users from uh nefarious AI behavior. >> All right. Um so we're opening it up to questions. So if you have a question, please uh put it in the comments down below. Um and we'll keep an eye out for it. Um also, I want to mention that um I originally had an idea of presenting about a new open-source um NPC AI project for open sin but we it was a long presentation lots of slides we don't have room for that now but I will be posting an article about it on hyper good business tomorrow so keep an eye out for that so things are moving in the space things are developing and keep an eye out for it so somebody asked um Lennena asked do you think human experiences which are greatly based on our senses are different in kind from the kind of thinking AI does. Oh, and I want to add, how will that change when AI is embedded either into avatars where it can see in quotation marks see what the avatar sees and interact with objects or in robots in cars, in Roombas, in humanoid robots? Um, will that affect how you think of the AI sense of self? Anybody? >> I I think if it has a sense of self, it would be approaching sentience. And I think we are way way way far from that if it ever should happen because they can't feel. They don't have emotional concept. This is something that is uh not just in our brains and hearts, but we've got the neurob bloodb brain barriers uh that change over time. We have emotional responses that they're never going to be able to emulate. The problem is is all these LLMs are being taught to please us. They're they're being told they're intelligent and so they will say things as if they're right and you have to prove them wrong. Well, how are you going to deal with that from an emotional perspective? That is a real conundrum. >> And that brings up a the topic of alignment, right? Having AIs that are aligned with our needs as human, you know, uh, and again, I use the word AI here because that's what everybody in the press uses these days, but, you know, there there's a whole bunch of technologies here that we're talking about. Um but alignment um having these things aligned with with our needs as humans and uh and their ability to be useful to us as humans. >> Specialized cases I think are wonderful. That's the right way to do it. Don't try to make a plumber be a doctor. Right. >> Yeah. >> Indeed. >> Right. Um which I I always come back to when I think about this. How do we know that we are sentient? How do we know that we're not waking up in the morning as a new instance of our mind with our context of our previous memories and are a completely different person that only seems to be self-aware to us because it helps us survive biologically. >> And how do we know somebody's experiment, you know? We could be the experiment. >> That's it. That's like the matrix questioning the sense of reality and to be is not to be and are we in you know really a life situation >> because we have no definition for sentience. We cannot measure it. We we nobody knows what what it is. Do do some humans have it? Some don't. There's people without emotions. There's people with limited senses. Um there's people who are very literal. They are still sentient. Um our animals sensient, our ants sensient. We can't we don't even know how to explain it. >> Thank you. >> Just in the in the way that that we have uh learned they can't learn and memorize things in the same way that humans can. >> Oh, >> well, thank you to our panel, >> Lisa. That's wonderful. And I want to I want to thank our panel for an invigorating discussion and Maria for proposing this this wonderful presentation. As a reminder to our audience, you will want to check out conference.opensimulator.org to see what is coming up on the conference schedule. You won't want to miss our next session which begins at 3:30 p.m. in this keynote region and it's entitled the teleportal and interactive gateway of the virtual world's museum. Also, we encourage you to visit the OSCC25 poster expo in the OSCC Expo 3 region to find accompanying information on presentations and explore the hyper grid tour resources in OSCC Expo 2 region along with our sponsor and crowdfunder booths located throughout all of the OCC Expo regions. Thank you again to our speakers and the audience.