In this episode of Your AI Injection, Deep and the Xyonix crew explore the potential of virtual concierge services, discussing their role in transforming organizations and implementation challenges. The four Xyonixians discuss the capabilities of conversational AI technology, the importance of context and personalization, and the challenges of controlling AI chatbots. The discussion also touches upon the dangers of bias in chatbot platforms and the need to understand the unique value proposition of virtual concierges. The episode concludes with predictions for the future of virtual concierge services and the potential for more comfortable and personalized interactions.
Check out some of our related content:
Deep Dhillon: Hi there. I'm Deep Dhillon. Welcome to your AI injection, the podcast where we discuss state-of-the-art techniques and artificial intelligence with a focus on how these capabilities are used to transform organizations, making them more efficient, impactful, and successful.
Welcome. These are, we got a bunch of Xyonix here, um, Carson and Bill, who y'all have heard from before. And Andy's here he is, uh, new to the podcast but not new to Xyonix by any means. So why don't we get started a little bit. Today's topic we're gonna really be digging in on, um, on this concept of a virtual concierge.
And I'm gonna start by just kind of describing some actual human concierges, cuz I don't know that everybody has interacted with one, but I mean, the one that most people probably know of is of course the hotel concierge. And so all of us have probably been at a hotel and I remember I was, I had just graduated and got my first real job and stayed in my first like, fancy hotel.
And um, I had a friend with me and uh, and they're like, why don't you just ask the concierge? I was thinking. Who's that? So I go up and I talk to this, this person who had like very well manicured nails and was just this very, you know, distinguished gentleman who knew everything about everything. And I thought, wow, that was cool.
So hotel concierge is one, um, a couple of other ones. You know, everybody's seen a movie with the, you know, with the fancy apartment building in New York that's got, you know, the concierge that does all kinds of stuff for you, from organizing events to coordinating deliveries, maybe even, you know, helping you with your housekeeping services or whatever.
We're seeing this concept in healthcare, you know, a lot more where you've got like personalized assistance to patients and maybe their families, um, helping with medical facilities, transportation, lodging, and meals. So, you know, you see it in cruise ships, in retail, airlines, events, fancy clubs in LA have these folks.
And I think the, the one thing that you know, everyone thinks about is it's like, it's somebody who really knows a lot, they're. Lovely conversationalists. Um, but they're also, um, able to perform actions for you. So they're not just talking, but they're often able to like, make things happen. Most recently, you know, I was in, in a nice hotel in Delhi and um, I thought it was kind of random cuz like when you, when you fly into India, you know, usually your, your sim cards don't work that great and uh, it's just not worth the effort of having a foreign sim card.
So I, you know, got some for me and the family. At the airport and in cla you know, it was just complete mess to get these things activated. Cuz we flew in at like two or three in the morning. I'm like, ah, whatever. We'll figure it out the next day. So a bunch of us are pretty tech nerdy. We're, we're like, what the heck?
We cannot get these things activated. So then I thought, I'm just gonna ask the concierge, go down there. He is like, oh yes, no problem here, blah, blah, blah, blah, blah. He's like, just leave your stuff with me. I'll figure it out. And then, so we're like, okay, fine. Whatever. Messes around, figures it out, gets it all done, hands 'em back.
So that's the topic today is we've got this new Chad g p t like stuff that we've kind of talked about a bunch on the podcast. We've got this, which, you know, so we've got this capability to handle the highly conversational part. What do we think about this idea of the concierge in general and, you know, the future of it in, in technology.
So I'm gonna just kind of leave it there and let somebody jump in.
Carsten Tusk: So my, my first question is like, you go to the concierge when you go to the hotel, right? Why didn't you just Google what you wanted to know? Did.
Deep Dhillon: You've never been to India,
Carsten Tusk: so you, but it, it's, it's more of a general question, right? So let's say you're not in India.
Let's say you're in the US where, you know, you could actually read a Google something. Like, I'm in the hotel, I don't find the restaurant. I think the main
Deep Dhillon: reason you don't, I mean like, I don't know. It's a good, it's a great question is like why, when do you use a concierge versus like seeking it out and solving the problem yourself?
I think that's the meta question,
Carsten Tusk: right? Right. Because that, that will tell us what is the difference between that, right? Between this like completely computerized, automated answering server that is unpersonalized and just gives you like some objective answer to your, to search and like a concierge. So what.
Before we figure out how to build a virtual one, we have to figure out what it offers above that, right? Or
Deep Dhillon: or just why people use it. Yeah, so what do, what do, what do y'all think?
Andy Skalet: I think there's a couple different dimensions to it, and one is curation, right? So if you're a business person, you're heading somewhere, you wanna figure out somewhere to go to lunch, you don't wanna sit down and research all the restaurants in Manhattan or whatever.
You just talk to someone who already deep has made this point before that they can take a look at you and maybe get a good idea what kind of recommendations they can make and very quickly draw from their banks of curation. On the technical side of the sim card thing, you and I might be inclined to figure that out ourselves.
I think a lot of people probably won't, and if they can get help with something and not worry about it, they're gonna be happy with that as well.
Deep Dhillon: So convenience seems to be like a key, a key driver. I was,
Bill Constantine: I was gonna say the number one thing, when you think about you, and all of us here are very technically savvy.
When I think about some of the folks that I know that older and not so technically savvy, certainly comes down to convenience and maybe a, a former communication that they're more used to, you know, actually talking with a real human being. And so that's the other thing about these, uh, I would say as almost like a qualitative note, that these concierges that we create virtually, you know, they're, they're probably gonna have to be designed to be effective in a way that, that facilitates different forms of communication, the style of communication that people are, are used to.
Carsten Tusk: I think the curation that Annie mentioned is, is really important. And that's kind of like why we currently see search engines, uh, fretting about, you know, this, this new technology because you can actually execute your search. Let's call it search in something like g p t by asking a question. And it does like the sifting through the results and the consolidation to you to give you like one precise answer.
Now that can be good. Sometimes that's bad, but you know, it's convenience. Uh, that's for sure. When I go to a concierge, uh, especially in a hotel, right? There needs to be a certain level of like personal skills, right? There needs to be empathy. I, I want kind of like I ask him, uh, and typically I get an answer back, well, you know, what do you like, what kind of food do you like?
And so you build like this, this personal relationship with your concierge and then what you're looking for if there's a match there is that that person's personal recommendation. So I feel like that individualized or that that, um, um, that person's personality comes through in the recommendations they give you sometimes.
Deep Dhillon: Yeah. So what I'm hearing is like two things. One is like convenience, which we covered Carson's. Talking about here, I, I'm sort of reading as personalization, like a heavily personalized rendering. I
Andy Skalet: was gonna make a related point that someone who a hotel is gonna hire as a concierge is going to have a very high degree of social acuity and skills.
They're really gonna put a new stranger talking to them at ease, uh, navigate the conversation very smoothly, show interest in the person. Um, I think it's a pretty high bar for social interaction actually.
Carsten Tusk: And just, just as an example, let's say you go to a hotel and there's three different concierges, right? One is an older gentleman, the other one is a young woman, and the middle one is kinda like a middle-aged man with a fishing head on, right? And I wanna know where to go fishing. Which of the 3:00 AM I gonna ask if
Bill Constantine: you feel that you have some sort of comradery, some with someone in some way, shape, or form?
You do, you do trust them a bit more and you're, you're more open to their suggestions. You,
Carsten Tusk: you, yes. Yes. You're, it's not necessarily trust, but alignment between what you're looking for and what you think that other person, you know. Yeah. Is also
Deep Dhillon: like, sometimes it's definitely trust, like, you know, so like the example when I'm in Delhi, you know, trust is a huge factor.
If I want a cab or a driver, I trust that like this high-end hotel has vetted the drivers. That's maybe less of issue. Or if you're gonna
Andy Skalet: hand over four phones, yeah,
Deep Dhillon: that's less of an issue, you know, I think in a lot of the West. But it's a huge issue if you're flying into like, you know, other places that have high crime rates or whatever.
Even, even in the west, like, you know, like we've got plenty of cities that are, are anything
Carsten Tusk: close to safe. So I was gonna ask you earlier, did you still have money in your bank account after you left your phone there?
Deep Dhillon: I did, I did. But that's where the trust is massive because, you know, it's, it's important.
Yes. This, this exact hotel, like, you know, just, this is a bit of a side, but like a few years ago, maybe 10 years ago, we were there and my wife, um, they have safes in the room and my wife had left like, I don't know, five, 10 grand worth of jewelry in the safe. And, um, and totally forgot about it. And then, and then we went about, and like maybe a month later she's like, oh my God, I think I left the majority in the safe at the hotel.
She calls them up and they're like, oh, no problem, madame. Your stuff is, you know, it's, it's perfectly fine next time when you come through on your way out, just, you know, we'll, we'll have it all for you. And, you know, fast forward 15 years later, I think it's been, it's been quite a little while. We never go to any other hotel in Delhi Bluff, but Lu Meridian, I'm gonna tell it to the world cuz they are, I mean, it's an amazing hotel.
It's like this French hotel in Delhi that's just awesome. But like, but that trust is huge. So we're, I'm hearing like personality, I don't know, empathy slash social connection or something. Convenience is one.
Bill Constantine: I wanted to make a comment here deep about something that Carson had said. I agree with him wholeheartedly in terms of people I communicate with at hotels.
I was recently in Hawaii and I absolutely avoided the person that was at the desk. After hearing them speak, because I knew they weren't gonna get me really the information that I truly wanted. What, what these virtual concierges have in their advantage is actually taking on different personas for different people.
You know, you can have multiple personalities built into
Carsten Tusk: this thing. How do they know which one they should take on? Right. You know? Right. This is all, this is all the game of detection. They have to ask and how do they ask politely?
Bill Constantine: Yeah. So we could talk about that, but you could have the user sort of select the style they like, or you could try to infer the style.
Carsten Tusk: Humans determine that, kinda like how they're approached. You read a person. Right. So if you're the question, well,
Deep Dhillon: yeah. You've got clothing, like you mentioned with the fishing hat. Right, exactly. And how somebody's. Just kind of speaking to you and interacting, and maybe even before they say a word, there's a lot of social cue reading that
Andy Skalet: they do.
There's way more context in the first interaction. But you can imagine if, if a hotel chain or another business implemented a concierge service, maybe you have a business traveler who stayed in that hotel a hundred times before. You have a lot of context about that person and what they've asked for in the past and what they need.
That's right. That the person in the Chicago Hilton isn't gonna have when they walk up. So it's different context. I mean,
Carsten Tusk: the best thing would be you to sign up with your Facebook, uh, id, right? And then they know everything about you anyway. In the first place. It's a different story. Um, yeah, basically, um, you have to make that analysis that everybody online on the web is trying to make about you.
Like, who are you, what are you interested in? And you would have to make that as the introduction to your interaction with the agent, cuz it plays a role in, in how you interact
Deep Dhillon: with that person. What about the visual aspects? So, you know, we've got a number of new. Really powerful, um, capabilities coming out in the machine learning world from the ability to create a quote avatar with your, like a really realistic sound version of your voice.
You've got, um, you know, the ability to like, make a pretty darn good kind of look and feel around you. I don't know if you all have seen synesthesia, but those, their tools are pretty amazing. What, what do we think is the role of this kind of voice and image capability in like future virtual concierge services?
Bill Constantine: So, one of the things that I'd like to talk about here is a concierge in terms of health, and particularly in mental health. There was a study done not long ago in the military for soldiers who had P T S D and they were interfacing with a counselor slash psychologist. Sitting on the other side, you know, of the conversation in a remote room, maybe even in a remote part of the country, whatever.
But instead of having a face-to-face conversation, which we are having right now, they actually had a digital characterization of that psychologist. I mean, for all we know that that soldier was talking to a giraffe, you know, cartoon version of that psychologist or whatever. But what they found was, is that the soldiers were much more open to sharing very personal information because they felt number one, less judged and they felt a certain sense of anonymity.
And freedom to explore those kind of very sensitive topics. So I think that these virtual concierges that have this visual form, in some ways, it helps people to feel more relaxed and open up. They don't necessarily feel as judged, uh, as they would be maybe in face-to-face with a, with a real human.
That's a, that's a very inter interesting aspect of all of this. I would say conversely, I would say conversely though, that we as humans are very judgmental about images, uh, more so than we are about, say, text or something. We, if we see distortions in these images of these digital characterizations, I mean, I personally, it drives me crazy.
I think it's easy for me to pick up on, and I, I don't know if I would be, uh, so, so happy to interact with something like
Andy Skalet: that. Bill, do you have any further context about what the soldiers were looking at? Uh, you mentioned giraffes, but was it like a cartoon image? Well, they, they really
Bill Constantine: weren't, they really weren't looking at giraffes.
They were just looking at like an avatar of the psychologist. And actually, I think in some cases they were given a choice of the type of avatar that they would prefer. You know, very much like we were talking about. What's the type, like Carson alluded to, you know, what, what are the, the different types of concierges that he might approach, uh, for different types of information?
And this was, you know, these. People got to say, Hey, you know, I want somebody who looks like a military man because they know what's going on in the field. As opposed to just some guy with a bow tie who's never, who's never touched a weapon or been in combat, you know, sort of sim simple psychological things.
I think it's actually gonna have an incredible impact. The visual aspect will have incredible impact, but I also think that we're very, very picky as human beings. So it's gotta be good.
Carsten Tusk: Yeah, I think I would agree with that. Um, I mean, the first thing you eliminate is, is kind of like bias, right? Because in any human interaction, the first thing we do is we see somebody based on their appearance.
We have certain perception of them, right? Uh, can be good, can be bad, can be trust, can do the opposite. Um, and I think you eliminate that if you have like a more neutral, more standardized appearance, right? Uh, the second thing is trust. Of course, uh, people are. Kind of hesitant to share their personal feelings, emotions, problems with like a random person because they're also afraid of bias.
They wonder, what does that person think of me? What does he do as he walk out of the room and laugh about it? And so we see that people have much more open and honest conversations with, uh, electronic agents and computer programs. And it's not just like in, in a visual scenario, also like chatting with a chat bot.
Then they would with another human being. Because that, exactly, like I said, that feeling of being judged goes away. Do they, they really. The perception is there that the machines are way more objective and they are, you know, they don't have an opinion, they just have rules basically,
Deep Dhillon: coming from the person who, who said that, uh, by definition they're biased.
Carsten Tusk: yeah. And, and so that's, that's actually something else I wanna say. If you reverse it, because we started this conversation by saying that the concierge must have like an opinion of the person that is interacting with them in order to personalize their interactions. Right? If you would reverse that and say there's a camera recording, the person that is talking to the concierge, and you kind of try to judge, um, their personality based on their appearance, well then you're un biased.
Andy Skalet: right.
Deep Dhillon: Well, you're biased. No matter what. No matter what, all these models are gonna inherit the bias of their
Andy Skalet: training data. The concierge, uh, human concierge is potentially problematic in that regard. That they're, they're looking at someone being like, oh, this looks like a successful business person.
Oh, this person looks like a casual traveler. Well, maybe the casual traveler is a, you know, person of color who's a. You know, crypto billionaire, who the hell knows, right? But they are making a
Carsten Tusk: judgment. But that, that's what makes us human, makes us like talking to a concierge. Because usually what happens is you walk up to the concierge, they have a first impression of you.
That's how they greet you. Then within the first couple minutes of your conversation, they might actually, if they're a good concierge, figure out that, and he's just, you know, he's looking for this. And so when you ask you for recommendation, then you base your recommendation on your perceived. Personality impression from that person after a bit of conversation and
Deep Dhillon: it changes all the time.
It's also, I mean, it's also in their job des description. If they're like really judgey and they're not, they're not gonna make very great concierge. Like part of what makes them sort of good at what they do is their ability to connect with anyone and help that person feel heard and then and also know a bunch and tailor their, their responses need help with computer vision, natural language processing, automated content creation, conversational understanding, time series forecasting, customer behavior analytics.
Reach out to us at xyonix.com. That's X-Y-O-N-I-X.com. Maybe we can help.
Bill Constantine: I wanna make a comment about Carson, about what you said earlier on about empathy. If you think about the concierge in a medical clinic, when I walk up to to check in, I'm usually greeted by people that are overworked, very tired. They always have to wrestle with things they don't really particularly like, like computers and insurance information and blah, blah, blah.
And really the visual vibe of that conversation. It really sucks. I mean, I don't feel very cared for or respected. I don't feel like I'm getting a lot of empathy, even though I might be feeling terrible at the moment. You know, here again, we have the opportunity with these virtual assistants or virtual concierges to change that game completely.
So I think of more of as a positive thing. So in a way, I'd rather, I'd rather talk to them than some of these
Carsten Tusk: people that I've talked to in the past. Then here's, here's a real question, right? It's like even if you talk to a human, there can be real empathy and there can be fake empathy. That human cannot give a damn about you and still like react with you or interact with you as if they're empathetic to your problems, right?
Sometimes you feel that with a bot. It's always that way. You know, a bot doesn't have any empathy, even though the way it acts with you kind of might pretend to be that. I honestly don't think that with any kind of virtual thing, we can get that emotional empathy that we're looking for here, but we don't need it.
Right. For the consumer.
Deep Dhillon: I don't know. I mean, like, if that were true, then an actor or actress would be incapable of moving you, but they can, like when you see an an actress or an actor faking it, if they're really darn good at what they do, you will feel whatever it is they're trying to get you to feel.
Yeah. But I feel like it's, feel like you can get bots to fake it. I mean, we can. We'll get good at that one day.
Carsten Tusk: I don't think so because it, you know, it's never real, you know? Yeah. Well, well
Deep Dhillon: I think that that's true, but that sort of presumes that humans are not capable of being diluted. I mean,
Bill Constantine: I might say that what I'm probably looking for in that experience is not that they necessarily have tons of empathy for me, and that I feel like I have sunshine blowing out my ears after I've talked to them.
Carsten Tusk: But I absolutely don't want a negative experience and I usually get a negative experience from these people. And maybe also, maybe also, empathy is the wrong word for, for like a virtual concert or a chatbot, I think. I think like reading your personality is the important thing and acting accordingly.
Right? Cuz you are asking them about recommendations. And recommendations should fit. What you are looking for, and that is different from person to person in order to give you a good recommendation about what movie to watch. I need to know what the heck you like in
Deep Dhillon: movies. Yeah. And that's, you know, good conversationalists do that all the time.
Like if somebody, you know, somebody asks you, Hey, you know what, good sh you know, what shows should I go check out? I mean, obviously my first thing is like, well, what kind of music are you into? Because I, I mean, there I can, I can give, recommend some stuff you will absolutely hate. And I, you know, and I may be able to recommend something that you'll actually love, but I can't know that unless I know a little about you.
So I wanna shift the, the conversation a little bit. So it's, you know, it's April, I dunno, what is it, April 20, uh, 2023. Like what, why are we having this conversation today? Like what is different today technologically speaking? And let's kind of geek out for a little bit. What are some of the raw ingredients or, or tooling that's now available in 2023 that makes it even feasible to, you know, to imagine, you know, a really sophisticated virtual concierge, whether it's working in an airline or a ship, or a hotel, or a residential or healthcare or whatever.
But what is it that makes this stuff so much easier and more feasible to get to, like a new level of capability than we've sort of seen in the past?
Andy Skalet: Deep. I have a little story to add for context to lead in and my research for this. I found that IBM Watson had done a robot for Hilton, I believe, back in 2016, which looked like a little short version of Azimo that would stand on top of the desktop and was supposed to be.
Making cons. I mean, literally was a robot concierge that was supposed to be making calls in the backend to the, was making calls in the backend to the, um, Watson IBM to do speech to text, but also, you know, was intended to be a conversational agent. And obviously, you know, a, the history of AI agents goes way back, be, be before that.
Connie was her name, didn't take off the, the, uh, Watson Hilton robot. And so it's a, it's a great framing of like, what, what is different today? And I obviously ChatGPT is a big part of what we're experiencing. That feels different, right.
Bill Constantine: For me, Andy, I appreciate that sort of historical view because this has been coming around for a long time.
I mean, even, even, we even can think about this in terms of Star Trek or Andy's sci-fi show we've seen where they're talking to a computer screen or something to help them out in, in setting Facers or whatever they're going to be doing. I would say, you know, personally, the day that I got access to GPT-4.
Was a day, I'm always gonna remember. I really do feel it's changed my life, how I actually work. There isn't a day that goes by since then that I haven't used ChatGPT or from Open AI or GPT four to do some aspect of my job. The one of the reasons why is it's become something that's moved from like migrated from like an entertainment tool or something that's kind of fascinating to something actually useful conversations that you have at ChatGPT seem more natural. They seem to be more constrained. They're not g hallucinating as much. They seem to be very relevant. They seem to be picking up almost on my personality. So just in terms I, it's hard to sometimes quantify that, but to qualify it as it just, the conversation seems a lot more natural with these large language models now and wrapped into these conversational bots.
They are fascinating and they're working and now the whole world's turned onto them and we, we talked about this stuff a year and a half ago. Now the whole world is turned on to them and using them, the whole game has changed. Now it's a matter of trust.
Deep Dhillon: It also goes beyond conversation, right? Like conversation's important, but it's not, it's not everything it takes to get to a concierge level.
So what, what else is going on besides conversational capability? Cuz I, I totally agree that, that the LLMs and the reinforcement layer is having a profound difference in our ability to talk to these things. But like, what other capabilities beyond conversation, you know, are, are sort of relevant here?
Because it feels like there's a lot of movements in other areas. Like, you know, voice rec is at a, is is at a whole new level now due to, you know, advances there. This ability to like, have my voice synthesized, you know, and, uh, and, and create like a, a realistic avatar experience. That's something that. You know, we couldn't do a couple years ago at a reasonable cost point.
And then there's just all of this like ability to take actions. There's just like the overall a p I world where you've got so many businesses and restaurant like the ability to like book a restaurant, you know, there's, there's platforms to jump in and grab an p i endpoint and actually, you know, close the reservation, which, you know, maybe 12 years ago you couldn't do all this stuff.
Seems like it's converging to make this kind of vision possible.
Carsten Tusk: Aside from being a more conversational interface and doing like one aspect that I think is good for search and we elaborated on that earlier, which is the consolidation, right? It really kind of like zeros in on your question and, and prevents you, um, removes the necessity of going through like source material yourself to find the solution, which has positive and negative aspect.
I'd like to ask the question, what do you think is enabled by it? That was not possible before? If we disregard the more human interface and the convenience kind of like that it offers.
Bill Constantine: So I think, uh, one answer to that is to take it a little bit away from our world versus highly technical and involved in ai.
Yeah. General, pretty fan general for the world. I would answer that question with, in how my children reacted to this technology and how my wife did. I think they were blown away by ability to, to bootstrap and, and generate creative content for them that they could then iterate on. So, good, good point.
It's is a matter of, it's generative content. You know, I think there's, part of it is that you and I say if we're asking it to do, say generate a program, well, we're very, very experienced in the, in the language that they're asking. So we can, we can be very sort of critical of that output. But for someone who's writing a paper for the first time, or maybe a business plan for the first time, or, you know, sort of doing something, uh, In an introductory way, even if they're an expert, they still were gonna use this tool to bootstrap their work and get a lot of stuff done.
I mean, I think it's just getting better and better.
Carsten Tusk: What do you think this technology enables virtual agents to do that it was not possible with virtual agents before? Obviously I
Deep Dhillon: think the conversational aspect is huge and I, I, it's so good at so many things. I
Andy Skalet: mean, the, we, we've all had the experience of going through a phone prompt, customer service thing where it's like, state your problem and then it's just like so canned and painful to go through that interaction.
You're not gonna do that if you want a restaurant recommendation and you're on your way out of the hotel. So having something that, that is conversational, that feels rewarding to speak to that. Feels like it has creative and customized recommendations. Even if it's not really creative, it's really collaborative filtering or something under there.
I think, I think that makes a big
Carsten Tusk: difference. It feels more human-like the interaction becomes more human-like, uh, more what we're used to. Less canned, less, less mechanical.
Deep Dhillon: I mean, I use it for everything like bill, uh, e everything. Even proposals that I send out, I mean it, before it would maybe take me three hours to write a proposal.
Now it takes me 15 minutes. I'd go back and forth. I prompt I get a bunch of stuff. I like edit it. Uh, obviously I don't send out something like that without editing it. That would be asinine. But like the, but it deals with a blank slate problem
Carsten Tusk: specifically for the chat bots. I agree with the more human human interaction part of it, and that possibly enables, uh, it to be applied to customer interactions where previously you didn't want to or shouldn't use a chat bot, right? There's one problem with it though that, uh, we can all agree on that one.
It's really good at talking about anything you ask it, but how do you actually correl it to like, be specific to a, a domain and have expert knowledge within that domain? This is the key I,
Bill Constantine: this is what's our biggest challenge. We all listen to a recent, um, interview Jeffrey Hinton, Hinton, you know, sort of this godfather of ai and he viewed these LLMs as idiots of VAs.
You know, they're gonna kill you in a. In the local pub trivia competition, but you know, really they don't have the ability to reason like we do as humans. We have a lot less knowledge actually, and a lot less training they do. And they have this vast, literally trained on almost the world's width of information at their fingertips.
So I think that we're gonna spend a lot of our time, you know, at Xyonix and with customers in the future in developing so-called custom models, where a base model of an l o M is used to, to train these customized models, to do very, very specific tasks very, very well. And maybe even do, you know, various aspects of things like, uh, better sentiment analysis, detector, better, DataTree, better, blah, blah, blah.
I mean, there's
Deep Dhillon: gonna be, but one of the things that, that I'm seeing Bill to riff on that a little bit, is it's both frightening and amazing, is like the stuff that people are doing with this AutoGPT idea. So they're giving it like really high level. Goals, like, get me a pizza. And then the system is sort of breaking it down and figuring out like, Hey, where are some pizza joint?
Like, you know, where are some pizza joints? Okay, how do I actually make a phone call or book something online? It's like a multi multiple sequence of asking, thinking, action, coordination, going out and committing an action. So you have to like, you have, and for, for
Andy Skalet: anybody who hasn't seen that it, it's set up as, at least the demo that I saw is set up as like a browser plugin.
So it has the ability to load websites, click on stuff, fill in form data, that sort of thing. So the interfaces. Creating these steps and then actually taking these actions on a user's machine without, uh, the user intervening within those steps, which brings up a lot of interesting questions and it's concerns and it's,
Deep Dhillon: and it's also leaning on ChatGPT four for the brain part, like for the thinking part.
And that's the part that's fascinating and frightening. That's simultaneously, well, that's, it's
Carsten Tusk: also the problem, right? ChatGPT is very good at, uh, content generation, and it has developed some sort of abilities of, well, maybe it's reasoning, it generates text. You still have to convert that text to specific actions you should take.
Right? And that, that's like one aspect that you need to do even for a chat bot or a concierge service because,
Deep Dhillon: well, that's the part that's fascinating is that, that it's doing that like the auto, the auto G P T process is formulating the. The prompts it's going through and formulating the list of actions
Carsten Tusk: because we doing it.
Yeah. Because it has rule engines that are working in there that, that trigger on certain things. Chat PT says that then execute these actions. Right. And so you have to come
Andy Skalet: up with, execute the action of clicking on a button or filling in a form or something like that. But exactly, we could get really good at writing those actions.
I, I don't think that's a big stumbling block. The, the
Carsten Tusk: real challenge is control. Yes. Agreed. Because not just do, you have to have, make the judgment call that what it does is meaningful and not wrong. You also have business constraints. You don't want the concierge at the Marriott to sell. You should really not run the room here, go to go to the Hilton, you know?
Um, so, so business constraints. And taking action are like the biggest challenges of interpreting the output of chat G p T. Um, and then the fact that of course it has a mind of its own. We, we don't control it. We actually do not know why it's reasoning the way it's reasoning. And we have very limited ways to actually control that.
And I think these are the biggest hurdles that we are still facing before we can make that reality that we are just talking about, that we seek some glimpses of technology making the reality. And in
Andy Skalet: terms of control, like the concierge at the Marriott is gonna have an internal control that kicks in if you, before they order you a thousand pizzas.
So they're gonna be like, Hmm, wait a minute. But auto, G P T, I'm not so sure, you know.
Deep Dhillon: Well, and I think that's part of the process of building a product, right? Like if bef, if you're building a product that relies on this ability to have more agent like behavior, Then, you know, you're not gonna start off with an unlimited a w s budget.
You know, you're not gonna start off like you're gonna put it in a box. Like we're, you know, we're able to put things in boxes and incrementally scale up our experimentation. Like, we're not gonna start off with the moon and, and des design like a stock trading system that goes back and forth for Chad g Bt Ask sits for some code, runs the code, makes the decision, and then just like trades on an, on, like an exponentially rapidly increasing budget.
Like, that's not gonna happen without human. But honestly, but I want you to
Bill Constantine: be honest, if you were to set aside couple hundred bucks and have it go just crazy, I would be totally up for that experiment. Like, I'm only, let's do it, it's called gambling. It's in entertainment. When I, when I looked at, for example, Microsoft has their office suite of tools and it's called I think, co-pilot where they've integrated basically, you know, G P T four along with these products.
And, you know, my mouth dropped open because this stuff that they're gonna be putting out is, is so amazing. Now, it's a little bit annoying to me that it really corrals you in this world. They really reinforce in their advertising of this pro of these products, how you as an individual have control, but you really have control in this corral that they've built for you.
And I have to say, people will use those things because they're very, very convenient. I personally wouldn't wanna write my daughter's graduation speech. Have it done automatically by GT b t or some chat bot. But you
Deep Dhillon: might do that because it's, are you kidding? I would immediately start there, like, I never write anything anymore without starting on.
I wanna, I wanna
Bill Constantine: make one more point in general is that now we've given access to tools that can be actionable. They can order pizzas. Yes. They can pull up graphics. They can do stuff for us. It's super dangerous, but if controlled and sort of bled out, it might be super powerful, super
Carsten Tusk: convenient, super powerful, super convenient.
Also, please unbiased, right?
Deep Dhillon: We already have the big players all have their own. Bard is already out. I, you know, I've started using it a lot more just cuz open AI is chat GTS often blocked. So at a minimum you're gonna have an oligarchy of these. But then you have all of the private renderings where, you know, where folks who have very secure private data but wanna build kind of concierge like services like this or chat services.
They on very sensitive data are gonna have to like, Fine tune and tailor those models. So I feel like we're going towards an ecosystem more than central only. And I agree it's getting a lot harder cuz those systems are so much more expensive to run now than they, than where we were two years ago. But I think it's unlikely that we're gonna have like one uni.
Uh, I think if there's zero chance we're gonna have one unified situation, it's likely that we might wind up with a defacto oligarchy of them, of the big
Carsten Tusk: tech companies. Well, but also it's kind of like, Sure there's more than one, but at the end of the day, great. The LLMs of the world controlled by the tech industry.
The tech industry has its own as a, as an entity, has its own opinions, et cetera, et cetera. They decide what their feet into those models and the content that is generated is extracted from that. So that completely controls what's comes out of it and how these models are thinking. Sure, there's more than one, but it still takes millions of dollars to build and run them.
It's, it's still a strong price. I
Deep Dhillon: agree that it's frightening, but is it that much more frightening than the fact that they already control how we talk to each other? I mean, the same, the same companies control exactly how we communicate with social media. Uh, and I, I guess it is, I mean, to your point, it's frightening cuz we almost like, you know, we've brought down democracies, like we have all kinds of problems with social media and I feel like we're just getting started with
Carsten Tusk: ai.
Yeah, look at it. I mean, for example, again, we're getting sidetracked, but for example, the last election we had to spend thousands and thousands of dollars to hire all those people in Eastern Europe to write these fake Facebook things. Right? Now we don't have to do that anymore. We have Chatt p d to generate all that fake political information for us.
And I don't think we'll face that in the virtual concierge business at all. Right? That, that's a more outlying, I just mentioned that because Bill said how great it is to generating content. And I think there's a danger in that if you suddenly let these models think for you instead of thinking yourself, but in these highly personalized virtual concert scenarios, all that goes away.
There's no problem anymore. We just love the ability to generate language and act on a more human level. Um, but here we have the problem of that control that we hate in the generic sense for, for content generation with, and it's bad because it's bias In this isolated worlds, we wanna be extremely biased.
We wanna be supervised because now these are little business applications. They're supposed to represent my business rules. Um, and so what we don't want in the big scenario, we need to have into the extreme in the small virtual concierge world,
Deep Dhillon: have a data, have a hypothesis on some high value insights that if extract automatically could transform the business.
Not sure how to proceed. Bounce your ideas off one of our data scientists with a free consult. Reach email@example.com. You'll talk to an expert, not a salesperson.
Let's talk about that a little bit. So let's talk about some of these chat bot platforms, you know, that you can use that have, not just conversational modifications, but the ability to like take your conversation out into different vectors like Slack and other places. They have the ability to like wrap actions up.
And so this is like tools like RAA and bot press and dialogue flow. And then you have, you have like a Google and, and Microsoft and, and others that have their home assistance and their paradigm of like how to build applications in this kind of chatty world. Like what are we seeing a, like how are we seeing these platforms?
Evolve in the last, you know, maybe year or so that's different from maybe 10 years ago when they got started and they were just bad at parsing, like multiple choice answers. Like they were, there was all kinds of problems with them before. Like h how are we seeing these platforms helping us realize this virtual concierge vision or not?
Carsten Tusk: I think we're involving all these different platforms, come out with their own or, uh, creative approaches of how to domain constrain these large language models. We see them try to basically fine tune them on corporate so that their answers are more alike. The conversation they would want them to have.
I mean, chat, G P T did that, right? It came out of G P T with human feedback. We see them putting like really hard constraints on there with the knowledge base behind it where we trying to corral them to only answer if the content is actually found in the knowledge base. And we have come up with really sophisticated tricks of like feeding the knowledge base, using semantic search back into the models and tell 'em to constrain it because the models are still one shot models.
They have no memory, they have no real history. Even if it makes you think that when you interact with chat, G B T, every single request to these models sends one block of context in. And based on that one limited block of context. It makes, generates its response. Right? I'm not sure where we're going from here.
I think these approaches will like become more sophisticated. People will come up with new ideas to do this. I think people will work on actually having memory for these models. Uh, so that's kinda like where we're currently at.
Deep Dhillon: Andy, what do you think, I know you spent some time with some of these platforms.
Andy Skalet: Well, I mean, I think, I think the bar is way higher for a lot of, you know, anyone who has a customer service interface that's a chat on their website or whatever, they're going to need to have a much higher quality of interaction in order to compete in that space. I do think the challenge of keeping them on topic, if you have a business that you know, you need to kind of keep them within the knowledge base that you have is a big one, um, as we've seen in some of our work.
But, you know, I think it's a space that's moving quickly and we're gonna see more and more agents in different aspects of our software interactions.
Deep Dhillon: So let's say you're, uh, changing the topic slightly. If, let's say, um, you know, you guys want to talk directly to the product managers out there who maybe have products that are doing all kinds of stuff, like totally different things that they, maybe they've never even thought about this virtual concierge concept.
Like what might you say to them in terms of how to think about their, whether or not these sorts of capabilities make sense in their product and how to, and if they do, like how to go about shaking a tree and figuring out how to take the first steps?
Andy Skalet: I would think about what interactions your customers have with the product today.
Do they essentially need training that you can't afford to supply from a human training force? If that's the case, because maybe you have a complex user interface in your product or whatever, then an agent like this could provide a relatively low cost lift to get training out to people. That's one, maybe one dimension to look at people interacting with your existing low tech chatbot and then quitting after a couple of interactions.
Are they achieving their goals? You know, improvements could be made there
Deep Dhillon: likely. I imagine there's hundreds or thousands of companies where executives are screaming at their teams who have these dumb chat notes on their website saying, Hey, I use chat g PT four and it's amazing, and your, our stuff stinks.
What help, you know, fix it. And uh, and then they're like, but we have private data, but we can't just send all of our stuff over to this external party, but we don't even know how these things work. Like, you know, there's like a, I think there's gonna be a, a ton of action there. Bill, what are your thoughts on this?
Bill Constantine: Yeah, speaking about this from sort of a technical perspective, I would think if I was a product manager and trying to implement one of these things, I would be very jealous of GTP four and chat G P T right now. First of all, as we all know, we don't know what the heck is going on behind the scenes. We have a sense of what.
Is going on. We know that chat two b t, for example, is comprised of multiple models and it combines the responses from multiple models to give you an appropriate response. Whenever it interacts with you, it's taken into account your history and that chat session, it's somehow being summarized, et cetera.
So, but it's working. What they're doing is working. It's got people involved, it's got people excited. And if I'm app, you know, if I'm a product manager sitting up there, it's like, man, I wanna have something like that from my company. But there are some barriers that we've already discussed. There's one of issue of trust, there's one of keeping these things on the rails and do I wanna pay giant corporation X, Y, or Z to be able to accomplish those things.
So there are some technical. Hurdles. I think, like for example, not to sell zonic, but we, there is this idea of doing customization and sometimes it's referred to as fine tuning. Well, you literally can go to God, you can go to GTP four and have it create exactly the type of interactions that you want. And you can use that data for training, so-called lower level models.
And those low, lower level models could be proprietary, uh, but they're much cheaper and they're much faster. Or they could even be something that's based on an open source model. So you can, you know, we're always gonna wrestle with this idea of containing these things. Yeah. But you can at least address.
This idea of data privacy, you know, you can, you can lock it down into your own company. You can talk about how are we gonna make this faster? How are we gonna make it cheaper? Because I really don't wanna, you know, and you don't have
Deep Dhillon: to send all your customer data to some third party if you don't want to.
Bill Constantine: party. Right, exactly. But I also think that speaks to the customization and personalization aspects of this. So I think there's just gonna be an explosion, uh, as we've talked about in terms like a concierge, you know, the, the types of conversations that people feel comfortable with and those, those types of conversations.
If they're not com, if they're comfortable, people are gonna come back. They're gonna wanna maybe interact with that chat body even, or talk to, you know, it's gonna. I think the more comfortable it gets, the better. Certainly more risky because we're opening ourselves up to worlds that, that we, you know, it's a level of trust, but I think that's gonna be a, a huge concern going forward.
Have you ever
Andy Skalet: used a chat interface in an app and enjoyed it? There's a real opportunity here. If you have a customized assistant within the Marriott app or whatever, the WIN app that is consistently helping you out, that honest, you know, it could be a differentiator for that. For that business or, you know, at some point it's prob maybe they all have a pretty good one, but the one that you've been using has your history and it makes that customer more sticky.
So I think it's worth investing in, in the concierge
Carsten Tusk: space. Just to wrap that up, uh, sure. I, I'm very data driven. So my first bit question would be, um, why do you want the virtual concierges, right? What problems do you think the virtual concierges could help you with? Uh, that kind of requires data, right?
So if you have the idea of putting a chat bot on your webpage measure demand first, would people actually use it? Um, and also learn a little bit about the, the interactions that people would have with it. What do you want it to answer? What problems do you think it can solve? And then I think it also depends on the ratio of what is the complexity of those interactions you anticipate people to have with your chat bot, right?
And if you have a high ratio of trivial things that you think a virtual consumers could automatically answer with, with security and safety, it's
Deep Dhillon: great. Let's fast forward five to 10 years out. Um, all your greatest fears or, uh, desires have been fulfilled with this virtual concierge concept. Everything you're looking at with respect to technological evolution.
Give us your best three to five sentence rendering of the future, uh, and and, and what it looks like with respect to, um, this concept of the virtual concierge. Uh, I'd like to
Bill Constantine: kick that off. I, I wake up in the morning. I'm going to brush my teeth. I talk to the mirror. Hey man, how's it going today? Hey, how are you doing?
Doing great, boy. I sure missed, my kids are at college. What can you load up? Uh, the, uh, the Constantine Kid chat bot? I want to have a, like a little conversation with them, with them get, uh, that experience and, oh gosh, I would really need to order some food today. What do you think? Yeah, that sounds good.
How about what would you like to order? Okay. Well, let's do that. By the way, bill, I've noticed that you haven't been running a lot lately. How about I schedule some time in your schedule? Did it go running? Okay. That would be great. Thank you, man. I really appreciate it. Okay, I gotta go now. I'm gonna go talk with my toaster.
I don't think that, I really do think that that world is actually coming, whether we like it or not. Yeah. And I also think it could be very
Andy Skalet: fun. Bill did a good job illustrating the kind of, some of the maybe good things. I'll probably take, uh, more, more of a negative view here because I think we would be remiss in this conversation not to bring up the replacement of human labor with the concierge.
Not only in a hotel concierge scenario, but potentially in lots of different spaces where you have customer service tasks that are kind of rote and very repeatable. The interaction, the complexity of interactions is not high. So we could see the chat bots replacing a lot of human jobs in this kind of space and.
Our interactions maybe are with these AI bots, which also maybe funnel a lot of the commerce to big business and are less likely to be connected to local businesses and things like that. Hopefully the laborer finds more rewarding jobs anyway, but I think that's a larger societal question that we'd save for another podcast.
Deep Dhillon: I, I feel like there's. A good potential future outcome, a moderate potential future outcome, and a downright horrific potential future outcome. I don't know exactly which ones are gonna play out, but I would say as far as what I think is definitely gonna happen, what Bill, what you're describing. That capability is going to be built by just about everyone under the sun.
Like we are gonna be having more conversational interactions with things that just do a lot more for us. And even if it's not conversational, that other parts of it are still gonna happen. I think it's gonna most likely be visual. You know, when I look at companies like Synthia and the rate at which they're growing and the costs being driven down, you know, people are gonna be able to take texty stuff, put it in, have this great avatar 3D rendered thing, talk back to you.
I think as much as we might love or hate that, it's gonna be both. We're gonna have these bots, we're gonna be interacting with them in a very sci-fi featurey sort of way. And I think the most common thing will be to swear at them at the same time. I mean, if you're, if you're an engineer behind the scenes at Amazon and or, or Google looking at Alexa or Google Home, I mean, that's.
People treat their bots terribly and that stuff's also gonna happen. But at the same time, I think we're gonna go from a world where you use your Alexa to check the weather and play music to a world where you're using those things to do a lot more, like quite a
Carsten Tusk: bit more. And I think, I think we have a visionary, it's Ionics.
Um, his name's David and he has a
Andy Skalet: flip
Deep Dhillon: phone. Now he sold it or he lost, he finally got shut off by the cell phone company. He had a two G phone. For our audience's sake, it's this little tiny,
Andy Skalet: oh, was 3g actually, they finally shut off the 3G that it was using. Oh, I thought it was
Deep Dhillon: two G. Oh, okay. Yeah. So they finally shut, they finally shut him off and then they, I think, didn't they send him a free phone?
They said like, just to get you off our
Andy Skalet: network, they gave him like, look, he couldn't refuse. They gave him a smartphone for like $50 or maybe it was free or something. Yeah,
Carsten Tusk: that's how they get Damn
Bill Constantine: it. Dam. It's my man. I mean, he can, he can, he can actually change, fix his own car. I imagine that. You don't have to look
Andy Skalet: through.
No, I think that's, he replaced his car's automatic transmission with a manual one. So he, he doesn't need a bot to, to do too many things, I guess.
Deep Dhillon: Alright everyone, thanks for coming on
Bill Constantine: By the way. I just wanna say this is not real Bill. This is virtual bill and you've all, you have all been fooled.
Deep Dhillon: Perhaps you're not sure whether AI can really transform your business.
Maybe you don't know what it means to inject AI into your business. Maybe you need some help actually building models. Check us out at xyonix.com. That's X-Y-O-N-I-X.com. Maybe we can help.
That's all for this episode. I'm Deep Dhillon, your host, saying check back soon for your next AI injection. In the meantime, if you need help injecting AI into your business, reach out to us@ xyonix.com. That's X-Y-O-N-I-X.com. Whether it's text, audio, video, or other business data, we help all kinds of organizations like yours automatically find an operationalized transformative insights.