Building AI Agents That Understand Humans with Marissa Chacko, Google

I feel like the term bots is probably a bit dated and should be retired. I feel like AI agents or conversational assistants is probably more apt for where we are. I see as like back then we were like at the cusp of the kind of conversational agent time period where we saw the rise of Alexa, Siri and Google Assistant. And they were the starting point for really this voice interactions. And I think a lot of that helped spur a new way to interact with computers. But we saw they were pretty brittle, right? You could only speak in a very specific syntax. And we saw users converge on things like playing music or setting alarms, but they didn't really do a wide range of topics. But what was interesting is you see the first sort of native assistant voice users, if you think about kids, the first technology they interact with is like asking their Alexa to play baby shark, right? And so you've got this whole generation of kids that are now being raised with a voice interaction. So I think that's a really interesting part of that, that time. And now we're in the next step. I feel like we're halfway to AI agents. I see we're in this place where we're having dialogues with AI.

This episode is brought to you by Boast AI. Each year, the US and Canadian governments give out billions of dollars in R and D tax credits and innovation incentives to fund businesses like yours. But the application process is cumbersome, prone to frustrating audits, and receiving the money can take up to 16 months. Boast AI gets you access to research and development tax credits and innovation funding opportunities without the headache and red tape. Join thousands of North American companies leveraging boast AI software to maximize cashback. Check out www.boast AI.

Welcome to another episode of our Traction podcast. This is the Next Wave. The Next Wave is a different take on our podcast, where we're bringing back some of our most favorite guests that have either graced our stage or joined our podcast 10 years ago. Five years. And we're going to revisit their topics to see do they still hold up in today's world of AI. Traction is brought to you by LaunchAcademy CA, one of Western Canada's main tech hubs and a international accelerator program that provides you access to the Canadian startup Visa as well as Boast AI. Your go to partner for R and D tax credits. Now I know that it's no way for you to build a unicorn. This week's guest was originally on stage attraction in Vancouver in 2016. Eight years later, we've got Marissa Chaco. Marissa is a product leader with over a decade of experience in the tech industry. Spanning a variety of different domains and product stages, Marissa has played pivotal roles in shaping innovation of products at companies like Google, Lyft and Foursquare. From spearheading the development of AI powered voice agents like Google Assistant to leading a massive recovery effort at Lyft during the pandemic, Marissa brings a wealth of experience and passion for solving complex problems. Outside of tech, she's a competitive equestrian, show jumper and a passionate ceramicist. Marissa, welcome back to Traction.

Great, thanks. It's wonderful to be here and a lot has changed.

Yeah, eight years feels like 20 with the pandemic in the middle. But we're here today to talk about bots. Your original conversation attraction was the bots are in 2016. Bots are the new app. So I want to quickly show a video and we'll get right into it.

And just to make sure we're all on the same page, when I talk about a bot, I'm really talking about any two way conversational dialogue between a bot and a piece of software or a software and a human being. Regardless of if there's a human back there, I'm really looking at software, human interfaces.

So bots eight years ago, hot topic today. How have they evolved over the last eight years? Can we even call them bots anymore?

Yeah, I feel like bots now is a dirty word. It's used to represent fake humans posing as humans, spreading disinformation. So I feel like the term bots is probably a bit dated and should be retired. I feel like AI agents or conversational assistants is probably more apt for where we are. I see as like back then we were like at the cusp of the kind of conversational agent time period where we saw the rise of Alexa, Siri, Google Assistant, and they were the starting point for really this voice interactions. And I think a lot of that helps spur a new way to interact with computers. But we saw they were pretty brittle, right? You could only speak in a very specific syntax. And we saw users converge on things like playing music or setting alarms, but they didn't really do a wide range of topics. But what was interesting is you see the first sort of native assistant voice users, if you think about kids, the first technology they interact with is like asking their Alexa to play baby shark, right? And so you've got this whole generation of kids that are now being raised with a voice interaction. So I think that's a really interesting part of that, that time. And now we're in the next step, I feel like we're halfway to AI agents. I see we're in this place where we're having dialogues with AI. They're not quite agentive yet in that they aren't able to do a lot of tasks yet. I think that agentive AI is where everyone's trying to get to, and that's really the AI that can take actions on your behalf. You can say, help me plan a trip, but it's not booking your tickets for you, it's not booking restaurant reservations, it's just giving you ideas. So I think we're not quite there yet at that next phase, but that's we're very close and that's where we're headed.

We're actually probably seeing mockups more than anything of what is eventually going to be reality for us with those AI agents. You mentioned a baby shark. I'm like, to the chagrin of many parents, the kids are trained on these bots. But these bots, can we call them AI agents? Are they interchangeable or are they totally separate things from what they were before to what they are today?

I think the backend technology has significantly changed, but I think that realistically it's part of the same continuum. It's really like a conversational way to interface with computers and get things done. And so I would consider them part of the same evolution. It's just the bot was at that stage a very clunky version of it, and now we're getting to a much more advanced version that has much better natural language understanding, much better perception and context recognition. So I feel like it's a very souped up version. But it was still based on the same initial human computer interaction.

So the intended purpose and objective and outcomes are the same. It's just the back end technology, whatever the technology is fueling it is 2.0, 3.0. It's so much further than Exactly.

Yeah, absolutely.

So on that note, let's jump to the next clip.

The second reason that I'm so bullish about bots is that bots can deliver a seamless, rich user experience that was previously fragmented and complex. Let's take a simple example. So I have a friend named Eve. And when Eve gets home from work and she's tired, she wants to put on some James Brown. So what does Eve do to listen to some James Brown? She opens up her phone, she navigates to Spotify, searches for James Brown, presses play. But she's not done yet because then she has to go to her settings, has to Tap on Bluetooth and then finally tap on her speakers to connect. Simple, right? That was seven steps. And that didn't even touch on unlocking the phone or navigating within Spotify. Now let's contrast that with what would happen if Eve hadn't Amazon Echo, she just say, Alexa, play James Brown on Spotify and that's it. And that's why bots are so powerful. They can take a complex task that previously required interacting with multiple devices or multiple apps and turn those into a simple voice command.

Previously, you just said that the AI agents are evolving, but they're not quite there yet. But can, from a developer's perspective, entrepreneur's perspective, can you really view an AI agent as a product? Can you build a company around an AI agent or what it can eventually be?

Yeah, absolutely.

I see different companies at different stages. If we look back, we started at more of a web native company, like the Googles and the Facebooks, they were web native. Then we got the Instagrams and WhatsApps and Ubers that were mobile native. And now I think we're getting into the AI natives from like a company perspective. But I think it could be at all places. It could be a company, it could be a product, or it could be a feature. I think if you're more of a web or a mobile native company, it's probably going to be a product or a feature because you have that legacy set of products that you also want to keep. I think if you're a developer now, you're probably looking at more of an AI native company that you're building, and the whole company is based on a foundational AI model that you're using, and everything is springing up out of that. And so I think that's like OpenAI Xanthropics, but as well as some of the companies that are building apps and services. On top of that, I think there's still a question of where the value is going to be and some concerns we can talk about later as to will ChatGPT or OpenAI consume a lot of those potential opportunities for developers. So there is still a question there, but I think the other thing I think about as we get into that agentive future is how the ecosystem is going to develop. Because one of the problems we saw with conversational assistants was that they were hooked up within the 1P ecosystem and that Google Assistant was hooked up with Google One P apps like YouTube Music servers that they have, but not really a lot with 3P. And we're seeing the same thing. Happening. As you look at Apple intelligence, all the things they preview are like, hey, you can do all these things with Apple apps or things that Apple can control on their phone, but they don't necessarily talk to Uber or Expedia or all these other apps. And I think that's the problem that still exists is how are these agents going to actually interact with some of the broader ecosystems. And those apps have developed a lot of, they've invested a lot in building that app, building the relationship with their customers, and they don't want to be disintermediated by a chatbot that's going to come in front of them and say, I'm going to treat you all like dumb services. If Marissa wants a car to the airport, I don't care if it's Lyft or Uber or Waymo, I'm just going to book that for her and she's never going to talk to that app. I think that's the problem that we're still seeing as how do we get the ecosystem to come along so that way apps can be, or these agents can be more agent that they can take action. They're going to need connections to things outside of those like 1P ecosystem. So we haven't seen that happen yet.

This kind of brings back that Alexa series type scenario where I as the owner of that Alexa unit or Siri wanted to act on my behalf. And so these same things with these AI agents, if I have an agent trained on me to help me less of my burden or make my family life better or make my work life better, I want it to be my representative. So I think some of these legacy companies need to accept that you're going to have to deal with my agent because it's putting me in 10 different places at once as opposed to eating up 12 hours of my day to do my tasks.

Yeah, exactly. And so it's how do they preserve that sort of relationship with you as they also then are interfacing with your agent instead of you directly? And how do they make sure that they maintain that preference if there is one? And I think that was something that was never really figured out with the conversational assistants. And that's why some companies integrated with them. But a lot of them didn't see much success and they fell away. The only ones you see that did well were like a Spotify, which makes sense that people are always playing music on their devices, but things like an Uber, it's. You're not ordering an Uber off of your HomePod that much. It Just doesn't make that much sense. And so a lot of those integrations fell away and wasted and developers didn't get much benefit from them. And so I think they need to. Developers need to see a benefit from them and not feel like I'm losing my preference with the customer and my relationship with the customer.

But there's also now, what do you call it, selfishness, but like a prioritization. I personally building these agents to represent myself and make my life easier. Companies are building their own agents to make their workflow easier and their operations more efficient. What's the scenario where you can have an agent talking to another agent and it's basically bots and bots, agents and agents talking and facility and transactions without either party really being part of it?

Yeah, I think it's a. Some people think it's dystopian and in some cases, maybe if all of our agents are in meetings instead of us. If your agent is hosting the podcast and my agent is talking on behalf, is that still valuable? But I do think there's a lot of use cases where it makes sense.

If, like dmv, yeah, nobody wants to go to the dmv, but yeah, the.

DMV has an agent and my agent knows all my information. Like, great, talk to the DMV on my behalf, fill out the forms, wait in line, whatever you need to do. I think there's probably a use case where it's like, my agent is talking to a business agent, like an expedia, like a DMV on my behalf. I think there's also a world where we have personalized agents where it's, hey, you and me are trying to schedule a time to do this podcast. Our agents can coordinate that, figure out the right time, or figure out, hey, I'm hosting a dinner for nine people tomorrow. What does everybody like to eat? What's their dietary preferences? As long as their agent knows their personal context and preferences and my agent can interact with them, like we can be taken out of that equation. And I think that's actually going to be great. So I look forward to that world. I don't think it's going to be that dystopian, but there's a little ways we need to get to, especially around personalization.

That's a perfect segue. Let's get to the next clip. So just for the audience, these are clips from 2016. Marissa has been a thought leader in this space for quite a while.

And while we've come a long way in making these interfaces more usable, they still have a steep learning curve. If any of you guys have switched from a PC to a Mac or iPhone to an Android, you've experienced this pain in trying to figure out where things are. And the reason being is that we have to translate our goals into what the system allows. But with bots, we can flip that paradigm. We can allow users to just tell us their goals directly. I don't know if you guys watch Seinfeld, but there's this example from when Kramer's pretending to be Mr. Moviefone and George is trying to type in the digits that correspond to the movie he wants. And in the end Kramer's just like, why don't you just tell me the name of the movie? And that's so much easier. Right? It's this ability to simplify those complex tasks and allow users to talk in terms of goals rather than navigation that gives bots a huge advantage over traditional apps.

I had to sneak that Seinfeld clip in there. Seinfeld is. It's a classic. It doesn't go. It doesn't go bad like that. It's usable in so many different scenarios. But that is a great point of. We went from bots taking full on direction to again, dystopian view of AI agents taking the initiative to do things on your behalf. Did they know what your needs are? They know what your dietary restrictions are. So if you're invited to a dinner, it knows right away to make sure that those requests are put in or send a note to the kitchen. Where do you think we are in that timeline? How far are we from that being reality?

Yeah, I think that there's like a couple of steps that we still need to see. One is starting to talk about like personalization. We're starting to see memory being incorporated into a lot of the large agents like ChatGPT. Starting to store your memories, I think both in terms of a structured and unstructured way. But using memory is pretty hard. Back at Foursquare, we were thinking a lot about ephemeral versus permanent. Right. You may say, I'm a vegetarian and I'm always a vegetarian. But then you might also ask for, hey, I want kid friendly options. And that might just be because there's kids coming with your friends and that's like a one time thing and you don't want it to store that and say, always give kid friendly options. And so it can be hard to know when to use these memories and how to use those. So I think that's something that people are still working through and there are certain frameworks, but it hasn't really been used that well yet. So I think once we get the memory down, then we start to get the personalization better of how do we use that memory in context to make better decisions. And we start moving from just predicting and suggesting to actually taking action and being proactive. We'll move into a world where we say, hey, here's some suggestions for dinner. Would you like me to put these on your instacart list of things to buy soon? Hopefully once you've approved that flow a couple of times and feel confident in what it's suggesting, eventually you'll just be like, hey, just order all my dinners for the next week. And it just arrived and I didn't even see what my agent ordered for me, but I trusted it and we built up this rapport and it's personalized so well that I now give it, I can delegate it to do all of my grocery shopping for me. And I think that's where we're going. But you have to first like the, the personalization, the memory and the human in the loop, like approving until you get to that proactive step. But we're not that far off. I think it's still like that memory and preferences piece that it's a little bit tricky to actually nail down. It's always of a technical thing and more of a human problem because it's always, it's not even easy for us.

To understand that overcoming that stigmatism or that barrier of this is now something reality. I, I just watched a podcast with Mark Cuban and he was mentioning how there is a Costco being built in the center of like an apartment complex. And so the refrigerators in the apartment complex would actually back up themselves to the Costco, get stocked up and then go back into your house. And so that is like your household agent speaks to Costco, say, hey, we're low on milk, I need to get a refill. Okay, back yourself up at 2pm and we'll stock you up. So we're getting there. This is being built right now and so these agents are getting there. But it is that stigmatism or can you picture a boomer accepting this when their Gen Z kids are doing it on a day to day basis?

Yeah, and it'll be a learning curve. I think with everything with sort of adoption of AI, we're seeing people's preferences and openness changing rapidly. I think as they're seeing the utility of this. Even Roy looked at this eight years ago, people were pretty wary about having Mics like always on in their homes. And now they're like becoming more open to that and having cameras always on their home. Especially if these models are running locally and their data isn't being shared in the cloud. They're like, you give me a lot of good value for it. People are becoming more open to that. And you've seen Amazon try things like, hey, let me open your front door and leave your packages inside for you. Which is feels a little bit crazy, but it's also will eventually get there, I think. And so you've got to keep trying these things that push the boundaries and if people find value out of it, they're willing to give up what feels like potentially a little bit of privacy invasion there.

Yeah. The perfect example is for anybody that travels, the big no was, hey, why am I giving my biometrics to these governments and foreign countries where I'm walking into? I don't want them to have my facial ID and stuff on record, on file. Now it's, oh, I can just whiz through customs. Like I go to uae, I've gone to Japan full on facial recognition. I'm in and out of the gate in less than five minutes and that's worth my privacy. It's like, hey, it can get me through this 200 person lineup in five minutes, no problem.

Yeah, exactly. So I think like we have to have that right value exchange. But it feels like people are just becoming more accepting of it. And I know some people worry about that, that like we're becoming too accepting of this exchange. But I feel like that there continues to be good like push and pull. There's companies that are much more privacy leaning and some are less. But I think it's good to continue to have that debate.

Yeah, there's always going to be nefarious activities and the more technology gets evolved, it's supposed to be user friendly, easier and easier to use. But with those conveniences come challenges of protecting your data, protecting information and seeing how those agents or even data sets are leveraged by other third parties. This is actually a great transition. I love this. We're transitioning to all the different video clips that we have here. We'll go to the next one.

Let's talk about anthropomorphism. It basically means giving human characteristics to animals, events, objects or software. And we do it as humans because it allows us to understand complex things. And in the realm of technology. We've already been treating computers like humans for a long time. They have a memory, they catch viruses, they go to sleep, they speak a language. And with bots, we're really taking this human computer relationship one step further.

So talking about being comfortable with technology, one of the things that's coming rearing its head is AI agents or aur avatars. These are actual replications of human beings or fake human beings. There's been a big trend around AI girlfriends and boyfriends. We've got deep fakes going on left and right, which are very concerning. But we're also getting into like robotics and these animatronic robots that look exactly like human beings and it's hard to tell them apart. Now you see them as customer service agents at hotels at the front of the desk. What are your concerns around this? Is this coming scary? Is it? Hey, this is again just the next level of evolution for this type of technology. What are your thoughts here?

I think the scary part for me is that the technology moves faster than our human brains are evolving and our ability to analyze and criticize this. I worry about my parents, for example. I had to tell them, hey, if you get a call from me saying I've been kidnapped and asking for money, ask them a question that only I would know because I'm worried about this sort of like deepfakes spoofing my voice and using it to have these types of attempts on people. So that's, that is concerning. But I do feel like it also is opening up this way for us to again connect emotionally. While I didn't think deep fakes would become so big so quickly because I didn't think it would get this good, I do feel like this ability for us to connect with software as if it was human and knowing that it's not human, we still choose to connect emotionally with that. That was something that we saw back then and is still true now and even more. And so I think my optimistic side says, you know, this will give a lot of people a way to connect and potentially deal with like difficult topics through like therapy AI or feel supported when they're going through something and they don't have anyone to talk to. And I think those types of relationships with AI and that sort of like anthropomorphic use is actually good because there's a lot of people who need support and can't necessarily get it all the time, or having coaches or things like that, I think that's great. And my hope is that will also cause us to value human to human interaction more and that will seek out these IRL real world experiences because we'll know that we need both of Them I think my pessimistic side or more cynical side would say like, maybe this will make us increasingly isolated. And I'll be like, I don't have to talk to humans at all. Just talk to my AI friends and they tell me I'm wonderful and I could just stay in my basement and I'll rely on AI solely for connection support. I hope we don't go there. And I feel like some of the younger generations are finding a little bit more balance where they're pushing back on how much technology is involved in their lives. And I think trying to find a bit more balance. But I'm guessing that we'll probably overshoot this, go overboard and spending way too much time with AI agents and then have to balance back and realize that actually there's nothing like being in person with humans and we need both of those to be successful. So I do think there's a place for it, but we just have to be really wary about it and use it in those cases where people really need that support and connection. But a human isn't always available.

Yeah. And it has been documented that for introverts, even people with traumatic experiences, these are tools that can help ease them along to that path of being more ingrained in public. But as you said there, it does run the course of I don't need to go and make friends because I've got my AI friends. Yeah, very helped me along. So hopefully the companies that are building this technology will work it into the algorithms to wean these people off of it eventually or get them to have real world experiences that enhance the overall experience for them.

Yeah, yeah. Well, we don't want to see is like the her version where you become increasingly extended relationship to the point where you shut out everyone and you're in love with your agent and yeah, that's what we need to prevent that dystopian world. But. But yeah, I say things to an AI therapist that I wouldn't say to my normal therapist. And it's just things where I'm like embarrassed about it and it's nice to have something where you're like, okay, another human doesn't have to hear this and I can try talking about it and destigmatize it even in a safe setting like that. So I find that really helpful.

Yeah, I don't want to be just googling WebMD. I want to avatar of a pharmacist or a physician helping me diagnose myself, self diagnose. Not that I advocate that and, but also not that I'm an expert on this. But the AI girlfriends, AI boyfriends, that whole trend, it's happened for a reason. And younger generations today are getting more introverted and it's getting harder and harder to meet people. The whole online dating world, it's a whole other beast. And hopefully these tools can help push people back into that IRL and realize how in person is so much better. Real people are so much better.

Yeah, it's interesting too, the trend of like schools taking away phones. And I feel like that might be also helpful. And okay, there's a time and place where you can spend time with your AI boyfriend and girlfriend or friends, but then we're going to take away your phone and you're forced to socially interact. And so there may be some of that as we have this transition stage. So make sure we balance it when. When it's easy to fall on too hard on the AI side.

Yeah, I think it, like you said, it is a pendulum. It's going to swing far and then it's going to have to try to make its way back. But hopefully there's enough social pressure or social validation that this is the right way to go.

Yeah, yeah.

Getting to the next topic, I want some advice here. So let's watch this clip.

How do you build an awesome bot? What should you be thinking about? So my first tip is fail gracefully. I love this video from DARPA Robot fails. If any of you guys have seen those just like you provide feedback mechanisms in your app when users do something they shouldn't, you should be providing the same feedback mechanisms when you're bottom does something it shouldn't be doing. Users often mistakenly think bots can do a lot more than they can because speak like humans. Because humans write the dialogue for bots, but bots can't understand that way.

I remember watching those DARPA fails. They were hilarious at the time, but now it's, oh, sorry, I didn't mean to laugh at you, my AI overlords. Those are crazy good robots, optimists from Tesla. They're doing tasks that are just eight years ago. There's nowhere, no way anybody would assume that they'd be able to accomplish those things today. But just like bots, the AI agents today are doing things that nobody thought bots were able to do. So what are some of the challenges that you've seen companies face with things that they need to overcome? Like where are some of the warning signs or pitfalls that you would advise people to be wary about as they're developing their own AI agents?

Yeah. Something that we're thinking a lot about for Google Assistant is there's some like very deterministic things that Assistant needs to do, like turn on the lights or lock the doors or things like that. And those things work pretty well with a very specific syntax. But if you open those up to an LLM, there's a chance that it's going to hallucinate and it's going to say, oh yeah, I set your alarm for 6am and then the one, and if it didn't, but it said it did, that becomes a big problem. You just lost a lot of trust with users. So like the hallucination risk I think is fairly well known and being tackled. I think that's going to be a little bit more of a problem. If you're using open source models, you might have to do more of that work yourself. But then there's, okay, we can go towards, you can tune down the amount of hallucination you want, but in doing so you'll reduce the amount of creativity. And so there's all these trade offs that you have to think about. If you want it to be more deterministic, even in a, even though it's not really deterministic, but you want it to be more accurate and be less likely to make up things, then you're going to trade off creativity. And for some people who have very narrow use cases, that might be okay. But if you wanted to be able to come up with 20 fun podcast names and turn on your lights for one agent to do both, those are on different sides of that like hallucination spectrum. So I think that's a hard thing for developers to deal with. I think the other challenge that we did see with conversational agents is this blank page problem. And you've seen Gemini OpenAI. Here's some sample prompts, here's sample follow up prompts. Because people open it and they look at it and they're like, maybe it's blank page problem.

I don't know, what do I do next? Yeah, yeah, how do I start?

Yeah, because you could put anything in there. And so they're having to like prompt people and give them ideas. And this continues to be difficult when you get to a voice only modality. Right. Because you just like stare at a speaker. I could ask you anything, but I have no idea where to start. And so we had to keep trying to train users to like understand how to use that technology. And I think that challenge still exists. I haven't seen it done very well for any voice agents. Alexa Does a lot of. By the way, did you know you could do this? But it's really hard to discover that way. So my hope is we'll teach people in a text based modality where you have the ability to use discovery and UI to help people along and then they'll translate that naturally over when we go to voice only modality, when we don't have that UI to help guide people. But that's what I see happening. There have been a lot of failures and I think it's hard for people to back out of. Like, I know for a while I was trying to create images that had numbers in them or like a specific piece of text. And God bless. Those models could not keep the text in the same. Like the word the same. It said traction, it kept misspelling it. It would just say T R A C. It would just keep making up things and it just couldn't. I would put quotes around it, be like, spell it exactly like this. Those things are getting better. But it's hard when you're a user because you're like, am I doing something wrong? Is there? And those areas are going to be hard for developers because it's a lot of long tail problems and it's not like the same problem happening again and again for people. And it's not always clear why the model is failing and how to get people out of it because these models are still largely a black box. So we don't always know how to help the user out of it or if there even is a solve for that.

And then I just got a vision of what happens in our personal lives, but also in our work lives. We're not going to have just one AI agent. So going back to the discussion of AI agent, talk to another AI agent. You're going to have these loops that could be death. Loops of them just going back and forth, having hallucinations on each other.

Yeah.

Saying, oh my. My owner went to dinner with friends and family with some kids and the other agent saying, here's a playlist with Baby Shark 20 repeated 20 times and your owner's gonna like it. No, they don't. And so that's the other oblig obstacle I see coming down the pipeline is now, okay, how do you make them function in real world scenarios where they're not operating in isolation?

Yeah. And you could see there could be failures at any point within those agents. And you could say, okay, I want to book a trip to Amsterdam. My agent talks to the Expedia agent. Somehow I get a ticket going to Rome and for the wrong day. And I'm like, who do I blame? Do I blame my agent? Do I blame Expedia's agent? Like, how will we debug those problems? Because again, it's like a lot of black box being passed through. And so how do we feel confident in what each agent is doing? And a company like OpenAI obviously has a lot more resources, a lot of researchers that are like really good at building those AI models and Expedia and no shade on them, but they don't have that depth of research and AI expertise. And so you're probably going to see them moving at very different speeds and potentially not being in the same place that a user may be in terms of their confidence of working with a ChatGPT as they would be with like an experience Expedia AI. So I think that's something that may also be problematic if those things are at very different speeds. Of course, if Expedia uses like the same underlying models as OpenAI, then maybe that bypasses that problem and things stay on the same level. But for these companies that are integrating with these models, it'll be increasingly hard to keep up with every model change, especially if the models start forking as we're seeing with the reasoning models forking from like kind of the mainstream models. And so that might make it difficult too for users to have the same expectations across agents in terms of capabilities.

Yeah, let's skip to the next clip.

And my final tip is obvious, but I've seen folks make this mistake already. Don't rebuild the wheel, leverage your existing strengths. For example, don't build your own bot platform. They already exist and they have hundreds of millions of users. So take advantage of the growth opportunities there where your existing and prospective users are already spending tons of time.

So we just talked about the challenges and obstacles and pitfalls, but that's on a, and this on a positive note, where are some of the opportunities? What are you seeing here? You've had a front row seat for over almost a decade here of what's happening with these automations and where do you see the opportunities moving forward?

Yeah, I think just like then there are these big companies that are building a lot of the foundational work and like use that, use open source models, don't reinvent those wheels mostly because most companies aren't gonna be able to afford to do that. Right. Where I see a lot of value is in like private data and existing customer relationships. So if you're already a company that has data about how users Interact what their preferences are, leverage that private data to like really train a very customized model that's gonna be great for your specific use case. And I think that's gonna beat out like a generalized model that doesn't have access to that data, like A, an OpenAI scanning the Internet anthropic is not going to have detailed information about customer travel booking preferences, for example, and the same way an Expedia does. And so Expedia's travel booking AI agent should be way better because it has a lot of that data and it has an existing relationship with the customers that it can leverage. So I think for any existing company I would be really looking at like, how do I build a custom model using my private data that will really surpass what anyone can do just with publicly available information. And then if I was a new company who didn't have an existing product, doesn't have existing data, then I would be looking at like, how do I take advantage of being AI native and look at these incumbents who are holding onto their existing moats and say I can move a lot faster, I don't have an existing moat to protect, I can undercut what they're going to do. So I would look at taking advantage of those existing models and tech that's being built and saying let me move faster and undercut because I don't have that baggage. And I think that's where you're seeing a lot of people going after search, for example, because Google has to protect its search moat and has to protect its ad revenue. And they are trying to evolve, but you see a lot of people saying that's a great way that we can like undercut them. And so it'll be interesting to see if it ends up working out. Obviously I previously worked at Google and so I have some vested interest in search continuing to work there. But, but I do think it opens up a lot of opportunity for these new AI native companies to move a lot faster and deliver product way quicker, get feedback and iterate in a matter of weeks, vers, months or years that we're seeing before. So the time to market is really shortening.

And I think it's like a whole case to be made for people that think that they don't really have a role in AI in the sense that no, you have data, you have information that can be extremely valuable, whether that's a corporation with insurance data or something about earthquake regions or whatnot. But my own company, one of my portfolio companies, is working in the sports space and we're Trying to help athletes monetize their data. Their athletes generate so much data, whether it's from their wearables, from their on field performance to their financial data, their health data, their travel data, they've got a wealth of information just floating around them that is so unorganized. Can you put that into a data set that can then be leveraged for their own purposes, but also take their health data and make it accessible to universities that are doing research on concussions or whatnot or fitness recovery. And one of these doctors that we were talking to was saying you can take data from athletes on their personal recoveries and apply it to seniors and help them recover from a surgery or an Achilles injury or whatnot. Because who has healed the fastest, who has had the best resiliency as athletes? And so how do you extrapolate that for seniors? So there's so many different ways to leverage your data in this new world of AI and AI agents?

Yeah, absolutely. And I think that's exciting too from a sort of government or global health perspective because there has always been this long tail of data that we've never had like the power to really analyze. There's some small percentage of population where like drug X, if you have conditions Y, Z, A, B and D interacts really poorly. But like we never had a big enough data set and the ability to analyze catch those cases. And I think that's where there is going to be a lot of power if we find ways to combine these data sets too from like a global benefit perspective.

And I saw another study, I don't have these exact numbers. I'm just going to split some percentages out. But they're relatively close to where they were. But they had doctors with access to AI and doctors without access to AI to diagnose conditions. And I believe the Doctors without the AI were about 60% accurate. The doctors with AI were 70% accurate. So it's 10% more, it's not that much. But the AI on its own was 97% accurate. So what that indicated was that the doctors that actually had access to AI didn't want to listen to the AI. They felt they were, they knew more on their own without the AI. And then the numbers just proved. So the AI was accurate.

Yeah, that's incredible. You think about. I think there's a study a while ago where I think Google did it. They were looking at like retinal scans or eye scans and the AI was able to detect cardiovascular issues from eye scans. And humans have never realized that there was any connection because we weren't looking for any features of, like, cardiovascular problems in the eyes. Yeah, maybe bastardizing the study a little bit, but basically it was like, for.

Our audience, we are by no means saying that we're experts on these health studies, but we're just anecdotally sharing what we've seen or heard.

Exactly. But it was cool because AI was uncovering features in eyes that we never knew existed that were predictive of certain conditions. And so you think about that in so many different areas where you're like, what other data could predict earthquakes or climate events or illness that we just never looked at it that way. But AI is able to analyze things at a very different level, and I think that's, yeah, just hugely impactful. So I'm really excited to see what other kind of strange connections there are that we never notice, but that exist in the world to help us better understand and potentially predict and prevent some of these problems.

Yeah, it's going to be a fascinating world that we're going to see ourselves move into in the next little while. Let's end this off with something that you ended your talk with in 2016.

I'd like to end with one of my favorite quotes by Nicholas Negroponte, who is one of the founders of the MIT Media Lab. And he said, it's not computer literacy we should be working on, but human literacy. Computers have to become human literate.

Computers have to become human literate. Are we there yet?

I think we're really close. I was just seeing the other day, there's now OpenAI released their real Time API that's doing speech to speech. Right. You could always say that these LLMs are really good at understanding text. Now they're going to be able to understand sort of the emotion, the context of our voices. And soon you'll imagine that they're also going to have video footage so they can understand our facial expressions. And so that's starting to get the totality of how we communicate. Right. It's visual, it's auditory, and it's in the meaning of our words. And so we're so close to that that the AI will be able to understand the full context of what we're saying. And I use the word understand, it's not actually understanding, but in a way that it's able to use that context to give me what I want to understand my intent. I think we're very close, and I think that's amazing. And I think I'm really bullish about that future because it just continues to cut down the barriers and continues to cut down the frustration. If my AI knows that I'm frustrated or my AI knows that I'm happy, it can address me in a different way. And that's really, I think that's just going to be helpful for making sure, like what we get out of those agents is really aligned to what we want as humans and less to this is what programmers have said exactly the computer should do for us. It's much more personalized and it's much more aligned with our behavior. So I'm pretty bullish about that future.

Yeah, technology is supposed to make our lives easier and our lives happier, so hopefully these AI agents are going to get us there faster. Marisa, it's been an amazing topic. Love the conversation. I got a feeling we can come back in a year and we'll have even more progress in the last eight years than we had since your last talk at Attraction. Thank you for joining us and hopefully we'll hear more from you in the future.

Yeah, thanks so much for having me. This was great and yes, very excited to see what happens in the next month to a year. It's going to be a crazy ride.

Thank you for tuning into this episode of the Next Wave podcast by Traction. If you enjoyed our show, please subscribe to our YouTube channel at YouTube.com artuptraction and give us a five star review on Apple Podcast and Spotify. This episode is brought to you by Launch Academy. Launch Company is Western Canada's main tech hub with over 6,000 entrepreneurs being incubated from over 100 different countries. Our companies have raised over 2.5 billion in funding and they come from a variety of different backgrounds from blockchain, AI, cybersecurity, quantum computing, agtech, fintech, health tech, you name it. If you're interested in learning more, check us out at LaunchAcademy CA so that I can see what's working best. Not trying to be a hypocrite, no work done, some silly focused on.