How Hume AI Is Teaching Voice Systems to Understand More Than Words

23:30The inverse might be true where you've solved a very great use case for a customer service agent talking about billing for one particular product. But the minute you start to introduce a new product or something that you have another family member tied to your plan, right, or you start to ask another question, things now start to get confused and it doesn't hold up, right? And so that's really where we've got a lot of work to do in the industry and where we're very well positioned working with many of the large model providers out there on this and the agent providers to ensure that as they start to look at these use cases that they want to optimize for, they fundamentally don't have to expect users or customers or enterprises to have to figure out what the tradeoffs are.

24:12I often joke, right, if penicillin was the right drug for you to be told, but they said, Grant, we want you to take penicillin, you're just like, that's weird, right? Like, I might take it, but, like, I'm not going to feel comfortable coming back to this, like, automated thing from my doctor. And you're pretty educated and, you know, pretty forward-leaning with that, right? That's actually a real story that we had with a real customer. And even they were kind of laughing. They'll be named nameless. They're just like, you're not going to listen to that, right? And you won't see that in text because it's flat.

24:45And you won't be able to measure that in a transcript because it's probably spelled the right way. And there is no word error rate assigned with, like, you should take penicillin. Make sure it's not on an empty stomach. It's like, ah, cool. It's the right number of words and it said the right thing. But that's not how it performed.

25:02Corey Noles:It reminds me of the problem with generalized models where you push the capabilities really far in one way. We've kind of seen this with the Frontier lately. Like we push really far in coding, but then all of a sudden they are really bad at writing and like, you know, Opus 5. Something over here is less good. Yeah. Yeah. Opus 5 would famously tell you like this is most insane stuff like, oh, like this load. This is the load bearing feature over and over. And you're just like, I can't read what you're saying because it's so like jargony. It makes me wonder, is this bullish for, you know, everyone making specialized voice models and actually just like leaning into, hey, maybe it is better to specialize as opposed to try to make one.

25:39Corey Noles:perfect general model that's incredibly difficult. Like a medical care model, like a legal model. Yeah. Or even down to your company, like a Cori company model. Or to use the medical example, this hospital's model versus that hospital's model. What do you think, Andrew? Yeah, I think these are great ideas, and I fundamentally agree with you all. I think obviously the large hyperscalers are pouring billions of dollars to make sure that they get this right at massive scale across all of those languages and provide the tooling and the infrastructure for people to be able to go build on top of that and not have to find other things.

26:18And they're doing a very nice job at that. Don't compete with scale is what you're saying. Well, but the flip side, right, is if you think about it, right, in all of the vertically enabled AI solutions that are out there that primarily took back office legacy human workflows and put those into LLMs just to make those things easier, you're seeing the same thing happen with respect to voice. And you're seeing that happen on a language by language and country by country basis. I mean, one of the most fun things that I have is waking up in the morning and going to my Slack channel to see who wrote into our website, because invariably it is someone in a country that you cannot believe that they are building something so specialized and so powerful and so impactful for their particular use case that it really is just such an exciting time, right, to be alive.

27:07And you are seeing a lot of this, whether it is in Africa or South Africa or in the Far East or in Australia or, you know, in the Netherlands or even Canadian access. I mean, you're seeing it literally all over the place and it's really exciting. And I think in the end, right, there's not going to be one thing that rules at all and there's going to be a multitude of options. So I guess to your point, there is an opportunity just like there is in these vertical AI solutions with those workflows. There's opportunities where you have an advantage by understanding how humans should perform with respect to voice in a particular function.

27:43Providing the nuance that scale can't, I guess. Absolutely. Absolutely. Absolutely. And look, at the end of the day, access to the network, compute, storage, data, and underlying math is getting commoditized, right? And so it's now about the abstraction level above that and how do you apply your intellectual capital and your horsepower and your domain-specific knowledge against that to drive true competitive differentiation. And we believe this harness and this infrastructure we've built is going to allow people to have this private eval infrastructure and effectively compound all of the institutional knowledge and capital that they've got across all of this with respect to voice to deliver those outcomes in an outsized manner.

28:31And that's where we believe it's really fun because, you know, if you just are publicly releasing all of the data for benchmarks, people are just going to go benchmarks and get models to climb the leaderboard. and you're just going to go rely on those and it's not really accurate. So private evals is really where this is all at, which also supports the folks that have this institutional knowledge are going to be able to create very impactful models on a global scale by language, by use case, by vertical. Wow. So let's say company A calls you and says, we're getting into voice or we're in voice and we don't like where we are.

29:07what does that look like when when you come into the picture yeah well the first question is why don't you like it sometimes they know sometimes they don't know uh sometimes there's a cold start problem where they just can't get a voice to even sound real before they even kind of get to how does it perform in real life so we help with the cold start problem given our heritage in these things. We also help them identify, assess, and understand where the model is or is not performing with respect to their peers, with respect to their use cases, and with respect to how they're fundamentally expecting it to perform.

29:45We're able to run that through our system and deliver back to them a very comprehensive analysis and set of analytics that says, hey, look, these are the types of things that you need to really understand. Here is how you would evaluate them. Here's how you would fix them. And most importantly, here is what you would do to understand regressions and how this performs continually time over time, because it's not enough to go through all that and get it into production. It has to stay that way consistently. And you have to be able to measure and account for that. Right. And so we work with people to design that on a holistic basis so that they can ensure it stays that way.

30:25Because then what happens? You get a new customer, you get a new use case, you get new feedback, right? You realize that that was great, but now people are asking for other things because you got it to be so good. So at some point, you can be a victim of your own success where it's like, I nailed that. Corey and Grant can chat for five minutes and this thing sounds literally like it was on their podcast, right? But now you start to go to the next topic, the next thing, the next language, and all of a sudden, right, I don't want to pick on one of you, But one of you starts to veer off and you're like, oh, wow, I didn't account for the fact that when I tried to go do that.

30:58So now how do I go back and see where I was at and start that process all over again to take sort of the one plus one, right, equals three approach? Yeah.

31:07Corey Noles:That is interesting how like your frame of reference always changes. This is like a classic thing in AI where we get used to a certain level of capability and we think, wow, this is amazing. And then we, you know, we get bored of it in a month because we're trying to push it to the next capability frontier. And I've heard that, you know, a lot of software engineers in particular are very susceptible to this because, you know, when you don't have a specific like, okay, this is the end date. this is the feature we're shipping and then we're moving on. You just get in this loop where you're like, well, I can make it even better and I can make it even better.

31:42Corey Noles:And you have to be really careful with that in the age of AI when everything is so easy. How do you handle that? How does your team handle that at Hume and how do you keep control of like what's most important and don't get like featureitis and go off on all these different tangents? Featureitis. I like that. Featureitis. It's a bad disease. Yeah, it's a bad disease. Look, we're in a really fortunate position because we've built this infrastructure for ourselves over many years. And we've delivered models against this. I mean, we're no longer competing or delivering those models on a consumer basis.

32:19So our infrastructure is very hardened across hundreds of millions of hours at large scale. And we've been doing this for the last year with large model builders and many others. And so for us, it's really about repeatability, scalability, and consistency, right? And I think if you just go back to the workflows that we talked about earlier and making sure that those workflows are best represented into vertical AI solutions so that they work for the end users there, the same thing is true with respect to voice, but it's harder because it is multidimensional. And so how do I make those workflows very easy to repeat consistently?

32:54And so a couple of good examples is you've got large model providers that are doing checkpoints on their model every couple of weeks. Right. And every couple of weeks, they want to run the regressions and make sure that the regressions didn't change with respect to where they were last time. So they didn't make one change and break another. At the same time, they rolled out five, 10, 20, 30 new voices, new use cases, new languages. Right. Any combination in there. Right. And so you really just need to provide a consistent, reliable and scalable framework for them to go do that. And that's what we perfected because we had that for ourselves.

33:24And now we're just opening it up to effectively be the Switzerland of voice AI so that everyone could use this and provide this type of research infrastructure for folks so that they don't have to go build it themselves.

33:35Corey Noles:Yeah. You just got very, very clarity of focus from so many years of doing it. That's cool. Yeah. Yeah. It was a lot of fun. How far do you feel like voice AI has come and how much farther do you feel like there still is at the horizon to go? it's come a long way but we're still in the early days there's no question about it right for all of the reasons that i had mentioned and i think primarily right you have to look at this on a global scale and you have to look at as i mentioned earlier the different dialects the different vocal bursts the different mannerisms and the way in which everyone interacts by the way last time i checked let's just take you know what everyone calls english or the united states of America or where you have English speaking countries.

34:21We have plenty of folks that live here that like, you know, weren't born here and speak many other languages, right? Or married significant others. And so they grew their family up, you know, speaking in a different language and different culture, even though they live in Texas. Right. And so you have to account for all of those things that were very early, right. And voice performing right at that level with all of these, you know, multi-dimensional factors, right? That even seemingly in the English language, you would think are easy. But when you sort of start to double click on all of this, it's a very hard problem.

34:53And there are people investing a lot of money and a lot of time and a lot of effort to go fix that. And so we have a long way to go, which is really exciting because at the end of that, you're going to see a better customer experience for all of us. You're going to see new paradigms exist. You're going to see, you know, the intersection between humans and robots and wearables actually manifest in a way in which it's a core and central part of our life. And you're going to see voice really, truly become that primary interface as it starts to really understand us and work with us as we expect it to.

35:24And we got a lot of work to go get there. So it's early days and early innings, but the promise is absolutely there and people are working really hard to get there. So, you know, over the coming years, you're going to continually be surprised month over month as we all are with everything that we see in AI. Yeah. Are other languages as far along as English or farther? Or I just think when I think of the planet and then the hundreds of languages or thousands even maybe that it encompasses, I'm curious, are we just lucky to be in the U.S. and at the forefront of this? Or are there other languages that are doing that are performing really strong as well already, too?

35:59Yeah, I mean, I think, look, I think the large majority of CapEx that's been deployed into AI systems holistically, like U.S. has the lion's share of that. So, you know, we certainly are doing a fantastic job there and are certainly ahead. I think when you start to think about other languages, like obviously if you go to India, right, and you think about the number of dialects, right, it's actually exponentially more complex, right? And then you have dialects speaking to other dialects, and then you have dialects and you have villages even inside of the dialects. So you really just have this complex set of dimensions in many other countries.

36:36And that's not to say India is not doing a good job here. They actually are. And they're investing heavily, and there's a lot of innovation happening there. My point there is there is a set of complexity across the world that many are dealing with and many are trying to tackle. And sort of holistically, there's just a long way to go. But U.S. obviously is doing a fantastic job here. But voice is still a multidimensional problem, as we had discussed, and you got to be able to measure it appropriately and solve for it in the right ways.

37:02Corey Noles:What do you think we're the most behind in voice? Like, is it long running duration? Is it the dialect problem you just talked about? Like, where would you want to see the most improvement in the near term? Like, where's the weakness, kind of? Yeah, yeah, from your perspective. Yeah, it's in the depth of the conversation, right? So, again, you know, one, two, three, five, seven turns, like, okay, fine. But like, you know, how many times have you talked to a customer service agent and looked up at your phone? You're like, oh, wow, I was just on with I don't want to pick on an industry for 32 minutes.

37:32Right. And whether they were helpful or they weren't helpful, it was a 32 minute phone call. Right. You know, we've got to look up your information.

37:39Corey Noles:It just goes on forever. You're on hold for 10 minutes. Read through a guide. Correct. Exactly. So you've got duplex models with tool calling and reasoning, right, is kind of where that technology stack is going to address those problems. And that's where people are working and investing really heavily right now. And we definitely have a long way to go because of a lot of the ways in which these systems fundamentally just have complexities. But that's what people are going to be solving. And I think it's going to be fantastic once you start to see that come to the forefront. And a lot of people are working on that in a very deep way right now.

38:14And so when you start to look at things like that, there's a lot of work to do, but it's going to be very powerful when it comes back. Absolutely. Absolutely. I've always felt like, not specifically, well, yeah, call center industry and places like that definitely have a lot of motivation to see a technology like this reach point where it is reliable and performable. Because like you said, there's this whole element of just calling customer service. It's inconvenient. You're on hold. You're waiting. They're literally looking through a script to figure out where your issue lines up. To be able to make that a five-minute issue would sure be a better experience, I would think.

39:00There is. And, Corey, I've done the research. there are tens of billions of hours of audio in that industry locked up that are sitting in storage bins, S3 or otherwise, that folks are fundamentally not unlocking human emotions, human understanding, and all the metadata associated with that as the real ground truth to go build their voice agent experience on top of. Have they not seen the billions of dollars that they could make if they licensed that data?

39:32Corey Noles:Why are they not making licensing deals? Like, what's wrong? Then there's that. Then there's that for sure. Like to the extent that they actually have the consent to go do that, of course. Yeah, sure, sure. Yeah, if they don't have the consent, yeah. But even then, right, the inverse is also true. Many are deploying voice agent solutions, right, without trying to mimic the human behavior and understanding that they have from being in business for 30 or 40 or 50 years and all the recordings that they have. And that just is unfair to us as consumers and as loyal customers of a lot of these companies.

40:07And that's what we believe is a massive opportunity to help in that space. And we're the only folks in the world that do that. And we're really excited with a lot of the work that we're doing in that space right now. It's really cool. You guys really saw an opportunity and took a risk. I mean, the fact of the matter is you had great models. as when we were testing them the thing that stood out to me the most about the human models was that i felt like like the human element was the most there of a lot of other ones i was listening to so so i uh i had i admire the the guts it takes to shift so fundamentally and and go this direction but it makes all of the sense in the world yeah well thank you uh we got a lot of work to do but we, to your earlier point, saw the fact that we truly believe voice will be the primary interface inside of AI and that in order for that to happen, it has to align with humans and that we just so happen to have a research infrastructure built that scale to go serve that and we thought that would be a much larger market opportunity particularly at the time in which the models were going to get commoditized over time.

41:17It meant more people were going to build more systems than we had discussed and that we were probably better situated in the infrastructure business than the raw model business. And so far it's played out really well and we're excited about the future. Yeah, because in the model business, you're very much more competition every day, a race to the bottom in price. There's a lot of reasons to shift. It makes perfect sense.

41:41Corey Noles:So you have this voice replication leaderboard, right? Is that correct? Yeah. I think that's really interesting. So a model there can sound extremely natural without actually sounding much like the person it was supposed to clone. So my question is, is the industry optimizing too much for like, oh, wow, that sounds real, instead of measuring whether these systems are actually doing the job to be done correctly. So does it actually sound like the person? Yeah. You hit the nail on the head, Grant, and fundamentally what we have built is this scenario builder, which is, hey, look, when you want to evaluate a model, you shouldn't worry and optimize for what's measurable, which is what everyone does today, i.e.

42:23the transcript. You should optimize for what's most important. And to your point, Grant, right, it's not most important that you just sound real, right? Like, do you sound real in this context? Can you do that across a multitude of questions? Can you do this in a multitude of environments of acoustic quality? Can you do this across languages? And you should be able to just define those scenarios. And we have an orchestration layer that says, got it. Thank you for defining that scenario. We are now going to programmatically give you the plan by which you can most effectively and efficiently get that measured, get the analytics back and understand where that was at.

43:01And that is exactly one example, Grant, and one leaderboard. And there's a multitude more that we will continue to publish and continue to publish more evals. And then all of these become composable templates and workflows in our library for people to be able to go run and be able to go measure and choose what models to bench against and choose what scenarios and what languages so that over time this becomes very easy and you're never going to be guessing how you're going to perform in a real-life production scenario. Makes sense.

43:29Corey Noles:AI safety, right, has been a big buzzword the past two weeks for all sorts of reasons, which we won't go into. Boy, it hasn't. Unless you want to, Andrew. I'll leave it up to you if you want to get into that. But I do wonder, do you have any benchmarks around voice AI safety? Like how do you think about that? Because like Voight-Navis is a whole can of worms, right? It does. It does. I'm not taking the bait, Grant. Good try. I'm not taking the bait. I know. I know. It might be good for ratings, depending on what I said. I don't know. Or distribution. But I'm not going to do that. Look, I think safety is really important, right?

44:05And that's fundamentally why you have to measure this against real-world scenarios with humans, right? And back to my silly, very simple example, I'm fine, is actually a safety hazard, right? If a healthcare voice agent thinks you're fine and doesn't help you and you actually really needed something and you actually were scared, you just didn't know, right? That's a big problem, right? And that could be a human safety issue, right, just on one end of the spectrum. And so we're trying to work with all of the right guardrails in place to be able to measure those effectively. We're actually also doing a lot of research right now into the safety aspect and partnering with a number of the world-leading organizations on this.

44:47So stay tuned for some more research to be published from us in that respect because we think it's very important and we'll continue to invest heavily in that area. Because now that you sound natural, now that you can do the right things, now that you can do it at the right time in the right place, we need to make sure, right, that you are who you said you are and you're performing in the way in which you are. And the right things aren't linking and all of those things. So that's why we're very excited about the position that we're in from an infrastructure layer and continually to add value for folks over time.

45:16Corey Noles:I was wondering about this. Is there like a synth ID equivalent for a voice? Like if someone was to impersonate, I don't know, to pick a person, Joe Biden, right, on the phone and say like, hey, I'm calling my bank. Can you wire me$5 ,000? This is Joe Biden. Is there any way for the industry to have any automated tools to say, oh, that's not him. Like we know that that's built into the model. Yeah, it's very challenging, but there are a lot of folks working on that. and it's an area of massive investment in the startup community and it's a really hard problem. A lot of really smart people working on it and I do expect many people to solve that at large scale and it's going to be a fantastic win for the voice community and I think you'll continually see a lot of advancement come out of that.

46:04So absolutely, yes, it's well on the path. I think many people will tell you they've already solved it and many might have. I'm not saying they haven't, right? But you're going to continue to see additional innovation there. Nice. Awesome. Yeah. Nice. Something I'd like to touch back on is we talked a little bit about voice as an interface. And that makes me wonder, do you see voice as a model in the traditional sense or as more of an orchestrator of other models, like maybe a large model, maybe specialty models or various tools? Like, I keep thinking of this as voice, as a gateway to real agentic work pretty quickly.

46:48It's all the above, Corey. And I think let's take the call center example, right, where you would have a duplex model with a whole bunch of reasoning, right, which basically means there's two of us having a large set of conversations. And action has to happen. So I have to go look at the billing system. The billing system gives me one thing. Unfortunately, the systems in a lot of these companies aren't architected perfectly. so they might have to go to a different billing system for the second piece because you came as a customer through an acquisition right uh or that product is in another system and then you might actually need to understand where the field technician was which is in your you know uh maintenance repair and operation system which might be a different vendor and you might be migrating all of those things right so now if you just think about that and i'm doing this from voice right i'm having to have the right set of conversations in the right way now you have a bunch of tools that are being called, right?

47:40And there's going to be reasoning and chain of thought reasoning to go through a multitude of those where you're going to have LLMs in the middle of all that. You're going to have perhaps other agents that are going to do other things, stitch all that together to deliver it back so that they could basically just tell you that a technician will be there tomorrow to go fix your issue. And they're going to credit you for your service that was down yesterday. Right? And that was a lot, but you just want that answer. Like, when will my cable be back up? Which I know is not life or death, but like, When you start to think about it and that and blow that example out, right, to your point, Corey, you're going to have a multitude of complexities that go into these systems at production scale to be able to account for this.

48:18And absolutely, you're going to have large models. You're going to have LLMs in the middle of that. You're going to have the internet in the middle of that, right? You're going to have buffering. You're going to have all of these other factors that go into quality and packet loss and background noise and speaker identification. And when I stop talking, am I actually done or am I confused or is this like a healthy pause? Right. Am I thinking? Right. Correct. Correct. Correct. And again, it's because this is multidimensional. It's a lot easier in text, which I'm not suggesting there aren't really hard problems in text.

48:51There are. Right. As evidenced by all the folks that are doing great work in coding and the legal tech startups that are doing a fantastic job and all of these other things. But like when you start to look at that domain specificity, right, that folks are solving in text and you start to apply that to voice, it's just, you know, a lot more variables are in play there, right?

49:09Corey Noles:When I think of voice as an interface, there's like multiple layers of interaction, too. Like just if you think about it in terms of let's say I had a perfect system. At some level, that perfect system, it would be me talking to my computer and I don't have to open another application. It's like my computer just has some sort of voice model built in. But then every other application that I use would also have some sort of voice interface. So how am I now navigating the voice on my computer, the voice, like to your point, like Google Chrome or whatever I'm using to access the internet, and then the voice model on the website that I'm actually going to?

49:47And your phone, your earbuds, your hardware as well. Right.

49:50Corey Noles:It's actually like a Russian nesting doll of complexity, the more that you zoom out on it. It is. And say what you want about the algorithms that target you on certain social media platforms that display weirdly relevant, right, sets of information for you to take actions that you weren't thinking about. But at the same time, right, there is a level of personalization that we do enjoy, right? Whether it's ad targeting, whether it's searching, whether it's context that remembers our last conversation and where we picked off, whether it are the documents that we uploaded to have the context for how my business runs, right?

50:28And all of those things. And so the summary here is that folks want the voice AI systems to be very well aligned to that same level of personalization. And again, I just keep coming back to everything that we spoke about and how difficult that is in voice and how we have to have the right level of systems and infrastructure to support the testing, evaluating and improving of those systems. And that's where we're focused because personalization has to be at the core and the forefront of how voice AI systems work because that's what we're used to in the rest of our life. And that's where other systems are advancing in dramatic fashion.

51:00Let's say everything you're working on is right and you're hitting on all cylinders. Where do you feel like the next two or three years take voice tech? Again, the primary interface in voice AI where that just becomes true, right? Where a large portion of what we do is with voice and we feel comfortable, we feel confident, it feels natural. And we feel like we actually have the personalization to deliver us the speed and the convenience and the level of accuracy by which we've come to enjoy text-based LLMs for a variety of things that we do today, whether it's planning a trip, whether it's running a simple web search, whether it's trying to find a document, whether it's writing an email or a multitude of ways in which we use all that today.

51:44You would have the same way and the same thing happening with voice. Look, you saw some of the large coding agents deliver voice mode and coding agents and a multitude of other things. And so that is absolutely where you're going to see it. And I spoke a little bit earlier about these duplex models with reasoning, and that's where you're going to start to see these things deliver far more complex, personalized interactions for you that are going to become so seamless. It's going to be that aha moment, like the first time you went to chat GPT and you're like, oh, my God, this is incredible. Right.

52:19You know, you're going to have those same type of feelings across, you know, a large multitude of voice systems very shortly. Man, and in a coding agent, it's something special. The exact same thing.

52:30Corey Noles:Yeah. Like I told Corey, like this when GPT-6 came out, it was the first time I really gave voice mode in Codex like a legit shot. And just the fact that you could be like, yeah, can you go add this feature for me? and then you're, you know, let's look this up on the web. And then, oh, yeah, can we start up this new Greenfield project that we're going to make from scratch? Like, it's just one chat the whole time. I'm sitting on my couch talking to my computer over here while I'm watching television. Yeah. And you've got it managing multiple different tasks at once. Like, I ran into a RAM problem.

53:04I had a RAM issue because I didn't have enough machine that I was trying to do for the number of projects. And but but what a what a way to deal with that. And it's so good at like, hey, check on all my projects. How are these going? And just coming back with a detailed report from all of those. And it's like, man, this is people don't know what they don't know right now. But that's the power. I code every single day. I'm not a coder. I've grown up more of the business operations. But yet I can't call a customer service desk. Right. Of my cable company. Right. And find out right when service will be restored.

53:42Right. Using the voice agent and any level of reliability. Right. But yet I'm rather sophisticated in how I think about these AI systems. Right. Yet I'm very unsophisticated when it becomes decoding. But I could just type and actually get a website spun up right in a matter of minutes. It's a big disconnect. What's the barrier there?

54:02Corey Noles:Why haven't more companies adopted this, do you think? Is it because they don't have the benchmarks like you've been doing? Is it that it's just they don't know they can do it? Where do you think is the barrier? Well, the benchmarks just allow you to keep score, but it's underneath of the benchmarks that is the research and the process, the optimization and the workflows that have been codified into a platform. And this decade worth of research at very large scale that allows all of this to happen. And again, fundamentally, right, you know, voice is multidimensional, as we had discussed, and text is flat.

54:38And so it just is far more complicated. There's a lot more factors that have to come into play. And that's why we believe an infrastructure designed to help improve voice AI at global scale is something that really was necessary in the world. And, Corey, to your point, that's why we made the pivot and are exclusively focused there and really excited about the future. Love it.

54:59Corey Noles:Andrew. One last question. Oh, go ahead. One last question. It's a fun one. We can keep it short. Do you think that – so, like, you know, we talked about talking to voice interfaces a lot here. On a personal level, do you think that at some point we're going to have to change the social contract a little bit to be more comfortable with people talking to themselves? How do you feel about this personally? Or do you want to know this? You know, Grant, it is the social contract. I hope you didn't trademark that because I might start using it. That's a great one. Go for it. No, I didn't come up with that.

55:31Yeah, look, our social contracts actually will have to change a little bit, right? But the job of the AI systems and the ones that will win are the ones that don't force us to do that. And that's the whole point of being natural and consistent and reliable and acting and aligning like we are is you should just be able to talk freely and not have to change your patterns in order for the AI systems to work. Like, sure, I can optimize and write a better prompt and follow the right skills and tell it to think like a McKinsey consultant and all of those things that you do where you don't get better results to go do that.

56:03But in voice, you should just be able to speak very naturally and very freely. But I think the social contract that will exist is because you're actually now going to be able to do far more things in a multitude of ways. And you might have to start to think a little bit differently, just like we had to think differently with how all of these things work in the LLM world. Right. And so I think that's where the real promise is going to come into play, because, look, you know, when you go to Thanksgiving, right, in six or eight weeks or whatever it is, And I'm pretty sure people are going to have, you know, their aunts or uncles or grandparents that are there are going to use the word chat.

56:38Right. Oh, I was using chat the other day and X, Y, Z. It's how I made this apple pie. Right. I mean, just think about that for a second. I have an 80 year old that just use chat to update like their pie recipe that's been in the family for generations. Right. You know, like that's wild. Like that's a new social contract. Right. of basically how I would go interact and kind of make something that I've made for 30 or 40 years. Right. It's kind of wild. Right. And so fundamentally, you will see the same thing in voice. It's a hell of a chef, by the way. I'll be honest. It is. It is. It is. I use it a lot in the kitchen and I am constantly amazed with with how well it knows for me to reheat a steak or something.

57:22Yeah. The problem is execution. Execution is tough. At least if you're challenged like I am in the kitchen. Yeah, I think we can all relate to that. I definitely relate to that. Andrew, thank you so much for joining us, man. It's been a lot of fun. Yeah, it's been great. Thanks for having me, Corey. Thanks for having me, Grant. It was a lot of fun. And keep up the great work. Huge fan of all the work you guys do. Thank you. Where can people go to keep up with Hume and what you all are working on and learning along the way? Hume.ai. Feel free to reach out anytime or feel free to reach out direct to me on LinkedIn.

57:56Very active there and happy to chat with anyone. Excellent. Well, I'd like to give one more quick shout out to our sponsors at SAS for making today's video possible. If you like today's video, please take just a moment to like and subscribe. And don't forget to check out our daily newsletter read by more than 700 ,000 people just like you over at the Neuron.ai. And that's all we have for today, everyone. So farewell for now, humans. Meow.