ボイスAIエージェントを信頼できるものにするには何が必要か

All right, Rowan, thank you so much for joining me. How are you? Doing well, and thanks so much for having me. No problem, no problem. Appreciate you joining us. Whereabouts are you now at the minute? I'm in our San Francisco office right now. Nice. SF. I've never been to San Francisco, but I hear good things. I hear similar... I don't know, I think it's Seattle actually, they reckon it's similar to the UK, isn't it, weather-wise? i would i haven't been to the uk too often so i'd have but i have lived in seattle um so i'd have to defer to you i think seattle was like it was very beautiful but it like rained a lot and like sometimes it was like gloomy so i don't know if that's that's kind of the weather you experienced sounds absolutely exactly like uh okay yeah yeah i'm sporting a nice little for those of you that are watching on the video i'm sporting a nice little smash on the front of my head here uh Play fighting with the kids. Oh, wow, okay. One of them happened to just run headfirst straight into my forehead with his teeth.

Oh, wow. I obviously came out a lot worse than he did. The kid's okay? Oh, he's fine, yeah. I mean, if that had been the other way around, there'd have been child services involved. But, I mean, there's no support like that for the poor old parent that just gets beaten up by the children, you know. But it's nice. Anyway, appreciate you joining me. came to chat around Voice AI testing a lot of interest in Voice AI this year seems to be having a kind of resurgence you know we had the Alexa boom in 2014 to 2018 or so that really kind of like put voice on the radar for a lot of people Covid came contact centre Voice AI started to kind of you know take off obviously it was all NLU sort of based at that point Gen AI sort of came around we've had you know initially it was all chat there were lots of murmurs about voice but I think you know we didn't really see a great deal of activity until 2026 feels as though now contact centers have woken up, businesses have woken up, voice is back on the agenda, and everyone's kind of going crazy for voice AI.

Is that kind of what you're taking from this year so far, or are you seeing anything differently? Yeah, no, I completely agree. I feel like voice AI has just been increasing over the last year and a half in terms of adoption, interests, enterprises beginning to deploy them, But I think also with all the personal assistants, I feel like voice is just kind of up-leveling. I think I saw Apple release a new version of Siri, which I haven't tried out yet, but I've heard they've been making progress there. There's obviously Muse, Instinct, Dots, all these components have voice kind of built into them. So I feel like we're seeing enterprise increase. Now we're also seeing consumer, I feel like, adoption and increase as well. So overall, just a really good time. Yeah, I mean, yeah, OpenAI obviously have been investing heavily on that side of things. Yeah. With the GPT Live 1 model and stuff like that. It was just on NatWest deploying kind of voice AI through OpenAI in the last kind of couple of weeks.

So, yeah, it feels like a good place. So in terms of BlueJ, in terms of like what you guys do, are you focused in any particular area? Is it kind of contact centers you care more about? Is it any voice user interfaces? Like, tell us where your sort of focus is and what it is that you do for those that don't know. Yeah, absolutely. I can start with maybe like a quick intro. And so, yeah, obviously, my name is Ron, one of the co-founders and CEO here at BlueJ. At a high level, we help companies test, monitor, and improve their voice and actually chat AI agents, so kind of conversational AI as a whole, working with brands like Google, DoorDash, cutting-edge startups as well. And primarily what we're focused on is kind of proving reliability in a lot of these systems. I think a lot of voice agents, a lot of chat agents that are being deployed are non-deterministic at their fundamental root level. And so the problem that we're trying to focus on at BlueJay is how do we enable these systems to be a lot more trustworthy?

That way, when they're deployed in production, they can be reliable. So we do this through what we call these really cool digital humans, so pre-production simulations. You can almost think of them as an army of fake customers that can call in and stress test your voice agent and your chatbot. We have production live monitoring, so once you go into production constantly monitoring the AI, seeing if it's going off policy, doing things like evals, automated issue detection, all of that good stuff. and then also trying to use that data to then fix the issues that we're surfacing. This ends up causing us to work with, as you mentioned, contact centers, a lot of enterprise AI teams, and so on. Interesting. And what are the kind of benefits of that kind of simulated testing on the voice channel specifically? yeah it's kind of funny

that's squarely how BlueJay started and the reason we built that first was honestly a personal struggle of myself and Faraz who's my co-founder we were actually building voice agents ourselves for automating customer interviews and also for the restaurant industry so we were exploring these two different use cases and I think the benefits of it primarily come from the problem that it's trying to solve which is like, dude, it's a pain in the ass to test these things, right? Like, I remember I was picking up my phone, and if you were like a shadow in my room listening to the types of calls that I was making, you'd probably be laughing because it was like not only random scenarios, but at some point I remember I was trying to like change my accent, I was like trying to speak a different language, I speak Spanish very poorly, so I was trying to speak the best Spanish I could, test it in Spanish, and it was just quite difficult to test, mainly due to, like, you know, obviously just being able to test a bunch of scenarios at once

in an AI agent is fundamentally difficult. But then voice was also really interesting because the language issue, the accent issue, the background noise issue, like there's all these different variants in voice, audio dropouts, kind of packet loss that kind of just compound on top of each other. And because it's real time and it's not like asynchronous like chat, that makes it even more difficult and complex. It's interesting because I kind of hear a lot about, you know, voice AI and, you know, you've got a few different components essentially, haven't you? So you've got like the speech recognition, you've got the synthetic voices, you've got the functionality itself, so like the actual agent behind the scenes. I know that GPT Live and some of these speech-to-speech models are sort of changing that typical kind of cascaded architecture, but still typically you've got a few different components to test. And what I've always found an interesting kind of question is, at what point do you test which components?

so for example you know speech recognition the model itself speech synthesis um the telephone line itself in terms of you know uh whenever you're hosting the actual phone call so what bits would you say in your experience are important to test and when should you do that testing yeah I think this is a really interesting question I can kind of talk about a customer that we work with and how they approach it you don't need to voice festivals can become quite expensive because if you're simulating a full voice call that one takes more money it also takes more time if I'm doing a full voice to voice call it's similar to how I have a three minute conversation so it takes more time it takes more money And to your point, you are starting to test different components of the stack.

If you take a purely test-cated architecture, you'd be testing not only the LLN responses, but you're also testing the SVT and the text-to-speech. Is it natural? Is it picking up on medical words, for example, in the healthcare industry and so on? The way we've actually built BluJet and what we also recommend to some of our customers is actually to start with text-to-text testing, where you're able to still test your voice agent and the logical underpinning of the voice agent really fast. Because text-to-text is super fast. It's also a lot cheaper. And we think that gets you around 60% to 70% of the way there because you're able to then validate the core functional pieces of your agent. Like, hey, is it calling the correct tools that I want it to at the right time? Is it responding with accurate answers Is it hallucinating Is it redundant Like you able to get a lot of those I would say like quirks and like functionality portions out of the way

And then after that, we still recommend, right, if you're building a voice agent, you need to test the system more similarly to how a customer is going to interact with it, right? At the end of the day, a customer is not going to call or like test your voice agent, right? They're going to call it and they're going to talk to it back and forth. And so that kind of then involves the speech-to-text layer, the text-to-speech layer, once again, in the cascaded system. And I think that is also equivalently important and kind of gets you to that 100%, hey, this thing is fully tested. I'm confident kind of pushing it into the product. It's interesting. So what about latency? How and where would you recommend testing for latency, if at all? yeah latency so yeah 100 i think like a lot of people are uh you you may have often heard this like a lot of people are on a quest to make their voice agents sound more natural right they're like hey i want more naturalness right and um oftentimes like people come to me and they're like i just really want to make it natural how does we take great naturalness for example and actually the

first thing i always say is like hey before we even get into like i would say some of the fancier ways to grade naturalness, there's actually probably two very important things that you just need to make sure work, which is one, as you mentioned, a low-latency voice agent, and two, a voice agent that's not interrupting at infract times. If you aren't able to have those two things nailed down, you can have a very poor conversational experience that, to the phrase natural, doesn't sound natural. In terms of testing latency, you can test, you can do a few things. You can test each part of the stack and what the latency of the SCT is, the OOM is, the TTS, right? And you can do that through traces, right? What you can also do is you can test what we call perceived latency, right? And we think this is actually a really good measure of what your customer actually experiences. So let's say I simulate a voice-to-voice call, right? And I have a Blue Jay, we call them digital humans, calling your voice agent. What you can actually test is like, okay, the moment my digital human stops talking or the fake customer on the line, how many seconds does it take for your agent to start talking?

So like the first bite of audio that comes out of your agent. And we can actually measure those gaps, right? And we can say, hey, that's by the way your perceived latency. That's like the latency that your customer actually feels on a call. And we think that latency is really important to understand and minimize, if that makes sense. And you actually probably want to be minimizing not only the average latency, but also the P95 and P99. Because if you're able to reduce the worst, like the tail end worst portions of your latency, presumably your overall experience across your entire customer base is becoming better. That's interesting. Have you noticed anything in particular that affects latency? and is it typically consistent or sporadic? I think there's a bunch of things that can affect latency. I assume you mean in terms of making a voice agent slower, for example, at responding?

Yeah, if I'm talking to a voice AI agent and I feel as though it's going a bit slower, or sometimes if it's going so slow I'm going to start talking again before it's had a chance to speak to me. or like, yeah, what are the kind of like, in your experience, what are the contributing factors to latency? And when you're doing this type of testing, is it that, for example, certain elements of latency are there in every single conversation or actually different conversations produce slightly different results, if that makes sense? Yeah, absolutely. I think it's really dependent on how someone's obviously built their agent, right, in terms of like how it responds, like the level, like the model that they're using under the hood, how much syncing they're doing at every single turn, how complex the actions are, for example. So if my agent has to go call a tool, and that tool is, I don't know, taking a few seconds to execute, and I'm making it such that the response is contingent upon the tool's results, obviously you're going to see huge levels of latency there.

In terms of mitigating that, I think there's really interesting things you can do. For example, you can put in, you may have heard this, like, give me a moment, or kind of like trailer phrases, so to speak, that allow you to sort of like artificially reduce your latency, even if you're kind of thinking about something in the background that's going to take a little bit longer. I think you mentioned like GPT Live 1 with like, I think they have something pretty novel, like a front-end, back-end architecture, which makes this process really cool and really easy to improve on, if that makes sense. But I would say, yeah, it really depends on how you've architected your agent, Is it a single prompt agent? Is it a multi-agent? Is it waiting on certain tools to execute? How big is the prompt? All that stuff kind of factors in. It's interesting. I spoke to Andre from a company called Seamly, fairly recently on the podcast, and they have kind of like a, not quite middleware, that's probably doing it a disservice,

but it's essentially like the voice AI component that if you're a text-based agent platform, you could use Seamly and essentially give it a voice and put it in the phone channel and stuff like that. And what he said was really interesting, which is that the latency that you will find from your ASR, your speech recognition, is typically fairly low. And the latency that you will find on the other end, which is your STT, your synthetic voice generation, that's fairly low. And the biggest contributor of latency, apart from obvious areas that you mentioned there in terms of like, if I'm making an API call, that's going to produce latency. But the thing that he said produces the most latency is the platform itself. And so I wonder whether if you've got any experience with that in terms of like, you know, is it the platform that causes the latency? Is it the API calls that cause most of the latency? Is it other kind of ancillary components? I suppose what I'm trying to get to is like, is there what levers can you pull as a business to combat latency?

I think that one is a good one, which is interject. You know, the GPT Live one's kind of like multi-channel type of audio capabilities are good. So it can, you know, it can do those little disfluencies and interjections and stuff like that while it's working. Is that kind of where we are in terms of, you know, we're going to just have to deal with that type of latency. and those are the techniques that we use to deal with it, or do you think there's other levers that we can pull to affect it? No, I think, I actually think we're going to do a really good job on reducing latency. I'm actually not, I think this will actually become like a, I'm pretty confident it will become a solved problem, if that makes sense. To the story that you mentioned, or the example that you're giving, yeah, I do agree. I think a lot of latency can be in that middle portion at the LLM layer. A lot of latency can also come from the orchestration layer. There's also, by the way, potential of just like, hey, in your telephony network, if you're calling across different regions, for example, that can potentially add to it as well.

A sort of network latency too, right? So there's all these different places it can come from. So in terms of, I guess, what we recommend when a customer or a partner may come to us and say, hey, I have all these latency issues. We actually wrote a really good guide at BlueJ, which was 12 ways to reduce voice agent latency. We'll go through all of them. I've already mentioned, and I can send that guide over here in our chat, but I've already mentioned, I think, one of them, which was this idea of thinking phrases that you can say that way the customer on the other line is feeling that it's a very fast experience. But another thing that you can do and something that we thought was really interesting was you can actually pull a lot of data at the start of the call. So let's say I call in and typically like, let's say I call in and I'm to a healthcare business. The healthcare business may be able to pull some data

in the background as they authenticate the user. That way, further into the call, when a long request comes up, they're actually not making the tool call then, if that makes sense. They've essentially preloaded the context that matters in the conversation. That way they can then immediately pull that context and give it back, right? So you've already done sort of the hard work in the background when the user's actually not noticing. It actually relates very strongly to, I'm sure you're familiar with Instagram. Everyone's using Instagram, right? It relates to something that Instagram used to do with their posts right Like a long long time ago in software where actually the time to post a set of images took surprisingly a long time right And the time to create a post took a long time back then. Right now obviously it's instant. But what they did was they started basically just recreating the post the moment you clicked start a post, right? And so as you were filling in the details of the images and the name of the post and stuff,

they were actually just uploading stuff That way, the moment you collect posts, it would be almost instantaneous. And so you can kind of use that similar framework or methodology within a voice agent or your agent responses too. That's really good. That's very interesting. I'll put the link to that 12 ways to reduce latency in the show notes. That's really good. So the latency one is an interesting one. The other one that you kind of hear a lot around is the kind of speech recognition. Now, there's some things that with the newer models, the speech-to-speech models, there's not a great deal you can do there because they are how they are. And, you know, probably word error rate-wise probably perform better than some of the more kind of traditional ASR models. But that's typically what you see the kind of next kind of bunch of errors outside of the agent itself is kind of like speech recognition issues. and it's interesting to understand at what point do you think

it is worth or should or recommended to test your ASR I think everyone should test their ASR right like especially if you're if your company if a core product of your company is your core business of voice agent you should be testing you know all portions of your that, right? Because that's your bread and butter. And I think there's a lot of vibe testing that goes on, but I think the only way that I've seen people build really good product is through really rigorous veg marks, really rigorous evals. That way, everyone at the company knows verifiably that this thing is really good, right? Which fundamentally involves collecting very good data. In terms of ASR testing, absolutely. And I think it matters more for, I think it matters more for, even more so for companies like in healthcare, for example, where the words may be like really hard to get and you may care a lot. Like for example,

if you're a health system, right? And you're saying like, you know, all of these like really hard to say names typically, and you want to get them right. And it's important that you get them right. Then absolutely. You should be, in my opinion, creating a data set behind that. You should be benchmarking it. You should be making sure that it goes up to par. All that being said, I am also very confident that as we collect more data, a lot of speech recognition, SCT models, all of these things I think are just going to increasingly keep getting better, and I think the word error rate will keep falling. And I think having really good SCT models will just become, maybe I'm wrong, but I think will become a solved problem, if that makes sense. And I think we'll just be able to plug and play things off the shelf that just kind of work. Yeah. I think what LLMs or what Generative AI rather does, I think, better than the traditional systems is that, for example, I don't know if this is the case, but I'd be very strongly surprised if it's not,

which is like if GPT-Live 1 on the input stream, it used to be that the traditional kind of neural network-based models would, as the audio comes in, it starts at the beginning of the audio and it starts transcribing to the end of the audio. So basically what it's transcribing, it doesn't know what's coming up ahead of it and it's not really looking at what's come before it. It's just looking at in the moment what's the next thing that I'm transcribing type of thing. Whereas I think what the generative models can do better presumably is take the full context of what's being said and start to build the context and fill in the gaps. so for example if I say something that is mistranscribed I can get away with it because the model itself knows that in this context this must mean that and therefore you can clean up clean up the transcript type of thing I think there is still some areas though where you're going to find a bit of a challenge and that's purely based on one the way people speak and two

the fact that you're working with an eight kilohertz audio signal because sometimes people fail to understand don't they you know even on calls like this where the audio quality is a bit better but still every now and then what did you say there or was that such and such or such so but i think again the generative ai models i imagine probably get better at that type of stuff as well which is in this context this person is probably talking about this or that i think i heard one of these two things and therefore the disambiguation sort of kicks in um but i was just curious because you know we hear a lot about you know the asr testing happening when the agent has been built and for me that's always i'll be interested to get your thoughts i think i think you kind of if i read between the lines you probably agree with this which is that that's kind of the wrong time to test for it because you've already made all your technology decisions you've already built your agent and then you slap the asr on the top and it's like you know oh we've got some problems whereas if you've got as you suggested a data set you know representative data and you can actually test it properly at scale before you get to the building of the thing.

At least you're going into that project with a bit of confidence. I just don't see any companies really doing that, though. I think they just take it for granted that it's going to work and that type of thing. Yeah, I think it's an interesting, like, it depends on the company. I would say, like, that's, like, kind of the true, like, real-world answer in my opinion, right? If you're a startup and you have to push to launch and grab customers and start delivering value and creating revenue, most likely the path you're going to take is like, hey, let me try to do as much benchmarking as I can and make sure the initial quality is as good as possible, but I need to release something that works. And so you're going to make decisions based off of that, and you're maybe time-frenched. And for what it's worth, the voice agent market is very competitive, right? So, you know, you have to do something like that. That being said, like, if you look at a really large, you know, health system, for example, or a bank, I absolutely agree. I think you do have the luxury to, you know, like, pause, sit back for a second and be like, okay, like, how do I want to do this work, right?

Do I want to, I want to make sure that I'm getting the best in class for each part of the stack, right? And by the way, now I think there's also a contentious debate of like, hey, do you do test data? Do you do speech to speech? Right. Like what what like direction do you even go in? Like, first of all. Right. And so I think that that's like a question that I would honestly spend time answering first. Right. Once you've chosen that, OK, like, hey, like what is the model going to look like? What orchestration layer am I going to use? Am I going to build like how much am I going to do on my own? What am I going to buy anything? That sort of, you know, questions start to get asked. But at the center of it, I always fundamentally agree that everything needs to be eval-driven and eval-centric, right? That is the only way you build really strong agents, right? There's no other way, right? In fact, actually, the time and cost to build has gone dramatically down. What has not gone down is the time and cost to trust what you have built, right? That is a very different type of question to answer.

And I fundamentally think that the only way to do that is through evals. I haven't seen a better way yet. But yeah. Yeah. Yeah, I agree, to be fair. I think it's totally sort of a fair position. Where do you stand on speech-to-speech versus cascaded? You said it's a decision worth kind of giving some thought to. Generally speaking, you know, where do you sort of stand? What are the benefits, pros and cons of each? Yeah, I think we actually have another resource. We recently released a benchmark that we call... So actually, BlueJay has a portion of it called BlueJay Labs, where we try to do basically a bunch of cutting-edge research. And we recently released something called MeeBus, which is basically a speech-to-speech benchmark. And the approach we took to creating this speech-to-speech benchmark is we actually built agents across different industries.

I think we started with healthcare, legal, and customer support, and we actually pre-built a lot of tasks for these agents to actually go and complete across these different industries. My co-founder, Faraj, who led the creation of this, also came out with something that we think is pretty novel. It's basically something that we're calling the conjunctive verifier, which allows you to try to deterministically verify through three different things like tool calls handoff adherence and database adherence whether your agent is accomplishing the task or not correctly. And this was primarily, as I mentioned, through speech-to-speech, and what we saw was really good results. Obviously, the technology is improving still, and it's getting a lot better, but we have started to see really good results across different industries where these speech-to-speech models off the shelf are starting to get better and make strides to becoming something that is production ready. And we've also seen this company start to adopt GPT Live as well and work with the GPT Live

team to put that into production. Now to maybe answer your question more directly, I think that the judge or the victor, so to speak, has not been firmly chosen yet, right? It's not like, hey, this is the gold standard approach. I think you'll still see companies building on cascaded, and they'll be taking that approach for having more control, guardrails over every single turn, and the stack, and stuff like that. But I do think, especially in the way GPT Live and other speech-to-speech models are getting built with that front-end, back-end architecture, I think it offers a lot of promise, and I think we at BlueJ, we're very bullish behind this type of architecture becoming a potential standard here. I mean, you can totally see how something like GPT Live and the fact that OpenAI being able to control that model plus the models behind it and almost from a stack point of view it's almost like OpenAI could potentially own more or less the whole stack

apart from the APIs or the MCPs that you're providing to the model and perhaps the data that sits underneath it or whatever. So it's interesting to sort of see that sort of world unfold. And it kind of makes me wonder a little bit. Maybe it's a bit of a divergence in terms of the discussion, but it makes me wonder about some of these platforms that are, you know, voice AI platforms. You know what I mean? There's loads of them that have emerged over the last kind of couple of years, purely pure play voice AI platforms. But, you know, if it's GPT Live 1, that's then, you know, whatever the model is underneath it, it's hard to see what value those type of companies end up providing over time because a lot of it just gets consumed within OpenAI, doesn't it? Yeah, I mean, I think there's a lot of things that a voice platform, like an orchestration platform may provide that I think goes beyond the model. For example,

managing telephony in an integrated environment, deploying these things, offering reliability. Also, as you may have five different speech-to-speech models, being able to plug and play the different speech-to-speech models, switch between them. I do think a really robust orchestration platform essentially gives control back to the user. And I think that's something that I've realized is control is really important to an enterprise, right? Because it allows them to honestly create more of that trust, right? I think naturally as humans, right, if you have more control, actually sometimes, maybe not in all the case for every human, but sometimes you have, if you have more control, you feel like you may have more trust in it, right? Depending on who you are. And so I think that controllability, that preciseness that an orchestration platform may offer can sometimes be very beneficial to those end customers rather than just kind of doing everything through one model company.

Yeah, that's a good point to be fair, yeah. I think the kind of quest for control is an interesting one because we've gone from one extreme to the other really, which is that it used to be that you used to have to map out absolutely every available dialogue path, every single thing that this thing could say. And in the process of doing that, I think what you kind of build an understanding of is like in an ideal world under ideal conditions, what would a great experience be like? And I think there was a lot of experience gained in that time in terms of how do you phrase a question to make sure you solicit the right type of answer? How should you use things like SSML to make sure that you pronounce things in the right kind of way? And as great as a lot of this voice AI technology has gotten, things like having to specifically mark up every element with SSML is just an absolute laborious task that nobody wants to do and you don't need to do anymore. But what that comes with is every now and then it'll get it wrong. Every now and then it'll phrase a question that doesn't quite sound right, doesn't quite have the same right intonation.

And so what you gain with the naturalness that we kind of spoke briefly about earlier, you do almost lose the control. So I think any area where you can instill a bit more control is always good at, I think. Not that you want to control a full thing, but you don't want just to leave it up to the generative gods, so to speak, and you get what you're given type of thing. Yeah, I think there's always questions that you have to ask. It's like, hey, what industry are you deploying in? Where are your customers? What volume is this going at? What's the impact if I get something? There's a calculus you have to do, which is like, hey, if things go wrong, what are the risks here? on essentially like, yeah, what are the risks here if I don't do things in an XYZ way? At the end of the day, though, I do just, I fundamentally think, and once again, maybe I'm like, I'm obviously very biased, but like, anything that gives someone more insight

into how the thing is working and kind of that like overview, that observability is just so crucial, right? Because as you trust it more and more, it's important that you are also getting flagged about, hey, these are all the things that are working, these are all the things that are not working, and you're able to act based off of that. I think that level of insight, regardless of how much you give away to something, is still very, very important. Yeah, agreed. Very good. Wicked. All right. Yeah, that's good stuff, that. I mean, I think there's... Is there anything we haven't covered? Is there anything that you think, you know, conversations that maybe we should be having today that perhaps we're not? I think I'm going to be like actually writing something about this soon, but I'm sure you may have read like Dario's essay about like pacing the frontier. And like, I think there was something by Satya that recently came out as well. And I think overall we have been so focused on what I essentially call the,

actually let me say that. I think there's two frontiers, right? And I think we've been very focused on the first frontier, which I call like the capability frontier and the capability frontier. frontier, I would say, is all of this great intelligence, the labs competing with each other to provide even better intelligence than the previous one. They're moving so, so fast that they're now saying, hey, we need to pace it. That's essentially what's happened. And I think that capability frontier is growing so rapidly. Now, the second frontier is what I call the deployment frontier. And I think that has not been pacing at the same level as the capability frontier. If I looked at the capability frontier, it's like an exponential curve, and the deployment frontier is like this. I think the fundamental bottleneck there is trust. It's like, hey, how do you eval these systems to know that they're going to work? But it also just requires a lot more work. It is to get the and this applies to voice and also just AI agents in general, which is

to actually make use of the intelligence that we are so abundantly creating, that is where the real work, in my opinion, lies. That's why you see, by the way, also investments from the labs to partner with these consulting firms, actually go and implement these enterprises and so on. That is where a lot of the hard work lies. And I think that deployment frontier is sort of where I'm excited to most work in. And I think it's also something that we as an industry, as entrepreneurs, as researchers, engineers, marketers, sales people, that is something that I feel like we need to be focusing a lot more time on. That way we can actually see the promise of this AI technology come true. Agreed, yeah. There's no point in it accelerating and getting better and better and better if no one's using it, is there? Exactly. Absolutely. Brilliant. I appreciate your time, my friend. It's been absolutely fantastic. Yeah, all the best. And we'll put those links to the seven ways to reduce latency

there was also one you mentioned in terms of the speech to speech testing yeah we'll stick that in there as well as well as a link to the BlueJay website your LinkedIn awesome cool we'll speak again soon thanks so much Ken really appreciate the time with you appreciate it nice one and thank you all for tuning in we'll see you again on the next one cheers now bye