In the sixth episode of Speaking of AI, LXT’s Phil Hall chats with Iain McCowan, Director of AI at Dubber about the entrepreneurial spirit, trends in Generative AI and data collection, and responsible AI. Iain shares his insights on working in both research and commercial environments and discusses some of the emerging trends within generative AI applications. You can also catch Iain’s advice for aspiring entrepreneurs. You won’t want to miss this episode!
Introducing Tech Entrepreneur Iain McCown, Director of AI at Dubber
PHIL:
Hello and welcome. My guest today is an industry veteran, a researcher, and a serial entrepreneur with more than 25 years experience in AI. He was principal research scientist both at CSIRO, Australia’s leading research and innovation organization, and at the Idiap Institute in Switzerland. He is the author of more than 50 research papers and holds multiple international patents. In the commercial domain, He founded Dev Audio, where he developed a speech-enabled microphone array, which was subsequently acquired by BiAmp Systems. And then he founded Notiv, which was subsequently acquired by Dubber, where he is currently director of AI. Please make welcome Iain McCown. Hi, Iain.
IAIN:
Hi Phil, thanks for the intro, looking forward to talking to you.
What are some of the emerging applications of generative AI that you are most excited about?
PHIL:
That’s great. Okay, so we’re more than a year into what I would call the post-ChatGPT era and AI, especially generative AI, remain hot topics. What are some of the emerging applications of generative AI that you are most excited about?
IAIN:
I mean, it’s a good question. There’s so much going on at the moment. It’s hard to keep up with all the changes and the different things coming every week, it seems. I think for me, I’ve been working in speech and language tech through my career, right? So that’s the area that interests me the most and is engaging me the most with what’s changed. And I think if you look at, you know, broadly, everyone’s aware, ChatGPT came out and that was kind of a step change in awareness and also a step change in capability of language models, to be fair. For the first time, human readable and output was available for different applications from AI models.
And so I think for me, what’s exciting at the moment is, you know, for the first time we’ve got these new tools that are able to actually understand conversation. And I think what, you know, a lot of the use cases are focused on human bot communication, if you like. But I think what interests me is the potential to unlock human to human conversations and, and extract insights and get more value out of conversations like this from this AI. So for the first time in, in my career or in history, you know, AI is capable of understanding, synthesizing, summarizing, detecting key points in a human-to-human conversation. I think there’s a lot of exciting opportunities around that.
PHIL:
Very cool. Yeah, I would certainly agree with all that. It is pretty exciting. I’m not from a technology background. I’m a linguist. But the idea that language technology is at the state of evolution that it’s reached now is very exciting to me.
IAIN:
Yeah, I know for like one, just one example, I guess the holy grail for us has been summarizing a conversation, right? And for years we struggled with extractive techniques where you try and get bits of transcript and piece them together and it just wasn’t readable. And whereas now, you know, you can get a condensed readable summary of a conversation. So that’s just one of the new capabilities there, which means that it’s unlocked new use cases for tech to actually give value to businesses, use cases for tech to actually give value to businesses.
Based on your entrepreneurial experience, what advice would you have for entrepreneurs who are considering this path?
PHIL:
Yeah, I can’t wait to skip meetings and read the summaries. I hope my colleagues aren’t listening to this. So, Ian, you’ve very successfully navigated from startup to acquisition twice.
Based on this experience, what advice would you have for entrepreneurs who are considering this path? And do you have specific advice on the challenges or opportunities inherent in addressing global markets from an Australian base?
IAIN:
Yeah, I mean obviously having gone through two startups, different journeys each time, you know, I could talk about that for hours or days, but I think what it comes down to is, you know, you’ve got to be creating something of value. Building value is what a startup is about. You’re starting from an idea or your abilities and trying to create value. in terms of delivering something to a customer, a new customer need that you think you can meet. Also building value for your investors and in your team, et cetera. So I guess if you bring that mindset to what you do is that every step of the way you’re trying to create value for either your customers or your team or investors, I think you’ll be successful at what you do. I mean, a startup can be an up and down journey.
So… the other advice, I guess, is always do something you’re passionate about and you love, and that’ll get you through the ups and downs.
I think in terms of doing it from Australia particularly, or I guess from outside the tech center of the universe in the US, my approach is to abuse a phrase, I guess, think global, act local. So, you’ve got to be globally minded from day one, wherever you’re starting a business from. You can’t launch a tech company and be just thinking about people around you. You’ve got to be thinking globally from day one. But that said, there’s smart people everywhere. There’s resources everywhere. Connect to people and networks and programs, investors, government support around where you are. And from Australia. Australia has been a great place to launch businesses from and that continues to, you know, that ecosystem and support for entrepreneurs continues to improve.
Do you have a preference for working in research or commercial environments and what are the specific challenges and rewards of each?
PHIL:
Very cool. So you’ve worked extensively in both research and commercial environments. Do you have a preference and what are the specific challenges and rewards of each?
IAIN:
I think my career as you touched on at the start has been, you know, kind of see it’s spanning that gap I spent the first part of my career in the research environment. I love that, but was frustrated that, you know, the endpoint was writing a paper or, you know, publishing. And then that enabled you to maybe get a grant for the next one and the next one. So you’re always taking ideas from concept to the point where you’ve validated it, written about it, but not actually take it into the reality of getting in the hands of someone who can use it. And so I think, you know, there’s different types of researchers, but in the research world, I was frustrated wanting to make, you know, not let it go when it got to a paper, but take it all the way through to reality.
But the research environment’s not often set up to do that. On the flip side, in large enterprise, you’re really set up to take products to market, but maybe less scope to be creative and come up with, you know, far-fetched ideas and test them out and things. So I like being between both. I guess that’s why I’ve done two startups is really taking some new emerging capability and spanning, if you like, that chasm between research and reality and taking it across.
As a product leader in the world of AI, what are your views on the role of regulations and its impact, positive or negative, on innovation?
PHIL:
Recently at the World Economic Forum, there was discussion of 2024 as potentially being the year of AI implementation. And there was a focus, a strong focus on ethics and safety. As a product leader in the world of AI, what are your views on the role of regulations and its impact, positive or negative, on innovation?
IAIN:
I think my approach to this is like as researchers in this area, we know what we should do. We know what to do, we know what the correct thing is to do, what’s right in terms of how we deal with data, particularly training models and practice around that. So I think regulation shouldn’t change behavior, but it does create accountability there to make sure that, hey, in the worst cases, there’s accountability. So I’m supportive of the regulation. Obviously, in realm of responsible AI there’s some new regulations coming out in UK, US, Australia at the moment that we’re continuing to watch, but from my seat, it’s not changing what we’re doing, it’s just providing that, I guess, that accountability and that structure to how we report on that and our processes around that, which is a good thing.
I mean, I think to touch on your question around innovation is always pushing boundaries, right? And maybe, you know, that’s important. And so you don’t want regulation that squashes innovation. So it’s got to be well informed regulation, not just knee jerk overreaction to things. So I think where it’s well informed, well- intentioned regulation, it’s good. You know, that. I’d say innovation should push boundaries but it shouldn’t be crossing, you know, there are lines you don’t want to cross and I think that’s where regulation comes.
Is there a solution to generative AI and it’s tendency to have, and indeed, act upon biases and to hallucinate? And are these flaws likely to prevent AI based on LLMs from becoming truly scalable?
PHIL:
I recently read on a related note, I recently read an article which said that the Australian government was very concerned about ensuring that AI was trained on unbiased data and it really got me thinking about what that actually means. And my next question, which is a fairly lengthy set up and question, is a reaction to what I read about that, because I feel like this concept of unbiased is quite risky. I mean, in a sense, all the data that’s out there is unbiased.
It’s a direct reflection of the real world. And in that sense, it’s unbiased, but it will produce technology with biases.
And so I’m slightly concerned that the Australian government might be using what feels like the right terminology, but it’s not scientifically informed. Anyway, to my question. Among the major flaws in generative AI are its tendency to have and indeed to act upon biases and to hallucinate. Biases and hallucinations in GenAI output are typically a flow on from deep seated biases in the data used to train the models. When discussing this, I’ve often encountered the view that what’s needed is neutral or unbiased data. My own take is that there is no such thing as neutral, unbiased real-world data.
So it seems to be that the main options might be to use huge volumes of data that are out there in unfiltered form and accept the inherent biases. And in so doing to accept that those who are underrepresented in the data will in fact be underrepresented by the products or to proactively introduce bias to address imbalances in a form of affirmative action, if you like. But when you do that, this neutrality becomes subjective. And then there’s the question of whose interpretation of neutral will prevail. So is there a solution to this? And is this likely to prevent AI based on LLMs from becoming truly scalable?
IAIN:
Yeah, there’s a lot in there, a lot to pick through, I guess. So in terms of is there a solution to this, so if the problem is, is there a solution to bias or hallucination from models? Yeah, I think there’s multiple ways you can attack that problem. Data, you know, the training data distribution being one of those ways that you can influence that. But you know, you’ve worked in data for a long time, so I’m not going to argue with your thoughts on the reality of actually collecting unbiased data, whatever that means. So I think, I guess the perspective I come at it from is a product perspective.
So we’ve got the problem of there may be bias in our system or it may hallucinate. How can we, knowing that our model may have those limitations or other limitations. How can we deliver a product that provides value to customers in a safe way? And so there’s different ways to do that. One of the ways we do it at Dubber, as you know, we’ve worked with LXT on this one, is, you know, not so much on the training data, but on evaluation data. Let’s collect evaluation data that represents particular population or demographics, such as languages or dialects or, you know, types of businesses, et cetera. And so let’s collect a smaller amount of data, which is probably more feasible and achievable and measure, quantify at least performance on different data. So at least then we understand, is there a bias and does my model work better for US Americans than it does for French Canadians, for example, if I’m doing transcription. And if so, what’s that difference and how can I address that? So I think step one is quantify and understand the bias in your product and that can be done with more reasonable amounts of data. And then, you know, the other way is old fashioned engineering, I guess, probably aligned with your affirmative action type thought line of thought in that, you know, I’m not a smart car, self-driving car engineer, but I imagine that inside there, there’s some AI, right?
That sophisticated AI that’s detecting objects, classifying them, making sense of the world that the car is traveling in. There’s probabilistic models at the heart of that, right? But they are engineered into a system that makes that a safe, as a whole system, makes that a safe driving experience for the driver and other people on the road. So I think that that same thinking applies to, you know, for what we do at Dubber, for example, we use generative LLMs, but we package them in such a way that they’re anchored to the context of the conversation so that less likely to hallucinate and show bias, for example. We constrain rather than letting it just generate whatever text it feels like and sending that direct to a user, we structure, get structured output. So only certain, I guess, categories, we’re looking for structured output from the LLM so that we can control what the user experiences in a safe way.
So I guess there. Like it’s a deep topic. And to your point, I’m also be wary if people thought that the solution was just unbiased training data, because as you said, you know, most of these AI models just need more and more data and the data they can get reflects the world and the world has biases and you know, that’s the way it is.
I think in terms of your final question there. Can we achieve, can these LLMs achieve intelligence with these limitations or is there a solution? You probably consider yourself an intelligent being, I do as well and yet I’ve got biases. I tend to hallucinate after a couple of glasses of wine, I hallucinate and rabid on answer things that haven’t been asked to me. So I don’t think intelligence may still have biases and… hallucinate as part of it. Who knows?
In the face of the combined wealth and resources of these mega-corporations, Apple, Microsoft, Amazon, Tesla, Meta, do you think that there’s still an opportunity for smaller organizations to thrive in the AI space?
PHIL:
That’s a very cool answer, yeah. And I like, to paraphrase, engineers can save the world. My next question. There’s an enormous concentration of AI firepower in the top seven corporations globally, the so-called magnificent seven. So between them, Apple, Microsoft, Alphabet, Amazon, Nvidia, Tesla and Meta account for 27% of the S&P 500’s value. This week,
The Financial Times reported that Apple, Microsoft, Meta and Alphabet had between them made more than 50 AI-related acquisitions in the past five years. So there’s a trend. In the face of the combined wealth and resources of these mega corporations, do you think that there’s still an opportunity for smaller organizations to thrive in the AI space? Or is it possible that startups are simply seen by these larger organizations as incubators for ideas that can be acquired and integrated later?
IAIN:
Um, such an easy one for me as a startup guy, I have to say there’s a space for startups and there always will be. One thing I learned early on in my, in my entrepreneurial journey is that, you know, every, every startup company is different, right? It depends on, you know, not every startup is out to become the next unicorn. So there’s a range of businesses and sizes and different objectives for why people start businesses. So I think there’s a place, a place for all sorts and all sizes, I guess specifically in tech startup land. I know there’s a number of smaller businesses that we work with because they provide the best solution for some of the things we need.
I know in the world of speech techs, for example, yes, the big tech companies like Microsoft, Google, Amazon have great platforms for that, but state-of-the-art performance is also available through smaller companies like Rev AI and Deepgram, for example. We’ve just spent two years building out our conversation intelligence platform at Dubber and at the heart of that we use an orchestration engine for our workflows called Temporal, which again is, I’d say probably a startup. Startup’s kind of a fuzzy term at what point you stop being a startup, I guess. So I think there’s always, there’s room for innovation outside the Big Tech, particularly as you get into specific niches.
So I guess come back to my early point around startups is what’s the value you’re creating and there’s, you know, there’s advantages to being a startup and a small company and that you can really focus, hone in on a niche or a particular piece of value that you want to deliver and really do that well and quickly and in a more agile manner than what a larger organization could be. But, you know, there’s no denying the environment is, scale, you know, those big companies are growing and there’s a momentum that, that causes and keeps building, right? But I mean, remember when I started my career, you’re probably the same. The big research, all the research papers came out of commercial labs, AT&T, Bell Labs, and, you know, IBM. You know, they were commercial labs. And then through the middle of my career, that’s why the pendulum swung back to academia, right? And all this deep learning and the current wave we’re in really came out of academic research labs initially and then has grown back into the commercial. So swings and roundabouts, everything’s in cycles, who knows where it goes.
What trends are coming up, or what shifts are happening in how data is going to fuel the next phases of AI?
PHIL:
Very cool. Okay, well, I have just one more question for you. If you were running this interview, what would be the one question that I haven’t asked you that you wish I had? And what’s the answer to it?
IAIN:
Um, so I think you’ve asked me a lot of questions and I’ve probably rattled on enough for this talk. So if I had the chance, I’d like to ask you a question if that’s all right. I guess, um, I mean, um, I’m on the, I guess the AI development product development side of things. I know you’ve spent 20 some years in the 30 years, who knows in the data side, right? And so AI, the current AI is, is largely built on data resources, right? That’s the fuel that’s driving this as well as processing power and other things.
So I guess from, I’m interested from your seat and where you’re at, we’re at an interesting point in the industry in that large reams of data are available. Models are now, you know, I remember back in the day we had the, the holy grail was unsupervised learning so that we didn’t have to hand label every piece of data. You know, we didn’t have to have the annotation expense. We could just give the model data and it would learn stuff. And so now we’ve got this, what do we call it, self-supervised learning is a big part of how these models are leveraging large data pools.
So I guess in this new world or current world, from your side of the fence, sitting on the data provider side of the fence, I guess what trends are you seeing coming up or are you seeing any shifts in how data is going to fuel next phases of AI or where the next wave is coming from.
PHIL:
Okay, good questions. And I will say you’ve mentioned two or three things in there and earlier in the interview, which I had to stop and think, have I actually published my previous interview already? Because they were red hot topics that we were discussing there with Roberto Pierroccini, who was formerly with AT&T, Bell Labs, IBM. And we discussed some of these very things and he referred to unsupervised learning. And we discussed some of these very things and he referred to unsupervised learning as the Holy Grail, which, yeah, I know you’re not blowing smoke when you say that. It’s definitely a widely held view. So yeah, where do I see the data field going? My view is that the need for data is not going away. My confidence level in my view is some fair distance short of 100%. So I say that with moderate confidence, but not high confidence that data is going to be required for quite a while to come.
In fact, one of my former customers in the automotive vehicle space, when I asked him how long he, how much more data he thought they were going to need, he said, he’s about the same age as me, and he said, well, Phil, I can’t quantify it exactly, but let me just say that when you and I are dead, we will still need more data. Yeah, and at that point we were generating data at a phenomenal rate. So, you know, his desired volume of data was unimaginably large. My experience with that is has throughout my 25 years working in this business, every time it looks like the data problem, well, every time it looks like the problems are solved and people start talking about, well, maybe nobody’s going to need data anymore.
What happens next is that we move on to try to solve bigger, more sophisticated problems and rather than the need for data going away, it 10xs. I’ve seen that 10x happen again and again and again. And so I still feel like if we are going to, if we’re going to build these super robust applications, the need for quality data and for human in the loop is still around for a little while. Am I answering the right question?
IAIN:
Yeah, I think so. Just your 10x comment made me think of back in 2005-ish, we collected the AMI Meeting Corpus when I was at IDIAP and we thought this was this huge, valuable, multimodal database. Then we recorded, it was 100 hours of transcribed speech in meetings.
And we thought that was, at the time, that was a significant resource. And now that’s a drop in the ocean. I guess my questions are, you know, we’ve got issues around, I guess, OpenAI and Wall Street Journal copyright of content that’s being used to train models. And then we’ve also entered a new world where a lot of the content on the web is, and probably newspapers and other things, is going to be partially generated by AI itself, right? So it’s kind of corrupt. It’s no longer pure human-sourced data. So I think my own take is that maybe we’ll swing back to quality over quantity. and just being more picky about the data. But I don’t know.
PHIL:
Interesting, yeah, actually that’s when you talk about pendulum swings. I’ve observed multiple times in my time the pendulum swing between just get me volume and don’t be too fussy. Get me volume cheaply. Don’t be too fussy about the quality, to okay we need the quality. We’ll pay for the quality. But that’s more important than volume. So we need to be very, very careful about the quality. And that one definitely goes backwards and forwards in quite tight cycles. Thank you for that. You’re the first person that’s actually turned the final question back on me and I’m very pleased you did. Thank you. Iain McCowan, it’s a pleasure talking to you as always, and I look forward to our next conversation sometime down the track.
IAIN:
Thanks again for taking part today. Thanks, Phil. I’ve enjoyed chatting to you.
