In this week’s SlatorPod, LXT Chief Growth Officer Phil Hall joins us to talk about the company’s journey, from providing high-quality Arabic data for a Big Tech company to expanding into 750 languages.
Phil shares his background teaching linguistics and leading business development for Appen before joining LXT. He discusses the key findings from The Path to AI Majority report, from the maturity levels among 200 executives surveyed to which industries are trailblazers in AI adoption.
Phil touches on search relevance ranking as a method to retrain machine learning and give more relevant results. He gives his thoughts on ChatGPT and why he considers it a step change in AI.
Phil talks about LXT’s growth strategy, with M&A a potential avenue to acquire new customers and improve the technology stack. He goes over some of the legal and security considerations when it comes to data, with permission and secure facilities taking priority.
Subscribe on Youtube, Apple Podcasts, Spotify, Google Podcasts, and elsewhere
The pod rounds off with Phil’s take on where the AI industry is heading, with quite a long way to go before it becomes stable, low risk, and deployed with absolute confidence.
Florian: Today we’re really excited to have Phil Hall on the podcast. Phil is the Chief Growth Officer at LXT, a leader in AI training data. What city are you recording this podcast in today?
Phil: I’m actually in Sydney, Australia. LXT head office is Ontario. We have people in the US, we have people in the UK, and we have quite a lot of people in Egypt, but I’m in Australia. I don’t know if you know my background. I was with an Australian company, Appen, and I was with them from 2001. I was employee number three, and I was there for 17 years. When LXT approached me to take on a role with them, I said, look, if I were you, I wouldn’t hire somebody that’s based in Australia and they said, trust me, we’re okay with that. Not for any good reason that I’m in Australia.
Florian: I wouldn’t go anywhere if I was in Sydney either. First, can you just tell us a bit more about LXT for the listeners of this podcast who tend to be more on the translation, machine translation side of things like when and was it started by who, and like, what are the key businesses that you’re in and some of the people that work for you?
Phil: The company was founded in 2010. It was founded to meet a demand from one of the Big Tech companies. They came to our founder, Mohammad Omar, and said, we need Arabic data, and we’re having trouble getting Arabic data at high quality and he said, sure, I can do that for you. And he built a really strong relationship with them with year on year growth throughout the next ten years, where they would just come to him and say, can you do this? And his answer was always yes and suddenly he went from doing Arabic data collection to the point where he was doing 250 languages worth of annotation of a wide variety of types, plus data collection. The company grew out of the head office in Cairo and then established a new head office in Ontario in Toronto. We currently have around 500 employees. That’s about 70 professional staff and then we have the rest are made up of annotators working in secure facilities. We are essentially a services company, but we’re a services company with ambitions to be more of a technology company into the future.
Florian: In your current role as Chief Growth Officer, what does it encompass? Is it sales, marketing? Or is it also kind of corporate development type of aspects?
Phil: It’s all of the above and I think the title itself gives you a pretty clear idea of what my job is. So when Mohammed approached me to come and join LXT, I’d been with Appen for 17 years, as I mentioned earlier, 10 years as head of sales and then the next seven years running the primary profit center within the company and through those 17 years I’d never gone backwards. We achieved growth every single year and he came and said, well, we’ve got a good business, it’s a good operational business, but we don’t have a huge customer base and we’d like to achieve some growth and Phil, I think you can help us with this.
Florian: LXT works across five main use cases like AR/VR, computer vision, conversational AI, search, relevant speech and NLP. That’s quite the journey from the original language center, so where would you say currently are kind of the particular strengths and what areas do you see are growing the most at the moment?
Phil: We’re still very strong in the speech domain. That’s where the company grew and the demand for speech-related work is still very, very high. But we are adding these other areas related to image, video, text, search. These are being added very, very rapidly and we’re seeing a lot of growth in them. I think that we will continue to see our growth for LXT being very focused on more of the language based end of the scale than the image end of the scale. Partly because I think the image and video market is quite crowded. The barriers to entry are quite low and therefore a lot of companies that want to be in this space are quite focused on this video and image end of the business. I think it’s probably the most competitive end of the business. I don’t want to use the term race to the bottom, but there you go, I’ve just said it, so it’s kind of a race to the bottom. Whereas the language area is a little more difficult to enter, the space is a little less crowded, and we’ve got 12 plus years of experience in that area.
Florian: Is that because the task is a bit more trivial to annotate like images or video?
Phil: There’s certainly that. I’m not going to say that a lot of the annotation that we do isn’t fairly straightforward as well. But in terms of image and video, you can set up one operating center in one part of the world and it’s kind of language agnostic. So once you’ve got a business set up, that’s it, you’re in business. Whereas if you’re going to provide somebody with a truly global service for language, that takes a bit of time to build up. Now let me ask you a question, Florian. We work in a lot of languages. How many do you think is a lot of languages?
Florian: There’s concentric circles at least in the translation business. Five would be the core that maybe expands to 30 and then it gets really difficult when it exceeds 30 and you get to 100 because then you need to rely on a much, much broader kind of network of freelancers or even subcontractors, so I think 100 would be a lot.
Phil: My experience is very much like yours. We do very high volume on five languages. We do significant volume on 30 languages. When I joined LXT, we’d worked in about 250 languages, but in 2022, we worked on 780 languages.
Florian: That’s also a challenge for the backend, for the payment system, for knowing even the names of all of these languages. That’s going very deep into the languages of lesser diffusion, I think is what people are referring to it now.
Phil: My background is linguistics. Before I joined Appen, I was teaching linguistics at a couple of the universities here in Sydney and I like to think that I know a fair bit about this, but really the languages that we’re talking about in this very long tail, I have no clue about the names of those languages.
Florian: They would be from where mostly? Africa or India?
Phil: There’s a pretty big proliferation across South Asia and across Africa, absolutely. To be honest, after we got past the 200 plus market, I stopped looking, so I’m not quite sure exactly where we’re going. But I know that these are languages that have a presence on the internet. Otherwise, it’s not simply… A lot of my linguist friends are doing their PhDs in languages of the New Guinea highlands or languages about back Australia. These languages, it’s a language preservation exercise. But in the case of the work that we’re doing, this is a language preservation element, but they’re all languages that have presence on the internet.
Florian: When you label, annotate, do the people that do the labeling and annotation, do they have to be familiar or fluent? Or what’s the degree of proficiency in a particular language you need to annotate or to the work that you do?
Phil: Generally speaking, the default assumption is native speaker. There are cases where you can work with non native speakers and be more cost effective. If you think about varieties of English, varieties of Spanish, varieties of Arabic, for example. If you think the work can be carried out by someone who is a native speaker of any variety of those languages, then you can save money. People whose native languages, Egyptian Arabic, come at lower cost than Saudi Arabic or UAE Arabic, for example, so you can save money there. But that’s usually a choice by our clients, not a choice by us. For us, native speaker is the default assumption.
Florian: One more aspect of the business is search relevance. Can you just tell us what this is, exactly? What is encompassed at that service?
Phil: Search relevance can be user intent analysis. So you’re given search queries, this is what it says, now, what did the user mean when they wrote that? Okay, so that’s a classic form of that. It could be search relevance ranking. So you have a search term, search phrase, whatever, and you look at the results that come up and you can mark off those and say, okay, this one should be at the top of the list. This is a good result. This is a spurious result. It’s not what the user meant and by having that relevance ranking, you can then retrain machine learning and get more relevant results. And you said that Lionbridge was looking at this a while ago. I’m sure they’re still looking at it today, or that Telus AI is looking at it today because it’s the kind of thing where there’s the need for data in terms of improving the algorithms is ongoing and you can’t just hang on to user data forever. It has a use by data on it, and it wants to be fresh in order to be effective. So even though the volumes of data they get… are quite large, they also get flushed and recharged, so it’s quite an ongoing business.
Florian: On the language side of things, speech, is there an element of retraining or labeling kind of connected to current events and just emerging concepts? For example, COVID, there’s a bunch of vocab that just emerged when this hit.
Phil: For people who are working in some areas of business. If you’re interested in the entertainment business, there’s new vocabulary emerging daily, so new artists, actors, performers, films. So there’s content related to that that’s new all the time. Emerging geopolitical events, so you do need to keep retuning for that.
Florian: Do you do any work related to translation or machine translation, or is that something kind of quite a niche area of the business?
Phil: I’s fair to say that work that we do is feeding into machine translation applications, but we’re not experts in translation and localization per se, and we’re unlikely to become that. It’s such a specialized business. We would be starting not just years, but probably decades behind the people who are the leaders in this field and I don’t think it would be a smart move for us to really refocus on that area. But having said that, over the years, there have been many occasions when a client has come to me, either in my role at LXT or prior to that, in my role at Appen, where they want an entire package where there’s elements of speech, there’s elements of text, there’s elements of translation, and they don’t want to source it from multiple vendors, and so we would quite happily carry out translation work in that context. But we don’t see ourselves as leaders when there’s some terrific technology, as you’d know, underlying modern translation work, and yeah, we would never catch up.
Florian: I want to talk about a survey you did about a year ago. You called it The Path to AI Majority and you surveyed 200 senior execs in kind of mid-to-large US organizations. Can you tell us what were some of the key findings and the findings that maybe most surprised you?
Phil: This is a self assessment, so these executives are rating their own organization or their own perception of things. But we went to a lot of trouble to ensure that we got the right audience. So 200 executives, but those 200 executives had to go through some pretty extensive screening to take part, including screening to actually check that they really did understand about AI and not just feel that they did. The rejection rate was high, so I think we rejected 800 and kept 200 people who… There were 800 people who thought they knew about AI but failed the screening tests to get in. But even so, it’s about self perception. One of the things that was surprising was that less than 40% were at the higher levels of maturity and then once we’d made that split between who’s at the more advanced stages of AI maturity and who’s at the experimenting stages, it was interesting. Some of the most interesting results were comparing the perceptions of those who’ve crossed that line into maturity with those who are at, well, I think we in the document we called them experimenters, but I prefer to think of it as at the aspirational stage. And when we looked at that comparison between those two groups, one of the things that really surprised me was that the aspirational group was under the impression that they’re going to do a lot with unannotated data. They’re just going to get their hands on tons and tons of unannotated data and they’re going to throw it in, and it’s going to do great things. They’re going to do lots of unsupervised learning. Whereas the mature group said, no, we need lots of annotation and supervised or semi-supervised learning. So that was one of the big contrasts was that… Obviously, the unsupervised learning, unannotated data end of things is the least expensive way to get into this business and the folks who haven’t done it yet, but are pretty sure they going to do it, they were all sure that they were going to do it without spending too much money and folks who’ve already done it were saying, yeah, not so much. But that were among the interesting findings.
Florian: For me it was interesting that the financial industry apparently is among the trailblazers in AI adoption. Why do you think that is? To me it would seem like a more conservative industry.
Phil: When we got the results back, we scratched our heads a little bit and thought what is going on here? But I think it’s probably driven by two things. One is that fraud detection is a natural fit for financials and it’s highly important, increasingly important, so they’ve got a very strong motivator to be in this space. There’s a lot for them to gain and then another thing that I think has led to this is that they have a lot of… A lot of the data that you’d be using in the financial sense doesn’t need annotation. So yeah, I mentioned supervised versus unsupervised, annotated versus unannotated. Another big distinction is structured versus unstructured. So when you’re dealing with lots of speech data, it’s unstructured and if it’s unstructured, you can’t do a lot until it’s been annotated. Whereas financial data is inherently highly structured. There’s a lot of metadata that comes with it and that would lend itself to applying machine learning with a much lower threshold of pain, a much lower barrier to entry. And so in the end, I think that they’re incentivized because machine learning can do a lot for the financial sector and they have opportunity that other sectors perhaps don’t have because of the highly structured data that they work with.That’s my hypothesis. I couldn’t tell you whether that’s fact or not, but that’s what I believe.
Florian: In terms of kind of general adoption into the enterprise, are we seeing roles like Chief AI Officer or what have you emerge? Who in the organization would be in charge of budgeting and then deploying AI? How is this getting deployed?
Phil: When I saw your notes, I looked up to see if I knew any Chief AI Officers and I don’t personally, but I do see the title emerging. I think that it’s probably… What I have seen is a proliferation of Chief Information Officer roles, Chief Data Officer roles, and Head of Machine Learning roles. I think they’re the people that are driving this and running what’s going on in this area. But Chief AI Officer may become a very big thing in the future. I think it’s the kind of thing that could change very, very quickly.
Florian: What are your thoughts on ChatGPT and the whole kind of buzz around it and Microsoft, OpenAI, and Google trying to catch up and all of these things?
Phil: I did ask myself that same question. Is this just a bit of showbiz or is it meaningful? I reached the conclusion fairly quickly then, really is meaningful. I think it’s quite transformative, at least in my superficial understanding. The gap between ChatGPT today and GPT-3, which I think was the predecessor, is a big gap, so they’ve made some rapid progress. Microsoft’s intention to use this to change the face of search is very compelling. I don’t know how much further it’s going to need to go to get there, but if anybody can make it happen in terms of resources, I would say Microsoft, the Microsoft partnership with OpenAI is well enough funded to make it work. And I think if you’re in any doubt as to whether this is showbiz or not, Google’s reaction at least has been reported in the last few days, indicates that they do not think it’s something to be taken lightly.
Florian: What I’m reading as well is that Google could have deployed this or something similar, that they were just more cautious. They have PaLM, that huge language model and they were a bit more afraid of like launching it or the blowback if it didn’t respond as people were expecting, would have been quite fierce. For OpenAI, it might have been a little easier as a startup to take some of that heat.
Phil: One of the biggest differences between the GPT-3 and ChatGPT is the way that it responds to hate and the unpleasant side of uncontrolled AI.
Florian: What I find fascinating, also speaking with you from the data annotation side, is when some of these major tech blogs would write articles like, there’s humans doing the annotation behind ChatGPT, it’s not all like magic AI. When they find out that there’s actually an army of people that have carefully trained these systems and fed it with data, like it was some kind of secret. These systems need a lot of human annotation where I guess you guys are coming in as well.
Phil: I think the articles that I’ve read seem to be focusing on the ethics of this. But I’m talking about our competitors here, not us, but I don’t see any genuinely unethical behavior going on. I actually see it as being all pretty positive.
Florian: Talking about competitors, what are your thoughts on these kind of fast-growing or well-funded new cohort of competitors like Scale.AI, SurgeHQ, or Snorkel? It’s one of the most recent one that came on my radar. Like, are they changing the landscape? Are they catching up? Are they doing something very different?
Phil: I’m always guessing a little bit with what exactly competitors are doing and my experience has been that there is often a gap, a substantial gap between the marketing and the reality. But I think that Scale has done a lot of positive things. I mean, they attract a lot of attention and clearly people are investing in them. But Appen also attracted a lot of attention and took a lot of investment on the basis that they were strong on technology and then turned out that in reality, they were still really a services company, and their share price dropped from 45 dollars to two dollars fifty. Not in one hit, but that’s a long way down. I don’t know whether… My impression is that Scale is quite strong on technology, very strong on technology, and that they’re doing technology partnerships with large corporations. At least that’s the word on the street, and that’s a pretty exciting thing. But I think my understanding also is that perhaps there’s still more of a services background in there than the publicity would lead you to believe.
Florian: I recently read a piece by Sridhar Ramaswamy, an early Googler, a partner at VC firm Greylock, and now co founder of a search firm called Neeva, and he basically broke down the current AI ecosystem into like five broad buckets. Foundation model players like OpenAI, AI frontend startups like Jasper, copilot like LLMs, like Lilt. They mentioned Lilt, that translation tool, aI agency in our view. Tooling companies like Scale.ai and LXT, and big compute clouds like Google, Azure, AWS. He said that data labeling companies were included under tooling companies and said they were “classic shovel providers in a gold-rush”. What are your thoughts about this framework and kind of the description of tooling companies as the shovel providers?
Phil: I’m not at all offended by it. When my friends have asked me what I do, and I explained to them and they still don’t get it, I’ve actually used the same metaphor to describe it. So I’m not offended by it, and I think it’s actually quite a reasonable position, and like a Gold Rush. I joined Appen in 2001, we were small, but we were still one of the biggest. We had three or four competitors around the globe. Even by the time that I retired from Appen in 2018, the landscape had expanded, but we probably didn’t have more than 20 competitors. But I read a report last year that said there’s probably 150 players in this space now. So, much like Gold Rush, there’s a lot of people coming and setting up shop close to the mines.
Florian: How do you approach maybe the buy versus build equation as a company? Has LXT done any M&A, or is that on your kind of growth strategy going forward or not at all?
Phil: It’s absolutely there in our growth strategy, and we’ve had some great conversations, but I don’t know if you’ve been through this yourself. I’ve been through a few. You have to have a lot of conversations with a lot of people before you actually find something where you can get it across the line. So, M&A requires a huge amount of legwork, going and looking at things, even if you suspect that they’re not a genuine opportunity. Let me take a step back. I’ve looked at successful M&A from a number of perspectives. It can simply be that they do the same thing as you, but they have a bunch of customers you don’t, and that will improve things. They could just be bringing in a monstrous amount of revenue. So at Appen we did an acquisition of a company called Butler Hill. It was a good deal. They had a lot of revenue, but really only had one customer, so they had a saturation problem. We brought them in and it improved our saturation problem, which wasn’t as bad as theirs, but by bringing the two together, things got better. The valuation of the two together was much higher than either of us as individuals and so you can work on that basis. But we’ve also done it for buying technology, and LXT, as I said early on, we’re a services company today, but a services company with ambitions to compete with the likes of Scale AI by being very tech forward. In terms of your initial question, build or buy, we have internal teams working on build right now, but we’re a small organization. The bulk of our effort and our payroll goes towards delivering for clients. So we’re definitely on the lookout for opportunities where we can buy somebody that’s got emerging technology, potentially somebody where the technology would benefit from our huge resources in terms of data. If you got somebody who’s got great technology and no money, they probably can’t actually take it to the next stage because they can’t get their hands on the data.
Florian: What are some of the legal and kind of ethical considerations when collecting data that you’re facing? Just some of the key points there.
Phil: We don’t run into a lot of ethical issues in terms of our data collection. We’re pretty careful about how we do it. You do have to make sure that you’ve got permission from people, but we’re pretty strict on how we handle that. One of the things that we… We do data collection, but we also do a lot of data annotation. A lot of that data annotation is live data, where it’s our clients, it’s their clients, their end users that are providing the data, and I think if you go back five years, Big Tech was pretty laissez faire about how they handled that. They’d send it to companies like us and have it annotated through crowdsourcing methodologies because crowdsourcing was a massive cost reduction methodology. What we’re seeing today, though, is that the Big Tech have been forced to become much more willing to spend money on having the work done in secure facilities, so that’s been one of the real growth areas for us. We have 100,000 sized crowd on our books at the moment, but we’ve seen a real shift towards the need to have people annotate inside secure facilities and we went from having zero secure facilities a year and a half ago to now having five.We’ve got one in Cairo, one in Montreal, and three in the Toronto area.
Florian: Wow, that’s a bit of a moat that you’re building around as well, so it’s not as easy for these 150 additional competitors to just replicate this overnight.
Phil: You have to have a lot of confidence to go out and spend the money on building the infrastructure and your clients have to have a lot of confidence to guarantee you enough work to make it worthwhile.
Florian: Also on data, what we’re seeing sometimes in the machine translation industry, that there’s like experiments going on with synthetic data. The machine generates the data that’s then used to train the machine. Is that something that you’re seeing? And almost like metadata, I know it’s used in a different context, but is that something you’re seeing and building that metadata that’s then used to train?
Phil: At least in my observation, that is very, very strong in the image and video space. Another reason that we’re opportunistically moving into that space, but we’re not driving hard towards it, so we feel that we don’t have that synthetic data capability. We probably need to develop it. We’re looking at that now, so it’s definitely meaningful, but I’m seeing it at largest scale in image and video.
Florian: What do you see in the next two to three years? Where is this all going? Are we at the cusp of like a major boom in AI or are we at the maybe the peak of current hype? Where would you position the industry at the moment?
Phil: No, I don’t think it’s peaked by any means. I think it’s real early days. I think when I look at the data business, the shovel business, if you like, I don’t see that going away. In fact, one of my clients a couple of years ago from a very Big Tech company was asking me to make a significant infrastructure investment and I said, look, if we’re going to do this, we need to have some comfort that there’s going to be a need for data to come, and he said, look, Phil, how can I put this? When you and I are dead, we’re still going to need more data than we’ve got. In terms of AI itself, I think that for major tech companies, it’s still something that’s hard to tame, so there’s some pretty exciting breakthroughs, but it’s a real hot potato. You’re juggling this thing that’s exciting. There’s a lot going on, but you can burn yourself with it. Exactly what you just described earlier with the Google example. So, I think there’s quite a long way to go before this becomes something that is super stable, low risk, and able to be deployed with absolute confidence. I could be wrong. I’ve got a good history of being wrong, so we’ll see what happens.