#293 The Straker Fraud; Astra Translation Performance

0:00Florian:We're moving into a world where these two areas like localization, translation, speech to speech, data, like it all moves in kind of all blends.

0:12Florian:Welcome back to Slater Pod. Hello there, Esther. Hey, Florian. This was our longest break. This has been our longest break since 2019, since we started this beautiful podcast. Not good. Why is that? I don't know. We were in San Francisco on the West Coast. We could have done a live podcast together. We could have done a live podcast together, but we were kind of busy recording, what was it? 10 interviews, 15 interviews, moderating, seven, eight panels, giving presentations, meeting 200 plus people. It wasn't a dull break at all. So, yeah, so again, we're back after a, I don't know, five, six weeks break here, but we do have lots of great podcasts scheduled for Q4.

0:58Florian:I guess for the first time, we're actually doing a solid scheduling job here. So I'm very excited about a bunch of great podcasts with guests coming up. Hey, I want to also make sure that people know that we have spent a lot of time and energy on improving our research subscription offering. so a ton of new features, much improved dashboard, much improved backend so you understand what subscription you hold, what features you can access and so make sure to go to slither.com slash subscribe and check it out and if you are a subscriber, first, thank you very much. Thank you very much. You make all of this possible and go and check out your new account center.

1:44Florian:Really, how long did that take us? couple of months was a long project. Yeah. Yeah. QAing, bug hunting, doing all kinds of stuff. You cannot, I'm learning also, you cannot vibe code this. You could try, but you can't really at the end of the day. There's too much complexity here. Um, so what's your, uh, initial thoughts or maybe not initial now they're a bit more kind of settled thoughts on Slitikon San Francisco and just generally the vibe, the location, the agenda. yeah you would think my thoughts are a bit more settled but I think there was a lot obviously to to process um what I can't remember how many hours but sort of eight hours plus of content from start to finish through the day um but no I mean fantastic location for one thing we had a really good turn you know in the heart of the city um which was fantastic that I the hotel as well I very impressed by the size of the room, all of the screen, the facilities, et cetera, which obviously made some really great networking.

2:47I think there was a real buzz around some of the booths with our partners. I personally went around and spoke to everyone, picked up some merch to bring back. So all of that was great in terms of things around the conference and the content itself. I mean, the agenda was very broad, but also not broad. I mean, it was, I think, you know, you said it when you were summarizing and recap, it was even more focused on AI than it has been kind of naturally. So in that sense, not broad, but broad in terms of the representation that we had on stage, you know, everything from interpreting to language technology to VC investors, enterprise localization and really startups, of course, as well.

3:39So really, I think, fantastic lineup. And yeah, kudos to our events team for getting so many brilliant speakers in one place at the same time.

3:49Florian:No, it was great and I mean, it was a busy two days. So first I joined a panel discussion at the XTM Live, which they hosted the day before. Then we went over to the St. Regis, it's just across the street, so from the Hyatt to the St. Regis to host our welcome reception with Ebru from Olang. I didn't know she was actually, that Olang is actually based almost like in the same neighborhood in San Francisco. So she started the welcome reception on Wednesday on that kind of positive startup founder kind of message. So wide open market, new opportunities, great start, great start. And then, you know, we didn't go too crazy on Wednesday evening and party and all that.

4:35Florian:So we wanted to make sure that we're set for the day, for Thursday, the next day. I started out with my presentation and then moved to a panel with Salesforce, Teresa Marshall and Sammy Altinaev from Freyze and from Salesforce. Salesforce with the big tower just down the block. So it was a good start to the conference and really spoke about how AI is now being implemented at these big globalization teams. And I mean, I guess Salesforce is about as big as it gets in terms of having localization teams. Generally, if you go on X now, I mean, Salesforce is all the rage because they just hosted, what's it called?

5:17Florian:Dreamforce, their conference. Also in San Francisco, they have all the big AI guys, right? Sam Altman and Dario from Anthropic getting interviewed on stage by the Salesforce CEO. Anyway, I'm digressing. And then what I liked about Crisp, the panel with Crisp in automated health systems, it's really, you have like this new, it's a new entrance into the speech to speech translation space. Like five years ago, the company barely existed. And now here they are at Sliticon basically talking to an at-scale client about how they're deploying their solution, right? So I really appreciated that. With ScaleAI, with Emily Hsu, I started to hit the limits of my expertise.

6:03Same. Yeah. I mean, it was...

6:05Florian:It was around, yeah, Data4AI and they launched, ScaleAI, they launched a new framework called READY, R-E-A-D-Y, like all capitalized, a new framework for deploying data into the enterprise, but super relevant, right? We're moving into a world where these two areas like localization, translation, speech to speech, data, like it all moves in kind of all blends. So, yeah, that was a good presentation. I need to rewatch it. And then we had Marco Trombetti from Translated. And I remember the first time I met Marco was like 10 years ago at a panel on NMT in Brussels at, blanking on the name, but there used to be these conferences in Brussels on NMT and well, we've come a long way.

6:52Florian:Here we are with a, you know. Yeah, and intuitively. Literally come a long way for us, you're right. Yeah, so he argued that a lot needs to be done on the quality gap, you need multimodal context, reasoning, self-evaluation. And then he has been speaking about this concept of the singularity for some time and he's now projecting it for the top 10 languages in 2030. But of course, the long-term languages with a lot less data will take longer. And then why don't you give us your impressions from the VC panel that you moderated? You know, we had Koshla Ventures, you know, top tier VC firm and Dell Tech Capital there on the stage.

7:36Florian:Very interesting conversation. So many, so many clips we could have clipped out from that and will clip out from that. For me, it was fantastic. I really enjoyed moderating that panel. It was brilliant really to speak to Kanu and Radhika as well. I did an interview with Canu, so you can watch that one on YouTube if you missed, if you went at SlaterCon. We spoke a bit about their investment. So Costa Ventures invested in Poly AI, London-based startup in the voice AI space. And Dell Technology Capitals also spoke a little bit about their investment in Cartesia. Cartesia. So both of these, you know, VC investors, as well as having that kind of broad AI enterprise technology experience and background, also have direct experience of investing in the voice AI space.

8:27I think that made it really relevant for the audience. And then just hearing their take on, you know, what they're looking for, what makes an attractive investment, what investors are ready to take bets on at the moment I think was was really fascinating yeah and kind of reflecting on I think it kind of echoed some of the themes that we'd had from from earlier panels so like on the voice you know we had the voice AI panel you mentioned with automated health systems and Chris so I think a lot of it was echoing back things that we'd heard maybe earlier in the day, but confirming some ideas to move forward.

9:09Florian:I think one thing that sticks to me is just the challenge these VCs have at the moment, because it's so easy to spin up a very impressive looking demo or kind of MVP. And so how do you really triage deals? How do you make sure that you pick, I mean, even companies that, I mean, it's VC, right? So you're not going to pick the winner in all the hundreds of investments you're doing, but at least like have a fair shot because like, you know, with these tools now, I mean, I could kind of, you know, brute force ChatGPT into spinning up a nice looking website. So probably kind of the track record of the founders is now even more track record or even like, you know, education or whatever, what these guys have done before is maybe even more important.

9:53Florian:So yeah, not an easy environment to be a VC these days, right? It can be quite overwhelming with the number, perhaps, of potential investments. I mean, it's their job, right? But like it's getting even harder than it used to. Okay. Okay. So look, it's a long agenda. I just want to make sure that we highlight all of the sessions. So then we had XTM and eBay, Marco Ratelli and Lorcan Malone, CEO of XTM and Marco from eBay, talking about basically, you know, also AI. And let me just leave you with a simple thought that Lord can share. Basically, you have to build a solution even in B2B enterprise localization that's kind of as easy to use as ChatGPT from a user perspective.

10:37Florian:But it's backed by centralized governments, traceability, risk controls, all of the good stuff that you can get from a highly sophisticated LTP. And then we had David Lee from IUNO, world's biggest media localizer. He spoke about a new product they launched called Chloe, which essentially adds a bunch more context to media localization or AI-driven media localization. He shared a hilarious clip at the beginning where you had like, there was a car chase or maybe, hang on, it wasn't a car chase. It was a car, well, it was like somebody was driving and then there was a woman that kind wanted to kill them by just grabbing the steering wheel and driving into like a wall or something and something like that wanted to cause an intentional accident and then the dubbing was completely emotionless.

11:33Florian:It was like, I will grab the wheel and I will now drive and the other guy was like, oh no, you're taking the wheel. The one I remember was, what are you doing? because the actual sentiment of her partner presumably was, what are you doing? You know, like this is crazy, but like you say, devoid of emotion, it was, what are you doing? Don't drive into the wall. So, yeah. So basically the point being like you need context, you need to understand the scene to make sure that, you know, AI localization works. So it makes a lot of sense. And then we had Coca-Cola, Europe Pacific Partners and Works on stage as well.

12:16Florian:And as I said in my LinkedIn commentary, if any client of Slater wants to give a joint presentation the way Eliza from Coca-Cola did with Works, please email me because that was great. So, they were on stage basically talking about the two-year rollout of works at Coca-Cola and how they brought people on board and how they overcame obstacles and how it works super well now. 40 ,000 users or something. Yeah, 40 ,000 users, right. So, Coca-Cola, Europe Pacific Partners is the bottler, I think, outside of the US, so it's a big organization with a ton of people. So, yeah, great joint presentation there.

12:54Florian:Then we had Marlon Duarte from UCLA Health. I think it's super important to bring in people from language access if you're in the US, from the interpreting side. And he again reminded everybody just how complex an organization and kind of a workflow state they run at these large healthcare systems that again, the European mind doesn't understand how big these hospitals are in the US and these systems. So yeah, we also have an interview online, check it out on YouTube. He also checked, he also mentioned the AmbienScribe technology, which is now increasingly being used in the US, kind of AI assisted documentation where doctors kind of record their interactions with patients and basically, he contextualized this in the context of language access and interpreting.

13:50Florian:So yeah, super fascinating presentation, very difficult to summarize in a few words. We'd go out and check the interview. I think he did a pretty good job. And then Spence Green from Lil closed the conference. So he argued that strong English benchmarks can hide weak AI performance in other languages and markets, which makes a ton of sense that even now with the Astra launch, which we're going to talk in a minute, the OpenAI Astra launch, it's very kind of English-centric still and companies like Lilt with a decade of experience in multilingual are going to probably be well positioned to make sure that this technology is work for the rest of the world, not just for English-speaking countries.

14:35One thing I would highlight as well, because I've had a number of people sort of ask me for recaps or impressions of Slatecon, et cetera, and the video recordings are available for subscribers. So if anyone wants to re-watch, you can in theory, go back and watch the whole conference, you can get that on the website as a subscriber.

14:55Florian:That's a lot of value right there. The entire conference online and London and once before, we probably have now a couple of hundred hours of video content in our replay function for subscribers. We haven't changed prices, no inflation on our subscription prices. So go and get it. Cool. All right. That was SlaterCon. We'll be back next year. Let's see if we're in San Francisco or back in the Silicon Valley or I don't know, maybe somewhere further north. Who knows? Let's see. Hey, so Smartlane, finally, the investment announced this week. They have a new majority shareholders. Shareholders. So congrats to Brian Murphy, the CEO, Jack, the founder, and of course, the battery team that exited.

15:45Florian:Interesting, actually, and we're going to get to the details, but I think it's interesting that the word AI translation is kind of front and center of all the press material they shared this week. Did they disclose valuation and purchase price, Esther? I would say unlikely and no. Unlikely. No, they did not. Unsurprisingly, not, no. So look, I mean, yeah, this is the announcement now. We don't know when the deal actually closed, but they announced it now. I mean, generally very positive to have an announcement like this for the language solutions and the AI market, kind of breaks that deal in investment drought that we've had for a couple of years.

16:22Florian:You know, we know that a number of the successful LSI and LTPs have been trying to get investment and well, here Smartling is and breaking that drought. I think it's always good to see an exit as in, this is not the first time that Smartling had a PE backer, right? So for battery to exit the investment and for a new PE to come on board, maybe signals to other potential PE investors that there is an exit path through in-language solutions. And if you need somebody to look at the due diligence and help you with that, you know where to go. Come to Slater. So, hey, so who invested? So it's a company called, it's a fund, it's a PE fund called Vitruvian Partners, and they are based in W1 London.

17:09W1, okay, that's very nice.

17:11Florian:That's what I read. Yeah, so the London-based growth investor founded in 2006 by veterans from a bunch of other PE firms, including Bridgepoint. Bridgepoint, obviously famous for their language wire investment. Big fund, about$20 billion under management, 90 investments, 10 offices, so quite large in terms of European PE firms. Their thesis, fast growing asset light businesses benefiting from industry disruption. So yeah, I mean, SmartLink takes some of these boxes, right? And so it's a fit. Investment size went on the website. Apparently it's between€40 million and€600 million per investment.

17:54Florian:So SmartLink's probably at the lower to mid end of this. Happy to get corrected if anybody wants to share the numbers with us. I'm quoting it euros if they're a London-based firm. Dollar and euro, it's now kind of similar, right? I won't quote. Pound getting similar too. It's all in the same neighborhood. Target valuation between 75 million and 4 billion, so same here. I'd say SmartLink's probably at the lower to mid end of this. They do a majority buyouts, minority growth investments. I think in this case, it was a majority buyout, right? I think it was. Yeah. Yeah, and then again, Battery Ventures exited.

18:30Florian:So some other companies they own, according to my deep dive on the interwebs is Wise, Skyscanner, Just Eat, Darktrace and Alphasense. Just Eat, is that something based in the UK? Are you ordering from Just Eat? Like an Uber Eats alternative. I think it was founded like before the kind of gig economy, but it's similar. And then in terms of strategy, they're supporting SmartLink with, you know, expanding iTranslation and agentic localization efforts, accelerating global growth and pursue M &A. So SmartLink is going to be on the acquirer side with that new investor backing. Again, congrats to Brian Murphy and the SmartLink team.

19:14Florian:Let's move to OpenAI Astra and talk about what OpenAI's new model called Astra can do for you in terms of translation. But I do want to share before that, I want to share an interesting observation that kind of put me back or took me back a couple of years in terms of AI usage. So I was speaking remote. Unfortunately, I wasn't able to make it to Sweden, but I was speaking to the Sprach Füretagens Branche Seminarium. I'm butchering the second one. I think I might've done okay on the first one. So basically the Language Association's industry seminar, the Swedish Language Association seminar, and I was given a presentation and I was looking to improve the presentation by having AI, the new Austria model, collect a bunch of logos from like Crisp and Phrase and Newphonic and some of the things that I put on the slide.

20:26Florian:And then it just went off and it kind of just came back with some logos and it was totally, except Phrase was actually the right logo, like a logo wall, right? I want to create a logo wall. And Phrase's logo was okay, but then the others are like, that's not Crisp's logo. That's not CaptionHub's logo. And then, yeah, it kind of just made it up. And then I pushed back to the AI. I'm like, hey, are you sure those are the logos? And I think they're wrong. And then I'm reading this now. Chachup Diostro said, you're right. I generated the image from company names without verifying the actual logos or supplying official references.

21:02Florian:The image model filled gaps with plausible looking designs. And I failed to catch that before presenting them. That was my mistake. I'm like, I'm not going to let this thing fly a plane. I'm just not. Okay, not tomorrow, no. Like I filled gaps with plausible looking designs. Like what's a plausible looking design for like CaptionHub's logo? It's like, come on. Yeah. I mean, with a company named like Crisp, you could have like a potato chip or something, you know? Yeah. It probably realized it's a language AI company, but like why not push back and say, hey man, I can't access those logos. Like it's four years into this and it still just goes off and does things and like wants to kind of please you but doesn't admit that I don't have the logos.

21:49It's like people pleasing or sort of like a bad employee who's like, oh, I was too scared to ask a question.

21:55Florian:Right, exactly. Like I read it just kind of, you know, BS here, which is interesting after, you know, four years of training and like all these sophisticated debates around AI and hallucination, all this stuff. Like here we are with the super powerful model. I'm not trying to criticize. Astra is incredible. I mean, I've used it for stuff now for the past few days. It's kind of like, it's incredible the power it has and the things it can do. But yeah, still massive. Yeah. For critical stuff, even in translation, you want to have some type of expert in the loop still, I think at this point, because there was big claims around when they launched GPT-6 Astra, which was a couple of weeks ago now.

22:35Florian:I mean, Chensung Hwang or something. We were there like, okay, thanks guys. So anyway, so Chensung Hwang from NVIDIA said basically, HEI has arrived and et cetera, so a lot of buzz. Now, we had Maria, one of our analysts, look at if they published anything around translation multilingual benchmarks. They didn't, but we collected a couple of independent evaluations on how Astra did overall. So there was RWS's M-Gate, which is a recent benchmark. And then there's this last translation benchmark that we also covered, which has humans tortured AI with really tough translation examples and see how the AI does.

23:23Florian:So let's look at RWS's M-Gate. It covers 30 languages, tests grammar error detection, and how well meaning survives in roundtrip translation, so this is from RWS's Train AI team. Roundtripping means English, another language, English. And so, Austral actually leads in MGate's translation benchmark, so it's significant new highs in languages such as Bosque, Tamil, Tagalog, but actually, you know which language regressed? Which? Finnish. Finnish. Finnish. Finnish, the final frontier for AI. Complicated. Those 25 cases throw you off, even the AI. I would love to speak Finnish. Imagine you speak a language that's just like, it's just generally tough.

24:13Florian:Everyone thinks it's tough. Everyone agrees. Yeah. Everyone agrees it's tough. There's like all this myth around how tough it is. Anyway, so okay. Finnish regressed. Some others didn't. but overall, Astra got the highest overall score, statistically tied with Claude Fable 5.1 and Gemini 3.1 Pro. I'm not using Gemini a lot. Do you use Gemini a lot? Occasionally, but it's not my default. It's kind of weird, huh? Sometimes you copy paste stuff out of a Google Docs into ChatGPT, even though it would have been one click away just to ask Gemini. All right, let's go to the last translation benchmark.

24:55Florian:Again, this is it crowdsources human author cases that have challenged leading translation AI systems and each includes rules specifying which errors a model must avoid. So Astra leads there too, but passes only around 45 % of cases. So still fails at more than half of these deliberately difficult cases and remains far below human performance. So that's not a failure rate for everyday translation. This is basically a stress test, right? The last translation benchmark. We need to get Willem on the podcast. I really want to understand more about this last translation benchmark. Cool. A stress test is also every time you need to announce your financials.

25:40Florian:And we want to just briefly walk you through Appen, AMN Healthcare, and AI Media's results. Maybe you take it. We have a few, maybe we don't need to go through this in great detail, but But basically these three public companies that have a close link to the language solutions and AI market have reported revenues and profits. So maybe let's walk you through this. Yeah, I suppose it's a mixed picture. There was Appen out of Australia that was reporting on its first half 2026 revenue. and so overall up revenue up 17.5 percent to roughly 120 million dollars with four-year guidance actually looking like performance on revenue level will be higher in the second half so they think they'll end the year between 270 and 300 million dollars driving the growth is China they said so China had a massive jump in revenue up 80 percent to 76.2 million dollars when it now accounts for around 64 percent of the total sales and which is it's a majority it's a fair whack so they say demand includes generative AI projects and support for Chinese technology companies that are expanding internationally.

27:10And the remainder of the business or the rest of the business, known as global, remains under pressure. So revenue in that segment fell 26.9 % to 43.7 million, with loss widening to 4.5 million. So, yeah, tough in global, although they did see some signs of recovery in that Q2 revenue actually rose 20 % versus Q1. And they had, they said new projects included coding, STEM, finance and robotics. Yeah, I mean, at that sort of high level, I mean, you can see that obviously the mix is sort of shifting now, China being a much more significant part of the business. They said it benefited from high margin Gen AI work and pre-built data sets.