#276 ChatGPT Translate and Weird Prompts

0:00Florian:The fact that the world's leading AI lab decided to pick the translation use case as a standalone offering, it's pretty big.

0:11Hey everyone, and welcome to another episode of SlaterPod.

0:15Florian:Hello Esther. Hi Florian. We are back with a quick catch up on the news. At least we managed to do it in January still. True. Yeah. Happy New Year. Happy New Year, 28 days in. All right, so we're back before January ends, and SlaterCon London is taking shape. We've confirmed some amazing speakers and panelists already. We're quite early here because it's happening on May 22nd, 2026 in London, so do go get a ticket. Really cool. Let me just go through a couple of the people that we recently were able to confirm for it as panels and speakers. So we have the co-founder and CEO of Gradium AI on a panel.

1:04Florian:So this is a voice AI company that raised$70 million right out of stealth mode, which is kind of insane. I mean, usually when you come out of stealth, you raise, I don't know,$1 million,$2 million. Now they raise$70 million, Paris-based. So Neil's going to join us. Arcadio's from PolyAI. He's the head of agent design and engineering. So Paul AI is an enterprise AI agent for customer service company. It's quite heavy there on the AI. But also we have the head of globalization solutions from Netflix, Alison Anders, joining us. And then Oculek's chief AI product and tech officer, Stefan Cinguino, is going to join us as well.

1:40Florian:And then a quick heads up, we're also doing a research track with posters in a separate room. Remember when we went to that, when was that? like 2018, we went to the EMTA. Was it EMTA? No. I can't remember. E-M-N. Was there an N in there? I'm blanking, but it was some, at the time, you know, it was called like NLP. So it was basically the machine translation track and you have these. In Brussels. It was in Brussels. We went there in person. Yes. And so you have researchers, PhDs, you know, standing there with one poster explaining the project they're working on. And so we thought, why don't we do that?

2:18Florian:So we have a track, I think it's 45 minutes during one of the breaks in a separate room, which I think about 40 to 50 people can be in. And yeah, research. So we have people from the University of Rome. We have somebody from the University of St. Andrews, from Dublin City University, and somebody from Meta, research scientists at Meta, who's going to be explaining what they're working on. So very excited. We also have a startup accelerator confirmed already. Anton Dvorakovich, CEO of Dubformer, is going to join that session. That's also happening in a separate room. So basically, get your ticket.

2:59Florian:This is going to be, again, the biggest Slittercon yet because we even rent an additional room so we can accommodate at least 250 people. So pretty excited about that. But today, let's go through what happened since we last spoke. When was that? Just before Christmas. Just before Christmas. Maybe a couple of weeks before Christmas. Lots of stuff happening. We must have picked up about 300, 400 things in our daily newsletter since then already. But yeah, today we want to focus on hiring a DeepL, AI labs and translation use cases. TAF, translation is a feature. One company putting something out that's quite interesting.

3:38Florian:Of course, we want to go briefly through the ChatGPT Translate launch, which they did very quietly. And then, yeah, you and I are going to talk about a bunch of finance stuff that our listeners will love. Tune out for. No, no, no, no. They're going to stick around for that. Cool. All right. So where do we start? DeepL? Yeah, let's start there. DeepL keeps hiring. Key execs. They also generally keep hiring. Not as quickly as they used to. They're checked on LinkedIn. So their two-year growth was 44 % and the six-month growth is 2%. So, you know, it seems like they're almost... In headcount, right?

4:19Florian:In headcount, yes. Yeah. Headcount, sorry. According to LinkedIn, which, you know, whatever that means. But so now they appointed a couple of new people to their leadership team. Detlef Krause joins his chief revenue officer, CRO, you know, 25 years of experience, enterprise sales, Microsoft, Salesforce, SAP ServiceNow. Now, he succeeds, DPJ, David Perry Jones, who we very much appreciated having on our panel in London back in 2024. So, good luck with what's next, David. The press release says he's actually retiring. So, let's follow him up, retiring from the tech industry. But these types of people never retire.

5:02Florian:So, good luck with whatever comes next. You're a CRO for life. You're a CRO for life. That's right. Right. Gavin Mee is appointed as Chief Operating Officer and he has a background with Oracle, Adobe, Palo Alto Networks, UiPath. Yeah. So CEO Jarek Kutlovsky says that these are very strategic hires for their enterprise growth and that's what it looks like. This is actually following a series of other senior hires, including a Chief Product Officer and CFO. So now seems to have their sea level ducks in a row for an IPO. Where are they all based? Are they based or do we know this? Because I think they had some senior people in or around London at one point, but then obviously it's more a German company.

5:53Florian:Let me look this up. Where are they based? I mean, the background suggests almost the US. Let me see. So Munich. So Detlef Krause based in München, Bayern. All right. That's nice. And then Gavin Mee based in... Oh, now you're going to have to help me here. I've got too many Gavin Mees. No, no, no, no. No, it's London. So Gavin's based in London. All right. We'll be in touch. Cool. So congrats and good luck with D-Bell and please go co-public. We're looking forward to that S1 filing. Now, these AI labs and Microsoft keep putting out data around the usage of AI for translation. So we published a piece on the most recent edition of Anthropics Economic Index.

6:50Florian:And I think we must have gone through this on a previous podcast. I'm just going to go through very, very briefly. It's also fairly complex. They're using a lot of very abstract concepts, which frankly, I haven't wrapped my head around, like, you know, verification, auditing, iterative refinement, and all kinds of other abstract terms about how people use AI. But I guess the most important data point would be that since September, the translation share of clot usage increased from 0.663 to 0.71%, so a modest increase in translation-related activity or of the proportion of which probably more people use clot, so generally a strong increase.

7:37Florian:And then our analyst looked at this and detected that AI's role in translation, unclothed, always say unclothed, which is their tool, is stabilizing around a specific pattern, which we, and I'm going to have to read this, said it's going to be, is automation for clearly bounded execution tasks. So actually doing just the kind of raw AI translation. Augmentation for review, validation and meaning level work. So check my translation, you know, maybe give me a bunch of synonyms and things like that. And human in the loop control where context, judgment and risks matter. So there I'm going to just defer to the actual article, go and check it out.

8:21Florian:All right. But thanks to Anthropic for putting this out publicly. very interesting data and for those who care, you can go and read it. Now, Microsoft also put out a report on Copilot usage. So what do we find from there? So Copilot, we're not using it, we're on Google, but I think a bunch of people use Copilot for translation related tasks. and they analyzed 37.5 million anonymized copilot conversations between January and September 2025 and apparently translation and language learning ranked fourth overall. So, I mean, I wish we could have split up translation and language learning. It kind of says a lot about like that they're putting them in the same bucket.

9:13It's like, ah, Duolingo. Translation, language learning. That's it.

9:19Florian:So, yeah, you could have probably taken them, kind of looked at them separately. Maybe it's because you don't really know the intention of, like you could use the same command for language learning that you would use for translation. Like, translate this for me. It's kind of like, well, are you translating it because you need to, I don't know, translate it? Or are you trying to learn the language? because obviously you don't know the language if you're asking them to translate it, I imagine. Good point. Yeah. There's so many other language learning. Yeah, means, tools. Yeah, tools. I mean, let me look this up because actually I think we're working on a story or did we publish it already?

10:04Florian:What was it called? Preply, Preply, Preply. Oh, okay. So, yeah, they raised a$150 million Series D company called Preply or Preply. Preply? Preply probably. Preply? Like Reply? I don't know. Preply. All right. Well, it's got some naming issues. Short name, but hard to know how to pronounce it. So, they raised a bunch of money for that. So, basically, translation language learning, go back to Copilot. Translation language learning ranked number fourth behind tech, work and career, health and fitness in co-pilot usage topics. And this excludes program languages. Thank you very much. Yeah. So this also, go and check out the original research.

10:51Florian:All right. So it must have been like 18 months ago when we had this, when we wrote this research report on translation as a feature. That was kind of the first wave of integrations of language AI into all kinds of products. And now, yeah, there was a big-ish launch with Adobe adding it. I mean, how many times in the past, Esther, have you, in your LSI past, how many times did people ask to translate PDF files? Many times, yeah. And then we'd ask for, I suppose, either, you know, like the underlying Word file or the InDesign file or whatever kind of source it came from because tricky. And the worst is, I suppose, sort of like scanned PDFs or scanned whatever.

11:38So, yes, many, many times.

11:41Florian:Now, the scan is still an issue, but one of the – so we're talking about Adobe launching literally translate this as a feature in their Adobe – I think in Adobe – what's it called? Reader, Adobe Acrobat Reader. So you open it up and you get a feature. Let me just open this up now. Yeah, and it says translate this, all tools. Yeah, translate this PDF is what it's called. And when you click translate this PDF, it opens it up in Adobe Express. So it kind of opens up a new SAS window, a new SAS tab. And apparently it does 40 plus languages. It can customize the tone, but it's very basic, kind of formal, informal, the usual kind of prompt stuff.

12:29Florian:But it has formatting issues, believe it or not. And interestingly, we found that it struggles more with editable PDFs than with scan PDFs, which is super weird. In terms of language or formatting? Formatting. Okay. Because that's always the problem, right? I mean, if it can extract the language properly, I think it's going to be a good AI translation, I guess. but the formatting is tough. All right, but the bigger angle here is that this is another development and it's kind of good enough.

13:12Florian:There's so many of these translation as a feature launches that now I think there's a bit of a tendency and we've had some conversations with LSIs about that that sometimes more and more end customers would just say, well, we're just going to use AI for this. We're going to skip the LSI. It's good enough. And there's a bit more of an acceptance of suboptimal AI translation now because it's everywhere, right? So people feel a little less guilty using it even in a professional setting. All right, let's move to something else to the company powering most of this is NVIDIA. So what did they do? Launch another framework-ish model?

13:57Looks like it, yeah. Double down on open speech AI.

14:02Florian:Yeah, and they focus on low latency, automatic speech recognition, text-to-speech. They push for open models and data sets. Of course, it supports real-time multilingual voice use cases. Now, I want to see those real-time multilingual voice use cases. Can somebody show it to me? We've been testing it and it's just not there yet. So I want to see how all these models are going to get used in actual applications. But the core here really, it positions NVIDIA as this kind of enabler of the SpeechEI ecosystem. NVIDIA is just everywhere. I mean, they fund a bunch of startups and they open source their models and their frameworks and all of that.

14:46Florian:Of course, they want to make sure that people stay in the larger NVIDIA kind of ecosystem that in turn drives demand for the chips. So, yeah, that's NVIDIA. Now, the big one, ChatGPT Translate. Have you tested it? Not yet. No, I mean, I'm sure I will, but no, not yet. It's just very basic. It looks like a, I don't know, 2016 NMT interface. It's literally like translate with chat GPT and then source on the left and target on the right. And then it has like four prompts below that you may choose to use, like translate this and make it sound more fluent. Or translate this and make it more business formal.

15:38Florian:Translate this. That's literally the standard prompts that they're showing. Translate this as if you're explaining it to a child. Oh, yeah. All right. Shall we unpack this? So first of all, translate this and make it sound more fluent. What an odd prompt. Sorry, I'm going to have to dwell on this for a second. Let me just... Go on. But you can also do your own prompts, I guess. Of course. It's not like you only have four presets. You have flexibility to instruct as you wish. Yeah, of course. Yeah, yeah, yeah. So, but. I'm going to have to add a but. Translate this and make it sound more fluent.

16:15Florian:More fluent than what? Well, I knew you were going to say that. Than the literal translation, I suppose. But it is odd. No, no, no, no. That makes no sense. Like, okay, you're prompting the actual translation. Yeah. So more fluent than the original? Ah, maybe. No, it can't be. This is not a post-editing prompt. This is a translation prompt. So you have German and then you copy paste it in this thing. Yeah. And then you make it more fluent than what? Like your, like, chat GPT, like your original translation? Isn't it like a two-part thing? So it's like translate it, that's part one. And then part two is make it more fluent.

16:58Florian:Yeah, but what's it comparing to? What's the more fluent? More fluent than what? Than that interim. Than it would have been if you just said translate it. translate it and make it fluent or make it sound natural or something, but make it sound more fluent. I'm just having fun because I'm detecting something that's absolutely not logical on the ChatGPT. We are splitting hairs as well. No, this is not splitting hairs. This is the AI revolution number one company. Okay, there's some four, three or however many predefined prompts. You don't have to use the prompts. The prompts themselves or the presets are like prompts to us humans about what you might want to tell ChatGPT about how to translate.

17:43Florian:I get it, but I want to split the hair because it makes no sense to say, make it sound more fluent. It makes no sense. Or make it more business formal. That also doesn't make sense. It just doesn't make sense. You're prompting the translation. you're not prompting you're not asking ChatGPT to make an existing translation sound more fluent. Put it in the comments on YouTube. Comment like what do you think? I think it makes absolutely no sense for this. It's not logical it's just a little example prompt. I get it but you know it's on ChatGPT.com slash translate and it's probably viewed by millions of people so why would your very first template prompt make no sense But all right.

18:28Florian:What does our audience say about ChatGPT Translate? So 45.6 % are calling it meh, so not very impressed. 28 % are saying it's an interesting launch and 26 % are calling it a huge deal. So I think it's a big deal. Yeah, I feel like the meh could be, well, the translation's not good. Great. ChatGP Translate is pretty good, especially on these kind of basic things. So 45 % of people think it's mere that ChatGPT has its own translate tool. It's like whatever. Yeah. Okay. Exactly. I would have expected more, I guess, yeah. Okay. We've got, what's that 50, 54 in total percent of people thinking it's either interesting or a huge deal.

19:18So.

19:19Florian:Correct. Fine. All right. We can, I can live with that. So we had some reactions on LinkedIn. So, Beer Works, head of marketing, Rodrigo Dimitrio, he said, I think they just vibe-coded this interface in a day. So, that's number one. They had some intern type the presets, obviously, Florian. Correct. And then from Blackbird, he said that he thinks this is not as core a product, but they launched not as a core product, but as the UI layer that resembles kind of Google Translate and DeepL. And he also highlighted how susceptible it was to prompt injection. So interesting post there. And then ServiceNow, Stephen Holmes from ServiceNow, he considers the launch as a reaction to Google's recent acceleration in AI and OpenAI needing to jump on the whole stack story.

20:17Florian:So the whole stack would be like, you know, in terms of Google, it would be infrastructure of Google Cloud and you have like the model hosting foundation models and applications and then like in applications you have to Google Translate stack so so yeah I mean our take is I think no matter how basic the UI currently is I think the fact that the world's leading AI lab decided to pick the translation use case as a standalone offering it's pretty big pretty big it just shows just how much demand there is for AI translation yeah I would I would tend to agree And I think it's a way to get more visits, you know, get more clicks, get more visits, get more time on the platform or on the domain.

20:56Florian:They need to make more money. They need to make money. Well, they need to make money. I mean, I guess they have, you know, they have all the trillions of the world, but at some point you want to show revenue. And yeah, seems to be something. Why else would they have bothered to add this as a standalone, you know, slash translate thing? Okay, so let's go to M &A, back to familiar shores. And you have a couple of acquisitions that you want to talk to us about in Italy and the U.S. In the U.S., yeah, just a couple that we picked up on that we've covered so far in January. So I think first out of the gate, probably in 2026, was Aglitec 14, an Italy-based LSI, which acquired Centrolingue SRL.

21:46So it's acquisition of another Italian LSI. It seems to be sort of on the smaller side because we're talking about sort of assets transferred, part of which is one employee, as well as translation memories, term bases, and a few other things as well. And yeah, so just to highlight that one, kind of a little first out the gate acquisition in Italy. And then we had something over in the US in a totally different field, which was more in the world of media services coupled with localization services. So this is a company called Cineverse Corp, which announced the acquisition of Giant Worldwide, a media services company specializing in content delivery, mastering and localization services.

22:39the idea here is that Cineverse has a product or a platform I think called Matchpoint and Giant is going to be sort of integrated into Matchpoint with its technical services the sort of client and studio relationships with the aim of expanding the AI automated media services of Matchpoint they describe this I mean I wasn't familiar with the platform previously they described Matchpoint sort of as a fully automated supply chain that does digital entertainment delivery so sort of end-to-end automation of the supply chain around delivery of entertainment but yeah they because Cineverse is listed on publicly listed on the Nasdaq there they I think disclosed a little bit about the sort of metrics well at least the financials of giant so they said it's expected that giant and so the acquired company will contribute 15 to 17 million in pro forma revenue um in fiscal year 2027 and also around three and a half to four million in pro forma ebitda so profitable huh tells you a little bit about this yeah sort of size size of the business exactly i mean i think yes um and then yeah looked up cineverse because it's not something I'm super familiar with.

24:03Looks like they have a market cap currently of around 42 million.

24:08Florian:Yeah, but like 25, 30 percent, you beat the margin. Not bad, not bad. Sure. All right, now let's move on to funding, which there's been a lot of activity that is not trivial to assess. So we struggled before the podcast to kind of agree on how we're going to be talking about this. but our friends at Synthesia, who, you know, who CEO was on the podcast, which feels like many years ago, many eons, many ages ago, they raised a giant Series E$200 million at a$4 billion valuation. So, yeah, why don't you give us the highlights there? And then we try to unpack this for, yeah, just a little bit how it matters, actually.

24:54Yeah. I mean, I think, yeah, you kind of identified the highlight, which is A, the amount raised and B, the valuation. So$200 million raised at a 4 billion valuation. it's well I would say from my standpoint one of the highlights is that it's sort of a UK and London based company and I think you know it seems like not only the investors but sort of London and the UK is somewhat embracing them in the sense that you know even within the the press release you've got quotes from several British politicians sort of singing singing the praises and saying how happy they are to have the likes of Synthesia in terms of the scale of the company and sort of how they're leading in AI as well.

25:45So I think they're trying to make a bit of a splash and equally you've got like the Mayor of London kind of commenting on how this is a big deal as well as the Chancellor of the Exchequer. So that's kind of, it's quite interesting and not something you see in a press release sort of every day, at least not the ones we're looking at. I'd say other highlights. I mean, in terms of the investors, you've got the round being led by existing investor Google Ventures. And then, I mean, speaking of NVIDIA being everywhere, you've got N Ventures, which is also another existing investor participating in the round along with many others.