[00:00:20] Nathan Wrigley: Hello there and welcome once again to the WP Build podcast. You've reached episode number 436 entitled. How AI is helping WPML translate your WordPress website faster and cheaper. It was published on Thursday, the 11th of September, 2025.
My name's Nathan Wrigley, and I'll be chatting to Amir Helzer from WPML in just a few short moments. But before that, a few bits of housekeeping.
Something that I don't ever speak of is my involvement in the WordPress London community. There's an event which is put on every month. It's called WPLDN, and it's a meetup for the WordPress folk in and around London. Well, it just so happens that this month's event coincides perfectly with another WordPress event called LoopConf.
You can find Loop Conf by googling it L-O-O-P-C-O-N-F. It's a one day event happening on the 25th of September, so that's a Thursday coming up soon. But WPLDN, which is an evening meetup, it's happening the night before. What more could you want? Head to wpldn.uk if you're in the London area and you want to find out more about that event.
But also, if you would like to make use of a 50% discount for the LoopConf event the following day, then you can go to their checkout page and use coupon code WPLDN. How nice is that? So you'd be able to come to us on the 24th, that's September the 24th, and then stay over the night in London and join us at LoopConf on the 25th. So that's really nice. I love the idea of that. So there we go. Check that out. wpldn.uk, and also you can Google LoopConf.
The next thing to mention is that if you like what we do at WP Builds, why not subscribe? Head to wpbuilds.com and subscribe on your favorite platform. Maybe that's X or Bluesky or Mastodon or YouTube. We publish all of our bits and pieces over there as well.
And maybe you would like to get more regular updates from us via email. There's an email form on there too. Fill that out and we'll send you two emails a week. One when we produce this, which is the Thursday podcast. It's typically an interview, but we also do our This Week in WordPress show. That happens on a Monday, 2:00 PM UK time, and we publish that the following day on a Tuesday as an audio podcast episode.
If you join us live for that, you can make comments. Many, many, many people join us and it's really fun. Normally there's a panel of three other guests.
If you'd like to be on that panel. I'm opening it up more generally. Head to wpbuilds.com/twiw. So that's the acronym for This Week in WordPress, which is the name of that show.
The very last thing is if you really like what we do, and you would like to advertise on this podcast and get your messaging out in front of a pretty large WordPress specific audience, head to wpbuilds.com/advertise to find out more. And here come three companies that are making use of that service and doing just that.
The WP Builds podcast is brought to you today by GoDaddy Pro. GoDaddy Pro, the home of managed WordPress hosting that includes free domain, SSL, and 24 7 support. Bundle that with The Hub by GoDaddy Pro to unlock more free benefits to manage multiple sites in one place, invoice clients, and get 30% off new purchases. Find out more at go.me/wp builds.
We're also joined this week by Bluehost. Bluehost, redefine your web hosting experience with Bluehost Cloud. Managed WordPress hosting that comes with lightning fast websites, 100% network uptime, and 24 7 priority support with Bluehost Cloud, the possibilities are outta this world. Experience it today at bluehost.com/cloud.
And last, but by no means least Omnisend. Do you sell your stuff online? Then meet Omnisend. Yes, that Omnisend. The email and SMS tool that helps you make 73 bucks for every dollar spent. The one that's so good, it's all most boring. Hate the excitement of rollercoaster sales? Prefer a steady line going up? Try Omnisend today at omnisend.com
And sincere thanks go to GoDaddy Pro, Bluehost and Omnisend for their continuing support of the WP Builds podcast.
Okay, what have we got for you today? Well, as I mentioned at the top of the show, we've got Amir Helzer from WPML.
People have been wanting to translate their WordPress websites for a long time. So you could, for example, have it in Japanese, German, Chinese, English, French, Spanish, and so on and so forth. But in the past, that process has been, well, let's say fairly manual. Plugins like WPML have helped you with that process, but it's often had quite a lot of human interactions. So you've had to send things off to translators, make sure those translations are correct.
And it turns out that now AI is in the mix. It is incredibly good at doing really robust, accurate, very quick, very cheap translations. And so guess what? WPML has integrated that into their product offering.
So Amir Helzer, the founder of WPML, is on the podcast today to talk about that. We talk about the role of AI in translation. We talk about how modern translation differs from what you might have been using in the past. We talk about how WPML has used new technologies to integrate it into their offering. How cheap it is. What is their new Private Translation Cloud, API or service, and how you might set this up on your website.
Honestly, if there was one area where I think AI is a really credible service, translations may very well be it.
So check it out on the podcast today. I hope that you enjoy it.
I am joined on the podcast by Amir Helzer. Hello.
[00:06:06] Amir Helzer: Hello, Nathan.
[00:06:07] Nathan Wrigley: Nice to have you back. It has been many like years. Right when we started WP Builds. I actually think, Amir, you were my first ever interview I think. You may not wish to remember that it was a, long time ago, but very nice to have you back.
Amir joins us from, the email address is On The Go Systems, but the, products that you probably know are things like WPML and Toolset and a whole bunch of other things.
I know it's a bland question, but would you just introduce yourself, give us your little bio, what you've been doing in the WordPress space, companies that you work with for products and brands and all of that?
[00:06:45] Amir Helzer: Okay, so I started with WordPress about 16 years ago. and our first product. Needed something like WPML. So WML was something that we developed as a small utility for a project that no longer exists for me. and, we've gone since then. So we created a tool set plugins also. And, recently a lot of our focus has shifted into translation, the translation itself.
So at the beginning, until recently until AI came along. Eh, we, were interested only in the software, in, in the tool that allows to build the multilingual side. And recently we've become very, we've become very interested also in the translation in the product. Not only allowing user translate, but being the translator itself.
we are interested in that and our clients apparently are interested.
[00:07:42] Nathan Wrigley: The, I guess the thing that's changed here, and you mentioned it there, is the word ai. it, barely feels that you can do anything these days in tech, without mentioning ai. But, I have my reservations about AI more. Broadly, but the, bits that I've enjoyed of ai, the bits where I see real utility is when the AI is walled in.
And what I mean by that is I'm not asking it to create something brand new. I'm giving it something and saying, here's the thing, tell me about the thing, or rephrase the thing or well translate the thing. And it would appear that AI is profoundly good at that kind of walled garden approach. I dunno what your thoughts are on
[00:08:27] Amir Helzer: So interestingly enough, I read a few months ago that this technology, that Transformers technology, which is the, basis for the lms, actually Google developed it. in order to translate, translate better, they had the Neur translation system, what we use as a Google translate, and it, works fine, but, people, they have researched how to translate more naturally, meaning to actually understand the text so that you translate after you understand, you don't try to translate the text, you try to translate what the text means.
Slightly differently. And, this is why smart people in Google developed this technology for their translation software because they were in the automatic, in the machine translation business. And I think that someone pulled the, plug from that project when they realize that this does not only translate.
[00:09:25] Nathan Wrigley: Okay. What? Because it's giving something adjacent, like it might change the meaning of it or
[00:09:32] Amir Helzer: they pulled the plug because someone sensed that it's intelligent, that
[00:09:36] Nathan Wrigley: Oh, I see.
[00:09:38] Amir Helzer: they, started the project because they wanted to translate from this language to another language, and they wanted to do it way better than Google translate. And then someone got spooked because he knew, he, had he, he got the feedback that it appeals to be an intelligent thing.
You can actually have a conversation. It's got opinions, it has plans, it wants to do stuff. Now a lot of this is imaginary. It's because it's so great at language. Then we think, oh, this, thing has plans, this thing wants to, com wants to predict the right answer. Okay? but this was as far as I know.
This was the, way they designed it, so how to translate properly. So actually us using it as a translation system, that's the original purpose why it was developed and this is why it works so well.
[00:10:26] Nathan Wrigley: I do remember, I, I dunno if this is the same story that you are describing that I'm about to recount, but I do remember a Google employee. Who, said exactly that? He went out and, I don't know, he published a blog post and he went on various interviews and things like that. And he was saying that he thought he'd encountered an intelligence of sorts, like a GI, we might now call it artificial general intelligence.
He thought that it was. Conscious for want of a better word. and then I think quickly everybody realized that wasn't the case, but Okay. That's really curious. What's also curious about what you are saying because I do, I speak to quite a few people on this podcast and I get the, I get an AI to make a transcript and it will faithfully transcribe what the people are saying.
So it does it word by word. Now, in the case of somebody who is a native English speaker, that's probably. Exactly what you want, because they're doing it correctly. However, I might speak to somebody who has an entirely different language, and they will just phrase things differently. the verbal will come in a different place and the, how it works.
The sentence won't be constructed in the way an English speaker would, but the translation engine that I have doesn't worry about that. It just translate whatever the heck it hears. But what you are saying, I think, is that more modern. Translation engines will take that and, be able to say, actually, do you know what?
That would be so much better if it was rephrased in this way. Is that what you're
[00:11:57] Amir Helzer: This is how an LLM translates when you give an LLMA bunch of text. It doesn't even try to map the words. It tries to understand what you're talking about and translate.
[00:12:08] Nathan Wrigley: Okay. so pres, presumably there's some close approximation to what you say and what it actually translates, but it's also, in my case, the ones that I have used for the podcast, it's in incredibly good, distilling that information. So you give it 10 paragraphs and you might say, I don't know. Cut it down to one
and, for reasons that I don't understand, because I don't understand enough about ai, it appears to be incredibly good at taking just that data and distilling it.
And somehow it makes a really good stab at just saying what the gist of it is. And and I guess that's what you are leveraging in your product. Yeah. Okay. So let's just reprise what you've said so far. So WPML, which is a translation plugin for WordPress. How long has that been around? it feels like probably the original one.
[00:13:05] Amir Helzer: I think it's 16 years old now.
[00:13:08] Nathan Wrigley: so 16 years you've had the plugin and it evolved with the technology that was available at the time, and by that I mean there was a definite need to have humans in the loop at the beginning. I presume that there was not much in the way of Google Translate when you began the whole
as time gone.
[00:13:27] Amir Helzer: recommended to people to use machine transition only as as a filler. Okay. So they don't stare at an, at a blank page. They have something which is written in, Italian, in Germany, in Hebrew. And then they can edit. And we, communicated them that they, would need to do a lot of editing, but it would save them some, time.
So they don't need to type in everything from scratch.
[00:13:56] Nathan Wrigley: Okay. So it was just used as a first pass,
[00:13:59] Amir Helzer: Yes, exactly.
[00:14:00] Nathan Wrigley: like a, you scribble it down and okay, it's not ideal, but at least it gets you part of the way and then, a human needs to be involved. but now I'm guessing the year 2025, the advice has now flipped. I'm guessing that you don't need a human, I hate to say the words, but I'm gonna say them.
You don't need a human being in the loop here.
[00:14:23] Amir Helzer: it's always nice to have a secondary. It's nice, but, in my case, for instance, I, I'm a native, Hebrew speaker. I speak okay English, and I'm native in Hebrew, and this thing translates from English to Hebrew. Way better than I. Okay, so I'm not an a professional translator, but I speak Hebrew and I know the topics.
So when we try to, when we translate our own website from English to Hebrew, it's gonna come out better. If we write it properly today in English and we let the LM translate to Hebrew, it's just gonna come out better than me writing it from scratch in Hebrew.
[00:15:05] Nathan Wrigley: Okay, so what we're basically saying is the AI has now reached. Human or better levels and I guess the big bonus here is that the AI can do all the languages, whereas previously, if I needed to go from English to Hebrew, I needed, I could have used the Google Translate, but then I needed somebody who was an expert in Hebrew to.
Double check it. and then if I wanted Spanish, there's another, and then Italian and the list goes on and on we needed lots and lots of points, but now presumably you more or less click a button and it does them all. not simultaneously, but approximating simultaneously.
[00:15:49] Amir Helzer: So we're, translating the fullest between 52 languages. I think the whole metrics. It doesn't matter which per we QA, the whole thing. we spent months queuing the whole, thing. there were a bunch of issues to deal with for each target and source language.
but this is why, me and others, we enjoy so much working with AI because once you get it working correctly. Then it works correctly. You don't need to repeat it, you don't need to overstate it. You don't need to remind it, remind, remind it to do stuff. when, you walk with a team of humans like me, okay, someone tells me to do something, I get it, and then over time I start forgetting.
I start being distracted. I start. the level is going to go up and down. How well I've slept at night, how many emails are waiting for me, what's my phone doing at that time? It's hard for me to keep my own level consistent, but once I qa the LLM, that's it.
[00:16:58] Nathan Wrigley: It is crazy to think about, isn't it? Yeah. So if you, put in an article at three in the morning, you don't have to wait until nine the following morning to communicate with the translator by 3 0 1. It's done 52 languages.
[00:17:11] Amir Helzer: then
[00:17:12] Nathan Wrigley: just, it's crazy.
[00:17:13] Amir Helzer: yes, and you teach it once. How to deal with a WordPress constructs like short codes, attributes, HTML and coding, little short WordPress constructs, placeholders, what have you. You teach it once. You teach it once. how to deal with the SEO attributes. You explain to it this little thing, which called the meta description needs to fit into this length.
It needs to preserve branding, it needs to do all this little stuff. And you, catch these things as you go. Okay? So you run testing, you feed it several thousands of pieces of content. You review the output exactly like the people who train the LMS do. Okay? You review the output, you, categorize the issues, you fix the problems.
Okay. You just fix them and then it remains fixed, and then you can fit it as much more content as you want. And it remains fixed. It doesn't revert.
[00:18:11] Nathan Wrigley: gonna forget to put that SEO tag in the, it's a hundred percent of the time, which I still forget, in my own English work. I forget to put those kind of things in. But the AI will reliably do it a hundred percent of the time. So let me just get the name of this product right, so anybody can go out and Google it.
what you've written in the show notes is, that you're calling this, you said it was w. WP ml's, new AI translation. Is that the product name?
[00:18:37] Amir Helzer: No, we called it, private ization cloud. PTC.
[00:18:40] Nathan Wrigley: PTC. So it's WP ml's PTC or Private Translation Cloud and the PTC bit. what's the p, the private in there.
What,
[00:18:52] Amir Helzer: because it knows your material, it isn't exactly like you would walk with a transactional, service like Google Translate, which provides the same answers to everyone in Google Translate or in Depot. you send it a string, someone else sends it a string. you'll get more or less the same reply here, it's customized to you, so it knows your context, it knows your website, your material, it knows with your target audience what is, your correct, tone of communication.
And it's like working with your own private expert translators, AI translators who know your stuff. So it's not a, gen, you don't get the same translation like someone else will get, you get translations, which are quick for your case Exactly like you would do if you, let's say you hire a translator, you train them, you, give them feedback.
you spend months mentoring them and integrating them into your business. That's the same experience here.
[00:19:52] Nathan Wrigley: Yeah, so what's interesting there is that there's something approximating humanity, in a sense there. It's trying to replicate your voice and does it do the same on the translation as well? So in other words, let's say that I want to go from English to Spanish. Does. How to describe it, will it retranslate things so it's not a, letter by letter or word by word faithful transaction.
So let's say that I rattle off a paragraph and it's very technical, and in the, olden days, maybe Google Translate would do that one sense at a time, one word at a time, whatever. Is this more a case of it's trying to capture the
essence of what you are saying and put it into the. The most appropriate Spanish way, for example, or Hebrew way of saying that exact thing.
So in other words, it's not a faithful translation, but it's, a more human, if you like, which is an odd thing
[00:20:46] Amir Helzer: It's the transition that you want. It's,
[00:20:49] Nathan Wrigley: Uhhuh.
[00:20:50] Amir Helzer: it appears as if someone who's very familiar with the topic wrote it originally in that language, for that audience.
for instance, a Spanish transition or Spanish text. Tend to be much longer than English because this is how society prefers to communicate in Spanish.
I go all climbing. Okay? yeah, I know I'm heavy. I shouldn't be doing this anyway. I go all climbing. Do not worry. The wall is safe. I haven't heard the wall yet. there's a sign at the bottom, okay? That says, warns you that someone's is climbing above you. Okay? in English it says, be well climber above.
Okay. Or just climb above. And in Spanish it says
[00:21:42] Nathan Wrigley: Okay.
[00:21:43] Amir Helzer: So it's, at least twice as long. And this is how it is. Yeah. Just this is just how it is. so when you would translate with an LLM. You'd get this kind of language. Okay. German and Spanish would be longer and typically more formal. in order for the Spanish transition to be informal, you need to tell it that, your, it's, let's say it's still describing a video game.
It's a website for video game. Otherwise, typically it'll figure out that it wants to translate a little more formal. Okay? it understands expressions. Okay? If you have an expression, like a word of mouth, it's not gonna understand it that way. It's not gonna think about it like word and mouth. Okay?
It's going to understand that people are spreading information from one another, okay? Because these are expressions that it knows. It knows it very well. This is why also, when we translate software with that, or we've got PT PTC for websites for WordPress and PTC for software now. So we've split it the same engine translates software and websites, in the same level.
So when we translate our own plug in WPML, okay, we learned that we, can reduce the glossary Okay. Just by explaining to it. That WPML is a ward post plugin for building multilingual sites justice. it allowed us to eliminate something like 200, 300 entries from our glossary. Yeah. because it knows that this is a, ward plugin, so it knows that when it says, something like, taxonomy, it knows how, taxonomy is referred to in Spanish for WordPress.
Okay. it knows that, what a post is. It's not a post office, it's a blog post. So when you. We previously, we had to use, a large taxonomy, large glossary when we were translating with human translators because it's very difficult to find a translator, okay. Who is intimately familiar with all the terminology of WordPress in different languages.
It's really difficult. The, terminology is long and it's complicated. It's detailed, I would say. You need to address all the ui elements as they're called in the CMS. Some things you can translate them several ways, but there's one way that WordPress project shows, right? so in PTC when we translate our own plugin, WPML, we've, we saw that the translations come out way better when the, when we TriMed out all the world plus terminology.
And we stayed only with our stuff. Okay. for instance, we are telling the, we're telling, the translate or now PTC we're telling it. Do not translate the thing that's called transition dashboard. It's not a generic transition dashboard. It's an element in our user interface and we want it to appear as transition dashboard in Spanish, in German,
[00:25:00] Nathan Wrigley: Yeah.
[00:25:00] Amir Helzer: know.
Okay. I, but we don't need to tell it anything about the word post terminology. It's fine. We just stated it's a word post plugin.
[00:25:11] Nathan Wrigley: And off it goes. It figures it all out. That's just utterly incredible. You, mentioned a bit ago, so I, just wanna unpick how bit aspects of this work, and I've got lots of questions. And the first one is, you mentioned that. it, it learns about you, I think you said brand and about your website and so on.
How does it do that, is that a process where at the very beginning of your relationship with WP ml's, PTC, you have to, I don't know, go through a wizard or something and tell it about you? Or do does it go out and scrape the website? how does it learn what you are and who
[00:25:44] Amir Helzer: started with scraping and we discovered that it's better to just go and ask. So we ask you three questions. What's the name of your product? Okay. Give a one sentence description of your product and a one sentence description of your target audience. So in the case of WPML? Yeah, our target audience, our people building world website, our website developers.
And why is this? Why does this matter? Because you'd want to use different tone of voice, different formality if you're translating something for teams or something for professionals building sites. And that's it.
[00:26:19] Nathan Wrigley: And do you get to, do you get to go back and modify
[00:26:22] Amir Helzer: You can if you want.
[00:26:23] Nathan Wrigley: ah, you know what? I didn't quite get that right on the
[00:26:25] Amir Helzer: so you can, if you want, we ask you in the setup whistle and, now we have a little project that is in progress. We haven't, it's not deployed yet, but this is gonna be in the next release that we will, we'll, working on so that, if you modify that, okay. you'll be able to experiment and modify these descriptions.
retranslate, but we will cache, we'll remember the previous transitions that you did with previous settings. So if you look at your, site and say, no, I prefer the previous one better, then we don't have to charge you again for changing the whole translation. Let's say you've got 20 pages and you just updated everything with a new description of your target audience.
So if you want to experiment, we will charge you for. different transitions, but we'll make it very easy and, free for you to go back to the previous transition that you had with, a d with a previous definition of the audience.
[00:27:30] Nathan Wrigley: Okay. Yeah. I can imagine in most scenarios you're probably gonna, because it's your business, you're probably gonna get it right on the first go. But I appreciate that there's a way to store a historical version of, okay, this was
[00:27:42] Amir Helzer: There, there's a learning curve. I wish, but there's a learning curve. We, are doing this because we discover that this is a new thing for everyone. And like, a carpenter needs to learn how to use his tools. It doesn't get it right on the first attempt, even though it's, a screwdriver.
or so our clients, we ourselves needed several iterations. To use these things really well. So the first iteration's not bad. It's, it looks very impressive. But then you, become better at this and you say, ah, maybe if I explain this and this, I'm gonna get exactly what I need.
[00:28:20] Nathan Wrigley: Okay. So there's the option to refine that over time. Okay. that's lovely. So I suspect there'll be many people listening to this who've never used a translation plugin. And so the whole ui ux of it will be a mystery to them. and I don't know if the, previous way that you were doing things and the sort of metadata and the way you were storing that in.
Blog posts, for example, were, is still the same? Is this the same? Do you specify languages in like metadata? H How does it work? how do I, let's say for example, I've, I'm typing in English and I have a, desire to type it in, to translate it into five different languages. how would I specify those languages and what would the process be to make it happen?
Each time I publish a post?
[00:29:02] Amir Helzer: so when you set up WPML, you go through a wizard and you provide answers. for instance, you pick the languages that you want, and WML would store this in the options table because it doesn't scale. It's one piece of information. and then the languages that WML adds to posts. Eh, WML uses its own table, not, post meta table so that we can index it and we don't bog down the whole system.
so WML creates, its own tables for adding language information and adding the connections between languages. but then the transitions themselves are stored in the regular WordPress tables. Because if you translate, if you write, let's say, natively in Spanish and you translate it to Italian and German.
Then you'd have one post, in Spanish, one post in Italian, one post in German, and WL sets the languages and connects between them. this is, how WL worked from the beginning. It's fine. We don't need to change that. the, and then your, user experience. What we recommend people to do today, 2025.
Probably gonna be different. 2027.
[00:30:13] Nathan Wrigley: Yeah. Yeah.
[00:30:14] Amir Helzer: I hope it'll survive 20 20 26. so what you do today, you finish this setup with all, and you go to a transition dashboard, which is, a bird's eye view of all the content on your site. Okay? And we recommend to people to pick one or two pages and send them to transition.
Use the AI transition, but just pick one or two and send them and look at the transition. often people discover that, many times it's fine. But when something is missing, you'll notice, oh, I've, I forgot to translate a menu. I didn't, select these custom fields as part of my text. Little technicalities, things that are fairly simple to resolve, but you want to notice them from the beginning, especially if you're using a plugin like a CF.
It. It's got statics set up, and if you've created your content using many custom fields, then you want to tell the mill which ones to translate and which not.
[00:31:17] Nathan Wrigley: Mm-hmm.
[00:31:18] Amir Helzer: which, holds an image or a video. You don't want to translate that, but if you have a field that is actually, that, the title of something, then you should translate it.
So you translate a few pieces of content. You have a look at the font and you say, I've configured everything correctly. I'm translating correct. Typically, you don't need to worry about the language these days, but you need to make sure that the technicality is right, which takes one iteration. And when you are happy with this, you return to the same transition dashboard and there's a toggle at the top that says, from now start doing back background transition.
[00:31:51] Nathan Wrigley: Oh,
[00:31:52] Amir Helzer: and that's it. And you're
done.
[00:31:54] Nathan Wrigley: so, from that moment on, after you've done your checks and balances the first time around, if you like, once you've got everything tuned in, from that moment on, if I write a piece in English, it will then just. Do it without. Okay. So it's autopilot basically from that moment on. Okay. That's fascinating.
One curious thing which occurred to me is, we, there seems to be a lot of discussion at the moment about, and I'm gonna call it AI slop, on the internet. And this is the idea that we are flooding the landscape with AI generated content and, and, all of the harms that can bring. But what's curious about this is.
That's not, this is it, because if you've written a good piece of content the first time around, okay, let's assume that you haven't created AI slop. It feels in many cases that, oh, AI generated more content. Okay, that's AI slop, but this isn't that. This is you translating, hopefully faithfully your high quality piece of content into multiple instances of high quality content, and it's like an SEO win from top to
[00:32:54] Amir Helzer: I think so. Yes. And this is the results that we're seeing on our site.
[00:32:58] Nathan Wrigley: so it's, working in that sense. You're getting surfaced in the Yeah, I've, I've used a few tools to check, AI content. I've, written something with an AI and I've checked it and the tools have been pretty good. working out is this ai and then when I've written something myself, obviously something of my self bleeds into that.
And, the AI. that's trying to spot the fake it, it says, yeah, that looks authentic. And so presumably the way that you've described the translation, that should work in the same way. it'll pick up the nuances of the language and it won't feel mechanical. It will feel authentic. It'll translate phrases in a, hopefully an authentic way.
So that's really interesting.
[00:33:42] Amir Helzer: I, haven't thought about this. this would be an interesting experiment to run a tool on the, on, on our transitions and see what it, determines.
[00:33:51] Nathan Wrigley: Yeah, definitely because I feel that people, organizations like Google, it, feels like they have to probably marshal their resources to cut down on the ai. I, don't suppose they wanna be surfacing AI slop, but if the original piece is interesting, you authoritative. I can't see why the AI wouldn't translate that given everything that you've just said.
It's trying to do a faithful and yet slightly modified for the language translation. So yeah, I'd be curious to know how you follow up with that. okay. Back to the technicalities then. Many of the tools that I'm using. When I'm, in invoking AI in some way, what I'm really doing is copy and pasting an API key from something like chat GPT and, off we go.
it sounds, I could be wrong, it sounds as if that's not what you are doing. It sounds as if you are taking on the, somehow the translation. I don't know whether you've got a, a big allocation at something like chat GPT or if it is an API key. Just explain where this data is going and how you manage that.
[00:34:56] Amir Helzer: Okay. So we are using several lms, actually we're not using Charge G pt, but we're using other lms, and we're using the API keys. And because of our volume, we need, redundancy fail failover. we need a bunch of mechanisms to make sure it actually works at scale. last year we translated I think about two and a half billion words a month. Yeah. This year it's way higher. This, year it's way, way more. So it's, a lot.
[00:35:31] Nathan Wrigley: Yeah,
[00:35:33] Amir Helzer: yeah, it is a lot. and, yeah, actually you should, look at our AWS charge.
[00:35:43] Nathan Wrigley: Oh.
[00:35:44] Amir Helzer: it's okay. I'm not complaining. it's a lot, it's a big operation. and, it's, a lot more than, sending the content to an LLM and getting it back, at a very, small scale.
You, you could just, create a, Python script, send the content, get the content, but then your rate of dealing with issues is going to be pretty high mo most of the time it's gonna work. But you'll find yourself dealing with exceptions on a daily basis
[00:36:21] Nathan Wrigley: Yeah. so is it a subscription service then? Am I, I've got my WPML plugin dating back many, years, and, we know how that works. that's a, that'll be a subscription, annual charge for the plugin and so on. Is this a subscription layer on top of that, based upon usage and the amount of tokens that you
[00:36:43] Amir Helzer: yes. But, there was a pretty, pretty generous, free tier. when we started all this, my, my understanding was someone just paid us and he wants to use the tool and he will be a little disappointed if 10 minutes later. He says, a message that says, now pull out your credit card. Again, I, don't like this kind of experience myself, so if you want to charge me, please charge me and then let me use it.
Please show me how much it's gonna work, it's gonna cost.
[00:37:14] Nathan Wrigley: Yep.
[00:37:16] Amir Helzer: I think that will including something like 90,000 free credit. Or included credits in the CMS accounts and our math based on the actual sites that Transl says that, this covers 80% of the sites that clients use WML four, for two, three languages, which means that at least 80% of the clients would never, see any payment request.
And if you're building a much larger site or a site with much, many more languages. when you exceed this quarter, which also gets renewed every year. Okay? So when you exceed this quarter, then you can either, buy prepaid, transaction credit packages, or you can put in your credit card and you can pay per use.
The pay per use costs less per word. It's the most cost effect efficient. eh. Actually web development agencies, they typically buy the prepaid packages, so they don't need to bother their clients with putting in a credit card.
[00:38:26] Nathan Wrigley: Oh, I
[00:38:27] Amir Helzer: And they don't definitely, they don't put in their own credit card. So they buy something and then they, resell it.
Basically, they, it's like a business opportunity for them. They tell their
[00:38:36] Nathan Wrigley: Yeah. Yeah.
[00:38:37] Amir Helzer: here, and I've included this in this amount, you can translate 50 pages. I, think, and I hope that they're doing a good bus, good business. I hope that they're reselling it for much more than they're paying us, because then everyone wins.
They're adding some value, they're making a profit. We make a profit. The client definitely makes a profit because for the client, the, person who makes the biggest profit here is actually the, end client, because whatever he's gonna get charged for this is pennies compared to the effort that it would take him to actually sit down and translate.
[00:39:11] Nathan Wrigley: Yeah. Okay. That makes sense. So I'm just trying to imagine how this works. I'm not plugging in an API key from a third party service. You, handle the AI interaction. I just, I set up subscription, annual whatever with you. pay as you go, or, you buy a package and then everything roots through you and.
so that the pricing of that all makes sense. Is there any sort of secret source, I, think you've described this already, but just to be clear, is there any sort of secret source that you apply on the prompt
[00:39:44] Amir Helzer: Yes.
[00:39:45] Nathan Wrigley: which somebody like me, if I just say, here's some text translate to Japanese, my result may differ from what you are getting out because presumably you are doing this a lot, you've probably got some intuitions as to how to best
[00:39:59] Amir Helzer: Now we don't have intuition. we hired, a team of linguistics. Okay.
[00:40:05] Nathan Wrigley: Oh,
[00:40:05] Amir Helzer: working full time for us. so the first task in making this work correctly was to, to pick many, sites. Go through them and translate them again and again. Again. And, see what's wrong?
Okay. Understand what's wrong. both linguistically and technically. Okay. And there were many things wrong. and then. Work with our engineering and figure out ways to fix them. So we're talking about localization issues, language issues. for instance here you write quotes with a different character in, French and
[00:40:44] Nathan Wrigley: Oh yeah.
[00:40:45] Amir Helzer: to take care of that. but you need to make sure that you don't translate the quotes, which are inside short codes. Yeah, there's those stuff to handle, that translate strings also, placeholders for variables. some sites use this kind of, character encoding some, another, some mix.
So you've got sites with a mixture of, unicode escape characters and UTF.
[00:41:11] Nathan Wrigley: yeah.
[00:41:12] Amir Helzer: all this stuff you need. There's no secret sauce. It's just a lot of engineering. and then we also use different elements. So you have different, small but important optimizations parallel, and then you've got, how do you encode it to the LLM?
How you get it. you need to make sure that, you get correct translations and you're not getting silly answers. I don't understand the sentence. You don't want, you don't want your homepage to say, I don't understand this, like what the LM just told you. Okay. Little things. Little things. think of it like you are hiring, translators now they know how to translate, but they're new to WordPress, they're new to HTML, they're new to a bunch of stuff, and you need to train them.
So you, you could do this yourself, obviously, but I think that the math works out better when someone does it very centrally for many hundreds of thousands of clients together.
[00:42:19] Nathan Wrigley: Yeah, but also I, I guess for all the languages, so you might be good at doing it for, I don't know, Spanish, English to Spanish or that way, round the
[00:42:28] Amir Helzer: this is also,
[00:42:29] Nathan Wrigley: not maybe into Hebrew or Dutch or all the others. And you've done
[00:42:33] Amir Helzer: we had to do this. we did not release any language. nobody checked it. We have to, and we, caught issues with all the languages that we, curated.
[00:42:46] Nathan Wrigley: Is there a, is there any part of the interface where, let's say that something pops up and I think, actually, do you know what there is something there that I would wish had come out differently. do you receive feedback? Is that like a support ticket thing? In other words, if I want to update you that something's not quite what I was anticipating, can I do that inside the PTC or the plugin or do I need to contact the support to make those
[00:43:09] Amir Helzer: Okay, so first, everyone is free to edit their own site. It's not like it translates and you can't touch it. So if you wish, you can translate, better than editing yourself. you can go to a segment and tell it in your language what you don't like about it, and then the, LLM will fix it for you.
[00:43:29] Nathan Wrigley: Right,
[00:43:30] Amir Helzer: I think many a AI tools do this today, so you don't need to know what is the right output. You need to know what's, bothering you. So you, could do that?
[00:43:40] Nathan Wrigley: Does the context of that stay with you, if So the next time I come and a similar opportunity for failure arose, does the, LLM maintain that context and, memorize that, okay, this is how it's anticipated.
[00:43:56] Amir Helzer: yet, but, we're working to integrate this now so that we will auto detect, we will not auto, we will detect your edits. And we will ask you, do you want to apply something similar to this in the future? The problem is that we're working with very little data because people are doing so little edits and the transitions are just good.
[00:44:17] Nathan Wrigley: Okay.
[00:44:18] Amir Helzer: No, really? Yeah, the
[00:44:20] Nathan Wrigley: Yeah. Yeah. I can
[00:44:21] Amir Helzer: so when we released this version of WML, which I think was about, three weeks ago, two or three weeks ago, we did something which is new to us and we were concerned, we're, we were offering a guarantee to the quality of the transition. Meaning if you get our AI transition and you're not happy with this, for whatever reason.
I don't know what the reason could be. just tell us and we're not only going to redo it or fix it for you. We can refund your money. You, can get us to fix it. You can get us to, re refill your credits and you can get us to just refund what you paid in yours.
[00:45:01] Nathan Wrigley: Oh gosh. Yeah. You're really standing by it. That's quite a promise.
[00:45:04] Amir Helzer: And, we're all engineers and.
when we were working on this, we, bounced dozens of imaginary scenarios in our head on how this is gonna go wrong. Dozens, since then, WL has already translated several tens of thousands of sites. From this new policy and so far we've had, zero requests for, issues. We, did, get three refunds for that.
And they were completely unrelated to the transition, actually. someone, yeah, just regular, refund issues, technical issues somehow leaked into that.
[00:45:49] Nathan Wrigley: Okay. Interesting. I, guess if, you were to rewind the clock, I don't know, let's go five years before any of this was available. Before you were just, you were advising people to get the, some kind of an ai, but. Not rely on it in the way that you are now. I think it, it must be a hard business that you're in.
you've got a fairly, it's, there's a lot of moving cogs. You've gotta get translators, humans involved. You've gotta wait, you've gotta approve things. And so on. This whole AI thing, whilst it might concern people out in all sorts of ways, this, must have been music to your ears. It must have been just Oh.
What the heck? this whole translation engine, the what the airs can do. That if, I could have built something, I would've had this built. it feels like this thing came along and dropped slightly perfectly inside your business, your clients for almost no money. I dunno what the costs are, but I know that it's cheap, right?
This kind of stuff is cheap. The, but you, it was just handed to you on a plate. Look, we've elsewhere, unrelated to you, we've built a whole bunch of technologies which enable you to make your plug in a whole lot better. You must have just been giggling to yourself. How the
[00:47:02] Amir Helzer: like a miracle, but,
[00:47:04] Nathan Wrigley: Yeah. Yeah.
And not many people have a, I suppose there's many people in the WordPress space who are trying to crowbar AI into their product. You can see it all over the place, right? they've dropped AI in, but you think, okay, it's clumsy at best. It never really belonged there. You had a product for which AI was the perfect answer.
[00:47:25] Amir Helzer: Yeah.
[00:47:26] Nathan Wrigley: that's a nice bit of serendipity.
[00:47:28] Amir Helzer: yeah, and you would think that this would be easy to market and of course it's not. And the hardest, it was, the hardest audience to market that was, internally to our team.
[00:47:40] Nathan Wrigley: Interesting. Just they thought, no way. It is never gonna work or,
[00:47:44] Amir Helzer: a change. And every change is met by, by skepticism, every change.
Again, we're engineers and we were walking out in our heads all the possibilities, how this could have failed. And when you look at all these com possibilities, it looks like there's no way it's gonna walk.
[00:48:03] Nathan Wrigley: And
[00:48:04] Amir Helzer: It
[00:48:04] Nathan Wrigley: It did. Yeah, it's amazing. really, amazing. Like I say, you couldn't have designed a better set of things. The, the LLMs that have come along, large language model, what are you dealing in large amounts of language. It's just, It's almost like you, you designed it.
just for yourself. yeah. That's amazing. I, I don't know that I have any more questions to ask you. I, do wonder, however, if there's some part of the system, some part of the process, some part of the plugin or the, the subscription that you are particularly proud of that I didn't mention that you wanted to get out there, that perhaps people wouldn't have known about.
[00:48:43] Amir Helzer: so I just want to clarify that, the PTC system that we built. we use it for, for software now and for websites. So sure enough, we get much more people know us much better for the, website transaction because W Mill has been out there for 16 years. but, we ourselves will seeing actually the bigger benefit to us as clients from translating software.
Because, it's actually harder. Translating software is harder than translating websites because websites is long, clear, meaningful content.
[00:49:25] Nathan Wrigley: Yep.
[00:49:26] Amir Helzer: User interface is disconnected. Strengths typically with poor English. Okay? Very, poor English. it's even poor English by design. Okay? It says click button.
Okay. That doesn't click, the bloody button. It says click button. it's not an English phrase. Okay. So it's, a harder task from the beginning. before we've had this, before we were using PTC ourselves, we spent roughly $8,000 to translate, to update the translations of our own WML plugin for every major release.
[00:50:00] Nathan Wrigley: Wow.
[00:50:00] Amir Helzer: yeah. $8,000 and we have many races. Okay. we come out with a major release, every few months. So our bill was approximately this, sometimes at le larger, sometimes a little smaller, but this was our bill because it's very slow walk. For human to do this work, a human must charge us a pretty high rate per ward in order to translate software correctly.
Or we can expect very sloppy work. We, can choose okay, but when, the walk is harder, you need to pay for it. And it was very hard to get the LLM to translate it also. But again, it's something that is teachable. So we taught the LLM how to do it. It's a different form, different explanations, different pre and post processing.
It's, there's work in it and we benefit the most actually, not only from translating our own website. It is a huge benefit for us, but, when we don't need to, pay again for translating our own software. So PTC translates, WN plugin and tools and everything else. That tool is, and now we've forgotten about translation.
The PTC is connected with our GI repository. Possession just happens in the background as, we continue as we go, when there's something, someone is working on a feature branch or, Or prepares it for tagging, it's just translated by in the background. The resource files just fly back, into GI and it, we eliminated this very expensive item for more checklist.
Previously we had this thing, do you remember the name String Rizz. It's a horrible thing. It means that when everyone's ready
[00:51:45] Nathan Wrigley: Yeah,
[00:51:45] Amir Helzer: yeah.
everyone's ready. You want to ship it now, you need to wait. Now you're waiting a couple of weeks. For a Dutch translator to get back from vacation.
we, don't have that anymore.
It's ready when, we're done testing we ship it out. That's it. No string freeze, no nothing.
[00:52:03] Nathan Wrigley: the button and off you go.
[00:52:04] Amir Helzer: there's not no button to press even.
[00:52:06] Nathan Wrigley: so that sounds to me like a new arm of your company in a way. Then, so are you marketing this out into the soft, so not WordPress in particular, but out into the software world in general? we'll translate, I dunno, we'll, organize your change log for you, but we'll also update all of the bits and pieces.
So that's quite an interesting pivot and, also feels Yeah, a real nice bright horizon there. There's probably a lot of opportunity there. I was wondering also if you have plans to, or already have opened the, translate, the PTC op as like an, I don't know, like an API endpoint that other software developers can hook into, or if that's just something that, we're doing this through the plugin,
[00:52:48] Amir Helzer: PTC is a system, as a standalone is PTC is solving WL. But as a standalone system, it does have, its API. software developers today already connect with this integrated to the, CI flow, either through the, our GIT integration or through our API. So they make a call, say I've got something that needs translating, and they listen to, the hook that says Transition is ready, and it is already integrated.
People are using it with our API.
[00:53:22] Nathan Wrigley: So the, software development side of things that you just described, that's where it's being used. It was just curious. 'cause I know there's lots of SaaS apps out there that are obviously making connections to, LLMs maybe to, I don't know, transcribe things or take meeting notes and things like that, but then might not necessarily do the translation portion on the other side, but there may be a need for that.
So I was just curious as to whether or not you
[00:53:44] Amir Helzer: This is our
[00:53:45] Nathan Wrigley: you were open to
[00:53:46] Amir Helzer: Yeah. You got it exactly right.
[00:53:47] Nathan Wrigley: Interesting. Okay. in that case, I think we've probably wrapped up everything that we need to do. just give us the URL drop. The URL we can find out most about. This is obviously WPML. We could search for that in Google or what have you.
Is there a particular page you wanna direct us to on your website that we can put into the show notes and people can hear you read into the
[00:54:08] Amir Helzer: So we've placed PTC project under ptc.wpml.org.
[00:54:13] Nathan Wrigley: Perfect. So ptc wpml.org. Go there and check out everything that Amir has been mentioning today. It's exciting times, Amir. I wish you all the best. I hope it's a success. It obviously is gonna make, make the translation process much simpler, much cheaper, much faster. that's a lot. Well done.
Thank you for bringing that to our attention today. Best of luck
[00:54:38] Amir Helzer: Thank you so much for having me, Nathan.
[00:54:39] Nathan Wrigley: You're very welcome.
Okay, that's nearly all we've got time for. I hope that you enjoyed that.
If you did, head to wpbuilds.com. Whilst there, head to the search option at the top of the website. Search for episode number 436, and leave us a comment there. We would really appreciate that.
The WP Builds podcast is brought to you today by GoDaddy Pro. GoDaddy Pro, the home of managed WordPress hosting that includes free domain, SSL, and 24 7 support. Bundle that with The Hub by GoDaddy Pro to unlock more free benefits to manage multiple sites in one place, invoice clients, and get 30% off new purchases. Find out more at go.me/wp builds.
We're also joined this week by Bluehost. Bluehost, redefine your web hosting experience with Bluehost Cloud. Managed WordPress hosting that comes with lightning fast websites, 100% network uptime, and 24 7 priority support with Bluehost Cloud, the possibilities are outta this world. Experience it today at bluehost.com/cloud.
And last, but by no means least Omnisend. Do you sell your stuff online? Then meet Omnisend. Yes, that Omnisend. The email and SMS tool that helps you make 73 bucks for every dollar spent. The one that's so good, it's all most boring. Hate the excitement of rollercoaster sales? Prefer a steady line going up? Try Omnisend today at omnisend.com
And sincere thanks go to GoDaddy Pro, Bluehost and Omnisend for their continuing support of the WP Builds podcast.
Okay. I'm really, really, really close to the end.
The last thing to mention is don't forget later in the month, if you want to join us at WPLDN in London, as well as LoopConf, check out the bits and pieces that were mentioned at the top of the show. We would really like to see you there. That would be lovely.
Okay, that's really all I've got. Except of course for the cheesy music, which is about to fade in. And I'll just say, have a good week. Stay safe. Bye-bye for now.
