WEBVTT
00:00:00.160 -->
00:00:04.080Hello, I'm Stephanie, and here for another episode of Field Notes.
00:00:04.080 -->
00:00:12.240And today it's kind of exciting because, as many of you may know, last week we just started our companion field notes series with Giulia Greco.
00:00:12.240 -->
00:00:19.120And today we're going to do a companion episode with Erik talking about the similar topic.
00:00:19.120 -->
00:00:22.640So this topic is something that we've been hearing about a lot.
00:00:22.640 -->
00:00:30.640And this is where teams are building their own translation solutions, sometimes without even involving localization at all.
00:00:30.640 -->
00:00:32.799So, Erik, from your perspective, what's going on?
00:00:33.119 -->
00:00:44.240Yeah, it was really amazing seeing Julia's input on this because she's raising something that I've also been noticing as well, which is this decentralization pressure that we're seeing.
00:00:44.240 -->
00:00:57.439So, because localization can be in embedded, this is not necessarily a brand new problem, but organizations would either have different teams doing their own localization vendors and handling things their own way.
00:00:57.439 -->
00:01:02.079This is a legacy of MA activity or different sort of business purposes.
00:01:02.079 -->
00:01:07.200You might have marketing and product working with different vendors, then they consolidate and then they break it apart again.
00:01:07.200 -->
00:01:10.079It's a it's a not that uncommon of a phenomenon.
00:01:10.079 -->
00:01:30.959But because we're actually seeing a new thing happening here where you can not only direct localization traffic via different channels, which we talked about last time, but also we're seeing this pressure to add this capability in a much faster, more agile way in different places within an organization at the same time.
00:01:30.959 -->
00:01:40.079So it's a very interesting phenomenon and it's hitting a much broader range of companies than I ever expected that would happen, that's happening faster than I would have expected.
00:01:40.480 -->
00:01:48.640So we have now teams are somewhat, you know, localization is escaping the localization department.
00:01:48.640 -->
00:01:53.920Why are teams doing this, do you think, instead of going through the typical localization routes?
00:01:54.079 -->
00:01:56.959Yeah, I think what we're seeing is that demand is there, right?
00:01:56.959 -->
00:02:03.840So in I think Julia pointed this out as well, but that there's a demand and the capability is there.
00:02:03.840 -->
00:02:07.200So the these teams are just reacting to these fundamental demands.
00:02:07.200 -->
00:02:11.680And so the marketing team, for example, has a need to get things out very quickly.
00:02:11.680 -->
00:02:12.400You know that.
00:02:12.400 -->
00:02:24.400Sometimes it can be very powerful to use AI to create content or help shape the understanding of what the market signal is and being able to bring that signal in and respond to it in a much more effective way.
00:02:24.400 -->
00:02:31.680Meanwhile, the product teams are sometimes being empowered to, or sometimes executives are demanding that they add a chatbot into there.
00:02:31.680 -->
00:02:38.240So now they have to add an AI capabilities like with some translation functions directly into the product.
00:02:38.240 -->
00:02:51.280And then you might have support teams that are dealing with auto-translating content, or which is sometimes end-user uh conversations with their users and and their support ecosystems, or sometimes their knowledge base.
00:02:51.280 -->
00:02:59.840And then, of course, we have on top of that this crazy vibe coding reality where teams are needing to implement lightweight scripts or plug and play tools.
00:02:59.840 -->
00:03:03.680But we hear people joking about vibe coding a CMS over the weekend.
00:03:03.680 -->
00:03:05.840Like it's a little bit of a joke, obviously.
00:03:05.840 -->
00:03:07.120It's not real realistic.
00:03:07.120 -->
00:03:08.639Please don't try this at home.
00:03:08.639 -->
00:03:25.120But the mindset is real that that now that we have these powerful automation tools and we have individuals who are building hundred million dollar businesses with no code experience, they're building an app themselves and then generating like a hundred million in revenue.
00:03:25.120 -->
00:03:29.919It's it's uh it's unbelievable how um how fast this is all changing.
00:03:29.919 -->
00:03:35.840So executives are definitely expecting fast and demand demanding this velocity.
00:03:35.840 -->
00:03:38.000The business is demanding that velocity.
00:03:38.000 -->
00:03:46.719So I think it's just realistically, it's not okay to create an arbitrary bottleneck if there isn't a legitimate business reason to do.
00:03:46.719 -->
00:03:57.280So I think yes, localization is escaping the localization department, but it's actually it's actually being driven by very meaningful business drivers underneath.
00:03:57.759 -->
00:04:00.639So then where does this become an issue?
00:04:00.639 -->
00:04:02.960So people are doing it, is it an issue?
00:04:02.960 -->
00:04:04.639When does it become an issue?
00:04:04.639 -->
00:04:05.439What do you think?
00:04:05.680 -->
00:04:19.199Yeah, I think it is when you have a distributed production system, then anybody who understands the consequences of poor terminology control will instantly see that as a major area of concern.
00:04:19.199 -->
00:04:25.759Um, you have fragmented systems, but I think these are technical things that we've been familiar with for decades.
00:04:25.759 -->
00:04:34.240But I think one of the key problems is the ownership and the governance aspects of it, which it's not clear who owns quality in this case.
00:04:34.240 -->
00:04:51.519So if you're thinking about this as a localization manager and somebody's vibe coding some new tools and then packing in maybe your content without even talking to you about it as part of some AI solution that they're building, and then it spits out some less than ideal output.
00:04:51.519 -->
00:04:55.360What happens to the localization owners of this?
00:04:55.360 -->
00:04:57.680How do they handle this ambiguity?
00:04:57.680 -->
00:04:59.759Who's accountable for fixing it?
00:04:59.759 -->
00:05:12.079So without that clear ownership, then you're gonna end up with inconsistent terminology and there's of course duplicated effort, but the big one that ties into what we've talked about before is the unmanaged risk.
00:05:12.079 -->
00:05:14.720They're not consciously managing that risk.
00:05:14.720 -->
00:05:21.040So you have this demand force which is driving towards less regulated, less controlled output.
00:05:21.040 -->
00:05:25.439And then you don't really know or understand how you're managing that risk.
00:05:25.439 -->
00:05:36.399So what ends up happening is that even though the technology could be owned locally or by individual teams, the accountability ultimately gurgles up to the CEO at the top, right?
00:05:36.399 -->
00:05:51.759So when we're trying to align the ownership and the governance around each of these kind of emerging systems, and you don't really have the clear feedback loops, the consequences are gonna pop up later and you're gonna end up with a lot of uncomfortable surprises.
00:05:51.759 -->
00:05:56.879So I think this is really what we're talking about, what Julia was calling shadow localization.
00:05:56.879 -->
00:06:01.360And I think it's not necessarily that people are trying to break the system on purpose.
00:06:01.360 -->
00:06:02.879None of this is intentional.
00:06:02.879 -->
00:06:18.720But because the market demands that we talked about earlier are forcing innovation essentially, trial and error, sometimes very just run fast and break things mindset, that you don't necessarily see the damage you're doing until later.
00:06:18.720 -->
00:06:31.519But I think what we're seeing is this kind of need to think about that responsive layer uh to support business need that's emerging.
00:06:31.759 -->
00:06:35.519So then as an organization or a company, what do you do?
00:06:35.759 -->
00:06:39.360Yeah, well, I don't think that it's gonna work to go back to a fully centralized system.
00:06:39.360 -->
00:06:40.639It's just not gonna happen.
00:06:40.639 -->
00:06:43.439That creates an arbitrary bottleneck.
00:06:43.439 -->
00:06:48.079But I think having a conversation about connective governance is like your starting point.
00:06:48.079 -->
00:07:01.839So I think each of the, and this is putting myself into the to the shoes of individual owners of different systems, they often don't feel empowered to really have that broad level of conversation.
00:07:01.839 -->
00:07:03.439So I just want to be realistic here.
00:07:03.439 -->
00:07:09.839This isn't this is an easy thing to say, but publishing standards, for example, centralizing knowledge.
00:07:09.839 -->
00:07:15.439One thought would be to make your content consumable by LLMs directly.
00:07:15.439 -->
00:07:17.279So that's bring your own LLM model.
00:07:17.279 -->
00:07:39.279So there's a new emerging thing here where you're basically trying to adapt or allow your content to be adapted to in this new AI ecosystem, in which case you as the team are saying, hey, here's the here's the markdown files you should use when you're running your model that we're gonna shape for you so that you're gonna have the best chance of getting better results from this.
00:07:39.279 -->
00:07:42.000There could be better, clearer rules of engagement.
00:07:42.000 -->
00:07:48.160I think some companies have said, go ahead and do whatever localization you want, but we will be an assessment layer.
00:07:48.160 -->
00:07:51.600We're not gonna be a gatekeeper, but we'll be maybe an assessment layer.
00:07:51.600 -->
00:07:54.879So if you're not sure what the quality is, we can do an assessment.
00:07:54.879 -->
00:08:01.600Maybe we'll plug in Taos or Model Front or something like that as part of kind of this monitoring layer.
00:08:01.600 -->
00:08:06.879Or we can have human reviewers look at the content and say, Are you gonna hit any red flags here?
00:08:06.879 -->
00:08:14.800You can also be an integrator, and playing an integration role means actively looking for ways of supporting those folks out there.
00:08:14.800 -->
00:08:21.040Like, how can you supply them with the best chance of being successful without creating this stuff?
00:08:21.040 -->
00:08:29.759So basically, you're offering yourself to be um kind of an advisor and accountability partner, and then being an expert in the loop.
00:08:29.759 -->
00:08:39.360And then, of course, that if you want to get it right, and that's again where the human arbitrage piece comes into it, that's where we want the human in the loop to be.
00:08:39.360 -->
00:08:43.840And the team needs to make sure that is getting the right answer.
00:08:43.840 -->
00:08:48.399That's for all the gold standard output, that's all for all the training corpora.
00:08:48.399 -->
00:09:04.240That's another piece of metadata, by the way, that I think we've had or should be having in the loop here, which is is this content consumable by any learning efforts, or is this an output that we don't want to learn from?
00:09:04.240 -->
00:09:07.840So empty output is generally not something you want to train on.
00:09:07.840 -->
00:09:17.039So the accountability layer is about delivering high quality content that you can stand behind as training corporate.
00:09:17.360 -->
00:09:17.679Okay.
00:09:17.679 -->
00:09:24.000So you've gone over like several points throughout this, but what would be your top takeaway?
00:09:24.000 -->
00:09:25.200Remember this one thing.
00:09:25.519 -->
00:09:29.840Localization isn't going anywhere, but it is becoming distributed infrastructure.
00:09:29.840 -->
00:09:33.200And as we talked about last time, there's these two tracks.
00:09:33.200 -->
00:09:39.919One is the facilitating the automated translation to get the best available option with a human in the loop accountability layer.
00:09:39.919 -->
00:09:41.279And then you've got just both of them.
00:09:41.279 -->
00:09:43.200There's the speed and there's accountability.
00:09:43.200 -->
00:09:45.120So make sure you have both of these supported.
00:09:45.120 -->
00:09:46.320Think differently about this.
00:09:46.320 -->
00:10:01.039And then some companies are making a rubric of their content with maybe even three or four layers deep of different content types and different sort of accountability layers, and they'll create intentional workflows for each of those work types across the different teams.
00:10:01.039 -->
00:10:04.879So you're you're not saying everything needs to go through this process and get to 100%.
00:10:04.879 -->
00:10:09.759You're mitigating the risk with the right approach that balances cost and quality outcomes.
00:10:09.759 -->
00:10:20.639So you're managing risk, you recognize that shadow localization is going to happen if you create unrealistic barriers to managing that content, uh, and then you end up with a governance failure.
00:10:20.639 -->
00:10:31.039So accept the reality that this is the direction things are going and think about ways of shifting towards connecting systems, deciding decision ownership, and anchoring accountability.
00:10:31.440 -->
00:10:31.759Okay.
00:10:31.759 -->
00:10:33.200All right, Erik.
00:10:33.200 -->
00:10:34.240Well, thank you very much.
00:10:34.559 -->
00:10:35.919It's always a pleasure, Steph.
