On translation, machines, and meaning
Translation is older than computers, older than printing, older than most of what we call civilization. It is what happens when one mind tries to reach another across a gap it cannot fully cross. Every translator knows this gap. Every machine, until very recently, pretended it didn’t exist.
We have spent the last three decades inside this problem. We started in the 1990s localizing software for Apple, when “internationalization” meant resource files and double-byte character sets. We watched translation memories arrive, then statistical MT, then neural MT, then large language models. Each wave promised to solve translation. Each wave moved the problem somewhere else.
LLMs will not solve it either. They are the most powerful tool we have ever had, and they are useless without the people who know what to ask them.
We hold this position as engineers, not as romantics. A generic model trained on the open web has no idea that your product calls a particular UI element a “workspace” rather than a “workbench”. It cannot tell that your Japanese users expect a different register than your German ones. It misses that the word “submit” in your procurement software carries a specific legal weight that “send” lacks. These are not edge cases. They are the work.
What we do is build systems that know these things. We curate the data. We write the prompts. We design the retrieval. We test the output against criteria that come from thirty years of watching translations succeed and fail in production.
The models redistribute the gap that translation has always been about, rather than closing it. Where the translator used to carry the whole weight of the crossing, the system now decides which parts to delegate to the machine and which to retain in human hands. That decision is the central question of our craft, and we believe it can only be answered by people who have spent years on both sides of it.
We argue, internally, about whether a model can be said to “understand” anything. The argument is far from settled. What we do know is how to work responsibly with systems whose epistemic status is uncertain. If we are honest, this is also what translators have always done with language itself. Language is a system whose workings escape us, and which we nevertheless use with professional responsibility every day. The machines are new. The condition is old.
We are also, deliberately, small. A client asked us last week if we could add image-to-translation to their pipeline, so they could drop in a screenshot and get a draft back. It took us five days. At other companies, the same request would generate a statement of work, three meetings, and a delivery date in the next quarter. We have linguists who code and engineers who translate, and that is most of the structure we need. When you ask us for something, you are talking to the people who will build it. When we ship it, we have already used it ourselves.
The team reflects this. Our founders started the company in the 1990s, one from computational linguistics, the other from coordinating localization at Apple. Our lead engineer holds a PhD in experimental sciences and has spent two decades building AI systems for regulated industries. Our senior project managers ran localization for Apple and other exacting enterprise clients for fifteen or twenty years before joining us. Our designer thinks about localization the way a typographer does, as something that happens to images, layouts, and visual systems, beyond the level of strings. Several of us teach at UAB, UVic, and other master’s programs where the next generation of translation technologists is being trained.
The localization industry is full of vendors who will sell you AI as a black box. Push button, get translation, ask nothing. We offer something different: a partnership with people who have thought hard about what language is, what machines can and cannot do with it, and what your specific product needs. We will show you the inside of the box. We will explain the trade-offs. We will tell you when AI is the right tool and when a human translator is, and we will mean it.
Translation is older than computers. The good news is that we have spent our careers learning how to use one to serve the other.