Writing
Why translation isn't a settled problem.
If translation were close to solved, the books in the bookshop would prove it.
There's a version of the story that goes: machine translation has gotten very good, the remaining gap is small, and the remaining gap will close on its own. It's a comfortable story. It is not what the books look like.
Walk into a bookshop in any non-English-speaking country, look at the translated section, and ask yourself a different question: of the books on offer, how many were translated by someone who read the original cover-to-cover, more than once, and made deliberate choices about every recurring word? In the best cases, the answer is most of them. In the average case, the answer is approximately none.
The gap between an adequate sentence-level translation and a literary one is not a 5% gap. It's the difference between knowing what the words mean and knowing what the writer meant. The first is a problem with a tractable answer. The second is a problem with a person at the end of it. Translation is settled in the same sense that cooking is settled: yes, you can run a recipe through a machine and get something edible. It will not be the meal anyone remembered.
We talk about this not because we want to argue against the model — we use the model — but because the framing matters. If you believe the gap is closing on its own, you build differently. You build to push humans further out of the loop. The tools you produce are calibrated to a future where the human's role is review-and-rubber-stamp, and they fail badly the moment the human has to actually intervene, because the workflow was never designed for them.
If you believe instead that the gap exists, will continue to exist for the kinds of work that matter, and is held open by something the model genuinely cannot supply, you build differently. You build for the case where the human stays in the loop, where their judgment is the point, and where the tool's job is to make that judgment cheaper to apply at scale.
We're in the second camp. Not on principle — on observation. The bookshop tells us where we are. We're trying to build for the place we are, not the place a press release says we're about to be.