@elilla it was never very good to be fair… adequate at best, but awful if there is any slang or nonstandard sentence construction. Quality as always is very dependent on what language you use to and from, Indo-European languages usually mostly work between each other, but there’s often quirks.

It’s used an LLM for like a decade and a half now (people think LLMs are new… not even close to new). I think their newest tweaks have made it… better in some ways, worse in others? Mostly a wash.

It does seem way more mid at being a dictionary, which is what I used it for, and it sucks they seem to have gotten rid of the multiple suggestions feature.

Ah well.

@elilla yeah, google translate has been changing meanings entirely on me recently, like, I'll be using it to polish/grammar check an email or whatever, and it'll just change a sentence to a somewhat similar sounding sentence with an entirely different meaning.

For example, I was refining an email to a doctor for a prescription renewal (still the same quarter, so I can do it without showing up and save us all time), and it just totally butchered it into talking about first appointments.

If I didn't know the languages on both sides, that would be a real issue.

@elilla Something I've noticed lately to add to this list of Bads:

5) Translate results are more superficially fluent and plausible

that's the output smoothing from the LLM. earlier versions of translation would fail visibly (whether you knew the target language or not you could see syntax or word choices that didn't make sense) and informatively (you could tell something was off, and if you knew the target language even a little you'd probably see it). you could never count on translation results being suitable for copy-paste, and so—the inevitable corollary—you knew NOT TO COUNT ON THEM

now however translate results look "good," i.e. plausible, and unless you know the target language *better* than a little that veneer of plausibility effaces your ability to gut-check the results (and not just in the moment, but by way of steady erosion the more you engage with the software. in effect you're developing a muscle memory that tells you you CAN and SHOULD count on the results)

machine translation is every bit as unreliable as it always has been and *will* be: anyone who's actually done translation knows that it's not a solvable problem (nor in fact a "problem" at all in that compsci-ish sense), any more than "intelligence" is. but you can make it LOOK solved, and if you've bamboozled enough of your user base into playing along then who's to say it isn't?

meanwhile you've boosted your AI numbers and given yourself a whole nother category of deskilled labor you can show the axe to; and the money's very happy with you