Hacker News Reader

Comment by autumnstwilight | Hacker News Reader

I learned Japanese by painstakingly translating interviews and blog posts from my favorite artist 15+ years ago, dictionary in hand. I also live and work in Japan now. Today I can click a button under the artist's tweets and get an instant translation that looks coherent (and often is, though it can also be quite wrong maybe 1/10 times).

In terms of the artist being accessible to overseas fans it's a great improvement, but I do wonder if I had grown up with this, would I have had any motivation to learn?

View on HN

I am learning Japanese (again) now and it's such a stark improvement vs when I first tried. When I don't understand something, LLMs explain it perfectly well, and with a bit of prompting they give me the right practice bits I need for my level.

For a specific example, when 2 grammar points seem to mean the same thing, teachers here in Japan would either not explain the difference, or make a confusing explanation in Japanese.

It's still private-ish/only for myself, but I generated all of this with LLMs and using it to learn (I'm around N4~N3) :

- Grammar: https://practice.cards/grammar

- Stories, with audio (takes a bit to load): https://practice.cards/stories

It's true though that you still need the motivation, but there are 2 sides of AI here and just wanted to give the other side.

Exactly this. LLMs make learning faster and easier for those who _want_ to learn, but conversely make it harder for those who don’t.

> When I don't understand something, LLMs explain it perfectly well

But my man, how do you know if it explains perfectly well or is just generating plausible-sounding slop? You're learning, so by definition you don't know!

Because at the beginning I didn't trust it and verified it in many different ways. I've got a fairly decent understanding of what LLMs hallucinate regarding language learning and levels, and luckily/unluckily I'm far enough for this to be a concern. e.g. I asked recently differences between 回答 vs 解答 and it was pretty good.

I also checked with some Japanese and my own notes contain more errors than the LLMs output by a large margin.

GPT was dismally, consistently wrong about 101-level French grammar. Isn't that info all over the internet? Shouldn't that be an easy task?

Yeah, my experience with AI and Japanese is quite the opposite. I used to use the Drops app for learning vocabulary, until they added genAI explanations, because the explanations were just wrong half the time! I had to uninstall the app!

Similarly, I used the Busuu app for a while. One of its features is that you can speak or type sentences, and ask native speakers to give feedback. But of course, you have to wait for the feedback (especially with time zone differences), so they added genAI feedback.

Like, what’s the point of this? It’s like that old joke: “We have purposely trained him wrong, as a joke”!

Sounds like the other user tolerates the flaws and works around them, while you don't and expect to not be fact checking every statement. I guess people's approaches to this can be very wide.

I am more on your side, personally. When learning, I do not want to spend half my time scrutinizing the teacher. At least not on the objective fundamentals. If the fundamentals are broken, how do I trust anything built on top?

And that's the scary thing: LLM's don't "build on top", they more or less do the equivalent of running around the globe first and come back with an answer.

I'd say a 10% error rate is quite drastic for spoken communication. Can you imagine if you lost 10% of the context for the comments here on HN?

I don't know, I feel like one of the worst things to do on the internet is to miscommunicate. And that can be hard enough with 2 native speakers (I just had a response completly blow up on me because they assumedly took my comment the wrong way).

>but I do wonder if I had grown up with this, would I have had any motivation to learn?

Machine translation has come a long way where I can at least get the general feel of say, a translated article. It is still far from the point where I feel like I can read a machine translated article and not have it come off as clunky. That last 10% or so is just uncanny enough for it to impact my reading experience.

As a fan who can read both languages, the ~10% error rate does concern me, since it's coherent-looking enough that many fans don't seem to seek out or share fan translations anymore, but still has a definite potential to blow up in a misunderstanding. However, there's not much I can do about it.

Reply on news.ycombinator.com