I'm working on https://nuenki.app/, a language learning tool. It teaches you a language while you procrastinate by inserting translations of appropriate-difficulty sentences into webpages as you browse HN etc.
Currently trying to reduce costs by switching from using DeepL (high quality, low latency, high cost) everywhere to a hybrid that also uses Claude (high quality, high latency, low cost) for text that is far from the user. Also experimenting with Gemma 2 9B via Groq to go in between them, but it's bad at following instructions and I don't quite trust the quality numbers I'm seeing for it (they're benchmarked with gpt-4o as a judge).
I'm also trying to work out marketing. I'm not good at it, and I dislike it, but I need to get good at it. Currently considering Reddit ads for awareness, some content marketing going over the technical details (there's some fun language processing and performance optimisations), and... I feel that's not enough, but I'm not sure what to add to that.
I'm running on very little budget (I just left school and I'd rather not go into my limited savings over this), so I can't afford to just throw money at ads.
Toucan has a similar concept, but it works on individual words rather than sentences. Beyond a smaller scope, it's also quite difficult to accurately translate single words without context - Nuenki has it disabled by default.
The tradeoff is that Toucan is free, and translation is quite expensive.
Would be really useful to have some way to hear the translated sentences so that you can also learn how it's supposed to be said - would be very useful for learning Japanese.
Some feedback: I was about to try now (paying) but currently I can only install it in my desktop (no iOS support), and I am very worried about the idea of giving full access to a close-source unpopular extension with not-very-clear ownership.
I think it would go a long way to create some trust to show who are you in the webpage and maybe open source the extension too (the backend is not as importa nt).
I'm considering open sourcing the extension. I don't want people to be able to copy it wholesale, but realistically that's unlikely and I'm going to write some blog posts on the technical details anyway.
Please do! i would love to contribute, ive been looking for something like this for some time but didnt want to install toucan since its closed-source and i dont trust it
awesome! i am also exploring “low effort language learning” space and making an app. would be interested in bouncing some ideas, got a social handle to share?
looks cool,it would be useful to have transliterations for non latin alphabet languages,or at least pinyin for Chinese, since learning characters is somewhat orthogonal to the words.
I'm also learning Chinese, and I honestly wouldn't care for Pinyin all that much. (Once you get to an intermediate level of Chinese, you'll want to learn characters anyway, because that's how Chinese is written in the real world.) That said, a hover option to transliterate and translate individual words (instead of the whole sentence at once) would be great.
For a more practical complaint, in the Nuenki browser extension, the settings window has a primary-colored (green) button in the bottom right that says "Log out", but has no "Save" button. I almost clicked "Log out" when I just meant to confirm the settings I had made!
Edit: Also one question, if I may - how large is the context window for the translations? I noticed that the Chinese it generates sometimes feels a little unidiomatic, but I'm not sure whether that's due to too little context, the fault of the translation engine(s) used, or just my feeling as a non-advanced Chinese learner being off.
Edit 2: Some of the translations are definitely just wrong, too. Again, might be the fault of the engines - after all, machine translation, even in the age of LLM, is not perfect.
I can't really translate individual words (translation is done on a sentence-by-sentence basis, so I don't have that data), but I could do transliteration. Perhaps hovering could transliterate, clicking could translate the whole sentence, and clicking again could reset it. It's difficult balancing the amount of things people want to do with the text (listen to audio, translate the word, translate the sentence, and transliterate the word) with how few input methods there are + text shifting.
The context window is only that sentence. It would be technically possible to increase the context window, but it would make it impossible to cache and increase translation costs even without that. Unfortunately the development of Nuenki has been defined by mitigating the cost of translation - before I had my many mitigations my prototype burned through six euros of credit in a day - so it's not practical to expand that context window.
I'll definitely recolour that button. Thanks for mentioning it!
The incorrect translations are the engine's fault. The version you're using uses https://www.deepl.com/en/whydeepl . Looking at my quality data for the new update (which adds Claude and Gemma as translation sources in some circumstances), that'll improve a little soon but ultimately not by much.
Currently trying to reduce costs by switching from using DeepL (high quality, low latency, high cost) everywhere to a hybrid that also uses Claude (high quality, high latency, low cost) for text that is far from the user. Also experimenting with Gemma 2 9B via Groq to go in between them, but it's bad at following instructions and I don't quite trust the quality numbers I'm seeing for it (they're benchmarked with gpt-4o as a judge).
I'm also trying to work out marketing. I'm not good at it, and I dislike it, but I need to get good at it. Currently considering Reddit ads for awareness, some content marketing going over the technical details (there's some fun language processing and performance optimisations), and... I feel that's not enough, but I'm not sure what to add to that.
I'm running on very little budget (I just left school and I'd rather not go into my limited savings over this), so I can't afford to just throw money at ads.