Hacker News Reader

> Even if it's a success, the likes of DuoLingo can copy that in a month and have so much more marketing and money to win the market

Only theoretically; in practice, Duolingo have a specific brand — IMO aimed at helping younger kids with schoolwork — they don't want to risk damaging, so if you target your language course at people who chafe at the style of art, animation, voice, and/or content of Duolingo[0], you still have a business opportunity even though Duolingo is also already working with GPT-4.

[0] This is why I've stopped using Duolingo entirely, while also having paid versions of Clozemaster and Babbel, which are also very different from each other.

>aimed at helping younger kids with schoolwork

IMO duolingo is aimed at peopple(adults) who say they want to learn a language but aren't ready to take it seriously

IIRC their demographics are heavily skewed towards younger people.

In US/Japan etc. the largest age-group of learners is 13-22

- Found the source: https://blog.duolingo.com/dear-duolingo-how-does-language-le...

I was really puzzled by Duolingo for ages - I just found all the animations etc. painful and other apps seemed to much better for language learning. And then my kids (10 and 12) started using it and absolutely loved getting streaks, lingots and so on and were choosing to use it for fun. It doesn't really matter that it isn't the most efficient way to learn French if they enjoy it.

Any sort of LLM-based tutor is tricky with children as you can't guarantee that it won't come up with something inappropriate, even with the various guardrails in place. As a parent, I'm happy to take that risk especially if I am around to explain anything, but if I were a teacher for example then I probably wouldn't be and I imagine businesses are going to be in a similar situation.

That explains it. I sometimes try it out again (another year, still no stories for Chinese) and the amount of animations and popups you have to deal with for finishing a lesson felt like they take longer than the lesson! "You completed a lesson, you started a streak! Set a goal now! You completed a quest and got a streak freeze! You moved up in the league! You earned points for the event! Invite your friends!"

> the amount of animations and popups you have to deal with for finishing a lesson felt like they take longer than the lesson!

Same.

For over a year before I finally gave up on it, I was force-quitting the app and reopening it for the next lesson because that was faster.

I watched the video, but I don't think this is practical. You can't realistically learn Japanese with gpt 3.5 as it makes ton of mistakes, gpt 4 is better, but still wouldn't use it. Even in this video it doesn't explain meaning of word 動産, but instead only breaks it down to kanji. Also you made mistake in your answer, you don't say おほしい and gpt didn't even correct it.

I've rewrote the "grammar prompt" for the helper tool on the right since the video and it does behave much better now. Although, as you say, you're limited by what 3.5 can do.

But for GPT not correcting me, it's "a feature not a bug"©. Specifically, he's supposed to act like a Japanese person and if he can understand what I'm saying, even if the conjugation is wrong etc. it should work with it. However if I speak in complete giberish or english etc., then you have a "game over" where you fail the scenario.

It's supposed to be a "japan simulator" so to speak, and in real life, people make effort to understand you, you can use gestures also, even write stuff on paper if necessary.

However I'm currently adding a "validate your answer" button where he should tell you if it's grammatically correct etc. and allow you to take it back.

Ultimately, it's not going to be perfect, especially not with 3.5, but at my level of Japanese I think it's more about forcing me to be creative when I speak, in scenarios that matter to me. Not just following a script from Genki, but needing to explain stuff specific to my situation etc.

> What we do is fairly easy, the whole "value" is in some fairly simple prompt on top of openAI APIs.

We are seem a lot of these apps nowadays. People come up with a prompt, send to OpenAI with the user data, hide it behind a certain web interface and it's an app they can charge for.

I've thought about this a million times for scratching my own itches but I just talk to ChatGPT and it's fine.

Wow, that looks slick. I wouldn't mind giving it a go!

I've created a mostly empty repo for now, you can subscribe to it to know when I've uploaded the code: https://github.com/johanlajili/watashi/

I would be interested in an OS project like this!

Anywhere I can find out more or be notified?

Hi, I've created the empty project here, I'll add the code soon enough, just subscribe for notifications in that repository if you'd like to get updated.

With the link it might be easier: https://github.com/johanlajili/watashi