Bringing AI into Bunpro

Bunpro
#123

I wouldn’t call AI a shortcut if you’re using it to bring valuable features that are meant to benefit people. I’m not sure what I have to say for people to stop assuming that the Bunpro team using AI means they would ship products that don’t work. They have even said they aren’t implementing it now because from their experience it isn’t up to the quality they would like at this time.

The only implementation of AI I think people who are asking for it are talking about is one that brings value and aid to the users and is what most people would consider useful overall. Nobody wants AI slop and half baked goods and I would be the first one to call out Bunpro if that’s what they made.

Similar to what someone brought up earlier Bunpro is a Business and it’s getting easier and easier for someone to build learning resources similar to Bunpro. They might not get to the same level in a weekend but give someone a year with those people also offering more features and possibly a faster path to learning and I don’t see it looking very good for Bunpro in the long run. There are Bunpro supporters who would stay regardless but I find it hard to believe that’s enough people to support the scale that they want to work at. Along with that I like using Bunpro and support the team and the goals they have but it would be hard to to switch to something else if it was a better learning tool.

It might not be today but I find it really very unlikely as the AI tools continue to get better and better for Bunpro to just ignore such useful and effective resources for the sake of AI = Bad. I understand the issues now but I think given a year or two those won’t really be around or at least not to the degree that they are now for things like GenAI.

#124

Isn’t gen AI hitting a plateau due to lack of data?
I am not sure it will improve much more that where it is right now.
Also mostly all the other japanese app release lately tend to be rushed and vibecoded so I think Bunpro is safe for a while as it is a very niche product.

I am stil not sure how gen AI could be implemented but I am all for an SRS algo rework and a better way of drilling your weak point.

Though the team clearly said it’s not on their to do list for now so we have to wait and see how genAI evolve in 1 or 2 years and if it could finally be trusted without having to jump into hoops with prompting which is the biggest grudge I have with AI.

I never know if it’s telling the truth so I rather not loose time with it. I think apps entierly without AI can thrive just as photography and digital program didn’t kill traditional drawing.
But that’s just my 2 cents, I’ll be watching how the tech evolve and sadly all the bad way it’s implemented.
All the Citroen car in Europe come equipped with Chat GPT so you can talk to it while you drive or it can tell story to your kid in the back seat… yeah sure and of course there is no way out of it so it listen to you 100% of the time… great.
I hate the car industry.

6 Likes
#125

LLMs are getting more not less expensive and are currently only as cheap as they are through massive subsidization. Per token rates which is what I assume BunPro would end up using can quickly get ruinous. Without a really strong use case (with measurable ROI) I really would hold off for now.

6 Likes
#126

Early on, I gave it limits to the types of corrections it would make. At the very beginning, I told it I only wanted corrections that would limit the ability to understand what I’m saying at all. Then, as those mistakes dwindled, I eased it up.

If you think it doesn’t understand the context, you can ask if "x context’ would change it any, because Japanese is a very contextual language.

#127

The evidence I’m curious about is any reason why AI will revolutionise language learning other than some thought of “AI is very powerful and revolutionary technology.”

Hypothetically, suppose AI was as good as the best Japanese teacher and it was a “teacher on demand.” It sure would be amazing, but that wouldn’t speed up language learning dramatically. We already have schools and teachers. And even then what would that add to Bunpro that would dramatically change Bunpro as a product? People will still need textbooks, flashcards and standard resources.

It’s simply incorrect to assume AI features will result in an extinction of previous learning techniques like cars were to the horse and cart. Certainly without any ideas of what would be so revolutionary.

6 Likes
#128

The reason is that AI can adapt to what you specifically don’t understand better than basically any teacher. AI specifically is strong in the area that textbooks are most commonly criticized. “it isn’t clicking for me this way”

1 Like
#129

AI can already do that as a separate product and it’s not clear what Bunpro integration would do to make that feature or Bunpro significantly better, let alone revolutionise language learning to the point of leaving everything else in the dust.

4 Likes
#130

Bunpro has access to all of the data from your current learning so in theory you could build around hundreds or thousands of hours of reviews and history to work with the user in a way asking a separate model couldn’t. In the ideal situation you improve Bunpro by enabling users to get the help they need that can’t be provided by a textbook, without the cost of a tutor which is significantly higher than paying for Bunpro several times over.

Maybe for some people Bunpro is perfect and everything clicks and that’s fine and I don’t think anyone is trying to change Bunpro in a way that would change that experience at all but there are also a lot of people that still struggle to learn even with great tools like Bunpro. So I don’t think it’s such a bad thing to be able to help everyone out in a way that makes learning more enjoyable.

I don’t mind the idea of implementing AI in a way that makes it it’s own section Then if you are fine with using Bunpro as is you can continue and maybe there is another tab for the AI features that you never touch.

1 Like
#131

I don’t see the point. Honestly if you want to add AI into Bunpro, with for example a “ask a question about this grammar point” you can easily build yourself a web extension, throw a Openrouter key and have a complete AI feature. I made a similar thing for another tool yesterday for jivx.com (writing practice) to ask follow up questions and the spend is so low it’s downright ridiculous, here’s for around 10 questions about “why so-and-so wouldn’t work in this sentence” or “why does this pattern uses this particle” etc etc:

image

Any serious AI can probably oneshot the task, or alternatively you can ask the LLM/agent what data it will need (probably one or two copy(document.outerHTML) of the website you’re working on.
Doing it this way sidesteps the sterile debate on whether it works, whether it is moral to use it, whatever culture war issues people-pattern match on, and people who do want it have it, and people who do not want it do not have it.
I highly recommend people to leverage the tools we have now to personalize and improve their own tools and workflows. There’s no reason to wait for the dev team to implement a new software feature if future compatibility is not an issue.
[edit: unbearable typo & somewhat ambiguous phrasing]

5 Likes
#132

Imagine the AI translates “This is a pen” to このペンはあるです。” As someone who has literally 0 knowledge of Japanese, how would they know whether it’s right or wrong?

To me, when learning something with AI, you need to be, at minimum, just under the level of the question you’re asking, or you won’t be able to verify the results.

Thus, AI features take away from features that elevate people’s learning to a higher degree.

Side note, for those of you practicing politeness levels and sentence writing, use Google Docs or Microsoft Word. Its auto-correct is far more accurate than AI.

6 Likes
#133

This doesn’t help me with negotiating politeness levels within 上下関係 tho just whether my sentence was grammatically correct 🥹

#134

The only AI I care about in relation to Bunpro is Bunpro team’s 愛 for the platform. :smiling_face_with_three_hearts:

24 Likes
#135

My main use case for AI with Japanese is sentence breakdowns of native material when I can’t parse something correctly, and there have been times where it has looked at the native material and gone ‘Here’s a more natural way to express this’

It exasperates me and I think really highlights the ‘correction first’ approach of a lot of these models

4 Likes
#136

Wow, a lot of activity in this thread since I last looked.

If you want to get into the nitty gritty of SRS and AI, this is a good read. I don’t necessarily agree with all of the points in it, but it is an interesting summary of SM-2, SM-5+, and FSRS.

https://supermemo.guru/wiki/Neural_networks_in_spaced_repetition

#137

Alright, that was clever.

@onlyelliot I built out a personal supplemental/companion app for Bunpro that reverse engineers the frontend API. I have a dumb brain so I wanted more customizable Cram functionality, especially where troubled items are concerned. No AI on this part.

But I did add a chat/conversation practice module that uses AI that ingests custom decks you build from your Bunpro content (or you can just do random).

If you’re interested, I can tweak some things so it’s more user friendly for someone other than me and get it into a public repo. And I’d be happy to add to it based on feedback.

You would need an OpenRouter, OpenAI, or Anthropic API key to use the chat functionality. Admittedly, I’ve only tested it with an OpenRouter one. It also needs your frontend token from Bunpro to sync your data which is easy enough to find. The only write functionality it has is to add ghost reviews to your queue. And everything stays local (token, API key, synced data). Open source so you can read exactly what it does. Let me know if you’re interested.

1 Like
#138

No one uses the word AI to mean the kind of machine learning that FSRS uses except AI boosters and people who’ve become resigned to the fact that their field (machine learning) seems to have been renamed AI by AI boosters.

The only reason why the public bought the idea that this stuff is AI is because LLMs generate sentences that seem ‘real’. They think it’s a guy. They think it can reason through problems when it’s just building statistically likely sentences. The general public literally thinks anything called AI is the same thing as chatgpt. They think it’s one superintelligence that can do all this stuff.

AI is already an extremely imprecise and misleading term (just like the word ‘hallucination’ which is meant to confer consciousness on something that cannot even reason). Using ‘AI’ to mean anything but LLMs and stable diffusion models is essentially misleading.

8 Likes
#139

Absolutely not. Bunpro is fine without AI and doesn’t need it. If you want an AI language coach that’s riddled with mistakes (because AI is just a statistical model and gets things wrong, like, all of the time), there are plenty of apps out there already. If Bunpro ever goes AI I am leaving.

7 Likes
#140

So many opinions here, wow.

I believe AI could identify the optimal interval for when to show users their grammar reviews. Especially since the current method isn’t as optimized as it could be.
AI could be able to identify especially hard points using data from all users and even find patterns showing someone who struggles with point A, also struggles with B and schedule accordingly.

It could be a opt-in feature, where you’re always free to use “classic” scheduling.

Is it “worth” the time to implement, or the computing power it would cost Bunpro? I don’t know.

1 Like
#141

Just another way for AI to take jobs from people. I use bunpro to help me study, not replace something else. If you really need the extra help, consider hiring a teacher. Support working class people. Ultimately, language is a tool to talk to other PEOPLE, not machines. Don’t take the human experience out of a very human activity.

10 Likes
#142

To be fair, I never made that claim. Instead I said that “if there is a good way to use AI to help language learners someone will find it.”

I happen to suspect that there is such a way, if only because AI is sufficiently flexible that it would be more surprising to me if there is not.

And a product need not be revolutionary to edge out the competition. Machine-made cloth, for example, did not result in a new product, it just reduced the overall cost, even (according to some accounts) at the cost of quality. Nobody makes it by hand anymore, at least on any scale that matters.