checked out the tool and think it's a cool idea! one piece of feedback though - I actually feel like the inverse product would be more helpful for me. What I mean is replacing ~95% of english text with words (Chinese in my case) that I can understand, and leaving the remaining ~5% (words I definitely don't know) in English.
At least for me, there's large value in consuming bigger volumes of Chinese to get me used to pattern-matching on the characters, as opposed to only reading a smaller amount of harder characters that I'm less likely to actually encounter
That makes a lot of sense, it really highlights the diffences in learning stages.
My current tool if primarily designed for intermediate language learners who have already learned some basic words, but still in the 'accumulation phase' - their main bottleneck is vocabulary size, so they need to see new words frequently.
it sounds like you are at a more advanced stage of learning Chinese, you have moved past simple vocab building and are focusing on flow and fluency reading. For your use case, that 'inverse' approach (Chinese with English safety nets) is definitely superior for pattern-matching, it's a different problem set, but a very valid one.
That's a really cool concept. Naively replacing words might work, but sometimes the context is needed. Maybe a model like gemini 2.5 flash lite would be fast enough but still maintain better context awareness?
I have personally had success with using Kimi for Chinese creative writing making the same assumption that Moonshot, as a Chinese company, has more/better Mandarin language pretraining data
As a fellow Mandarin learner - this is super cool! Intuitively makes a lot of sense for the "full immersion" component of language. I love to see exciting uses of AI for language learning like this instead of just more slop generation :)
I haven't dug into the github repo but I'm curious if by "guided decoding" you're referring to logit bias (which I use), or actual token blocking? Interested to know how this works technically.
(shameless self plug) I've actually been solving a similar problem for Mandarin learning - but from the comprehensible input side rather than the dictionary side:
https://koucai.chat - basically AI Mandarin penpals that write at your level
My approach uses logit bias to generate n+1 comprehensible input (essentially artificially raising the probability of the tokens that correspond to the user's vocabulary). Notably I didn't add the concept of a "regeneration loop" (otherwise there would be no +1 in N+1) but think it's a good idea.
Really curious about the grammar issues you mentioned - I also experimented with the idea of an AI-enhanced dictionary (given that the free chinese-english dictionary I have is lacking good examples) but determined that the generated output didn't meet my quality standards. Have you found any models that handle measure words reliably?
good question - however I don't think these are necessarily mutually exclusive.
I have repeatable workflows that harness the benefits of multiple agents. Repeatable workflows drive consistent results for single agents. Using multiple agents allows you to fully explore the problem space.
An example of using these concepts harmoniously would be creating a custom slash command that spawns sub-agents that each have custom prompts, causing them to do more exploration. The commands + agent prompts make the flow repeatable + improvable
I've been exploring the "AI as conversation partner for immersion" use case for a project I'm building and find it pretty helpful for a few reasons
1. Effectively infinite engaging comprehensible input at your level
2. Fantastic way to practice new vocabulary and grammar patterns (AI can provide correction for mistakes)
3. Somewhat fun - if you view chat as a choose your own adventure, the experience becomes more interesting
I just opened chatGPT's voice mode and mocked the worse accented english I could muster asking for tips on pronunciation.
chatGPT just told me that my pronunciation was perfect over an over. It's transcribing audio into text and has no sense for details needed to improve conversational skills.
I've tried speaking danish to ChatGPT and asking it very simple questions. I even tried using complete words and pronouncing them properly (inb4 kamelåså)[1], but it didn't help. I didn't manage to have it transcribe a single sentence properly.
I believe you, but I'm surprised it doesn't do Danish. It manages Cantonese though, which I think is fairly niche (Google translate doesn't support it).
I’m pretty sure the point is to have a conversation with someone (something) who is speaking correctly.
As another poster here noted, the effect of error correction is nowhere near the effect of having correct input. (See the “comprehensible input” poster.)
I've been using these fundamentals (calm, non-gamified, emphasis on focus & flow) for building a Mandarin language learning via chat with AI. My goal was to give the user a focused tool (i.e. chat with an AI at your level) and let them experiment & play at their own pace.
However, due to the more user-driven approach to this learning method (output-focused, user has to put in effort to chat with the AI and get feedback), there is more friction with using the tool. This isn't necessarily a bad thing - in fact, more friction can lead to more meaningful experiences. That being said, I believe the market will push tools to be low friction and low effort (i.e. gamified apps) that are focused on consumption rather than tools that require more user effort.
just my 2c from a fellow builder. if curious, check it out here! would love any feedback
I'm working on character.ai for learning Chinese, you chat with characters at your level, and get instant feedback on your writing. It's a way to get a wide amount of comprehensible input in an engaging way that also practices output.
This is really cool, I'm interested in this as I'm also a chinese learner and I thought about doing sometihng kinda similar (just locally)
I like the UI, really cool project.
I think the prompting might need more work to make it natural though. I just tried a "hungover chat with 996" worker, and the responses seemed to be lacking a little too much context
At least for me, there's large value in consuming bigger volumes of Chinese to get me used to pattern-matching on the characters, as opposed to only reading a smaller amount of harder characters that I'm less likely to actually encounter
reply