I’ve been a regular user of the “Explain” feature. It’s been greatly beneficial to me and my learning. Even though I know it can be inaccurate, it’s been a huge value add for me. However, it seems like the responses have been recorded some time ago with legacy models. Today’s new models from Anthropic / OpenAI / Google are way better at these tasks and re-running this would greatly improve quality.
Additionally and alternatively, I’d suggest including an option to customize your prompt (so many responses start with “Sure!” and have all kinds of fluff) and let users hook up their own API key for a selected model. That way, you can offload the cost and provide the necessary customization.
As far as environmental concerns, our usage of ChatGPT is quite minimal, and we will continue to minimize our use of it
would no longer be true if either of your requests were implemented:
re-generate explanations for tens of thousands of sentences every time the generating models are considered legacy
instead of 1 generated explanation per 1 sentence that all users share, enabling $n+1$ generated explanations per 1 sentence if $n$ different users each want to use a prompt of their own, and presumably also first try out $m$ different prompts on 1 sentence before they like the explanation
It’s great that Clozemaster has worked on saving responses to minimise the number of times that they are generated, which is so much better than almost any other tool trying to use AI. I do agree though that an update with time is needed if better sources become available (same applies to generated audio for pronunciation). In addition, these features are also needed for custom collections (which are key for those who use Clozemaster to supplement a language course with their own required vocabulary per lesson). If Clozemaster can support saving each response for everyone that would be fantastic, but even saving locally just once for the user/device would be a big gain. I’m afraid otherwise people will use other tools where every interaction is based on generating AI contents afresh.
If GenAI weren’t burning our planet, I’d just say, “Let the people chase their shiny new toy.” If the GenAI bubble wouldn’t be crashing our entire economy once investors realize they will never get their money back, I’d just stay silent and let the people get aboard the hype train.
But with Europe becoming more than 5°C hotter than pre-industrialization levels before the year 2100 due to new data centers, this affects me personally. I will have to flee my home country Germany before the next decade is over because it will be too hot to survive in Germany. Because shortsighted people demand ever more AI.
Clozemaster doesn’t “need” an update, same as your iPhone 17 Pro doesn’t “need” to be replaced when the iPhone 18 Pro is released.
GenAI produces mostly “slop”. Let those people go. The more intelligent ones will quickly come back to Clozemaster.
And many, like me, won’t leave in the first place.
I guess ‘needs’ has a subjective component - at A2/B1 I certainly need explanations for some sentences, even if indeed Clozemaster is not the only possible source for that (and for most useful custom collections it is currently not anyway).
Even more than the explanations I need good audio. For Dutch the current voices are are best mediocre (and one is quite bad), especially for elements beyond individual sounds, i.e. with respect to stress (at sentence level and sometimes even at word level), intonation and rhythm which for Dutch are all rather important. It did not bother me much at A1 level, but now starting B1 when there is more emphasis on prosody, I really notice it. I can now generate much more ‘correct’ audio with current tools but unfortunately can not link it with Clozemaster. Of course, ideally there would be a real native speaker audio, but that does not seem realistic for all languages even for the standard collections, not to mention custom ones.
PS: as for uses of AI, I’m not suggesting it for sentence generation or something like that, just to improve the already existing AI elements for better ones.
Your post seems to come from a quite self-centered perspective.
Have you understood my post? Considered any of my arguments? The demand for AI and the consequential construction of additional data centers will kill us all.
That is fine. So do I.
If only Clozemaster would provide some place for exchange where we can pose our questions and point out which part of the sentence we don’t understand—almost like a prompt for an LLM—but instead of a hallucinating LLM an actual native speaker or advanced learner would explain the sentence to us.
I don’t think seeking out further these points of disagreement will help this forum.
PS: I use ‘I’ as I speak for myself and can not presume having a mandate to speak for everybody (which would have been very self-centered)
@davidculley please stop attacking people. Your points may be valid, but arguing with people as above doesn’t create a positive or helpful learning atmosphere and is furthermore unlikely to actually convince anyone. Honey vs vinegar etc.
We are working towards updating explanations and audio as would be beneficial for learning, ie replacing bad explanations and TTS.
That’s an interesting idea!
Work in progress!
Thanks for letting us know! For which collections have you noticed this issue?
I apologize. Attacking people is not and was not my desire. And not what I think I did. As everyone reading this topic can hopefully tell, I care deeply about Clozemaster. And I’ve seen enough products that were run into the ground by ensh*ttifying them through the senseless addition of GenAI. Or decisions where the product manager wasn’t able to say No. I don’t want that to happen to a product I like. I don’t want that to happen by giving in to every hype or all the requests that its users make, whether they are reasonable or not.
We’ve had multiple threads in this forum already where people complain not just about the AI explanations themselves, but also about the spam to which they lead when mindless people just senselessly copy/paste them into forum threads. (For example, this topic.) I also care deeply about our planet and GenAI is just not sustainable environmentally and economically, an aspect that very few people seem to understand about GenAI or take into account. GenAI is also the favorite tool of neo-nazis trying to create a parallel reality to lure people into fascism. Just take a look at the Twitter accounts of Donald Trump or The White House.
In this year 2025, GenAI seems to have become religion for some. Even the slightest criticism or disagreement is taken as blasphemy against their god Sam Altman. Insomuch as they won’t even discuss factful objective arguments that disagree with their worldview or preferences. I’m very open to having an objective discussion that covers both pro and contra arguments, even when there’s disagreement. And I think I stayed perfectly on topic. I can’t say that though about those who didn’t want to discuss both the pro and the contra arguments or those who switched to ad hominem attacks against me rather than arguing against the objective, perfectly on-topic arguments I brought up.
I can see why some people disagree with me. GenAI seems like magic and many people might ask, “Why should we not make use of that magic when we were granted it?” (Does that remind anyone else of Boromir in The Lord of the Rings?) But I don’t see why my previous answer should be flagged as off-topic. The title and original post is whether the existing explanations should be re-run with fancier LLMs. My answer was an argument as to why that shouldn’t be done. Completely on topic. In my answer that was flagged as off-topic I even brought the discussion back to topic, to the arguments that were made but that nobody has considered or argued against thus far. In order to discuss all sides of the topic at hand. Whereas other comments thus far …
I don’t see how that’s not off-topic or less of an attack than what I wrote.
Anyway, I’ve made my points multiple times thus far, in multiple topics. I think it’s clear by now where some users (including myself) stand regarding GenAI. To be completely transparent, I think that GenAI should be smashed to (digital) bits and pieces. I fully agree that it’s unlikely that I will actually convince anyone if I make more comments. Not due to my tone or the soundness of my arguments or anything I could say to these people or have under my control … but because people who don’t want to be convinced can’t be convinced. See my remark about religion. I thus don’t see much value of chiming in another time in topics like this. So I will refrain from creating more comments in discussions about GenAI. At this point I think I will be more helpful if I restrict myself solely to contributing to sentence discussions for the languages I speak. Again, my words are not supposed to be attacks but arguments and observations based in reality. As the German learners in this forum can confirm, I’m all about being helpful and providing good explanations for the subject matter they chose to learn.
Thank you, that’s clear and very much appreciated, as are your contributions to the forum in general and sentence discussions.
Agreed! Also flagged. Now we’re all back to baseline.
We’re not religious devotees and we’re not going all in on AI to your points, but there are some very cool and very helpful applications to language learning that we are trying to take advantage of in reasonably responsible ways.
… I didn’t mean any specific person or any specific person who participated in this thread. Especially not you, Mike. I meant it generally, when I look at our society as a whole. For some in our society it’s true. I also didn’t mean to create a “us vs them”/“we vs you” rift in this forum. We all just try to become more fluent in our target language. Some out of fun, some out of necessity. We all share the same goal. I wouldn’t be part of this forum if I weren’t interested in helping others reach that goal as much as I appreciate them helping me become more fluent. And we are all very thankful for the Clozemaster team and the nice app you and the team are providing us.
… I’m not entirely against AI. I studied Computer Science and specifically deep learning. Why would I have done that if I weren’t fascinated by AI? I agree that there are indeed very helpful applications to language learning. And as long as it’s responsible use, you should absolutely go for them. I like the AI-synthesized audio a lot. And it doesn’t take up a lot of resources to generate them. I also use machine translation (for example, DeepL) every day. Without machine translation, I would often be at a loss. It’s only Large Language Models (LLMs) that I find incredible wasteful. Whereas with Machine Translation there is a clear correlation between input and output and where you can minimize the distance between generated output and expected output, with LLMs they are just word generators. Where nobody can guarantee that the generated output is close to the expected output. (That is, if you receive utter disinformation.) LLMs are by design not meant to produce fact-checked information. I’m only against LLMs, not other applications of AI.
I believe you when you say “reasonably responsible ways”. As you’ve assured us multiple times by now.
It’s precisely because I am an expert on this subject matter that I feel that I need to speak up and take a stance. Same as I would expect a carpenter to speak up when a house owner decides to use asbestos for its excellent insulating properties without understanding how harmful it is.
Thank you very much @mike for spotting these posts! Great to hear about the plans for more features for custom collections.
In Dutch all voices seem to have bigger or smaller problems at sentence level (while most are fine for individual words). Overall, the three voices used for fast track collections seem a bit better, but when Clozemaster audio is available in custom collection, then of the two voices used there one (very similar to Lotte on naturalreaders.com) seems to have wrong stress/ emphasis particularly often.
E.g. on elevenlabs I can generate audio with what seems to be more correct rhythm, emphasis etc. For Dutch, as example, voices “Ido - friendly narrator” or “Thomas - Professional Dutch Voiceover” sound decent, even if potentially not perfect (Í’d have to consult my tutor on that). I expect the voice generation may improve even further in the future, therefore user being able to choose him-/her-self the platform and voice for audio generation (once, then reusing it during the training) would be ideal. Being able to link individual audio file to each sentence (as upload to Clozemaster, as local file or the device or maybe even as a link to a file in cloud storage, preferably not limited to google drive as only option) or, even better, some automated audio generation solution with a single button per collection or per sentence (maybe with own API key as suggested by @young_mete ) would be great in my view.
PS: the same applies to sentence explanations. And also there I think things are already improving - while I note accuracy differences between sources, some are quite accurate (e.g. I have positive experience with Le Chat by Mistral AI)
