Hey, we all have opinions and the opportunity to post them but I’m sure mine will not “effectively shut down” any further discussion. In all the time it took to write and read all this, reckon I could have cracked the Italian congiuntivo: -) (Just gentle humour). In bocca al lupo!
I see you like to use the word “spineless” quite a lot. In Italian “smidollato” sounds pretty close. Anyway it made me smile to think of you posting your thoughts on the now defunct Duolingo forum; you would have come across some “interesting” mods there, for sure.
Moving on though, which of your languages are you enjoying the most?
yes, but, lol
For English to less popular languages, the Clozemaster database has thousands of sentences per language. For English to more popular languages, tens of thousands apiece… with zero overlap. (Lots of the source sentences are the same across languages, but that’s immaterial here— if you translate one sentence into 23 languages, you’re still going to need 23 separate explanations, each one written individually from scratch.)
Soooo. In total, that’s at least several hundred thousand explanations, and possibly even millions of explanations—of which a skilled human could only hope to write a few per hour, with nowhere close to the word-by-word treatment that chatGPT outputs.
Even at minimum wage levels, that’d be millions of dollars’ worth of work. Realistically—given how few people are actually capable of writing clear and correct grammar explanations (most native or otherwise fluent speakers of any given language rely completely on experience and intuition, so only a tiny minority of them will even have any explicit idea WHY each piece of a randomly chosen sentence is correct… and out of that already tiny minority, far fewer still will consistently be able to formulate written explanations that are clear and cogent enough to be understood by readers who are studying alone)— you’re looking at tens of millions of dollars, given the kind of wages you’d have to pay for work of that quality. Imagine the Pro fees we’d have to fork out to cover that😵💫
… and that’d be the cost of human-produced explanations just in English, ahhahhaha. Not even touching all those other source languages.
I mean. There are other corrections I would love to see on here—especially for pairings of two non-English languages, which contain huge numbers of actual errors in the base Q&A (not just in explanations).
E.g., Russian from Italian, and Italian from Russian:
These databases—at least the great vast majority of them—have clearly been constructed by just daisy-chaining data from the English-Russian and English-Italian databases. (This is the best that can be hoped for out of all automated procedures, since there’s no such thing as direct auto-translation between non-English languages—ALL existing auto-translators from western tech companies route everything through English, which would introduce even more errors and ambiguities, like way way WAY more.)
The problem is, because English is the intermediate stopover point of this process, absolutely everything that’s baked into Russian or Italian but not English grammar—and absolutely every instance of distinct forms in Italian or Russian whose closest English translations look the same—is lost going TO English, and therefore has to be randomized (basically just “guessed” by the system) FROM English to Russian/Italian.
The three most common of these are:
• Russian past-tense verbs are gendered; English and Italian verbs are gender invariant.
• Possessives are gendered to match the possessor in English and Russian—not in completely matching ways, but crudely enough to make a go of it—but do not change with the gender of thing possessed.
Italian (like every other Romance language) is the opposite on both counts: “her”, “his”, and “its” are all the same word in Italian—but that word takes on different forms depending on what fills the blank in “his/her/its _____”.
• English only has one “you”—whether singular or plural, format or informal.
Russian has one word for informal singular “you”, and another that doubles as formal singular “you” and plural “you” (also how French does this).
Italian has one word for informal singular “you”, a second word for informal plural “you”, a third word for formal singular “you” (which is the same word as “she”!!) and a FOURTH word for formal plural “you” (which is the same word as “they”!).
Just from these 3 differences alone, in either direction of Russian <—> Italian there are loads of wrong translations in the database.
• Loads of stuff ends up gendered the wrong way.
• EVERY instance of “you” is always completely randomized between formal/informal in both Russian and Italian, very often in non matching ways (since English has nothing even remotely close to formal/informal register). Singular vs. plural “you” is also, separately randomized in both RU and IT, unless it’s fixed by some other word(s) in context… and the "you"s even occasionally cross-pollinate with “she” and “they”, too.
• not even getting into the hot mess with possessives.
where the point is, if you were going English to Italian or English to Russian, then you’d constantly have to guess these things (and add all the other possible forms into the “Alternative correct answers” box)—but at least you’d know you had to guess.
In Russian to Italian or vice versa, on the other hand, plenty of the translations are unambiguously WRONG for reasons that are rlly nobody’s “fault”, and that are unfixable except by a person with professional-level competency in both of those languages.
These would be nice to fix first, before addressing peripheral features like longform explanations—but again, that’s laughably unaffordable for a small developer.
Welcome! Great name, by the way. I read this with interest, some very good points, well made. I’ve learnt over the years never to rely on just one Explanation anyway so I always test the water elsewhere and often ask kind madrelingua in our Italian Club if in doubt. Happy learning to you; -)
Thank you for the kind words, and I wish you the same! ︎
I skimmed over this thread very quickly before but I came back to it now to give my full support to David’s original point.
As I’ve gone through 3,500 Ger-Eng sentences since January I have noticed a pattern that almost every time a sentence has 4+ comments it’s typically someone asking a question and then someone posting an “answer” that goes “chatGPT said: blabla” with zero criticism or context. Sometimes there are no answers from an actual native (/advanced speaker) and sometimes there is a human answer but someone has posted an AI gibberish response anyway, sometimes multiple.
But more importantly and why the thread popped into my mind again, I learned this sentence today and was flabbergasted by the comments. If it was up to me I would have banned Peter on the spot but that’s a quite harsh measure to take when the platform itself uses the same technology for getting further “context” for the sentences. All I can say is I am not in the least surprised that this thread is made by the very David nor that Peter has chimed in with further valuable remarks.
I get that the explanations can be a powerful asset when learning languages. I checked one before writing this for context and at least that one looked fine and comprehensive at a glance (not that I’d know if it was correct). That is why I would like to propose a compromise that lets you both keep the feature and improve on it:
First, when a new explanation is generated (which you thankfully only do once per sentence I noticed) have the actual language model used obfuscated and instead present a big and bold “This explanation was generated with a LLM and must be taken with a huge grain of salt as it might be completely wrong.” This should make people less prone to deifying the tech and the magical thinking around AI. We don’t need to know which model was used but we do need to know to be cautious of it.
Second, have the explanations available somewhere so that natives can check any new ones and either mark them as fine or edit them (or discard) if needed. For explanations approved by humans you can instead display at the bottom “Explanation approved by [user]” and even display the Explain button in a different colour to indicate that the sentence has a proper explanation to it.
I like these ideas.
That’s what I meant by this comment in that other thread in this forum that also raises concerns about coherently-looking but wrong “explanations”:
I’m sorry, but I don’t like the idea, although laudable. My problem with this is that going through even a subset of all the AI explanations is a gargantuan task, given their verbosity and their abundance of low-level detail. I’m afraid it would take very long for me to review an explanation to the degree that I would dare give it my seal of approval.
From what I understood there doesn’t exist an explanation for every sentence pair by default but instead one is generated the first time someone clicks the button. If this isn’t the case then at least it could be made so that an explanation only pops up to be reviewed based on such idea anyway. This would greatly reduce the number of explanations needed for approval. Moreover I would imagine that different users desire explanation for similar sentences, like when a new piece of grammar or an idiomatic phrase pops up the first time.
I will admit that I personally would never do any approvals simply because I don’t want to read anything from a chatbot but not everyone is as vehement. My proposal was that if the explanations are here to stay we could at least improve the concept. As I said I’m all for simply removing them as I have found the old system of “if you have questions about the sentence ask away and someone might answer you in time” much preferable because the other users might have asked or pointed out something I didn’t even think of and then I’ll get extra context (actual human-provided context) to learn if I just click the comments open when I notice there are some. Or at least that used to be the case before the answers turned into “chatGPT said: [who cares]”.
I agree with @morbrorper. This would be a gargantuan task.
The problem with @Tuutti’s proposal would be: Who would do all this work? I, at least, wouldn’t do all this work for free.
I second that. If the feature that generates text by drawing words from some probability distribution won’t be removed (which has been stated multiple times already—that it won’t be removed, at least for now—which I respect and accept), then I agree with the view that at least the efforts to verify the generated texts (or to make people aware of the fact that these texts could be utter nonsense) should be increased.
If, in the end, more people need to be hired to do that (paid) work because it would be a gargantuan task that nobody (so far) would volunteer to do, which would probably lead to an increase of the costs for the customers, then one could return to the question, “Is all of that really worth it? Just to rescue a feature from elimination that is correct maybe half of the time?”
I would just like to reiterate my post of way back when since I find it extremely useful alongside other means. Perhaps I’m too easily pleased …
If people want to use ChatGPT for learning checkout Ecosia’s chat https://www.ecosia.org/chat?q=ecosia
It’s a version of ChatGPT but is running on 100% renewables (also worth checking out Ecosia as a search engine as it too is running on renewables and uses its profits to plant trees).
Perhaps Clozemaster could use this too one day?
this feature is extremely helpful and an essential add-on to cloze master.
There needs to be a shift renewable energy especially on the backdrop of increased usage of AI models. Nonetheless, neither this forum nor the developers of a language learning at will influence this.
It is difficult reading such condescending nonsense from someone that is emotionally bound to a global warming (or cooling?) theory. I find the explain feature extremely supportive. I haven’t experienced an error (though the pronunciations can be wrong). Humans make more errors than I expect from ChatGPT as that tool quickly scours the web and summarizes to create a readable explanation. And only has to do this one time and store the explanation? Efficient!!
It is a major reason for using ClozeMaster. Without the Explain feature, I would resort to Glossika.
Thanks Mike for not punishing the thousands to satisify the few (who don’t have to use the feature).
Just to clarify and see whether I understand you correctly: Are you denying (or not believing) the fact that Earth’s climate is changing due to greenhouse gas emissions from burning fossil fuels?
I think getting into sensitive political issues will eventually spoil this thread which we have all enjoyed so far. Already I’m taking your advice and getting out now.
Here’s an open question for the people who like to use the Explain feature: Have you tried using other resources – especially Wiktionary – to obtain similar information? If so, how did it compare? Are you aware of the feature that lets you click on a word and look it up either on one of the sites that Clozemaster gives you by default (Wiktionary being one) or a site that you configure yourself?
For every language I’ve studied except Hebrew, I’ve been able to find anything I wanted to know about a word’s grammatical characteristics (number, case, etc.) and etymology at Wiktionary. For Hebrew, where Wiktionary coverage is not very good, I’ve been able to find the information at Morfix and/or Pealim. It might take two or three clicks rather than one, but my experience shows that I can trust the information without worrying about it having been hallucinated.
I personally stay away from generative AI for a number of reasons. I’ve often seen it make mistakes in subject matter I know, so I don’t trust it with material I don’t know very well. I know that it uses huge amounts of energy (even if there might be ways to offset some of it by saving results previously obtained). It’s not very good at giving credit to its sources (though the AI I automatically see when doing Google searches has gotten somewhat better in that regard). It forces me to relinquish some control about which results I investigate further that an ordinary search doesn’t.
In this discussion, I see some people saying they like using the Explain feature. What I don’t see them doing is comparing it to the alternatives.
Thank you for writing this, @alanf_us.
Exactly. This is called the Gell-Mann amnesia effect. Which means: When you know a topic well, it’s relatively easy to spot when somebody says something wrong about that topic. This applies not just to “AI” but also to journalism.
But when you’re a beginner or unfamiliar with a topic, you tend to take at face value and simply believe whatever you’re told by a perceived authority (like a newspaper or an “AI”), even when it’s completely made up or not based on facts. “It was in the news, so surely it’s correct! They wouldn’t print wrong information, would they? As a newspaper, they certainly have fact-checkers to get their stories right, no?” Except that humans don’t even (consciously) have that thought, they simply believe without questioning. Even when they just saw an article whose topic they’re very familiar with where the newspaper got it completely wrong.
I’m not saying that you’re stupid if this effect applies to you. I’m not saying that you fell for some trickery due to being too naive. Before I’m being called “condescending” by a climate-change denier again who thinks that all the facts I posted thus far are “nonsense”. This Gell-Mann amnesia effect is just how the human mind seems to work.
This is exactly the case when you use “AI”. Often, you don’t even notice when it’s wrong. Even when you did notice that the ChatGPT explanation got it wrong, some people never seem to learn the lesson to eventually not trust the ChatGPT explanation anymore. It’s as though each new sentence were a fresh new start where ChatGPT gets yet another chance. For some people, ChatGPT can be completely wrong as often as it wants and yet they still put their entire trust into ChatGPT again in the following sentence. This is the Gell-Mann amnesia effect. Some people brush each and every mistake aside or immediately forget that ChatGPT had just been completely wrong in the sentence they just saw. This is the ‘amnesia’ part. And I believe that many Clozemaster users are unaware that they’re under this effect.
@alanf_us might be onto something: Are all defenders of the Explain feature even aware of the existing features? After all, they’re all hidden in various sidebars, you have to know how to find and use them, et cetera. In the mobile app, it’s even harder to click or hover with your finger over something than it is on your desktop computer with a mouse. Whereas the ChatGPT explanation is directly under their nose, easy to find, and requires just one intuitive click on a simple big button in an obvious position in the UI.
I wonder what would happen if, hypothetically, a click on that “Explain” button would open Wiktionary instead of ChatGPT. Make Wiktionary as “under your nose” as the GenAI explanation currently is. The Clozemaster users obviously like the UI. The one button that gives the ultimate summary, instead of having to look up, say, 5 words 5 individual times. Hypothetical question: What if we replaced that backend behind that button (ChatGPT) with something else? Maybe the Wiktionary links etc. just need a better UI so that Clozemaster users no longer feel they’re “punished” by the hypothetical scenario of no longer having that obvious simple “Explain” button (and I specifically mean the UI element that is the button, not the ChatGPT explanation it represents). I don’t mean to remove that button from the UI, that can stay. I’m just arguing for a more sustainable backend.
Wiktionary is designed to search a single word at a time. I don’t find this an inconvenience because I almost always only have a question about a single word (a consequence of the fact that I only use Clozemaster for languages where my level is intermediate). On the few occasions I have questions about multiple words within a sentence, I don’t mind looking them up individually.
However, if people use Clozemaster for languages where they frequently have questions about multiple words in a sentence, that might explain their fondness for the “Explain” button. I can’t think of a way around that other than automatically preparing Wiktionary-type links for every word in the sentence at once, which would seem to be a usability nightmare.