and welcome to your German Word of the Day. And today’s word is:
Which of course means toy, but we’re not actually going to talk about the word.
I mean… it’s Spiel+Zeug, “play-stuff”. What’s there to explain :).
No, instead of talking about Spielzeug, we’ll actually use Spielzeug.
Because I have created a little toy for German learners.
And it’s with ChatGPT…
Yeah, I’m actually not kidding.
I’ve built an example generator with ChatGPT.
Now, many of you have probably heard about ChatGPT, but some of you have not so here’s a little primer. Feel free to skip it, if you’re up to speed about it already.
What is ChatGPT
ChatGPT, or the GPT-series is a very large and powerful artificial neural net by a company called OpenAI (which is of course completely proprietary). The model is built and trained to create coherent and sensible human language and it’s very impressive how far the tech has come in the last two years.
Lex Fridman and other people who larp as AI experts are even low key suggesting that there are some sparks of consciousness in it.
I think that’s complete nonsense and I have actually spent quite a bit of time probing the model and I have plenty of examples that show that there is absolutely “no one at home” inside this thing. Like, seriously… I can post you some screenshots of how it fails very simple tasks.
It’s just a very very very complex machine.
Still, it’s very impressive what we can do with it.
I actually had my eye on OpenAI since 2020 because I wanted to use it for language teaching and learning. It’s just logical, and Duolingo actually does that now.
Quick Duolingo Rant
They have integrated GPT4, the most recent and most biggest version of GPT, in their new, flashy Duolingo version called
Duolingo Max Price
Or Duolingo Max for short.
You can have dialogues with the AI now and you can ask it to explain why an answer of yours wasn’t accepted, which is cool. But the offer is 20 bucks a month. Which I find heinously overpriced.
I mean… Duolingo says in its blogposts that they’ve done some super duper engineering together with OpenAI, but I don’t believe it. I think that a couple of engineers just did some fine tuning of the GPT base model while 90% of the team was busy A/B testing what the better prompt is after a lesson.
And now they’ll be getting swamped in reports about wrong AI sentences that they’ll never get on top of. Because… GPT is not all that reliable when it comes to language learning. It does make mistakes and spit out unidiomatic sentences and phrasings from time to time. I’d say 80% is good, 20% is not idiomatic. So for millions of users, that makes hundreds of thousands of possible reports per day. And no, I do not believe that they retrain the model as they claim.
But then again, Duolingo has spent years making its users accustomed to weird examples that make no sense and seem random, so GPT fits right in there. They really played the long game there. Well done Duolingo!
Anyway, I think ChatGPT is not quite good enough yet to be integrated in actual reference material (like my dictionary for example), but I also think it’s a useful tool and I have a number of plans I want to use it for in the context of learning German.
Anyway, the thing that’s already done, at least for now, is… drumroll… my epic example generator
My New Example Generator
The idea is simple – you enter a word (or a short phrase) and you get an example in German along with the English translation. And because that alone is boring, I have added multiple style options and I’ll probably add more.
So you can generate an example for the same word in various styles :).
Here’s a screenshot from when I was testing the system:
I mean… that’ some real wisdom.
And this isn’t heavily cherry picked either. The Buddha quotes in particular are usually really good.
I find it really quite fun to play around with, but there are a few things you need to be aware of when using it:
1 ) It sometimes ignores your word:
Instead, it uses a relative or synonym. For instance, instead of leihen it might use borgen. And for Überheblichkeit, it kept using the adjective überheblich. The reason for that is that words are processed as concepts and those are somewhat fluid. If two concepts (or vector embeddings, for the techies among you) overlap a lot, there’s a good chance the neural net takes in one and puts out the other.
I could try and force the model to use the exact word entered but then it would probably not conjugate anymore or stop using verbs in various tenses.
2) It sometimes says things that are not idiomatic:
In one example, it used groß in the sense of great, awesome, which is absolutely not idiomatic German. So if you get something that strikes you as odd, chances are, it probably is. After all, the model has no concept what’s idiomatic and what isn’t.
3) It doesn’t always get the grammar right:
Especially when you enter a phrase like bestes Bier, it might use the phrase in a sentence without adjusting the adjective endings. I think this is something I can iron out over time by fine tuning the prompts I am using, but
4) It gets stuck on one phrasing
Sometimes, it gets hung up on one meaning or use of a word if you generate multiple examples with it. For instance, I tried out doch and once used it as aber in the first position, it then kept doing this again and again. If that happens to you, pick a new word or refresh the page.
5) It struggles with prefix verbs and fails at words like “doch”
Yes, the quality of examples for prefix verbs is comparatively low. Or let’s say, the failure rate is higher than for nouns for instance. And don’t expect to get anything idiomatic for words like “doch” or “halt”.
I will soon add an option for “context”, so you could tell it to make an example for ausmachen in the context of “light”, for example. I think that’ll improve it somewhat.
So yeah, these are the drawbacks of the system.
If you want to systematically collect examples for example for all the meanings of a prefix verb for example, then this is NOT the right tool.
It’s more of a fun to play around with, and collect some vocabulary and it works best with nouns and adjectives.
But enough with the words… time to try it out!
I’ll add this to a dedicated page soon, so you can always find it directly from the menu and it’ll be a members only feature then, but for now, it’s free.
And I have several ideas for additional features and I’m also working on making it into an actual app. I just don’t have experience with that, so step by step.
Anyway, if you have other ideas for features, let me know!
And also let me know how you like the tool and which style is your favorite. And of course share your examples if you have a really hilarious one.
That’s it from me for now, have a great week and see you next time :)
Oh, and if any of you have used Duolingo Max… I’d really love some impressions and how the AI is doing there. And while we’re at it… Duolingo is not the only company that has “closely worked with open AI”.
Also, if you want to play around with AI learning, check out Langotalk.org
It’s got the same features as Duo Max, plus some more, plus it’s cheaper and looks better.