*** Quick note ***
Some of you see the newsletter with BOTH sections in English (not German English as usual). This is NOT me. You’re probably on iPhone or some other Apple device and Apple, being the overbearing helicopter nanny it is, seems to think it needs to “help” you. I don’t know where to change this setting or where it comes from, but the newsletter I am SENDING has German in it.
***
Hi everyone,
And welcome to the first episode of the new, epic little summer special called
Update August
Yeay!!
I have recently more or less finished several updates and improvements for YourDailyGerman and also for my Prefix Verb Dictionary that deserve a proper introduction. And while I usually don’t announce small upgrades or changes, the ones I have now are fairly substantial, so I really want to tell you about them and get your thoughts and feedback about them.
And since it’s summer break (which in Berlin means you spend a lot of time at home because it’s raining) I have decided that we’ll do a little mini series where I introduce the new features to you.
And today, we’ll start with something that really improves the usefulness of the YourDailyGerman dictionary: more audio!
So… let’s check it out.
First up, a little milestone.
12.000th Entry
Yes, I recently added the 12.000th entry in my dictionary.
So there are 12.000 terms and phrases in there now, all with HANDMADE translations and usage notes.
I did not scrape an online source or import some random list or use AI, because then it’d be just another soulless, mindless dictionary and no one needs that.
Every single entry is handmade by me. So I decided which translations, how to group meanings if there are several. I added the usage note that you see on many of them, and I entered the value for each meaning and added the family relations (about 70% of words are part of a family tree now). And I did examples for many of them, but we’ll get back to that later.
So yeah, you can find most of the common German words and some less common ones too.
And if there is no entry, you can ask ChatGPT about your search term with just one click.
Which brings us to the next upgrade.
Improved AI Example Generation
When a word you’re searching is not in the dictionary, you’ll get some search suggestions, but you can also ask ChatGPT about it with just one click. This feature has been up for a while, but now it’s BETTER.
I have recently switched to the newest AI-model by openAI (gpt-4.1) and I have to say … the quality is definitely better in terms of accuracy and reliability.
I have also reworked the prompts and the formatting and I think the feedback the AI gives for terms that are not in the dictionary is quite useful now.
Here’s an example for what I got back when I looked for “reingießen”:
***
Translation(s)
reingießen – to pour in (again), to pour (something) in
Example
- Kannst du bitte noch etwas Wasser in die Vase reingießen?
- Can you please pour some more water into the vase?
- Nachdem ich den Tee verschüttet hatte, musste ich neuen reingießen.
- After I spilled the tea, I had to pour in some new one.
Thoughts and Trivia
“Reingießen” is a separable verb: “rein” (short for “herein”/”hinein”, meaning “in” or “into”) plus “gießen” (“to pour”). It’s a very colloquial, everyday term—think pouring liquids into containers. The “rein-” prefix adds the sense of direction (“in(to)”), so “reingießen” is literally “to pour in.” In spoken German, “rein” is often used instead of the more formal “hinein.” “Gießen” relates to the English “gush” and “gust,” both having to do with flowing or pouring. You’ll hear “reingießen” in kitchens, labs, or even when refilling a drink. In southern Germany and Austria, you might also hear “einschütten” for similar situations.
***
This is at least usable.
And this improved quality also goes for the example sentences for words and their meanings.
Example Sentences
I have a lot of handmade example sentences in the dictionary (about 14.000), but many words still do not have examples and I can’t add handmade examples for all of them myself.
So I had added an AI example creator a while back but it was hit and miss.
It did fairly okay for common meanings, though they are a bit bland but for more specific ones, it did sometimes create examples that were unidiomatic or even grammatically wrong.
This has improved significantly now because of the new model and I also updated and refined my prompts quite a bit. That doesn’t mean that all the examples you’ll see will be correct. It’s AI, so nothing is ever 100% reliable.
But so far, I feel like they’re actually good enough to be an actual valuable addition to the dictionary and I am considering adding the AI option also to entries that do have handmade examples. Those will stay, of course.
But all right… all those were just minor upgrades.
Now let’s get to the ACTUAL upgrade.
AI Audio for Dictionary
So far, part of the words in the dictionary had a recording of me for pronunciation and for the rest I used a simple browser based speech synthesis.
But that sucked and didn’t work properly for many of you.
I have now added up to date AI audio creation for all the words and it sounds great!
Here’s how it looks:
First click will create the audio, so it’ll take a few seconds, but then you can replay it instantly. And it should be spoken twice – once normal, once slowly syllable by syllable. That doesn’t always work 100% but it’s fairly good and the pronunciation is generally really top notch (glitches do happen).
But while single word pronunciation can be interesting or helpful in some instances, it’s nowhere near as important as having audio for the examples.
Well… I added that too!
Many examples do have a recording of me reading them, but the ones I added independently of lessons do not. And the ones that ChatGPT generated also did not come with audio.
But that changes now!
I am adding audio creation to ALL examples. Including the ones that you generate with ChatGPT.
Here’s how that looks:
- That’s the part that I do by hand. None of this is automated and never will be
- Here, you can add any note that you want. It’ll always show up here then, also when you use the dictionary app in the future… ahem
- Click here to create examples for this meaning
When you click create, it’ll take a few seconds and then examples will show up.
And what’s new now is that they’ll give you the option to… create audio!!! I don’t know… I find that pretty amazing actually.
Here’s how that looks:
Clicking on “create audio” creates a clip that shows up on the right. You can replay it as much as you want and you can ALSO click “create audio” again and get a NEW version. That’s helpful if the first one was glitchy, for example.
I’m paying for this though, so please do not spam the buttons.
Honestly, I’ll have to see how costs look once you all actually use these features. It should be okay, the price for creating is pretty low and I’ll probably make this a members-only feature anyways.
But yeah… the example audio creation is rolling out first for the ChatGPT generated examples, but like I said, I’ll add it to ALL examples as well, so chances are you’ll see it soon.
And that’s it for today :).
If you want to test out the new features right away, a good entry is the one for “hochgehen” because it has two meanings and no handmade examples yet. Here’s the direct link:
“hochgehen” in YourDailyGerman Dictionary
(opens in new tab)
I know quite a few of you are skeptical about AI and trust me… I am too. But I think it’s a good enough now to enhance the dictionary and I hope you give it a try.
I’m really curious what you think and if you find this useful, so please share all your thoughts and feedback in the comments. And also if you have any ideas about how to improve it. I really hope you like it so far, have a great week and I’ll see you next time with a proper lesson… or another update. Like I said… there’s more to come :)
Sorry for being so late to catch up on your August posts, but wanted to say that on first impression, I’m quite impressed with the example generation – what a USEFUL idea! I’m especially impressed with the pronunciation in the audio, because what I have heard has been so much better than the generic Google Translate audios. (Of course, I’ve been steeped in Swiss German for the past two weeks, so you may have to take that assessment with a grain of salt.) This new example generation function is wonderfully helpful.
Hey everyone,
I’m Elisabeth and just wanted to say a big thanks to all of you who chipped in a bit extra to support learners like me. It honestly means a lot!
I feel really lucky to be sponsored by such an awesome community. Learning German has always been something I’ve really wanted to do, and your support is making it possible. Thank you!
This all looks good. Is it worth it to have the AI create new content every time? Would there be a way to save the AI content after one person had used it for that entry?
I like the audio. The slowed down audio is a bit odd though. I don’t hear the second “h” in “hochgehen” in normal speech.
This really is a mammoth undertaking, with so much handmade! I should make more use of it.
Oh yeah, that’d be easy to do BUT… then it would look like the examples are “proper” because after all, I am serving them along the rest of the content. And sooner or later, the dictionary will be littered with flawed or wrong examples and I do not want that.
If an example comes from my database, I need to have seen it at least, better yet MADE.
Im totally on board with keeping the content legit on the UI, that makes sense. But like, if people are asking the GPT the same stuff over and over, why not just save those answers in your database and reuse them? Users would still know it’s from GPT, but you wouldn’t be hitting your API every single time. Seems like an easy way to not blow through your budget.
These calls for examples are unbelievably cheap now as the token load is very small. Two years ago, it was a different story and the first couple of months I paid over 100 bucks to OpenAi. But I was very early then and costs have dropped big time. The audio is a different story, I’ll keep an eye on that and see how that’s going . But that too is way cheaper than 11labs which I am currently using for my story generator.
And also, I don’t think many people create examples for the same entry in a short time-frame. Over the course of a year, people created about 20k examples and there was surprisingly little overlap among the words the examples were for. It was a pretty even distribution across the board. Those examples were littered with crap though. Like 20 were severely flawed and another 20% were super bland or odd sounding so yeah… I’ll see how things go now with the new models and the new prompts I am using.
But yeah, overall I actually think having the option to get new set if the first one feels off is more useful for learners overall.
Oh about the audio… it is a bit glitchy. Also a reason why I have decided NOT to save and re-use it. I’d have to review it first. If oyu don’t like a particular pronunciation, refresh the page and you can make a new one.
i feel the illustration is based on something i recognise but i can’t quiet put my finger on it.
Just to echo what someone else said i really appreciate that the dictionaries entries are “handmade” and the effort that goes into them.
This is a Kastanienblatt, good Sir, my favorite tree (next to Platane).
hmm, i meant the whole thing, not just the leaf. But good to know. Looks like it could have been based on some packaging or something.
Hello,
Here are a few typos:
“welcome to first episode” (welcome to the first episode)
“This is at least usable. A” (This is at least usable.)
“examples and can’t add handmade examples” (examples and I can’t add handmade examples)
“it’ll take a few second” (it’ll take a few seconds)
“use this features” (use these features)
Thanks for the new features, AI is indeed getting better and as long as you don’t trust it blindly, I think it can be a great and very helpful tool.
Bis bald!
This is actually borderline magical in my book: the generated audio actually sounds nice and neutral, and is perfectly clear, and while the sentences are nowhere near as fun as those you generate by hand –a fair bit *blander*, let’s say– they seem (so far!) to be totally unobjectionable. It’s pretty frightening just how fast these language generation tools have been improving of late.
For those on Chrome on a desktop/laptop who might be experiencing problems getting results from the “Create Audio (AI)” buttons, it’s worth doing the big flush:
Chrome => Delete browsing data…
Other browsers have similar features hidden away somewhere, but I’ve been a lazy Chrome user for too long now to remember exactly where they’re hidden. It’s a bit of a painus in the anus (because it’s pretty coarse-grained, and you have to re-login to a lot of sites, among other things), but that command flushes enough of what Chrome has cached to ensure that the new functionality is downloaded from the yourdailygerman.com site.
I actually played around and tried to have the AI do something more funny or inspired but the cringe was too strong. Like asking a dog to give his opinion on foreign policy and the dog is really excited and barking and wiggling the tail and seems overall engaged and REALLY willing to do whatever you want but… it can’t.
Thanks for the quick intro to browser cache. I would actually not want anyone to flush the entire cache though just to get this one script. It’ll download eventually (I think I set expiry time is set to a week for .js ). The amount of “stay logged in”s most people have would mean a very rude awakening when flushing the browser.
Lastly, I have to correct your Latin, good Sir. Us Latin-knowers have to hold each other accountable so… it’s Painus ānūs.
Fair enough on the prompt situation: the last thing LLMs seem to be endowed with is *taste*, let alone *wit*, so the more prosaic prompt is probably for the better.
My entire education in latin could be inscribed on the head of a pin stuck in the head of an angel dancing on the head a pin stuck in the head of another angel, usw, usf. She ain’t great. Bordering on the heinous, ehrlich gesagt… “We lives, we learns, but we lives more than we learns”, as Cicero tweeted last week.
I saw that Tweet!! It got a community note that it should be “We live, we learn…” and so he deleted it. Grammar won over truth that time sadly.
Slick. The new feature looks like it will work really well.
first of all, thank you for your manual and craft work on the vocabulary. it is invaluable. In my case (iphone with safari browser) the create examples button works but the create audio button does not. But I can hear the audio of the word. Sounds good.
I don’t understand anything about technology, but I wondered: why is it that when one of us clicks and creates the examples, they don’t remain fixed on the page for the next person? Wouldn’t it be convenient to make chat gtp work only once?
The thing is… the examples are MOSTLY okay, but sometimes, they’re not. While it seems to be better now, it’s still not reliable enough unfortunately.
I don’t want to fill up my database with examples that I have never seen and approved.
When they’re already there it makes the learner feel like they’re “legit” and proper German.
When it’s just a creation of the moment, people are aware that these might not be idiomatic or correct and if they feel weird about them, they can create new ones to double check.
It’s really important for me to have the information that is “hard-coded” is reliable.
Still, having this AI-option is better than having NO examples so that’s the compromise basically.
That makes sense and answers my question!
I had the same question about getting the AI to work only once for each entry!
Very happy for the updates.
Sadly, on both Chrome and Safari, when I click the Create Examples, nichts ist passiert. The cursor switches to a ‘don’t click on me again’ cursor, but nothing happens.
Might be an issue with script caching. If you’re on desktop, could you try ctrl+f5?
Yeah, clearing the cache did it. Sorry. Websies basics 101…
Thanks for all the hard work you’ve put in, Emanuel! The upgraded AI examples are way better (more specific, more context). And your examples and contextualized explanations continue to be awesome- it’s the part of the site I use the most!
Really happy to hear that you’re using the dictionary regularly :).
As to your ” ‘reingiessen” example. although obviously correct, I might have preferred “nachschenken” in the sense of “to refill”.
You were clearly, and I must say effectively, showing the various meanings or uses of “hereingiessen”, and so I’m merely observing rather than faulting you in any way!
Nice job, Manne:-)
Totally agree. It’s not the ideal explanation. This is just copy pasted what my ChatGPT tool gave me.
I don’t expect it’ll EVER be able to grasp things like her-, r-, dr- and hin properly.
Agreed! Thanks and hope I didn’t antagonize/ sich verfeinden you.
Nice work, the audio sounds great.
I’ve just begun trying to understand ChatGPT in some depth. So when I read about v.4.1 or that you feel it’s good enough to enhance the dictionary, I see that as coming from someone who’s systematically looking at AI and evaluating it for a familiar and important purpose. Thanks for the info about AI in YDG context.
I can give a bit more context actually, if you or anyone is interested.
When I first heard about GPT 3 in 2022 before the actual ChatGPT came out I was really excited and impressed and I think I might even have applied for a developers license because I wanted examples for ALL entries in my dictionary and there is no way to realistically do it myself.
But when I played around after the release of ChatGPT, results were so unreliable that I felt like it’s NOT the right call for a language learning site to use this. Things need to be ACCURATE. Not perfect, but 97% would be nice, and the reality was more like 70%.
Later, I found a way to prompt the model so it sticks to the specific meaning I want, but even with 4o-mini this spring, there were sometimes glaring grammar mistakes like using “angangen” instead of “angegangen”. When you “talk” to ChatGPT, you can easily be fooled into thinking you’re talking to “an other”, but when you ask simple grammar questions or examples, you soon enough get examples that leave ZERO doubt that there is in fact “no one at home” in this system.
The new version now seems to have fewer of such “neural network farts” but in my opinion, it’s an inherent feature of the technique and they’ll never be able to get rid of it.
It was a bit shocking to see how many apps came out for learners that are based on ChatGPT or another LLLM and that all have an “explain word” or “explain my mistake” feature without actually really caring about quality.
The fact that every learner can now pretty much voice chat with a willing counterpart and learn that way is great and justifies all these new apps. But I think there should be a bit more care as to what these LLMs are good for and what they aren’t. But that might mean that you have to NOT offer a feature and most apps are not willing to say no.
The issue with German I find, is that words such “angehen”, “anbringen”, “aufkommen” etc .look fairly straightforward to first-time learners, only to discover that one such entry might have up to five or more meanings, many of them idiomatic.
Herein lies a crucial difference between German and English. One German word may have nine or so different, if lexically related, meanings. English often reveals nearly five uses for one German entry!