This is a short blog post to share a fantastic use case I’ve discovered for conversational AI based on large language models (LLMs): making intelligent guesses of what someone said in a spoken sentence.

Recently I was in a cafe in Düsseldorf, Germany. I took one of those corner seats where I can be alone to enjoy a quiet morning when the waiter came up to me with my order. Black coffee and a croissant. I thanked her and stood up, saying “Bezahlen?”, which I meant as shorthand for “I’ll come with you to the counter to pay now.”

As is often the case, because I’m still learning German, I could infer what she was saying, but I wasn’t sure exactly.

My inference is that she said, “No problem, you can pay later. When there’s a lot going on, then we ask that customers pay first. But when it’s not so busy, it’s fine to pay later!”

I smiled and said, “Alles klar, vielen Dank!” (Which means I understood, thank you.)

As she walked back to behind the counter and I took my seat, I started to wonder about the sentence she spoke. I wanted to know how to say what she said. That’s how one learns a language, right? By uttering one mimicked, scenario-appropriate sentence at a time?

All I heard was:

“wenn … los ist, dann …”

Those ellipses stood the words I didn’t quite understand. Those were my gaps.

The bolded word los is the word I was curious about. How can that word mean “busy”? I knew what it meant, and busy was not one of the meanings. Hmm…

But from the structure provided by the words that I did know — the very common wenn … ist, dann …, which means “when …, then …” — I could infer that she was trying to tell me that when it’s busy, then we ask customers to pay first. Otherwise, it’s fine to pay later.

Nice to have a brain that can infer! But this unfortunately does me no good in learning how to say the sentence myself.

What can I do?

I first tried to translate what I thought from English to German, hoping to find the word “los” being used in the translated output. Doing that on DeepL didn’t yield a desirable result:

Conventional translation: Fail

DeepL is an amazing tool; it’s just not suited for this kind of reverse engineering of a spoken sentence.

So I turned to ChatGPT because I have a paid subscription. I’m no prompt engineering guru, but I had an instinct that as a LLM trained on a gynormous corpus of words in various langauges, it would be able to make intelligent guesses and fill in the blanks for me.

Here is the first attempt, which was already a successful attempt:

Contextual multi-language semantic search: Success

In case that screenshot is not loading for you, here’s my prompt in text:

A waiter in a german cafe said something about “wenn … los ist” and I’m taking it to mean “when it is busy, then we’ll ask you to pay upfront, rather than later”.

I need help understanding what a proper german sentence would look like with the word “los” to mean “busy”

And here is ChatGPT 4’s response (emphasis mine):

In German, the phrase “wenn … los ist” can be a bit tricky to interpret directly without the full context, since “los” has several meanings. However, based on your interpretation of the situation - that you’ll be asked to pay upfront when it’s busy - it seems like the waiter might be using “los” in a sense similar to “busy” or “a lot going on.”

A proper sentence in German to convey this meaning could be: ”Wenn viel los ist, bitten wir Sie, im Voraus zu bezahlen.” This translates to “When it’s busy, we ask you to pay in advance.”

The phrase “viel los ist” literally translates to “a lot is going on,” which is a colloquial way to refer to a place being busy or crowded. In German, “los” can be a versatile word, often used in informal contexts to describe a situation with a lot of activities or people.