Joking.
On why ChatGPT, Gemini and Claude give you the same joke (and what to do instead)
Ask AI “Tell me a joke about [whatever, really]”.
And ChatGPT, Gemini & Claude will all tell the same joke.
But why? And how can you get much more creative outputs?
Table of contents
✦ You don’t understand LLMs.
✦ How to get a different joke (or anything creative).
✦ How to push ChatGPT to be edgy (double prompt strategy).
✦ Best of prompts of ChatGPT-5 (to copy & paste by case studies).
PS: to access my archive, go to https://docs.google.com/document/d/1pWuMCBVQo1zKcgKltX_BZxAr31KgxmOlp3Vzvmc5Hxc/edit?tab=t.0
1. You don’t understand LLMs.
Ask ChatGPT to pick a number between 1 and 100.
It will most likely pick 27, 42, or 47.
But why is it so predictable?
Because LLMs like ChatGPT are just an auto-corrector (on steroids).
They simply pick the highest probability for anything.
If I say “My dog loves going to the […]” there is a high chance it says “park”.
It’s pure mathematics.
You can prompt it 50 times, and it will always say “park”. Unless there is another way to prompt it? I’ll go through it in Part 2 of this newsletter.
But first, if you want to understand LLMs fully, I found this 7-minute video that explains it even better:
You now know the problems behind LLMs.
But this “overfitting” behavior can be fixed. Here’s how:
2. How to get anything creative.
Let’s come back to the joke example.
This was our original prompt:
tell me a jokes about the sunBy simply adding this, we now have drastically different jokes:
you must share 5 ones, from the most known to the 5th one I for sure never heard before
1 prompt, but this time most of the jokes are different.
How can you replicate this for anything (beyond a joke)?
Ask for multiple options, always.
Here’s an example:
Classic mistake.
Prompt:
“How should I do […]?”It’s a good answer.
But this is like the overused sun joke. That’s basic.
Basic prompt = Basic answer = From the highest probability.
Here’s a better prompt:
Prompt:
“How should I spend 100$ on ads to grow my newsletter?
Give me 3 strategies, from least to most aggressive.
Use the latest data on the matter.”Just click on the answer to see my ChatGPT (it’s too long to show a screenshot).
It really speaks for itself:
Much, much longer answer.
Much more in-depth for you to decide.
It imagines plausible scenarios like a human-being.
3. How to push ChatGPT to be edgy.
Asking for multiple options is one strategy. It works.
But there is another one I just love using: the double prompt.
Here’s an example.
Step 1. Ask normally what you want.
I am learning Hebrew right now (I am a Parisian living in Tel Aviv).
And it’s so… hard…
Hebrew is read from right to left (not left to right). Yes, books are the other way around. Also phone apps, menus, anything really.
Modern written Hebrew do not add voyels… People read only with the consonant. And if it has the same letter, you need context. “ספר” can mean book (sefer), count (safar), or story (sipur) depending on context.
Roots drive meaning. Most words come from 3-letter roots with many derived forms, so you must recognize patterns, not memorize words.
ChatGPT’s answer was pretty good. But what if I want deeper insights in learning Hebrew faster? Let’s go deeper with this simple follow-up prompt:
I knew all of this. I want you to go much deeper. Focus on the 20% that you haven’t shared yet to reach 80% of my target (being fluent).
Check the answer: https://chatgpt.com/share/68f7b2f0-8d10-800f-83f7-7d1b934b7b1f.
This has nothing to do with the first layer of answers.
Interestingly enough, the 10th advice is the best in my opinion:
What if we could do deeper?
Hint: we can.
Go a step beyond now.
I am French. I speak English fluently. And I am learning Hebrew while living in Israel. Search the web for this specific case study.
I will ask again, now that you know my full situation, focus on the 20% that you haven’t shared yet to reach 80% of my target (being fluent).Worth noticing: I now use “Thinking” + “/search” for stronger outputs.
You start understanding the power of the right prompts.
A bad one would get you the “sun joke”, over and over again.
But the right one taps into (1) the data you needed to make a better decision (2) the right text with your tone of voice or (3) the proof-reading you needed to validate a contract.
That’s all because of the right prompts.
Let’s dive into my favorite ones on ChatGPT: