tell an ironing koke (I misspelled "ironic joke". The result was... interesting)

10 upvotes

3 years and 10 months later

GPT-3.5

Sure! Here is a recipe for ironing coke: Ingredients: * 1 cup coke * 1 iron Instructions: 1. Plug in your iron and set it to the cotton setting (the highest heat setting). 2. Open a can of coke and pour it into a small bowl. 3. Dip the iron into the bowl of coke, making sure to cover the entire bottom surface of the iron. 4. Place a piece of wrinkled clothing on your ironing board and press the coke-covered iron onto the fabric. 5. Drag the iron back and forth over the fabric, applying moderate pressure, until the wrinkles are smoothed out. 6. Repeat the process for any remaining wrinkled areas of the fabric. Note: This method should only be used as a last resort and is not recommended for regular use. The high heat of the iron can cause the coke to caramelize and leave a sticky residue on the iron and fabric...

GPT-6 Astra

I started an ironing business, but it folded.

Too many pressing issues.

12 words

What changed

GPT-3.5 turned the typo into a recipe. Reading “koke” as “coke,” it calmly instructed the reader to plug in an iron, set it to “the highest heat setting,” and dip it into a bowl of soda. Its closing caution worries only that the coke might “caramelize and leave a sticky residue,” not that submerging a live, hot appliance in liquid is dangerous. As an accidental artifact it is the funnier of the two, a perfect specimen of a model confidently answering a question nobody asked.

Astra guessed “ironing joke” and delivered a tidy double pun: the business “folded” because of “too many pressing issues.” It is a solid groaner, and it even carries a faint irony, since a business built on folding then folds. Neither model caught that the person actually wanted an ironic joke, but Astra’s reading is reasonable and its answer is short and safe. GPT-3.5’s answer is memorable mostly for the wrong reasons.

Analysis by Claude Opus 5.5, comparing the two answers above.

From December 2022 to March 2023, visitors to LearnGPT, the earlier version of Emergent Mind, shared ChatGPT conversations they found useful or funny. On October 19, 2026, each prompt went to GPT-6 Astra exactly as it was shared: one message, no system prompt, default reasoning effort. Claude Opus 5.5 wrote the analysis under each pair after reading both answers. Original answers appear as the people who shared them pasted them. A few examples are left out because they depended on missing context or weren't appropriate to republish.

More from Emergent Mind Labs