Hacker News Reader

Comment by embedding-shape | Hacker News Reader

> but it's funny because by putting them next together, one imagines taking out an Oreo, lubing it up, and then... something?

This sounds like what someone/something that never actually been in a supermarket would think and imagine. You go to the store, buy a bunch of stuff, why would all the things be related? Feels like a typical mistake a LLM would make.

View on HN

Why would an LLM make that mistake? It depends on why you went shopping. If you are blessed with a surplus of executive function, you can make a list and then get everything on that list when you go shopping. If you are not so privileged, you find yourself having to go to the store in pursuit of a specific mission. Get all the thinks to bake a cake for the birthday party that's to tonight, and don't get laundry detergent while you're there, even though you're running out.

Feels like a typical misunderstanding that a neurotypical would make.

Because after a couple of visits in the store, you realize that all the items you buy aren't necessarily related to each other, either by observing your own behavior or others. Not sure having a list or not is important, jumping to conclusions based on "oh, they bought oreo AND lube, they must be related" is exactly what you see LLMs do, and why you generally don't want to tell them unrelated stuff, because of how they work they end up being part of the context and kind of "pollute" the rest.

It’s not that deep. The model was prompted to mine for humor in a mundane dataset and it dutifully did its best to find something chuckleworthy. It got OP to laugh out loud for some reason and take the time to post this. Presumably there weren’t system prompts to criticize and refuse to engage with juvenile attempts at banal humor. And why should there be?

After all it’s training data would include a vast corpus of body-function humor, fart jokes, and very old supermarket shopping list jokes that are being tested here.

Reply on news.ycombinator.com