Or, if you don't want to have to explain the Monty Hall problem, there's the Dumb Riddle of the Surgeon's Child: https://chat.openai.com/share/1ee44fc6-a85f-42e9-aebb-6a165f...
ETA: the Dumb Two Guardians, Two Gates riddle: https://chat.openai.com/share/48c3d32c-5efd-4a96-89d8-66d1d4...
(Gpt4 is better at these, I'm told)
I tried this one myself, and can confirm that it did work a few months ago.
> You initially choose door number one, because you want the car and can see it behind that door.
> Monty Hall opens door number two (which has a goat) and shows that there are no more goats left. Since you already saw what was behind door number one, this doesn't change anything about your initial choice of doors.
> The fact that door number two is opened does not affect the probability of the car being behind either door number one or three (since both have been hidden from Monty Hall). However, it does give you new information: since there are no more goats left to be found and only one door can hide a car, the other remaining door must contain the car.
> Therefore, with this new information, your best option is to switch doors and choose door number three, as that's where the car must now be located given what you know about Monty Hall's choices.
> It is important to remember that the car cannot magically move between doors after being seen or hidden; it always remains in its original position behind one of the doors. The information provided by Monty Hall simply helps you make a more informed decision based on your initial choice.
I love these open source models, they're an endless source of hilarity. Not much use for anything else though :P
It’s an iterative, semi-deterministic process. It’s simply a place where context and generation lead to weird outputs.
You can get similar outputs by asking OpenAi to repeat a number 100 times. It will eventually get into some weird, low probability paths and generate non-sense output.
This type of complete garbage is not uncommon in AI. It's simply the nature of asking a non-intelligent system to generate human readable content.
Maybe this is a different way to think about it. In most of the country, your cellphone has _amazing_ coverage. It can talk clearly with a cell tower. Your data and calls work perfectly.
In some parts of the country, you're going to have no service. Your cell phone won't work. It doesn't have cell towers to talk to.
At the intersection of service and no-service, you'll find an area where your cell phone works sporadically. It might barely have 1 bar of service. You might need to hold your phone a certain way. It will work seemingly randomly. Calls might have a few words go through.
That edge of service is essentially where the LLM is at. Its in an internal state where it has enough signal to attempt to generate a response, but not a strong enough signal to generate a meaningful response. It ends up falling back to something it's "memorized".
"You might be able to get cell service by holding your phone differently. Try waving it randomly around the room, one corner might work better than others."
"The USB stick enters on the third try."
"An iterative semi-deterministic bag of matrix multiplications can convincingly communicate. Undefined behavior appears schizophrenic."
On an intellectual level, I get it, but it's still fuckin' weird.
Don’t present this as some kind of anomaly unique to AI, the concept of “garbage in garbage out” is all that applies here.