ryandvm

It seems that in the last few years the Turing Test has rapidly fallen off as a useful experiment for testing artificial intelligence. With the right guardrails, the latest frontier LLMs are more than capable of simulating perfectly believable human conversation partners.

Instead, perhaps the next best practical test for separating humans from AIs is essentially, "can it make you laugh?"

The idea here is that LLMs are largely trained to be able to produce statistically likely sentences based on a massive training corpus, and while this allows for incredibly impressive query responses, it is absolutely atrocious at statistically unlikely outcomes, that is, humor.

Humor is often defined by being mentally incongruous which would necessarily make it statistically unlikely - the very opposite of what LLMs are good at.

Thoughts?