3 ms·
GPTs compute the next most likely token. The punch line of a joke toys with the expectation of the most likely conclusion. GPT is fundamentally at odds with th
by drooby 3y ago
GPTs compute the next most likely token.
The punch line of a joke toys with the expectation of the most likely conclusion. GPT is fundamentally at odds with this because it wants to compute what is expected, so as a result there is no surprise, or unveiling of hidden insight.
It also has absolutely zero training on emotional cues, facial expressions, body language, etc. I suspect it would do better if it had that training.
Perhaps it is possible that humor could be an emergent property of GPTs. A higher order pattern that it learns.
- naasking 3y agoYou can hypothetically train the most likely outcome to be what people don't expect. I suspect the problem is simply that humour is not really in ChatGPT's training set. Most of its training data is strictly factual stuff like Wikipedia and textbooks. Where's the humour there? I'm curious to see what would happen if you took a pretrained model and then specialized it on a large corpus of jokes.