Hacker News Reader

Comment by robg | Hacker News Reader

Cool in concept, but shows how important comedic timing is. Example where the semantics requires the pauses and drawn out expressions to really make it work. Might be where the jokes could be written by AI but perfected by humans, especially as comedians use audience feedback to hone their material. Even the best don’t know how any joke will land until it’s delivered. How could any AI incorporate that real world feedback loop?

View on HN

I was pretty impressed with it, honestly. But I understand what you're saying - a comedian's delivery is different from normal speech. It may be that GPT-3 just wasn't trained with enough standup routines.

(I dread to think what a Sam Kinison AI would be like)

What’s the training if every comedian has their own style, in words, intonation, and timing? Do we need to build personalized AI comedians? Or write new jokes for specific comedians trained on previous their routines?

To be honest, given the complexity and context connection DALLE2 makes it doesn’t seem all that hard to learn some speech distinctions of a couple of hundred people. It’s really not all that complex.

From listening to too many comedian podcasts, the training is performing the jokes to assistances, and adapt from their reactions.

Again and again, year after year, until you've mastered the interaction.

Simple: AAAAAAA AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA!!!!!!

It also shows the importance of a laugh track. I think a lot of Seinfeld material would sound weird, even performed by Jerry himself, in the absence of audience laughter or other contextual background noise.

I actually think this should be possible to train an AI to do. It could alter timing and monitor for laughs, it will learn when and how to pause, and use intonation to maximize for laughter.

The trouble with jokes is they're only funny to people once. You'd need an adversarial laughing AI to train against. And boy would that thing be loony.

Then why do dads love telling the same bad jokes over and over, and laugh every time?

Because we forget that we said it already

There's lots of people in the world.

I guess if you could show them all the same CAPTCHA with the same 1000 jokes... you could develop a very annoying neural network?

Timing of what audiences and where? If AI heckler as the above, is training to entertain other AIs or humans with diverse tastes?

Make it a GAN with AI hecklers

AI Statler and Waldorf.

It would be such a force for good in the world that Disney should allow it to be used for free like Volvo did with their patent for the three-point seat belt.

I wouldn't be surprised if this routine was guided by the creator. As in, they edited it to make it work and that was probably after discarding a bunch of terrible stuff. Nevertheless it's not unlike how Jerry actually works, he would write a lot, and then edit it down later. He worked hard at it and it sounded almost joyless.

All it needs is a layer of contrastive loss at the top, to tell real Seinfeld and SAInfeld (don't bother checking, it's already registered) apart, and boom! Indistinguishable delivery.

Reply on news.ycombinator.com