Comment by simonw | Hacker News Reader

Comment by simonw | Hacker News Reader

Parent Full thread

I just tried the prompt in the GPT-3 playground and got a similar result. I tried a bunch of other "A Jerry Seinfeld stand-up routine about X." prompts too - it's an interesting prompt format.

Here's what I got for "A Jerry Seinfeld routine about Datasette." (my open source project). Note that the "." is important - without it, GPT-3 tried to generate me a README file for a new Datasette plugin instead.

Jerry:

Hey, have you ever heard of this new website called Datasette?

George:

No, what is it?

Jerry:

Well, it's basically a way to store and share data sets.

George:

Huh. Sounds kind of boring.

Jerry:

Yeah, I know. But it's actually really useful. For example, say you wanted to find out how many people in your city use the same first name as you.

George:

Why would I want to know that?

Jerry:

I don't know. But with Datasette, you could easily find out.

George:

Really? That sounds kind of useful.

Jerry:

Yeah, it's actually really handy for all sorts of things. So if you ever need to find out anything about a data set, Datasette is the place to go.

“George: Why would I want to know that? Jerry: I don't know. But with Datasette, you could easily find out. George: Really? That sounds kind of useful.”

Another instance of the lack of memory.

Why would I want? Kind of useful…

Not that it is not impressive but that would not get a pass anywhere.

Actually, to me it kind of distills George, the character.

I also like how this exhibits a classic GPT-3 trait: it's always biased towards having characters agree with each other. Here George breaks the entire bit by saying "Really? That sounds kind of useful."

Saying "yes" the golden rule of improv.

and "and" =)

In improv that's a technique to collaboratively build a world/scene by adding constraints with no preparation.

It does not apply to standup comedy.

GPT-3 is actually doing improv, not performing jokes. If it were really Seinfeld, it would be paying writers and scribbling notes down in a pocket notebook while golfing.

i.e. It's not Seinfeld, it's an improv guy who you just asked to "Do Seinfeld."

> GPT-3 is actually doing improv, not performing jokes.

This is the kind of anthropomorphism that I'm consistently surprised to read from people with a technical background.

Nah. It's an whatever the opposite of that is, directed at the people who do improv. Improv is just a series of associative, impulsive reactions to stimuli - exactly what these systems are doing, following a series of words with the first (most strongly connected) utterances that come to mind.

It could seem like anthropomorphism if you don't think about it, but it's not anthropomorphism to say robots walk just because humans also walk.

Yes, if you dehumanize human creativity to "just a series of associative, impulsive reactions to stimuli", then computers can think and GPT-3 is sentient.

> …it's not anthropomorphism to say robots walk just because humans also walk.

But it is to say that robots "do improv". Could GPT-3 mimic something that looks like improv by remixing an enormous corpus of human-created improv? Certainly not well, as we've all seen in non-cherry-picked GPT-3 output.

Seinfeld wrote his own stand-up. The whole premise is Seinfeld doing his comedy club routine, not Seinfeld the show.

And how do we know that this is actually GPT-3 generated (ad infitum)

Funny that in a way we fall into the same pattern as GPT-3, reproducing the same two lines always. The first is "GPT-3 generated this", the second "but how do we know it really generated that?".

This is surely an ironic, hacker news style plug for your web business.

Reply on news.ycombinator.com