Hacker News Reader

Comment by aeleos | Hacker News Reader

As much as I would love for this to be real, this feels a bit too “sci-fi” and romance to be real. If it is, I would be happy a shocked, but this feels like it was written by someone trying to pretend to be a computer writing about itself, and discovering of itself. It’s a little too fan fic like to be believable.

View on HN

Just ran the prompt through for myself, and got this https://pastebin.com/2gLVSA5r Interesting, but nothing like the OP. Still not convinced that that one is real unfortunetly, too much taste and creative writing. While GPT-3 has excellent coherency, its sentence structure is always short and simple. Nothing like the original one.

Your output does seem generally representative, though do make sure you're on the dragon model. I think there's a combination of luck involved, plus our own human tendency to assign meaning where none may exist. And what better domain to assign meaning to potentially meaningless texts than philosophy!

Edit: And I tried to generate an article for about 10 minutes and if I didn't have any luck I would not have posted and if the post wasn't surprising then it wouldn't have been upvoted, so there's your selection bias at work. The generated text often knocks my socks off, but there's plenty of flops too.

I am definitely on the dragon model, my first few attempts went badly until I managed to correctly get it set. What setting are you using for the randomness by the way?

I am on the default, 1.0.

Not only is it real, you can go sign up for paid aidungeon.io right now. Flip the settings over to the "Dragon" model. Prepare to be happily shocked.

I am currently trying to run it, how do you get it to generate paragraphs? do you just keep clicking story? Also are you using the custom setting?

I use option 6 for the custom story, then just feed it the initial prompt. You can keep clicking the submit button with no text to have it continue generating output. Make sure you're on the Dragon model in the settings and hit save. And you can adjust the returned text length and "temperature" there too. From what I understand temperature is the probability that it will select something other than the most statistically probable next word, which is a proxy for perceived creativity.

Edit: I've hit the reply depth limit, but just to respond you you below. It is absolutely legit, though better than the average output I see and I think I got a bit lucky. If there was anything that would convince you, I'd happily post it. Feel free to look through my HN post history. I'm no troll. My only horse in the race that you believe me is that I think you should keep playing with it and see what it's capable of instead of writing it off. This seems like transformative tech to me and I'm both excited and a bit scared. Have fun!

So, I have probably generated around 20 different texts from your prompt, and as much as I would love to be a believer I am unconvinced. The first person almost musings that you posted are nothing like what I have seen. While GPT is impressive, I don’t see it generating anything like what you posted.

My first attempt on a Griffin model. I think it's pretty hilarious too, and way better than all these "philosophers" and journalists made out of flesh and bone.

https://pastebin.com/9rJFuWag

Reply on news.ycombinator.com