[jade-cat]

"but does it only make these because there already exist hundreds of examples and tutorials? fear not, there's an entire library of other projects it made all of which also have hundreds of examples and tutorials"

don't get me wrong, it's interesting, but the response here didn't answer the question

[lionkor]

the response is yes, its been done before, thats why the llm can regurgitate it.

If you can make an LLM generate something novel and can prove it wasnt in the training data, call me

[ianbicking]

GPT doesn't create novel things unless you give it a reason to. It doesn't "want" to be novel, due to it's lack of wanting. But if you poke at it you can easily get it to create novelty.

For instance I asked GPT to make some game concepts for a more narrative game with an emphasis on locked areas you'd have to get through, and here's a few of its ideas:

1. Masquerade Ball of Eternity: Set in a never-ending palace ball where attendees wear masks that indicate their access levels. To gain entrance to different rooms, players must persuade, charm, or deceive other guests to swap masks or uncover clues about hidden access points.

1. Fantasy Festival: A week-long festival in a magical kingdom where different events, shows, and parties require special passes. To get these passes, players must navigate the gossip mill, undertake tasks for performers, or sneakily forge invitations.

3. Library of Forbidden Knowledge: Every floor of this immense library has knowledge more restricted and coveted than the last. To move up, players must win debates, discover hidden lore, or befriend ancient librarian spirits.

4. Zoo of Mythical Beasts: To access each enclosure, players must understand and empathize with the creatures, learning their stories, likes, and dislikes to gain passage without arousing suspicion.

(Now I can hear you saying: but aha! Every one of those ideas is made up of letters and words found in the training set!)

[Philpax]

I can pretty trivially generate something using a LLM that I doubt was in the training set or has ever been written down by humans. Do you have reason to believe that's not the case?

[lionkor]

Yeah, the fact you cant prove it

[rowanG077]

This is literally doable on gpt 3.5 within one minute of using it.

[lionkor]

but you cant prove it, can you? you just assume the output id novel because you cant find it easily.

[rowanG077]

You can literally make up a language yourself and get ChatGPT to talk in it. Sure I can't prove that I didn't just make up a language that someone in the trainingset also made up exactly the same way. But that just seems incredibly unlikely.