3 ms·

I genuinely thought creativity was something else until LLMs hit escape velocity and humbled me hard. After that I realized that creativity wasn't some magical

I genuinely thought creativity was something else until LLMs hit escape velocity and humbled me hard. After that I realized that creativity wasn't some magical quality that would be hard to reproduce mechanically. And that also made me a little sad.

shw1n 2y ago
I'm glad you said this -- I felt the same way after making this discovery through my method outlined in this post It similarly took the magic out of creativity and learning a bit, and made it all seem like work The main way I've found around it is the joy in being creative once basic autonomy is achieved in new skills Consciously discovering the heuristic is another fun part
- jimmySixDOF 2y ago
  you guys might be interested in the latest Machine Learning Street Talk podcast [1] which right from the start is all about how LLMs are great for creativity, as in novel combinations of trained data (from memorization) -- but are not capable of the reasoning skill needed to verify if any idea is actually plausible given a set of constraints. [1] Do you think that ChatGPT can reason? https://www.youtube.com/watch?v=y1WnHpedi2A https://www.youtube.com/watch?v=y1WnHpedi2A
sva_ 2y ago
But LLMs to date can't really differentiate well between a creative insightful answer, and a nonsensical one. The selection process is still done by a human.
- qingcharles 2y ago
  This is true. I hadn't thought about that aspect.
xanderlewis 2y ago
I have to say, you were naive to ever think that. Rolling a die can produce creativity — and I’m not being facetious. Creativity, at least in fields traditionally considered to be ‘creative’ like music and visual art, is 90% randomness and 10% retroactively attempting to ascribe meaning to the randomness. And this is coming from someone who hugely values and defines oneself via an ability to be creative. Also, LLMs haven’t hit ‘escape velocity’; that terminology has spread like a meme (because it’s an attractively dramatic yet easy to understand idea) but isn’t backed up by any science. Maybe one day we’ll have the fabled ‘AGI’, but ChatGPT isn’t it.

shw1n 2y ago

I'm glad you said this -- I felt the same way after making this discovery through my method outlined in this post It similarly took the magic out of creativity and learning a bit, and made it all seem like work The main way I've found around it is the joy in being creative once basic autonomy is achieved in new skills Consciously discovering the heuristic is another fun part

jimmySixDOF 2y ago
you guys might be interested in the latest Machine Learning Street Talk podcast [1] which right from the start is all about how LLMs are great for creativity, as in novel combinations of trained data (from memorization) -- but are not capable of the reasoning skill needed to verify if any idea is actually plausible given a set of constraints. [1] Do you think that ChatGPT can reason? https://www.youtube.com/watch?v=y1WnHpedi2A https://www.youtube.com/watch?v=y1WnHpedi2A

sva_ 2y ago

But LLMs to date can't really differentiate well between a creative insightful answer, and a nonsensical one. The selection process is still done by a human.

qingcharles 2y ago
This is true. I hadn't thought about that aspect.

xanderlewis 2y ago

I have to say, you were naive to ever think that. Rolling a die can produce creativity — and I’m not being facetious. Creativity, at least in fields traditionally considered to be ‘creative’ like music and visual art, is 90% randomness and 10% retroactively attempting to ascribe meaning to the randomness. And this is coming from someone who hugely values and defines oneself via an ability to be creative. Also, LLMs haven’t hit ‘escape velocity’; that terminology has spread like a meme (because it’s an attractively dramatic yet easy to understand idea) but isn’t backed up by any science. Maybe one day we’ll have the fabled ‘AGI’, but ChatGPT isn’t it.