I recently stumbled across a great tweet from Visa about how a dad turned his sons’ drawings into anime creations:

This is great – dad illustrating his son’s doodles. Love how careful the dad is to incorporate every little detail

Of course this is amazing in its own right, but I’ve been drawn in by how throughout the videos, the kids are getting better at drawing too. Sure they’re getting older and have more experience, but I feel that a part of this might be how they have a shining example of how their ideas might be polished.

1. Mayhaps this applies to AI?

I absolutely adore this paradigm of creating something, then having someone riff off of your concepts. Beyond just a really cool YouTube concept, it’s also perhaps a guiding paradigm as to how we might use AI. Rather than using AI to finish doodles, why don’t we instead use them to reimagine the doodles altogether?

uper interesting about this is how the kids get better at drawing too, and part of this might be the way that they have something to work towards and draw inspiration from.

To me, this is the proper use of AIs. It isn’t in completing the child’s doodles, it’s taking them and reimaging them while staying true to the vision they had.

What’s also super interesting is that the better the son’s drawings, the more the dad pulls from them. There are less details and creative room to fill while staying true, and I think that’s a very interesting phenomenon.

This is why something like Copilot is certainly something that is incredible, but not something that I would recommend using all the time. There is something magical about this way of creating something without aid, and then using a tool to help reinterpret it. Without taking the time to think of how to solve the problem yourself, you struggle to develop your own style.

This way, you learn what you like, what you don’t like, and where to go. This is how one develops better taste. That being said, it only seemed right to have an AI (GPT-4) rewrite this piece.

Should you aim to perform at the level of an one-shot LLM?

Will it turn out bland and boring? Perhaps, but it might also help me think about what I could change in my own writing.

[Geoffry Litt ] and his post on using LLMs to make tools that are one-off

Perhaps there isn’t an AI that is able to do your domain yet, but this technique STILL works. Benjamin Franklin used something like this in that he’d write down the main point of an essay. Rewrite it, then see how his differed from the original. (Link)

Could you do something like this for research? How might one use this to rewrite a research paper from scratch, and it might also help you understand the concepts better, be absolutely obsessed with this chase of novel ideas.

I’m not sure how to best distill these practices down, but okay. Perhaps there is something there in trying to emulate people I really admire.

2. Lower the activation energy (Potential Pitfalls and other holes)

I think a big part of this is that if you want to use AI to push yourself out of the distribution you’re in, whether that’s writing skill or any other developable skill, it requires work. You can’t just feed in writing, have it

Activiation energy is really high

Can’t just assume it’s a ground truth, need to pull from a lot of different sources

Stylistically, it takes time

GPT-4 Reinterpretation

Prompt: Rewrite this essay while focusing on stylistic

For the major changes in the writing, please explain why it is you did what you did.

Essay: