To strengthen my mind, among other reasons, I am trying to write more frequently. However, this seemingly precludes using AIs as a significant catalyst for the writing process. After all, the entire point of the exercise is to do my own thinking and articulation. For certain tasks like routine emails it’s probably fine to just enter the important details of what you want to convey and then have the AI polish it up to the appropriate standard. But how should you incorporate this in more substantial or independent work, and where should you draw the line?
So far, I have hardly edited my posts at all1 so I can concentrate on just doing the writing. In the hopes of getting at least some feedback before posting the initial draft of my first post, I ran it by the latest Anthropic model2. Claude made some very good points about the structure and organization which I think I will incorporate in later revisions;3 part of the reason I have “modified” dates is because I plan to periodically update these pages whenever I wish to change things or have new insights. Despite the suggested changes being better in some ways, the words no longer felt like they were my own.
Part of the problem is that I tried asking Claude for a bunch of feedback all at once shortly before I wanted to post, so of course the resulting product would look drastically different. Going forward I think the right path is to iterate with the model several times, focusing on smaller aspects of each piece of writing individually.
I also want to write down as much of my own material as I can before using AI assistance to make further progress. That way, I have more of a hand in shaping the final version. For my first post, I was actually planning on writing about AI as catalyst. However, what started as a caveat to the benefits of lowered activation energy — that things get easier but creativity and novelty in general may be harmed as the same few frontier LLMs get used for most tasks — ended up occupying more and more of the piece until I made that the primary focus and shifted the catalyst discussion to the next post. Such meanders as I put my thoughts down on paper keep my writing grounded in what I am interested in, hopefully making things more unique to me as well as more interesting for any readers.
These solutions aren’t particularly novel. The president of my PhD program recently sent us all this article which covers similar ground. I didn’t read it until I already had the outline of this post, at which point I figured I might as well see what this author’s takeaways were. It turns out, they closely match mine. Upon reflection, she now writes the first drafts of all of her posts with no AI assistance, which is exactly what I am doing here. It’s funny that two people trying to minimize AI’s constraints on creativity still come up with such similar ideas. Hopefully some among you4 find this articulation useful despite its potential obviousness, mostly as a reminder to be careful when using AI to assist with writing.