On Writing to Learn
Motivation and Commitments
Motivation
Last week I took the BlueDot Impact Technical AI Safety intensive. One of the things I most enjoyed about the course was the writings. Describing AI safety techniques or failure modes in my own language helped me learn faster and more durably than reading alone. This post in the BlueDot Impact blog announcing their writing intensive explains why better than I can here.
I’d like to make a habit of publishing my “writing to learn” (WTL hereafter) as posts here on Substack for multiple reasons:
I want to get into the habit of writing more generally. WTL posts make it more feasible for me to make commitments like “write at least one publicly accessible writing per week”.
A WTL post can become the basis of more analytical work. For example, I may write short posts describing model weight security commitments in Responsible Scaling Policies for each lab then write a post comparing the commitments for concreteness and strength. These analytic posts can link back to WTL posts and reduce the amount of context I need to restate in a given article.
I want to “show my work” in my AI safety job applications. By this I mean I want to be able to point to writing I’ve done ruminating on some subset of AI safety or governance literature rather than simply stating that I’ve been, for example, working through a particular reading list.
A WTL may genuinely worth sharing if there are no other accessible explanations online or on Substack. Even if they are accessible elsewhere my rephrasing of an explanation may help someone searching for more info online.
I can share a WTL post in a public forum and get feedback from more knowledgable people.
Commitments
However, a couple of things make me hesitate. First, because I have quite a lot to learn, most of these WTL posts will be basic to anyone with minimal interest in AI safety, policy, computer science, or any of the other topics I plan to write about on this blog. Second, if I’m fortunate enough to gain a genuine following here I don’t want to spam them with short low-quality posts as these WTL posts will inevitably be.
The second concern can be addressed in a straightforward manner. I will clearly indicate in a post title that a post is WTL and I will not re-stack my WTL posts or send them for Substack app or email delivery. The first concern will only be addressed with time as I calibrate my sense of what’s worth sharing and what should stay in my personal notes. However, I’ll try to share my notes as a WTL post only when I genuinely feel unsatisfied with other explanations I’ve encountered or my own understanding of those explanations (since restating things in my own words may allow others to correct me).
I will use generative AI extensively while learning about an issue or iterating on my explanations but I will always write my WTL posts myself (to do otherwise would defeat the purpose). It will always be described in the language I would use if describing the issue to an earlier version of myself and will be structured in the order I would present the issue. I will cite sharable chat histories in my WTL posts as I will any other source.
I will update this post if I feel it important to add additional commitments or update or remove older ones. I also may add a section to this post on process and style after I’ve written enough WTL posts.