Do I Dictate Good Posts?
Since I was a child, I have longed to write as fast as I thought. Failing that, I’d like it to be as fast as I spoke. For one of the first birthdays I can remember, I asked for a copy of Dragon NaturallySpeaking. This was not very satisfying, and I ended up dictating several chapters of a fantasy novel to my mother, who served as amanuensis.
The tech has since risen to the challenge. Wispr Flow in particular let me dictate several blog posts during Inkhaven. But, as Beŋ says, “Speech to text has a different ‘texture’ than writing; when I write, I’m forcing structure onto my thoughts as I go. It’s a clarifying act. When I talk, it’s discursive, less sharp, more branching and sprawling.”
Does the sprawl slap or not? When I dictate a post, is the result good or bad? NB: it might be that I tend to dictate posts that would be good or bad, even if the dictation doesn’t have an effect. For example, sometimes I dictate when I’m in a rush. As a guy who read the Sequences but not Toolbox-thinking and Law-thinking, I’ll solve this by asking an LLM to fit a Bayesian model to my posts, and regress their likes and views on whether or not I dictated them.
Of course, whether I dictated something is not completely binary. Yes, for some I just basically sat down, spoke into the microphone, and fixed up transcription errors. Others were written in silence. And some were in-between. Paragraphs written by hand, others spoken. So I’ll give them a score of 0-2 on how dictated they are.
I fit the model … and, the diagnostics suggested some posts were having an outsized influence on the fit. In particular, The Spectre, a dictated post that got linked in a few places.
So I throw in some more variables that might affect the fit. Did I cross post it to twitter? Was it highlighted in the Inkhaven Spotlight? Did taylor.town drop it in his links newsletter? Was it fiction?
And also, this AI didn’t make the best modelling choices IMO. I guess there’ll always be a role for human ingenuity and understanding after all!
So I tweaked those choices, and made promotion effects additive rather than multiplicative. I tried lots of things until the guardrails stopped complaining.
And after all that, we’ve got some numbers! So, do I tend to dictate better or worse posts?
About 10% worse! Though that credible interval is pretty wide. The problem, you see, is that I’ve only written about 30 posts. And the model-fitting machinery wouldn’t mind a few more likes and views, to be honest. So, I’m going to do my part by writing a few more posts for this substack.
I invite you, too, to do your part: read & like!