Designing for an audience of one
I love this book Odyssey Works: Transformative Experiences for an Audience of One — about making really intricate durational performances for a single person. They're creative, thoughtful, and highly personalized and engaging!
On my mind lately: social media recommendations, and how they can be better — things like recommended profiles, algorithmic feeds, and discovery more generally, which so often feels pretty mediocre.
How can we apply ideas like those in this book on high-touch theatrical experience design to social recs? How might we build an almost absurdly resource intensive and custom-tailored social recommendation experience for an audience of one?
Here are some ideas for what this could look like! I'm thinking about Bluesky in particular — a big network with open data that'd make this kind of thing possible — but interested if it'd be possible for other networks too.
I'd love to hear from atproto devs or anyone who's explored similar: is this realistic? Order of magnitude ballpark, what would it cost?
Recommended Posts
What would it look like if you analyze every post on Bluesky through the lens of a particular user (say: me) — combing through their entire history to extract all the interests you can, and cross-referencing with the entire network of all users and all posts ever?
Yes it'd be computationally intensive and maybe a crazy thing to try to do for everyone right now, but it seems at least in theory tractable to do for an individual.
This would basically be designing my ideal 'discover' feed — but I'd want an option to include all posts, including evergreen ones that may be months or years old but super relevant and I haven't seen yet. Maybe an option for 'best ever' vs. 'best recent'.
Recommended Users
Similar, but slicing things a different way — I want a way to see a big list of users similar to me, maybe with an option to view by interest clusters or filter / order in other ways.
Essentially I think very similar to above: look at all my posts; look at all posts ever, grouped by user; see whose interests seem to align, resonate, share affinities…and help us connect!
Right now Bluesky's recommended follows, like recommended posts, is super coarse and rarely on point. But pretty frequently I come across something shared by a friend of a friend that leads me to a post by someone else that makes me think: why didn't I find you earlier?!
Other interactions TBD
Maybe we can go further and take into account your likes, weight replies in specific ways, base things off your existing network…
Maybe it's possible to not only make a better default 'discover' feed, but make it easy for users to suggest and generate their own feeds, mold and co-evolve them, make them highly personal!
I'm sure the full gamut of the design space here goes way beyond what I've sketched in basic outline above. That all can come later!
Intensively algorithmic vs. tediously human curated
The main way to actually implement this, I've been assuming, is computational…not sure how exactly it'd work, but I assume something with LLMs and temporary appropriation of a beefy machine.
However, it strikes me that you could very well enact a useful version of this entirely by hand! That might look something like: visit the profiles of everyone I follow, go through the people they follow, visit their profiles, and pick those who you think I'd like.
Of course, this would work best if you know me to begin with, and would be made much easier by doing some automated filtering (like starting with people I've interacted with, filtering out their follows who are inactive, otherwise proactively winnowing a bit…)
I'd expect this to take many hours, but it'd also be keeping more in spirit with the Odyssey Works book that originally inspired this, so…interesting thought experiment, at least!
I will pay $100 for this!
If you'd like an actual challenge based on the above:
Give me a collection of great posts and users to follow, on Bluesky, based on my activity so far!
Seriously, if it's realistic to do, I want to see what it looks like. Consider this an open bounty for a prototype that does something novel and (subjectively) good. Whether fully computational, done by hand, or somewhere in between.
My starting point is: what is the ultimate version of this it'd be possible to do right now at non-insane cost / complexity?
Then ideally we can at least have this as a lodestar to aim for in the future for everyone.
And, if doing it the ideal way is untenable right now, are there simpler experiments in this direction that'd still be cool to try? Feel free to propose some kind of creative alternative!