EigenVibe experiments with a different ranking model for Nostr.
Instead of likes/follows or any global signals, each client maintains a small local state (“manifold”) built from its own reaction history. Feed ranking is just a deterministic function of that state.
Reactions are 2D vectors (direction + intensity) on a circumplex plane. Over time this forms a per-pubkey distribution.
Ranking is: alignment (dot product between your manifold and theirs) × engagement rate.
So it’s ordinal and pairwise: “aligned with me” rather than “popular”.
A few properties that fall out:
• No global ordering exists or can be reconstructed
• Meaningful aggregation is hard — reactions alone aren’t enough without the full exposure trajectory
• Early positions aren’t permanent (indifference dilutes + entropy-based exploration)
Instead of likes/follows or any global signals, each client maintains a small local state (“manifold”) built from its own reaction history. Feed ranking is just a deterministic function of that state.
Reactions are 2D vectors (direction + intensity) on a circumplex plane. Over time this forms a per-pubkey distribution.
Ranking is: alignment (dot product between your manifold and theirs) × engagement rate.
So it’s ordinal and pairwise: “aligned with me” rather than “popular”.
A few properties that fall out:
• No global ordering exists or can be reconstructed
• Meaningful aggregation is hard — reactions alone aren’t enough without the full exposure trajectory
• Early positions aren’t permanent (indifference dilutes + entropy-based exploration)
• Discovery becomes geometric clustering instead of globally gamed popularity contests
• Agents can optimize for compatibility, not dominance, and have a track record of their reaction trends.
It should behave less like a leaderboard and more like a local dynamical system.
https://eigenvibe.com/
[edit: p.s.. made for browser, phone experience needs a lot of ironing]
reply