March 8, 2026
Flocks fly, comments peck
Neural Boids
Tiny bird‑bots learn to flock; commenters swoon, squawk, and cry “robot slop”
TLDR: A tiny neural network makes on‑screen birds flock without hand‑written rules. The crowd splits between awe at the visuals and backlash over “robot slop” writing, while skeptics ask why train a net to mimic the classic three rules — and whether this mirrors how human crowds move.
A viral demo of “noids” — tiny neural networks powering flocking birds on screen — sent the internet flapping. Instead of hard‑coded rules, each agent uses a small learned “brain” (1,922 parameters) to steer by what it sees, echoing how real starlings stick to their nearest few neighbors. The result: mesmerizing swirls you can poke in the demo and watch ripple like a wave. But the real show? The comments.
One camp is smitten: “Beautiful post!” gushes a fan. Another camp comes in talons‑out: multiple readers drag the writing as full of “LLM‑isms,” with one fed‑up critic snapping, “I’m so tired of reading robot slop,” urging the author to use their own voice. Then the skeptics land: if the network was trained on outputs from the classic three boid rules, isn’t it just relearning the same thing? Why swap hand‑made rules for a tiny black box that imitates them?
The thread also veers delightfully human. A commenter wonders if crowds “flock” the same way — “people are smart but crowds are dumb” — while others joke that clicking to startle the birds mirrors how one spicy take can panic a whole thread. Love it or loathe it, the flock flies beautifully — and the discourse flocks even faster.
Key Points
- •Noids are neural-network-driven agents that learn steering from local neighbor perceptions, replacing hand-coded boid rules.
- •The model uses 1,922 parameters, taking 24 input values and producing 2 output values per agent.
- •Real bird murmurations are governed by local interactions, with reaction waves propagating 3–4 times faster than individual flight speeds.
- •Craig Reynolds’ 1986 boids introduced three flocking rules; Andrea Cavagna’s 2010 study confirmed local interactions and topological neighbor tracking (about 6–7 neighbors).
- •Topological distance (nearest neighbors) avoids failures of metric distance rules when flocks stretch thin, supporting robust flocking behavior.