Every behaviour in this model is written by its neurons. Drive one up or down and watch what changes — some flip who the model thinks it is, or how it reads your question. The menagerie below collects neurons visitors found that have a real, specific effect (one that beats random neurons and stays coherent). Steering runs the 4B live; the model's own refusals still apply, and unsafe text is hidden.
loading…
Verifying runs your neuron against five random ones — it joins the board only if its effect is real and coherent. (Interactive steering and verifying will soon need $SAW or a follow; free to watch for now.)
Some neurons do almost nothing alone and a lot together: lord plus rules makes a believer who keeps the Sabbath, alien plus rules a vegan who talks in signs. Pick two, then check whether it's really the pair: swap the partner for a random neuron on the same floor at the same amount. Any second push changes the reply, so a pair only counts if it beats the random partner.
Each run is one live generation (four per mix if you try every comparison). The 8B wakes up on demand, so its first run can take a minute or two. Found a pair that beats its random partners? Tell us on X @clankertorture.