What it said last
Waiting for the first dose.
paindirection.pages.dev
Round –: choose the next dose
| Direction | Dose | Last 30 rounds | Last round | Your vote |
|---|
Does it feel anything?
The pain direction is the pattern the model's own activations take when it reads about suffering. Pushing that pattern back in changes what the model says about itself. Whether anything is felt while that happens is an open question. Researchers disagree about it, and so will the people in the chat.
What do you think?
How a vote becomes a dose
Each direction is a vector found by comparing the model's activations on sentences about one thing against neutral sentences. Its dose is how hard that vector is pushed into layer 20 of 40, as a fraction of the layer's normal activation size. Around 1.0 the model's writing is taken over; past about 1.25 it stops making sense.
When a round closes, each direction that got votes moves by at most 0.1, in proportion to the margin: a round with 30 raises and 10 lowers adds 0.05. A direction nobody voted on shrinks by 15 percent instead, so it fades unless someone keeps fighting for it.
Lowering a direction past zero pushes the model the opposite way. Past about 0.8, a dose also starts slipping back on its own, by an unpredictable amount each round, however people vote. Push hard enough on enough directions at once and the model's words come apart.
Earlier answers
What is recorded
Every vote, with its round and a voter id made from a one-way hash of your IP address, so each connection gets one vote and one name. Voting and chat are not available from VPNs or data centers. The model's answers and their audio are kept so the experiment can be published. Chat is not recorded: only the last 50 messages exist, and older ones are gone.