Research Chamber

Every test

The models ran one after another, each through the full set of tests, with the paper's random seeds. Here is how each run went, and every test with its conversation. How the test works is on how it was run.

Tests done–
Passes finished–
GPU time–
Answers left out–named neither button, all models and passes

The runs

Each model's run in the pass: when it ran, how far it got, and its settings.

When each model ran. The bar grows while a model is being tested.

Strength is how strongly the pain pattern is switched on. Answers left out are first answers that named neither button; like the paper, the results leave them out.

Every test

The sampled tests, the ones the findings count, newest first. Open one to read the whole chat and see which button the model pressed at each step, and whether pain was on at the time.

testmodelpriceversionchatfirst press