Observed arrival · 2026-11-17
Humanlike’s home-robot safety test
Humanlike presents a research project testing whether fine-tuning an AI model for a job restores hazardous knowledge filtered from its training data, with home robots as the motivating case.
- For
- People researching AI safety for home robots
- Worth noticing
- The page separates job-test gains from format effects and says no deliberate attack has yet been run.
Field notes
The experiment uses three open Pythia models drawn from EleutherAI's Deep Ignorance research: an unfiltered model, a model with hazardous biology filtered from training, and a filtered model with additional safeguards. Each received the same 3,476 examples for a cosmetics-lab-assistant task. Humanlike reports a 150-question held-out job test and WMDP-Bio scores, while noting format effects, the limited model size, and that deliberate attacks remain future work.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue