Neural networks are clean
Goodfire published evidence that transformer models internally represent concepts as clean, interpretable directions in activation space, not distributed noise.
Read the note 1 min read
From the notebook
2 notes on this topic.
Goodfire published evidence that transformer models internally represent concepts as clean, interpretable directions in activation space, not distributed noise.
Steering vectors let you nudge a model's behaviour without retraining. They fell out of favour when newer models stopped responding to them. DeepSeek-V4-Flash brought them back.