x

LESSWRONG

LW

GabrielKS — LessWrong

GabrielKS

GabrielKS

Message

Hello, world! I’m Gabriel Konar-Steenberg, an AI safety researcher from Minneapolis, Minnesota. I’m currently conducting a LASR Labs fellowship extension funded by a grant from Coefficient Giving. My London-based team, mentored by Stefan Heimersheim, is testing the robustness of LLM interpretability techniques through more realistically trained model...

27

1

14d

GabrielKS

Hello, world! I’m Gabriel Konar-Steenberg, an AI safety researcher from Minneapolis, Minnesota. I’m currently conducting a LASR Labs fellowship extension funded by a grant from Coefficient Giving. My London-based team, mentored by Stefan Heimersheim, is testing the robustness of LLM interpretability techniques through more realistically trained model...

The Model Organism Lottery: Model Organism Interpretability Strongly Depends on Training Methodology

TL;DR Current model organisms (MOs) for interpretability benchmarking are typically constructed via a dedicated, “post-hoc” SFT step. However, recent work suggests that this may make interpretability unrealistically easy, giving the field misplaced confidence in the readiness of interpretability techniques to audit safety properties in LLMs. We show that across activation...