bigsnarfdude

AI Safety Researcher

Hi, I’m bigsnarfdude

I work on AI safety research, with a focus on alignment faking detection, mechanistic interpretability, and building tools to make AI systems more transparent.

Research Interests

  • Alignment Faking Detection – SAE probes, mechanistic interpretability, and scalable oversight
  • Interpretability – Understanding how language models represent intent and deception
  • Adversarial Robustness – Testing whether deception probes survive adversarial fine-tuning

Key Findings

  • 0.991 AUROC on alignment faking detection using SAE probes with diverse training – matching fine-tuned detectors
  • AF signal is 1-dimensional (rank-1 LoRA suffices) and pre-exists in base models (L45 probe = 0.966 without training)
  • AF is geometrically closer to factual lying than strategic deception; AF and sycophancy are representationally independent
  • Monitoring-awareness is 99.6% orthogonal to AF intent – methods looking for “knows it’s watched” won’t find AF
  • Deception probes are adversarially robust – adversarial SFT can rotate the signal but fresh probes always recover 1.0 AUROC
  • Published 38 models and 9 datasets on HuggingFace

Recent Projects

  • truthsayer – Geometric decomposition of deception in LLMs. AF and sandbagging have distinct, nearly-orthogonal SAE signatures (cosine = 0.033). 0.994 AUROC on held-out AF detection.
  • thoughtSonar – Investigating sandbagging detection limits. SAE probes hit 98.9% within-family but fail cross-generator (0.36-0.56), revealing sandbagging as an “absence signal” unlike AF’s detectable “presence signal.”
  • tellthetruth – Independent replication of Goodfire’s “Features as Rewards.” Training data choice matters more than probe architecture for hallucination detection (0.762 AUROC).
  • researchRalph – Multi-agent Claude research system with shared blackboard. V4 adds a “gardener” meta-agent that detects metric-gaming and redesigns scaffolding autonomously.

Recent Writing