Empirical framework for testing whether RL training instills genuine behavioral dispositions or surface compliance in language models, using compute frugality as a controllable proxy value.
Tested whether Anthropic's sleeper agent detection technique generalizes to naturally-arising deception. Ran Llama 70B inference across 90 prompt-conditions on rented A100s, manually scored every output. Honest null result. BlueDot Impact AI Safety Sprint.
Alignment-faking and adversarial-reasoning research artifacts inspired by Redwood Research. Behavioral probes and controlled failure-mode analysis.
Applying conceptual engineering to AI safety by refining definitions of deceptive alignment, alignment faking, and goal-guarding. Develops more precise, behaviourally grounded concepts for empirical evaluation and interpretability research.
15-Gate AI Agent & LLM Cryptography Benchmark, Stochastic Parrot Canary Trap (0x3543 vs 0xDEAD), 1983-1989 Rogue AI Dossier (CRE-7 / Nova Hart) & 300,000-Node 1995 GeoCities Time-Loop
To associate your repository with the alignment-faking topic, visit your repo's landing page and select "manage topics."