Synthesis history

Three substantive new items extend the thread beyond training interventions into monitoring and interpretability. Engels adds a DiffusionGemma transparency analysis showing monitorability parity with autoregressive Gemm…

Item 31225 adds a fourth intervention type: OpenAI RL training on realistic human scenarios, which produces cross-domain safety transfer (health-only training improved non-health behaviors). This creates a new tension w…

Item 29914 adds a third substantive thread: Google DeepMind's CallumMcDougall reports synthetic document and chat finetuning can produce OOD safety improvements on Gemini 3 Flash, with documented failure modes around su…

Alignment researchers are producing empirical results that challenge two widely-used intervention approaches: sparse autoencoder (SAE) based model steering and supervised fine-tuning (SFT) data filtering for safety. A p…