Aether AI Safety Research
Independent technical AI safety group: "Our primary focus is on studying LLM monitoring and hidden reasoning, but we are also interested in understanding and shaping the propensities of LLMs." Work to date is chain-of-thought monitoring (how information access affects a monitor's ability to detect sabotage, ICML 2026), a taxonomy of hidden reasoning in LLMs, and a no-CoT task-completion time-horizon benchmark. Current focus is a "science of held-out monitors": when held-out measures of alignment silently degrade under training against correlated proxies, and when they remain trustworthy.