AI safety
Notes on alignment.
- Introducing Atlas ResearchAn ASI alignment research organization working on one question. When a model keeps training, what keeps its alignment intact?
- Robust SuperalignmentHow long a model's alignment lasts may be a property in its own right, separate from how aligned it looks today. How I'd measure it, with correctability scored alongside.
- Latent Invariant AlignmentCan the internal structure that carries a model's alignment be shaped early in training so it stays in place, and keeps doing its job, as the model keeps training?
- Yoga for AIAlignment may last longer when a model's sense of who it is stays steady under pressure, and that steadiness may be trainable.
- Extinction is not erasureToday's fixes for misalignment may suppress a disposition without removing it. Learning science predicts when it comes back, and what would remove it.
- Automating alignment researchOnce agents can run a week of experiments overnight, progress depends on how many of their results someone can trust. How I'd measure that, and Crux, the reviewer I'm designing for it.
- Scalable oversight for AI-written researchOversight evidence is strongest where a judge can check something specific. AI-written alignment research is where it's thinnest. What I'd test first.
- Introducing CruxAn early concept for checking alignment research done by AI, before anyone relies on it.
- The Atlas research agendaRobust Superalignment first, starting with Latent Invariant Alignment. Then automated alignment research, then scalable oversight.