Align · Pillar
Artificial Alignment & Safety
Making powerful systems do what we intend — under distribution shift, deception, and scale.
Constitutional AI — Anthropic
Training assistants with a written constitution and AI critique loops — a public blueprint for scalable oversight without pure RLHF dependence.
Superalignment — OpenAI’s framing
OpenAI’s (evolving) public stance on aligning systems smarter than their operators — recursive oversight, interpretability, and empirical bets.
Frontier Safety Framework — DeepMind
Risk thresholds, capability evals, and escalation policies for frontier models — safety as an operational protocol, not a slogan.
Many-shot jailbreaking — Anthropic
A concrete failure mode: long-context prompting that overrides refusals. Alignment that works in short chats can fail at scale.