Anthropic employs a "Responsible Scaling Policy" (RSP) to manage risks linked to increasingly capable AI models[5]. This policy uses a framework called AI Safety Levels (ASL), drawing inspiration from the U.S. government's biosafety level standards for handling dangerous biological materials[5]. The ASL framework is designed to implement safety, security, and operational standards suited to a model's potential for catastrophic risk, with higher ASL levels requiring more stringent demonstrations of safety[5].
Anthropic's best current models are at ASL-2[4]. The company defines containment and deployment measures for ASL-2 and ASL-3, and it commits to defining ASL-4 safety measures before training ASL-3 models[4]. ASL-1 refers to systems that pose no meaningful catastrophic risk[5].
Anthropic is also invested in mechanistic interpretability, which involves dissecting and understanding the internal workings of AI systems, particularly deep learning models, in an effort to make AI behavior more predictable and understandable[5].
Citations:[1] https://www.prompthub.us/blog/using-anthropic-best-practices-parameters-and-large-context-windows
[2] https://www.techrepublic.com/article/anthropic-claude-large-language-model-research/
[3] https://aizi.substack.com/p/comments-on-anthropics-scaling-monosemanticity
[4] https://ailabwatch.org/companies/anthropic/
[5] https://klu.ai/glossary/anthropic-ai
[6] https://help.promptitude.io/en/articles/8892919-understanding-anthropic-models-a-simple-guide
[7] https://www.anthropic.com/news/anthropics-responsible-scaling-policy
[8] https://www.lesswrong.com/posts/vAopGQhFPdjcA8CEh/anthropic-reflections-on-our-responsible-scaling-policy