Anthropic's CEO, Dario Amodei, presented the company's Responsible Scaling Policy (RSP) and its AI Safety Levels (ASL) framework at the UK AI Safety Summit. The RSP, which Anthropic published in September, was the first such policy from a major AI company. It aims to address the challenges of measuring AI risks as capabilities advance rapidly.
Key Points
- The Responsible Scaling Policy (RSP) was initially suggested by the Alignment Research Center.
- Anthropic's RSP has two major components, including the AI Safety Levels (ASL) system.
- ASL-1 models, such as a specialized chess AI, represent little to no risk.
- ASL-2 models, which describe current systems, have present-day risks but do not yet exhibit dangerous capabilities that could lead to catastrophic outcomes in fields like biology or chemistry.
- For ASL-2 models, the RSP requires implementing best practices like model cards, external red-teaming, and strong security.
- ASL-3 is triggered when AI models become operationally useful for catastrophic misuse in CBRN (chemical, biological, radiological, nuclear) areas, as defined by experts.
- ASL-4 represents an escalation of catastrophic misuse risks from ASL-3 and introduces concerns about autonomous AI systems that could escape human control or become a main source of a serious global security threat, such as bioweapons.
- Deep executive involvement is critical for crafting an RSP; Dario Amodei spent 10-20% of his time on Anthropic's RSP for 3 months, and a co-founder devoted 50% of their time for 3 months.
- Anthropic's RSP is a formal directive of its board, which is accountable to its Long Term Benefit Trust, an external panel of experts.
Context
According to Anthropic, the rapid progress of AI, driven by computation growing at 8x per year, makes it difficult to predict when AI will acquire specific, potentially dangerous skills. This situation necessitates frequent monitoring of emerging risks and protocols for appropriate responses. The RSP attempts to meet this need, drawing a comparison to the internationally recognized BSL system for handling biological materials.
Why It Matters
This policy outlines a structured approach for AI developers to manage increasing risks as AI capabilities advance. It suggests a framework for internal governance and accountability, providing a potential model for other companies and informing future regulatory discussions.
What To Do
- Note the distinction between ASL-2 (current models) and ASL-3 (models capable of catastrophic misuse).
- Observe the measures Anthropic requires for each ASL level, such as model cards and external red-teaming.
- Consider the role of executive involvement and board directives in implementing such policies.
- Watch for further developments in RSP-style frameworks from other companies and governments.
Keep Exploring
/atlas/claude-family /techniques/system-user-separation /techniques/ptcf