November 2023 Summaries

Claude 2.1 is a significant update to the Claude model, delivering advancements in key capabilities for enterprises, including an industry-leading 200K token context window, reductions in rates of model hallucination, and the introduction of tool use. This new feature enables users to integrate Claude with their existing processes, products, and APIs. The updated model also boasts a 2x decrease in false statements compared to its predecessor, improving honesty and accuracy. Additionally, Claude 2.1 features improved comprehension and summarization capabilities, particularly for long, complex documents. The update is now available over API in the Console and is powering the claude.ai chat experience. The pricing has also been updated to improve cost efficiency for customers across models.

The US government has issued a comprehensive executive order on AI, addressing risks related to privacy, fairness, bias, and catastrophic risks, while recognizing the benefits of AI in improving government services. The G7 has produced an International Code of Conduct for Organizations Developing Advanced AI Systems, outlining responsible practices to mitigate risks across the AI development and deployment lifecycle. A historic summit on AI safety at Bletchley Park resulted in the Bletchley Declaration, a call for multi-stakeholder action to harness AI benefits while addressing its risks, with 28 countries represented. The UK has announced plans to establish an AI Safety Institute, while the US plans to launch a companion initiative via NIST, focusing on evaluating AI systems and developing guidelines for frontier threats. These events mark the beginning of a new phase of AI safety and policy work, with governments demonstrating unprecedented interest in evaluating and monitoring AI systems.

Anthropic's Responsible Scaling Policy (RSP) is a framework aimed at mitigating AI-related risks by establishing a system to monitor and respond to emerging threats. The policy, developed in response to the rapid progress of AI systems, has two major components: AI safety levels (ASL), which categorize models based on their risk level, and frequent testing for dangerous capabilities. ASL-1 represents low-risk models, while ASL-3 and ASL-4 represent increasingly severe risks, with ASL-4 triggering concerns about autonomous AI systems posing a significant threat to society. The policy emphasizes the importance of executive involvement, making protocols into product and research requirements, accountability, and collaboration between companies and governments to refine and improve RSP-style frameworks.