September 2023 Summaries

Claude is now generally available on Amazon Bedrock, providing secure cloud access to foundation models and tools for building generative AI applications. Every AWS customer can now build with Claude, with exciting new experiences including Agents for Amazon Bedrock, which orchestrate API calls using AWS Lambda functions. This feature allows Claude to take on a more expanded role in understanding user requests, breaking down complex tasks into multiple steps, carrying on conversations, and taking actions. Secure customization and fine-tuning of Claude models through the service optimize performance with expert knowledge and proprietary data. The general availability of Claude advances responsible AI integration for enterprises, with application layer solutions to be announced in the coming months. Leading enterprises such as Bridgewater Associates are leveraging Claude to augment analyst productivity and create novel applications like an Investment Analyst Assistant on Amazon Bedrock.

Anthropic, a company focused on developing safer AI, has announced a partnership with Amazon to expand access to its technology. The agreement includes an investment of up to $4 billion from Amazon, which will provide Anthropic's safe and steerable AI to AWS customers through the Amazon Web Services platform. This collaboration aims to develop reliable foundation models in the industry, combining Anthropic's frontier safety research with Amazon's expertise in secure infrastructure. As part of the agreement, Anthropic's products, including its Claude 2 model, will be supported on Amazon Bedrock, enabling developers to build on top of state-of-the-art models and incorporate generative AI capabilities into their work. The partnership is also focused on promoting safety best practices in AI development and deployment, with both companies committed to responsible AI technologies.

We conducted an experiment to evaluate techniques for maximizing Claude's recall over long documents, with a focus on improving its performance in multiple-choice question-and-answer (Q&A) pairs. The goal was to identify effective prompting strategies that can help Claude accurately recall specific pieces of information from a long document. We tested four different prompting strategies: base, nongov examples, two examples, and five examples, both with and without the use of a scratchpad to pull relevant quotes. Our results showed that using many examples, pulling relevant quotes into the scratchpad, and providing contextual examples can significantly improve Claude's recall performance on both short and long context lengths. Additionally, we found that there is a monotonic inverse relationship between performance and the distance of the relevant passage to the question and the end of the prompt for Claude Instant, while Claude 2 performance shows a small dip in the middle. The experiment highlights the importance of careful prompting and using relevant quotes to improve the accuracy of long-context Q&A prompts.

The Anthropic Long-Term Benefit Trust (LTBT) is an independent body designed to align the corporate governance of Anthropic with its mission of developing and maintaining advanced AI for the long-term benefit of humanity. The Trust has five financially disinterested members who will select and remove a portion of the Board, growing over time until a majority is elected. This structure aims to fine-tune corporate governance to address the unique challenges and opportunities presented by transformative AI. The LTBT is paired with Anthropic's Public Benefit Corporation status, which already aligns its governance with its mission. However, the Trust provides additional accountability and incentives for directors to balance public interests with financial interests of stockholders, particularly in key decisions where consequences reach far beyond Anthropic. The Trust's design is an experiment, informed by corporate governance scholars and practitioners, and its gradual "phase-in" allows for course-correction as it becomes more mature and has profound effects on society.

Anthropic's Responsible Scaling Policy outlines technical and organizational protocols to manage the risks of developing increasingly capable AI systems, focusing on catastrophic risks that could cause large-scale devastation. The policy defines a framework called AI Safety Levels (ASL), modeled after biosafety levels, which requires safety, security, and operational standards appropriate to a model's potential for catastrophic risk. The ASL system categorizes models into four levels: ASL-1 (no meaningful catastrophic risk), ASL-2 (early signs of dangerous capabilities), ASL-3 (substantial increase in risk), and ASL-4 and higher ( qualitative escalations in catastrophic misuse potential). Anthropic aims to strike a balance between targeting catastrophic risks and incentivizing beneficial applications and safety progress, with the goal of creating a "race to the top" dynamic among frontier labs. The policy has been formally approved by Anthropic's board and will not alter current uses of Claude or disrupt availability of products.

Anthropic has partnered with Boston Consulting Group (BCG) to bring their AI assistant, Claude, to more enterprises around the world. This collaboration aims to help BCG customers deploy safer and more reliable AI solutions by leveraging Anthropic's Constitutional AI techniques. The partnership aligns with BCG's focus on responsible AI, enabling them to advise clients on strategic applications of AI and deliver business results through the use of Claude models. Anthropic and BCG have previously worked together, including a recent collaboration at the United Nations, and will now extend their partnership to help organizations harness the value and impact of generative AI in an effective and ethical manner. The goal is to set a new standard for responsible enterprise AI and promote a safety race to the top for AI deployment.

The company is introducing a paid plan called Claude Pro for its chat experience, which has been popular among users since its launch in July. The new plan offers 5x more usage of the latest model, Claude 2, for a monthly fee of $20 (US) or £18 (UK). This allows subscribers to increase their productivity and access priority support during high-traffic periods, as well as early access to new features. The paid plan also includes additional benefits such as increased file uploads and conversations over longer periods.