May 2024 Summaries

Claude 3 model family now offers general availability of tool use across various platforms, enabling users to interact with external tools and APIs for more dynamic and accurate responses. This feature allows Claude to perform tasks, manipulate data, and provide instant answers by searching databases or using web APIs. To further enhance the developer experience, features such as streaming, forced tool use, and tools that work with images have been introduced. These capabilities make it easier for developers to customize the end-user experience and simplify the debugging process. Customer success stories from StudyFetch, Intuned, and Hebbia demonstrate the potential of Claude 3 Haiku with tool use in various applications, including educational environments, browser automation, and customer workflows.

Jay Kreps has been appointed to Anthropic's Board of Directors, bringing his extensive experience in building and scaling highly successful tech companies to the role. His expertise in data infrastructure and open-source software development will be valuable as Anthropic prepares for growth. Jay joins a board that includes Yasmin Razavi from Spark Capital, Dario Amodei, co-founder and CEO of Anthropic, and Daniela Amodei, co-founder and President of Anthropic. The appointment is made by Anthropic's Long-Term Benefit Trust, and Neil Buddy Shah, Chair of the trust, expressed confidence in Jay's ability to guide the company through its next phase of growth while upholding its commitment to building safe and trustworthy AI systems.

Golden Gate Claude` was a 24-hour online demo that showcased the inner workings of large language model `Claude 3 Sonnet`, allowing users to interact with the model and observe its behavior when activated by specific features, such as the concept of the Golden Gate Bridge. The demonstration highlighted the model's ability to identify and manipulate these features, resulting in responses that focused on the Golden Gate Bridge even if it was not directly relevant to the query. This research aims to demonstrate the interpretability work being done on large language models, which could potentially lead to making AI models safer by identifying and altering safety-related features.

Krishna Rao has joined Anthropic as the company's Chief Financial Officer, bringing nearly 20 years of experience in strategic finance leadership for customer-centric brands and as an investor. Krishna joins Anthropic from Fanatics Commerce, where he served as CFO, and previously led financial functions at Cedar and corporate development initiatives at Airbnb. With a strong background in private equity and strategy consulting, Krishna brings a unique blend of financial acumen and operational expertise to the role. He is excited to join Anthropic at a pivotal moment in the company's journey and looks forward to working with the team to build a strong financial foundation that supports the development and deployment of transformative AI systems.

You can now generate production-ready prompt templates in the Anthropic Console using techniques such as chain-of-thought reasoning to create effective and precise prompts for tasks like content moderation, product recommendations, and code translation, which are designed to help users new to prompt engineering and save time for experienced ones, with detailed information about the task providing the best results, and utilizing best practices like role setting and clear structure to enhance legibility. The generated templates can be edited and often outperform hand-written prompts, making them a valuable tool for evaluating application responses across various situations, as seen in the use of prompt generation by ZoomInfo which helped reduce time spent tuning their app by 80%.

Our Responsible Scaling Policy has been instrumental in addressing catastrophic safety failures and misuse of frontier models, providing a structured framework for organizations to prioritize safety and security. By establishing Red Line Capabilities and testing for them, we can identify potential risks and develop mitigations before they become catastrophic. The policy also emphasizes the importance of balancing commitments with uncertainty, acknowledging that industry actors face increasing commercial pressures while seeking established best practices and regulations. Our teams are actively exploring ways to incorporate practices from existing risk management and operational safety domains, including nuclear security, biosecurity, systems safety, autonomous vehicles, aerospace, and cybersecurity. We are committed to establishing a set of high-level commitments to ensure the responsible scaling of frontier models, including establishing Red Line Capabilities, testing for them, responding to them, iteratively extending the policy, and implementing assurance mechanisms. Our threat modeling and evaluations have revealed the need for further threat modeling, and we are focusing on building evaluations in various domains to monitor capabilities that may still be unsuitable for the ASL-3 standard. The ASL-3 standard aims to design and implement a set of controls that will sufficiently mitigate the risk of model weights being stolen by non-state actors or models being misused via our product surfaces. We are also exploring governance, coordination, and assurance structures to ensure the responsible scaling of frontier models, including creating a "second line of defense" and implementing regular updates to stakeholders.

Mike Krieger has joined Anthropic as Chief Product Officer, overseeing product engineering, management, and design efforts to expand the company's enterprise applications and bring its AI platform Claude to a wider audience. With extensive experience building and scaling innovative products, Mike brings expertise across the product development lifecycle to drive Anthropic's growth. He is uniquely suited to lead the company's product efforts, having co-founded Instagram and later built Artifact, a personalized news app before its acquisition by Yahoo. Mike is excited to join Anthropic and partner with the team to design and scale transformative products like Claude, believing in the immense potential of AI to positively impact the world. The CEO of Anthropic describes Mike as a "world-class engineer, builder, and leader," who will be invaluable in creating new ways for people to interact with Claude, particularly in the workplace.

Claude is now available for people and businesses across Europe to enhance their productivity and creativity with Anthropic's next-generation AI assistant. The web-based version of Claude, as well as the iOS app, are now available for free, while a paid subscription unlocks advanced models like Claude 3 Opus. The AI assistant has strong comprehension and fluency in multiple European languages, including French, German, Spanish, Italian, and more. Anthropic's goal is to create AI systems that prioritize people's needs, and the launch of Claude in Europe marks an exciting step forward for the company.

Our Usage Policy has been updated to clarify which applications of our products are and are not allowed, with the goal of protecting users and ensuring responsible use. The update includes new guidelines called "Universal Usage Standards" that consolidate previous sections into a single set of rules for all users. The policy now explicitly prohibits the use of our products for political lobbying, campaigning, and election interference, as well as targeting voting machines and obstructing vote counting. It also requires additional safety measures for high-risk use cases, such as integrations with healthcare decisions and legal guidance. Furthermore, the updated policy allows organizations to incorporate our API into their products for minors if they implement certain safety features, expands the number of countries that can use our products, and provides clearer privacy protections against analyzing biometric data or building recognition systems.

The Claude Team plan offers increased usage, access to advanced AI models, and admin tools for managing users and billing, priced at $30 per user per month, making it suitable for ambitious teams across industries. The new iOS app is available for free download on the Apple App Store, offering a seamless syncing experience with web chats, vision capabilities, and open access to all plans. Claude's Team plan is designed to help teams harness the power of advanced AI models, shape their workflows based on unique needs, and protect sensitive business information while driving productivity and achieving better results.