Anthropic

Tool use is now generally available. With tool use, Claude can, if instructed, intelligently select and orchestrate tools to solve complex tasks end-to-end. Start building today with the Anthropic API, Amazon Bedrock, or Google Vertex AI: https://lnkd.in/ewVXpBec. Tool use also supports message streaming, "forced" tool choice, and vision support—helping you build more natural, focused, and multimodal experiences. Hear from some of our early customers: "Claude with tool use is accurate and cost-effective, and now powers our live voice-enabled AI tutoring sessions. Within just a few days, we integrated tools into our platform. As a result, our AI tutor, Spark.E, acts agentically—displaying interactive UIs, tracking student progress in context, and navigating through lectures and materials. Since implementing Claude with tool use, we've observed a 42% increase in positive human feedback." —Ryan Trattner, CTO and Co-Founder, StudyFetch "Claude 3 Haiku with tool use has been a game changer for us. After accessing the model and running our benchmarks on it, we realized the quality, speed, and price combination is unmatched. Haiku is helping us scale our customers' data extraction tasks to a completely new level." —Faisal Ilaiwi, Co-Founder, Intuned (YC S22) "We leverage Claude 3 Haiku for generating live suggestions, automating prompt writing, and extracting key metadata from long documents. Claude 3 Haiku's tool use feature has unlocked capabilities and speed for our platform to generate reliable suggestions and prompts in real-time." —Divya Mehta, Product Manager, Hebbia AI

Our most intelligent model, Claude 3 Opus, is now available on Vertex AI. Alongside Haiku and Sonnet, Opus provides businesses with industry-leading accuracy, coding, reasoning, and vision capabilities.

We're grateful for TIME's recognition of our work: https://lnkd.in/gyHyxkde. We remain focused on the necessary work ahead to build frontier AI systems that are reliable, interpretable, and steerable.

Welcoming Jay Kreps to Anthropic's Board of Directors: https://lnkd.in/eATvCFpp

Welcoming Krishna Rao to Anthropic: https://lnkd.in/eMnCxxzP

This week, we showed how altering internal "features" in our AI, Claude, could change its behavior. We found a feature that can make Claude focus intensely on the Golden Gate Bridge. Now, for a limited time, you can chat with Golden Gate Claude: claude.ai. Our goal is to let people see the impact our interpretability work can have. The fact that we can find and alter these features within Claude makes us more confident that we’re beginning to understand how large language models really work. Read more: https://lnkd.in/ehaU7EbF

New Anthropic research paper: Scaling Monosemanticity. The first ever detailed look inside a leading large language model. Read the blog post here: https://lnkd.in/eyGAH4yF Our previous interpretability work was on small models. Now we've dramatically scaled it up to a model the size of Claude 3 Sonnet. We find a remarkable array of internal features in Sonnet that represent specific concepts—and can be used to steer model behavior. The problem: most LLM neurons are uninterpretable, stopping us from mechanistically understanding the models. In October, we showed that the technique of dictionary learning could decompose a small model into "monosemantic" components we call "features"—making the model more interpretable. For the first time, we’ve extracted millions of features from a high-performing, deployed model (Sonnet). These features cover specific people and places, programming-related abstractions, scientific topics, emotions, among a vast range of other concepts. These features are remarkably abstract, often representing the same concept across contexts and languages, even generalizing to image inputs. Importantly, they also causally influence the model’s outputs in intuitive ways. Among these millions of features, we find several that are relevant to ensuring model safety and reliability. These include features related to code vulnerabilities, deception, bias, sycophancy, power-seeking, and criminal activity. One notable example is a "secrecy" feature. We observe that it fires for descriptions of people or characters keeping a secret. Activating this feature results in Claude withholding information from the user when it otherwise would not. This work is preliminary. Whereas we show that there are many features that seem plausibly relevant to safety applications, much more work is needed to establish that our approach is useful in practice. Our research builds on prior work in sparse coding, compressed sensing, and disentanglement in machine learning, mathematics, and neuroscience. We are also pleased to see work from many other research groups applying dictionary learning and related methods to interpretability. There’s much more in our paper, including detailed analysis of the breadth and specifics of features, many more safety-relevant case studies, and preliminary work on using features to study computational "circuits" in models. Read the full paper here: https://lnkd.in/eaYkgSNM.

Last year we published our Responsible Scaling Policy—a set of technical and organizational protocols to help manage the risks of developing more capable AI systems. As we continue to iterate on our empirically grounded framework, we're gaining valuable insights. Today we share reflections on our progress: https://lnkd.in/eTnJFpMW

Welcoming Mike Krieger to Anthropic: https://lnkd.in/eg63wQQp

Claude is now available in Europe: https://claude.ai. The Claude iOS app is also available on the App Store: https://lnkd.in/gFV-hubK