Arcus

At Arcus, our mission is to bring intelligence to complex data and processes. We are building compound AI systems that leverage LLMs and multi-agent systems to understand complex data, perform nuanced reasoning, and take action. If you’re interested in solving hard problems, building intelligent systems, and defining the next-generation of AI capabilities that can do real work, join us in our mission! We’re hiring across platform, ML, and product! Apply at https://lnkd.in/eVqZWicY or email us at [email protected]! Hope to see you in our inbox! 🚀

🚀🌆 Arcus is taking over Times Square! Our mission at Arcus is to build intelligent systems that can understand, reason, and take action, working in concert with humans to solve hard problems and do real work. If this mission resonates with you, we’d love to chat! Check out careers at arcus.co/careers. Huge thanks to our friends at Brex and Kasper Koczab for making this happen! #GrowWithBrex

We’re hosting an ML Hang with our friends at 8VC for AI Researchers & Founders in SF! We’re hosting all-star AI founders and researchers from Stanford, OpenAI, Anthropic, DeepMind and early-stage research-focused startups at 8VC’s SF HQ on August 13. If you’re an ML Researcher or Founder interested in joining, DM our CTO Arda or email him at [email protected].

Arcus’ Planet-Scale RAG significantly enhances agents' ability to understand complex data and do real work! The problem with RAG is that splitting a document into chunks for retrieval results in each chunk having a minimal representation of the original content. This leads to a loss of context and essential information, and finding the needed chunk gets difficult for agents since each piece is taken out of its original context. Moreover, as the volume of data increases, the noise in each retrieval increases, leading to more frequent matches with incorrect data. Arcus tackles this issue with a proactive dual-layered strategy. 1️⃣Multi-Tiered Approach: This approach breaks the search into stages for efficiency. First, our retrieval system finds relevant documents and then searches within those documents for the most pertinent chunks. This method filters out irrelevant data better than traditional methods. Adding more tiers refines the search, progressively narrowing down the data to find the best match for the query. 2️⃣Semantic Tiering: Our retrieval system uses Semantic Tiering to refine search by categorizing documents based on meaning. In a multi-tier system, we group documents first by topic, then search within those groups, and finally within individual documents for specific chunks. This addresses the challenge of extracting and clustering the correct semantic information to ensure accurate and efficient retrieval. Learn more about how Planet-Scale RAG massively increases the capability of AI agents: https://lnkd.in/ehCceSR3 

Top LLMs are wrong about 40% of the time when used as agents to perform tasks that require calling multiple tools. According to the Berkeley function calling leaderboard, the highest tool calling accuracy is 90.18%, which is fairly high for a task requiring only one tool, but tasks can often require multiple tool calls to reach the desired outcome. For instance, if a task requires 5 tools, the probability reduces to 90.18^5, which is approximately 60%. As the complexity of a goal increases, the number of tools required to achieve that goal increases, dropping the probability of success exponentially. At Arcus, we address this problem by building robustness through systems around the tool-calling capabilities of agents to achieve real-world, production-worthy performance. Read more about two main design patterns that allow our agents to handle high complexity: https://lnkd.in/enDkQwmC 

🤖 Most agents plan one step at a time for complex tasks, which is error-prone and unreliable as these agents might not be able to consider possible obstacles nor adapt mid-episode to achieve the desired goal. 🔨 A possible solution is designing agents frameworks to have long-term plans that start with breaking down longer goals into sub-tasks and assigning them to specialized executor-agents. This decomposition into sub-goals allows for a bigger-picture, higher-level plan of the steps that need to be taken. ⚡ We believe long-term planning can increase the scope and difficulty of the tasks that agents perform and result in robust, production-worthy AI workflows. 🚀 Read more about how we at Arcus think about the future of LLM-based agents: https://lnkd.in/enDkQwmC

Tools allow LLM-based agents to call functions and interact with the external world. Agents can use these tools to make API calls, browse the web, execute code, etc., and perform actions that can be stitched together to do more complex tasks (as we do at Arcus). A tool-use workflow generally involves the following steps: 1. The agent decides whether to use a tool or return an answer directly. 2. If required, the tool is executed. 3. Agent analyzes the result to determine if the task is complete or needs further iteration. If it doesn’t, then the agent returns the result.   4. The agent repeats steps 1, 2, and 3 until the desired result is achieved. At Arcus, we build advanced AI systems that leverage LLMs, agents, and tools to perform complex, real-world tasks with high accuracy and performance. Read more about tool-use and agents in our blog post: https://lnkd.in/efRUk8xw

🍻 Another Friends of Arcus Happy Hour! We really enjoyed the company of some of the best founders and builders in the NYC tech and AI community! It’s always special to host such a crowd! Thank you everyone for coming! Good conversations and great vibes! We’re super excited to host more happy hours in the coming months - stay posted for the next one!

Join us at Arcus for our second Friends of Arcus Happy Hour on July 10! Meet our team and some of NYC's top tech talent and founders in the AI community. Come by our Flatiron office and unwind with snacks, drinks, and great conversations!  DM or comment for an invite link!

We are thrilled to welcome Shaurya Jain to our team at Arcus as a Summer Business Operations Intern! Shaurya is a student at NYU, pursuing Computer Science and Business Studies. He has a strong passion for startups and Artificial Intelligence. We are excited to have him on board with us. Welcome, Shaurya!