Databricks Mosaic Research的动态

I’m as bullish as you can get with agenetic systems.

As #GenAI projects move from POC to production, Databricks customers are shifting from deploying single models to leveraging AI agent systems. @NaveenGRao, Matei Zaharia and @pwendell describe how modular system design has led to a new development paradigm for intelligence applications:

From Simple Conversation History to Event-Driven AI: How State Management Is Evolving I’ve been exploring how AI Agents manage memory and state, and I recently came across a great article from MotleyCrew.ai that dives into four key implementation patterns shaping the future of AI: 1. Conversation History: - Basic approach that appends the full chat history in prompts - Quality can degrade as the conversation grows - Costs scale quickly with length 2. External State Storage: - Leveraging vector databases, relational/graph databases, and retrieval-augmented generation (RAG) - Enables efficient lookup of relevant context or structured data - Keeps prompts lean and reduces token usage 3. Stateful Tools: - Persistent tool instances that maintain context - Sharing state between different tools/agents - Managing things like authentication tokens and global state 4. Event-Driven Systems: - Loosely coupled messaging with persistent event logs (e.g., Kafka) - Asynchronous workflows and cross-agent orchestration - Scalable architecture for complex AI deployments Why It Matters: - As AI systems mature from prototypes to production, robust memory and state management are becoming critical. - These patterns offer AI Engineers a roadmap for deciding how to store, retrieve, and leverage data in a cost-effective, scalable way. - We’re seeing parallels with the evolution of distributed systems in traditional web development, suggesting that best practices from that domain can guide us in building resilient AI systems. If you’re working on AI Agents, consider starting simple (conversation history + RAG) and gradually adopting more advanced patterns (stateful tools, event-driven architectures) as your use cases demand. The key is balancing complexity, performance, and maintainability. Read more here: https://lnkd.in/e9iDAXJB #AIAgents #AIEngineering #LLM #StateManagement #MachineLearning #DataManagement #DistributedSystems #EventDriven #TechTrends #FutureOfAI #Kafka #Scalability #Complexity #RAG #AI #GenAI #Memory #DeepLearning #Agents

Come and Use the advanced MLOps features of xpresso.ai xpresso.ai MLOps platform features:? 1) Multi-model Single Endpoint Inference and? 2) Multi-container Inference Important points Here are the key points for the features Multi-model Single Endpoint Inference and Multi-container Inference offered by platforms like xpresso.ai: ??. ??????????-?????????? ???????????? ???????????????? ?????????????????? This feature allows multiple machine learning models to be deployed and served through a single endpoint. ?????? ????????????: ???????????????? ????????????????????: Reduces infrastructure overhead by serving multiple models via a single endpoint. ?????????????? ?????????? ??????????????????: Allows the inference system to route incoming requests to the appropriate model dynamically based on metadata or user-defined logic. ?????????????? ???????????????? ????????????????????: Simplifies operational complexity since only one endpoint needs to be managed. ????????-??????????????????: Saves costs on infrastructure by consolidating multiple models into a single deployment instance. ??????????????????????: Supports seamless scaling of the endpoint for high throughput while managing multiple models efficiently. ?????? ???????? ??????????????????????: Particularly useful for applications needing multiple models (e.g., multi-language models, different classifiers, etc.). ??. ??????????-?????????????????? ?????????????????? This feature allows deploying multiple containers within a single inference setup, where each container can host a different model or service. ?????? ????????????: ???????????????? ?????????? ??????????????: Supports diverse frameworks, versions, or environments by allowing different models to run in separate containers. ???????????????????? ??????????????????: Containers operate independently, enabling concurrent processing and inference across models or services. ???????????????? ??????????????????????????: Enables highly customized pipelines by combining models, pre-processing, or post-processing steps as separate containers. ?????????? ????????????????????????????????: Supports heterogeneous deployments where different models or services may rely on distinct runtime environments. ??????????????????: Ensures that the models or services are isolated, preventing conflicts between dependencies. ???????????????? ??????????????: Individual containers can be updated independently without affecting the entire inference service. ???????????????? ????????????????: Useful for multi-step inference workflows where pre/post-processing or additional AI services are part of the inference process. These features are essential in modern MLOps as they provide the ability to maximize infrastructure efficiency, support diverse model requirements, and offer scalable, customizable deployment options. UST #MLOps #BigData #Microservices Arijit Chowdhury Lal Babu Sah Anushree Sarkar MLOPS Solutions Private Limited

Data-First IDP: Driving AI Innovation in Developer Platforms https://buff.ly/3Py1Pn8 Traditional internal developer platforms (IDPs) have transformed how organizations manage code and infrastructure. By standardizing workflows through tools like CI/CD pipelines and Infrastructure as Code (IaC), these platforms have enabled rapid deployments, reduced manual errors, and improved developer experience. However, their focus has primarily been on operational efficiency, often treating data as an afterthought. This omission becomes critical in today's AI-driven landscape. While traditional IDPs excel at managing infrastructure, they fall short when it comes to the foundational elements required for scalable and compliant AI innovation:

?? Exciting Milestone in My AI & Automation Journey ?? Over the past few weeks, I've had the opportunity to dive deeper into LangSmith, LangGraph, and Azure integrations to build a more robust and observable AI-driven automation workflow. I successfully executed an Azure query to track resource changes in my environment, incorporating LangGraph for better observability and state management. The real highlight of this project was using AI agents to analyze and interpret the data from the Azure query in real-time. Here's how I utilized AI agents: ?? AI Agents for Workflow Automation: The AI agents were integrated with LangGraph to enable dynamic decision-making. They interpreted the results of my Azure query and guided the flow of information, ensuring that I captured the right details about each resource change. I used LangChain to bind the AI agents to specific tasks, such as making the query, interpreting the results, and determining whether the workflow should continue based on the data received. This helped me automate and streamline what would otherwise be a tedious manual process. Through the use of LangGraph, I added memory and persistence, allowing the AI agents to recall and use previous context to guide subsequent steps in the workflow. This made the system smarter and more adaptive. With this setup, I was able to query Azure Resource Graph for detailed information about recent resource changes, including who made the changes, what changes occurred, and how they were made. Here's a sneak peek at the query results: ?? Recent Activity: User: [email protected] Changes: Create and Update operations Client: Azure Portal By integrating LangGraph’s state management and LangSmith’s tools, I could not only retrieve granular data but also capture it in an organized, observable workflow. However, there’s still much to learn and explore as I continue this journey: ?? Challenges & Areas for Improvement: Reducing latency when using AI agents in real-time queries. Expanding the prompt to capture even more granular data about resources. Understanding LangGraph more deeply for complex workflows and decision-making processes. I’m eager to continue expanding my knowledge in AI and automation, optimizing workflows, and enhancing observability for more scalable and efficient systems. ?? The future is bright – I’m excited to dive deeper into these tools and see how they can help us build smarter, faster, and more reliable systems! #AI #LangChain #LangGraph #Azure #Automation #Observability #CloudComputing #DevOps #MachineLearning #AzureDevOps #TechInnovation

Checklist of the technical requirements for an #AI application platform: ???????? ???????????????????????from #enterprise and extraprise systems, and sensor networks ??????????-?????????? ???????????????????for cost effective elastic scale-out compute and storage on private and public clouds ???????? ???????????????????for low-latency local processing and AI predictions/inferences ???????????????? ?????????????????for continuous data processing, temporal and spatial processing, security and data persistence ???????????????????? ???????????? ?????????????to provide a consistent and comprehensive object model representation across a business ???????????????????? ???????????????????????????to provide a catalog of AI based software services ???????????????????? ???????? ?????????????????to provide robust user access authentication and authorization ???????????? ?????????????????????using AI and Dynamic Optimization Algorithms with full lifecycle support including development, testing, and deployment ???????? ?????????????????with support for multiple programming languages, standards-based interfaces (APIs), open source machine learning and deep learning libraries, and third-party data visualization tools ???????????? ?????????????????for Collaborative Development between Software Developers and Data Scientists #enterpriseAI #businessoutcomes #AIatScale

Dive into the world of AI Workload Design with Siddhesh Naik, Country Leader of Data, AI & Automation software at IBM India & South Asia. Gain invaluable insights as he discusses the pivotal role of Automation by Design in shaping AI systems with Dataquest. Join us as we explore the critical factors and optimization strategies essential for designing systems tailored to AI workloads. Don't miss this opportunity to glean expertise from Siddhesh Naik and revolutionize your approach to AI implementation. ?? #AI #Automation #Data #TechLeadership #IBMInsights #AIWorkload

Get a closer look at Sumo Logic Mo Copilot! Built on AWS Bedrock and designed for #DevSecOps, Mo Copilot is your #AI assistant that turns natural language into Sumo Logic queries and visualizations for faster investigation and response. ?? https://ow.ly/mi4230sIAoz