Eyal Lantzman

🚀 NEW on We ❤️ Open Source 🚀 The Python Software Foundation has halted all grant funding for the rest of 2025 after exhausting its annual budget early. This isn’t just a PSF issue—it’s a signal that open source ecosystems are underfunded and underprotected. In this article, Shirley Bailes examines the ripple effects: burnout, delayed security patches, reduced diversity outreach, and the slowing pace of innovation—especially in AI, where Python dominates. If your business depends on open source (hint: it does), this is your wake-up call. Read the article: https://lnkd.in/erimtJX3 #WeLoveOpenSource #Python #OpenSource #RiskManagement #TechLeadership #AI #Security #PSF #Sustainability

25

1 Comment

Visibility into distributed AI workloads isn't just nice to have - it's a competitive necessity. That’s why last week, we launched two powerful observability tools for the Anyscale platform: The RayTurbo Data Dashboard transforms how you debug and optimize data pipelines: ▪️ High-level pipeline visualization with both Tree and DAG views ▪️Bottleneck identification with actionable guidance ▪️Advanced log view that automatically finds problematic data patterns ▪️Persistent debugging that survives cluster termination The RayTurbo Train Dashboard tackles distributed training complexity: ▪️Clear visualization of training progress and job state ▪️Detailed error context when things go wrong ▪️Unified interface for logs and metrics across your entire cluster ▪️One-click profiling for CPU, GPU, and memory optimization Stop spending hours manually piecing together logs. Get clear, actionable insights instantly. Read more on our blog: https://lnkd.in/gWkp4tBE Or try it for yourself with $100 in free credits: https://lnkd.in/gmApC5AJ 

23

our co-founder and CTO Paul Groth talks about bridging the worlds of academia and industry on the Data Engineering Podcast.

3

Think about it: the mouse revolutionized how we use computers. 🖱️ Now, AI and LLMs are about to do the same for how we interact with software. 🤯 In this new episode of #thedatachief (hosted by Cindi Howson), Navan’s Ilan Twig explains how we’re moving from endless menus and buttons to smart, context-aware interfaces powered by AI. The next UX revolution is here — and it’s intelligent. 🎙️ The next big shift in user experience is already happening — don’t miss Navan’s insights on The Data Chief. Tune in now! https://lnkd.in/eVz9KckX

6

Seamless PyTorch compatibility in ROCm means developers can build once and deploy across both cloud and client. Alex (Wei) C., CEO of Nexa AI, shares more!

26

Huge thanks to our co-hosts at RisingWave and Aiven for an incredible event in San Francisco! We were thrilled to see such a high-energy turnout from the streaming and AI communities. It was a privilege to have our CEO, Sijie Guo, join Yingjun Wu, Filip Yonov, and Dmitry Kan to discuss the "Future of Kafka + AI." The session made it clear: the intersection of real-time data and agentic workflows is the next frontier for intelligent systems. Looking forward to the next one! 🚀

12

"Multimodal AI integration?" Everyone's talking about it, but they're missing the mark. 1. Combine models effectively by starting with open-source libraries like PyTorch and its versatile support for integrating vision and speech. Experimentation is key. 2. Use vibe coding to swiftly prototype integrations. For instance, linking a speech recognition model with a text generator can be done in a lunch break. 3. Avoid getting tangled in model complexity. Start with simple tasks, like audio transcriptions paired with image classification, and iterate based on your findings. 4. Build compatibility layers using transformers like BERT for text and ResNet for images. They share architectural similarities that make integration smoother. 5. Try asynchronous processing to handle diverse data streams. Python's `asyncio` can help you manage real-time data input seamlessly. 6. Leverage AI-assisted development to clone functional modules quickly. This approach has cut my integration time in half. 7. Remember that models aren't one-size-fits-all. Tailor each model to specific task requirements—precision in speech doesn't always translate to image data. ```python import asyncio from transformers import pipeline async def process_data(image, audio): text_pipeline = pipeline('text-generation') vision_pipeline = pipeline('image-classification') text_result = text_pipeline(audio) image_result = vision_pipeline(image) return text_result, image_result asyncio.run(process_data('image.jpg', 'audio.wav')) ``` What are your biggest challenges when it comes to integrating different AI models? Let's exchange insights! #AI #MachineLearning #GenerativeAI #LLM

LLMs are powerful tools that can help you with many tasks. But they're reactive, and they wait for your input before doing anything. Here, Balajee goes over how to use LLM agents, which have memory and autonomy, to take your use of AI to the next level. https://lnkd.in/gSZeEdHt

124

As LLMs push toward 100‑trillion‑parameter models and agentic AI workloads explode, compute performance is advancing ~20X every two years, while memory bandwidth improves only 2–3X. If you’re working on AI silicon, HPC, or data‑center architectures, this is a must‑read on how HBM4 is breaking down the 'memory wall'. https://lnkd.in/gxUAeiQ3

64

1 Comment

Announcing Claude Opus 4.5 on Vertex AI: Today, we’re excited to announce that Anthropic’s newest model, Claude Opus 4.5, is generally available on Vertex AI. As Anthropic’s most advanced model to date, it excels in coding, agents, vision, computer use and office tasks — at one-third the cost of its predecessor, Opus 4.1. We remain committed to providing Vertex AI customers with even more model choice, a core tenet of our platform. This new model adds to the latest Claude Sonnet 4.5 and Claude Haiku 4.5 models already available on Vertex AI. For customers already building with Claude Sonnet 4.5, Opus 4.5 serves as an upgrade delivering frontier performance across existing use cases.  When building with Opus 4.5 on Vertex AI, you get a unified AI development platform to quickly build, deploy, and operationalize your AI applications and agents — backed by Google Cloud’s reliable and high performance infrastructure, secure-by-default platform controls and advanced AI protection, and Agent Builder stack. How Opus 4.5 helps businesses — from complex coding to more capable agents  * Agentic tool use: The model, paired with Anthropic’s next wave of tool use improvements, enables more capable agents with new behaviors. For example, programmatic tool calling lets agents execute tools directly in Python. Tool search lets agents dynamically discover from hundreds of tools without context window bloat.  * Coding: The model can turn multi-day development projects into hours-long tasks, working independently with broader language support and enhanced behaviors like more efficient code, better test coverage, and cleaner architecture choices. * Office task agents: The model has improved memory to maintain context and consistency across files, alongside a step-change improvement in creating spreadsheets, slides, and docs.  * Financial analysis: The model can connect the dots across complex information systems. This includes regulatory filings, market reports, internal data, making sophisticated predictive modeling and proactive compliance possible. * Cybersecurity: The model brings professional-grade analysis to security workflows, correlating logs, vulnerability databases, and threat intelligence to support threat detection and incident response. Build on Vertex AI, our unified AI development platform Vertex AI is the trusted, open platform for building, deploying, and scaling differentiated AI applications and agents with unconstrained choice. By combining cutting-edge partner models like Claude Opus 4.5 with our enterprise-grade tooling, you can get the most out of your agents without sacrificing control, safety, or speed.  * Get the most out of your agents: Rapidly build, scale, and govern agents using our comprehensive Agent Builder stack, which makes it easy to move multi-step workflows from… 🔗 Google IA #ClaudeOpus #VertexAI #AI #ArtificialIntelligence #MachineLearning

2

🚀 If you’re building with LLMs, your system architecture matters more than your prompts. Most people stop at: “Call the model → Get the output.” But real LLM-native systems need workflows—blueprints that define how models think, route, evaluate, and collaborate. Here are 6 core LLM workflows you’ll find in production today: 🧠 1. LLM Augmentation Classic RAG + tools. The model extends its abilities using: • Retrieval (vector DBs) • Tool use (APIs, calculators) • Memory (short-term + long-term) 🔗 2. Prompt Chaining Workflow Sequential reasoning in steps. Each output is validated → passed to the next step. Perfect for multi-stage tasks: reasoning, summarization, translation, evaluation. 🛣 3. LLM Routing Workflow Inputs route to different models/prompts based on intent. Example: classification → Q&A → summarization on separate paths. 📊 4. Parallelization Workflow (Aggregator) Run multiple models/tasks in parallel → merge results. Great for ensembling and multi-perspective outputs. 🎼 5. Parallelization Workflow (Synthesizer) A conductor coordinates multiple agents/models. Ideal for structured multi-agent systems that need harmony, not just aggregation. 🧪 6. Evaluator–Optimizer Workflow The underrated powerhouse. One LLM generates → another evaluates (pass/fail + feedback). Loop continues until quality targets are met. If you're an AI engineer, don’t rely on single-shot inference. Design workflows that scale, self-correct, and adapt. 📌 Save this for your next LLM system architecture review. Credit: Aishwarya Srinivasan Follow Buzz Data Science for more interesting updates. #LLMArchitecture #LLMEngineering #LLMOps #AIWorkflows #AIAgents #RAG #VectorDatabases #AIEngineering #PromptEngineering #AgenticAI

3

AI/ML developers face tough orchestration challenges, including data pipelines, GPUs, failures, and deployments. Temporal’s code-first approach makes it simpler and faster to build reliable systems. Over 90 AI companies already use Temporal Cloud. Check out our blog post to see why 👉 https://lnkd.in/gKxWwMXc

17

PlanetCassandra #11 brings a deep dive into the future of database operations 🚀 This session explores AI-assisted Cassandra operations using Claude Code, where Jon Haddad demonstrates how domain-specific AI skills can embed real-world production expertise directly into workflows. Learn how structured methods like the USE method, outlier analysis, and double-loop learning can help diagnose issues faster, optimize performance, and improve decision-making without years of hands-on experience. 🔗 https://luma.com/fjncw7py #Anant #AI #Cassandra #DataEngineering

2

Lee Briggs, director of solutions engineering at Tailscale, has been keeping an eye on the Model Context Protocol (MCP) for connecting LLMs to your data. Why? "It seems absolutely terrifying," he writes. See his thoughts, and example fixes with code, on Tailscale's blog:

20

Exciting to see one of the founding voices in AI research launching a new chapter with Advanced Machine Intelligence (AMI) Labs, aiming to push beyond the limitations of today’s large language models. Rather than focusing solely on text prediction, AMI Labs is developing world models that learn to represent and reason about the physical environment — a foundational shift toward AI systems that can plan, remember, and interact with the real world. This move could mark a significant evolution in how we build and deploy intelligent systems for robotics, industry, and beyond. #ArtificialIntelligence #WorldModels #Innovation https://lnkd.in/g9wAAnNC

4

Super bullish on the use of Turso for embedded/edge AI. They're finalizing their rewrite of SQLite in Rust with a supposed 200k writes per second and improved concurrency. Key improvements include an async-first architecture that addresses SQLite's main pain points such as limited write concurrency and "database is locked" errors. The team is implementing features like MVCC (multi-version concurrency control) and BEGIN CONCURRENT to enable multiple concurrent writers, going beyond SQLite's traditional limitations. Turso is designed cloud-native from the ground up with fast connection opening and efficient schema management. It works across platforms (Linux, macOS, Windows, browsers via WASM) and is experimenting with modern language and runtime support. All while allowing open contributions to the core engine. Key for industry. In my opinion it will grow to become standard for robotics, embedded replicas, and scaling Agent-to-Agent long term memory solutions (A2A). I explored using it in an outbox worker queue pattern with Mistral OCR here and I'm liking the ease of use: https://lnkd.in/d6K-smfm

32

5 Comments

Guest post by our own Developer Advocate 🤖 Jacky Liang, on the Pydantic blog, about how Tiger Data (creators of TimescaleDB) built Tiger Agent for Work—our production AI Slack bot that handles thousands of conversations using Pydantic AI and Logfire. Great deep dive into the technical architecture and observability setup: https://lnkd.in/eNXnRgkS #PydanticAI #Logfire #AIAgents #TigerData

6