NVIDIA’s Audio Flamingo 3, GoogleSQL’s new pipe syntax, MetaStone-S1, Fractional ReasoningAn Exclusive Look into Next Gen BI – Live WebinarDashboards alone aren’t cutting it. The market’s moving toward something new: data apps, live collaboration, and AI that works the way teams actually work.See what's driving the rise of Next Gen BI, how Sigma earned a top debut on the Gartner Magic Quadrant, and what’s next for our roadmap.Secure Your SpotSponsoredSubscribe|Submit a tip|Advertise with usWelcome to DataPro 142: Tools Driving Tomorrow’s Thinking 🔬📈In this edition, we spotlight the breakthrough tools, patterns, and practices that are reshaping research and production in AI and data science.From NVIDIA’s Audio Flamingo 3 pushing the frontier of multimodal reasoning, to Fractional Reasoning’s elegant solution to adaptive LLM compute, and MetaStone-S1’s bold performance claims, this week’s releases are not just incremental; they’re foundational. Meanwhile, Kiro is redefining the dev experience, merging agentic coding with production-readiness from day one.On the systems front, Amazon EKS now scales to 100K nodes, opening the door to AGI-class workloads. And GoogleSQL’s new pipe syntax is winning hearts in the SQL community for its clarity and composability. If you’ve ever loathed nested subqueries, this is your moment.For those making decisions about tooling, don’t miss our link on Foundation vs. Custom Models, a smart, grounded guide for teams navigating performance vs. control. Also featured: Amazon SageMaker’s new unified catalog, practical AutoML with AutoKeras/Keras Tuner, and a no-fuss walkthrough of deploying Streamlit apps to AWS.Lastly, we dive into deeper reflections: Strands Agents 1.0 brings multi-agent orchestration into the real world, and standout articles explore paradox pitfalls in metrics, and how data’s 40-year evolution is shaping AI’s next wave.Let’s get into it. ⬇️Cheers,Merlyn ShelleyGrowth Lead, PacktTop Tools Driving New Research 🔧📊🔵 nvidia/audio-flamingo-3 · Audio Flamingo 3 (AF3) is an open Large Audio-Language Model (LALM) by NVIDIA for research use, capable of reasoning across speech, sound, and music. It supports long audio inputs, multi-turn voice dialogue, and chain-of-thought reasoning, achieving state-of-the-art results on 20+ tasks through unified audio representation and extensive dataset training.🔵 Fractional Reasoning via Latent Steering Vectors Improves Inference Time Compute: Fractional Reasoning introduces a model-agnostic, training-free method to dynamically adjust LLM reasoning depth at inference. By scaling latent steering vectors, it tailors compute per input complexity, boosting accuracy and efficiency. Compatible with Best-of-N, majority vote, and self-reflection, it outperforms fixed prompts across GSM8K, MATH500, and GPQA benchmarks.🔵 MetaStone-AI/MetaStone-S1: MetaStone-S1 is a 32B-parameter reflective generative model that rivals OpenAI-o3-mini on math, code, and Chinese reasoning. It combines Long-CoT Reinforcement and Process Reward Learning for efficient, high-quality inference. MetaStone-S1 achieves deep reasoning while reducing policy model costs by 99%, enabling fast, accurate outputs across multiple benchmarks.🔵 Introducing Kiro: Kiro is an agentic IDE that turns AI prototypes into production-grade apps using spec-driven development. It auto-generates requirements, design docs, and implementation tasks, and uses hooks for event-based automation. With built-in test coverage, design clarity, and consistency checks, Kiro helps developers ship reliable software faster and with greater confidence.Topics Catching Fire in Data Circles 🔥💬🔵 Do You Really Need a Foundation Model? Not every use case needs a foundation model. This guide compares foundation and custom models across performance, cost, latency, and control. It offers a decision framework, practical examples, and hybrid strategies to help teams choose the right approach, balancing rapid prototyping with long-term scalability, privacy needs, and task-specific optimization.🔵 Automating Deep Learning: A Gentle Introduction to AutoKeras and Keras Tuner. This guide introduces AutoKeras and Keras Tuner, two AutoML tools that simplify deep learning. AutoKeras automates architecture and training, while Keras Tuner optimizes hyperparameters of custom models. Together, they streamline experimentation, reduce guesswork, and boost performance, ideal for tasks like image classification, tabular modeling, or rapid prototyping with minimal manual tuning.🔵 Amazon EKS enables ultra scale AI/ML workloads with support for 100K nodes per cluster: Amazon EKS now supports up to 100,000 nodes per cluster, enabling ultra-scale AI/ML workloads with 1.6M Trainium or 800K GPU instances. This breakthrough powers large model training, reduces operational costs, and preserves Kubernetes compatibility, paving the way for AGI-scale innovation through enhanced orchestration, resiliency, and open-source flexibility.🔵 Exploring pipe syntax real-world use cases: GoogleSQL's pipe syntax reimagines SQL with a linear, readable data flow using the |> operator. It simplifies complex queries, streamlines data pipelines, and improves log analysis clarity. By eliminating nested structures and enabling intuitive chaining, pipe syntax boosts productivity, maintainability, and accelerates insight generation across BigQuery and Cloud Logging workflows.New Case Studies from the Tech Titans 🚀💡🔵 How Metrics (and LLMs) Can Trick You: A Field Guide to Paradoxes. This article unpacks how paradoxes like Simpson’s, the Accuracy Paradox, and Goodhart’s Law mislead both data science and LLM evaluation. It shows how surface-level metrics can distort truth, urging practitioners to embrace contextual, nuanced measurement, especially in BI and Retrieval-Augmented Generation, where incentives, imbalance, and aggregation errors can derail decision-making.🔵 What Can the History of Data Tell Us About the Future of AI? This sweeping 40-year history of data explores how shifts in storage, architecture, and business models have shaped intelligent systems. By tracing personal, public, and enterprise data, from PCs to cloud to AI, the piece reveals how incentives, infrastructure, and data ownership will determine the trajectory of AI’s future.🔵 Streamline the path from data to insights with new Amazon SageMaker Catalog capabilities: Amazon SageMaker now streamlines analytics with new integrations: QuickSight for in-studio dashboarding, S3 Access Grants for secure unstructured data sharing, and automatic onboarding of Glue Data Catalog datasets. These updates unify structured and unstructured data, accelerating workflows from raw data to insights, governed, discoverable, and ready for ML and BI use.Blog Pulse: What’s Moving Minds 🧠✨🔵 Deploy a Streamlit App to AWS: This hands-on guide walks you through deploying a Streamlit app on AWS using Elastic Beanstalk. It covers preparing your code, switching from Postgres to S3 for data, configuring AWS infrastructure, and managing deployment steps. Ideal for developers needing scalable, secure alternatives to public cloud endpoints like Streamlit Community Cloud.🔵 Accuracy Is Dead: Calibration, Discrimination, and Other Metrics You Actually Need. This guide challenges accuracy as a primary evaluation metric, urging data scientists to adopt deeper, problem-specific tools. It explores advanced classification metrics like ROC-AUC, log loss, and Brier score, and regression metrics like R², RMSLE, and quantile loss, emphasizing calibration, uncertainty, and decision-readiness over surface-level model performance.🔵 Introducing Strands Agents 1.0: Production-Ready Multi-Agent Orchestration Made Simple: Strands Agents 1.0 is a production-ready SDK for building multi-agent AI systems. It introduces primitives like Agents-as-Tools, Swarms, Graphs, and A2A support for inter-agent communication. With session persistence, async performance, and flexible model integration, Strands simplifies orchestration, scaling from prototype to production for complex, collaborative, and distributed agentic workflows.See you next time!*{box-sizing:border-box}body{margin:0;padding:0}a[x-apple-data-detectors]{color:inherit!important;text-decoration:inherit!important}#MessageViewBody a{color:inherit;text-decoration:none}p{line-height:inherit}.desktop_hide,.desktop_hide table{mso-hide:all;display:none;max-height:0;overflow:hidden}.image_block img+div{display:none}sub,sup{font-size:75%;line-height:0} @media (max-width: 100%;display:block}.mobile_hide{min-height:0;max-height:0;max-width: 100%;overflow:hidden;font-size:0}.desktop_hide,.desktop_hide table{display:table!important;max-height:none!important}}
Read more