Mihir Wagle

Data quality is the hidden constraint holding back AI performance. When data is inaccurate, incomplete, or siloed, even the most advanced models produce unreliable outputs and perpetuate bias. #ai #dataquality #datagovernance #cybersecurity

3

My team is on a roll!! In our latest research, we dug into one of the hardest challenges in AI today: causal reasoning — and exposed major blind spots in how state-of-the-art models handle complex, real-world causation. 💡 Most evaluations force models to choose one answer. But real-life causes aren’t that simple. So we tested 8 frontier LLMs using multi-select (polytomous) tasks across 576 scenarios. The results? Both surprising and sobering. 🔍 Key insights: - Models adopt a conservative strategy when identifying causes but over-select confounders—revealing asymmetries in reasoning across task types. - Chain-of-thought prompting? 🚫 Not helping here. - How you score these evaluations can change which model looks best: Our results show that rankings shift significantly depending on the scoring metric used. - We uncovered behavioral signatures by model family: some are cautious and precise, others more liberal and prone to over-selection.     ⚠️ The implications for AI in healthcare, law, and policy are huge. Missing causal factors isn’t just suboptimal — it’s dangerous. 🧠 This research doesn't just offer a new benchmark. It offers a new lens for understanding model behavior under uncertainty and instruction complexity. 🔗 Full methodology, scoring analysis, and model comparisons https://lnkd.in/gvTEDhej Major kudos to the team: David Harper, Konstantinos Karageorgos & Abigail Thornton, PhD How are you measuring causal reasoning in your AI systems? #AI #LLMs #CausalReasoning #AIEvaluation #ResponsibleAI #WeloData #MachineLearning #AIRisk #ModelSelection #AIResearch

34

1 Comment

Today we announced that Maia 200 is now online in Azure. Maia 200 reflects a year-long focus on making AI inference more efficient and easier to deploy at scale. It delivers 30% better performance per dollar than our latest-generation fleet, with strong FP4/FP8 throughput optimized for large-scale AI workloads. In addition to performance and deployability, it also helps us make production-ready AI at hyperscale more sustainable by lowering energy per workload. Pretty exciting. 

21

The AI Research Team at Cisco Central AI Org keeps innovating how LLMs can be applied to help customers in managing large complex operational problems. This blog that the team just published is a great example of addressing the 'Large Context Window' bottleneck, delivering a scalable framework that slashes LLM latency and operational costs without compromising precision. https://lnkd.in/giFG45Cx

61

I am delighted to share that our paper, “CodeGEMM: A Codebook-Centric Approach to Efficient GEMM in Quantized LLMs,” has been accepted to NeurIPS 2025! It has long been a prevailing assumption that while codebook-based quantization (often referred to as vector quantization) can achieve higher accuracy than other formats such as uniform quantization, thanks to its more flexible representation for a given number of bits, the lack of regular structure makes it extremely difficult to accelerate GEMM operations. In this work, we challenge and overturn that long standing belief. By precomputing and storing inner product results in what we call a Psumbook, we demonstrate that codebook based quantization can indeed be made efficient for GEMM. We firmly believe that the future of cost-efficient AI services lies in combining compression algorithms with bold, innovative implementation strategies, and this paper represents one exciting step in that direction. ----------------------- Title: CodeGEMM: A Codebook-Centric Approach to Efficient GEMM in Quantized LLMs Authors: Gunho Park (NAVER Cloud), Jeongin Bae (NAVER Cloud), Byeongwook Kim (NAVER Cloud), Baeseong park (NAVER Cloud), Jiwon Ryu (NAVER Cloud), Hoseung Kim (NAVER Cloud), Se Jung Kwon (NAVER Cloud), Dongsoo Lee (NAVER Cloud) Abstract: Codebook-based quantization delivers superior performance in extreme low-bit environments but is hindered by latency overhead in existing dequantization-based kernels, which require loading the entire codebook into programmable cache memory. We propose CodeGEMM, a matrix multiplication kernel optimized for computing codebook-based quantized models by precomputing and storing inner product results in a Psumbook, thereby eliminating the need for traditional dequantization. The kernel supports a wide range of hyperparameter configurations, enabling exploration of trade-offs between latency, memory usage, and accuracy to achieve optimal performance. CodeGEMM obtains a 2.27× speed-up compared to the state-of-the-art codebook-based quantization method with comparable accuracy in the 2-bit configuration.

175

2 Comments

People are leveraging AI to build semantic layers. People are leveraging semantic layers to use AI. That just sounds too funny to me.

4

3 Comments

People see the results. They don't see the labor behind them — the relentless strategy work, the market deep-dives, the innovation bets that had to be right. It took deep focus, genuine team spirit, and nights where sleep wasn't an option. But we made it. And it was worth every hour. To the team I've had the privilege of leading — you are exceptional. Deep knowledge, lead with kindness, obsessed with customer outcomes, and quietly passionate about what we're building at Teradata. You don't talk about leading. You just lead. This is what vision plus execution actually looks like. Go Teradata. 🚀 Jefferson Uy, Pablo Escobar de la Oliva, Michael Riordan, Timothy Miller, Josefina Tamia van Geloven, Jessica Velez Perez, Danny R., Karthik Logasundaram, Doug Frazier, Faraz Ahmad, Pankaj Purandare, Mahesh Kharwadkar, Sukumar Burra, Kumar Sri Chandra Bhaskar Adabala, Subhadip (Sub) Bandyopadhyay

18

Attending Snowflake BUILD 2025 was such an incredible experience. It’s amazing to see how Snowflake is re-imagining the future of AI and data. Here were my top 5 takeaways from the conference: 1️⃣Building the agentic future: AI agents moving from concept to production. 2️⃣Cortex AISQL: natural language queries, no prompt engineering needed. 3️⃣ Cortex Agents: from queries to multi-step reasoning and automation. 4️⃣ Managed MCP Server: bridging AI tools through standardized connectivity. 5️⃣ Gen2 Warehouses: a data foundation built for real-time, enterprise-scale AI. Watch the full breakdown in my video 🎥 And if you’d like to learn more about Snowflake Intelligence, you can check out the link here: https://lnkd.in/gnRkJPv2 #SnowflakePartner #SnowflakeBUILD

303

31 Comments

Tuning Spark, Ray, or Dask for ML isn’t one-size-fits-all. In a new Domino Blueprint, Principal Solution Architect Sameer Wadkar shows how API-driven scaling complements autoscaling to right-size clusters on demand—cutting idle spend and speeding bursts for GPU/I/O-heavy work. Practical steps, a lightweight Python client, and Helm install included. If you care about FinOps and MLOps, this is for you. Read the blueprint: https://gag.gl/WwWso7

1

There are a lot of cool things about our new Q-CTRL AI Assistant, but one of my favorites is just how much EASIER it empowers people to work with quantum computers. We're starting with a deployment in Q-CTRL's docs, which has a wealth of technical information, enriched by our team's expertise, real-world examples , and client use cases. Now, anyone (literally) can truly experience the benefits of quantum computing. Not only do we provide the software abstraction layer, but we're also providing an AI-powered, expertise-driven configuration and interpretation layer for users and customers to get started with our products our dive deeper with more use case examples. It can not only generate code samples, but also the quantum circuits and inputs for the use cases you're interested in. Check out the video, and try it out by clicking "Ask AI" here: https://docs.q-ctrl.com!

17

#AI coding #agents are changing how I evaluate vendors. A strong argument for products is if they can be configured purely through code and #API / #MCP. Not because they have their own AI features - but because they let your AI agent do the work. Lightdash is a great example. It is not the most mature BI tool, but probably one of the most AI-ready - because they made all data and configuration accessible via API and gave AI agents the context to act on it, so Claude Code can end-to-end create dashboards by modelling data in the data warehouse and then build the visualizations. If your product is not API-first, AI cannot bring in outside context or act autonomously within it. That is becoming a real competitive disadvantage. https://lnkd.in/e6Gp3MGj

53

14 Comments

I met one AI development agency founder, and they said our 50% RAG apps failed in production last year. Because teams don’t know how to handle: chunking + embeddings + vector DB + reranking + hallucinations + evaluation + monitoring. So I'm sharing a RAG Playbook that shows exactly how to build production-ready RAG systems, with practical examples. ✅ What you’ll learn from this playbook This playbook helps you go from: “RAG demo is working” → “RAG system is reliable in production” You’ll learn how to: - reduce hallucinations using prompting techniques - use advanced chunking methods (not random splits) - select the right embedding model for your use case - choose the best vector database (based on real needs) - add reranking to improve relevance - follow clear steps to build an enterprise-grade RAG system - run evaluation before production using key scenarios - monitor and optimize performance after deployment - track RAG quality using 4 powerful metrics If you’re building a RAG-based feature for your product, this playbook will save you weeks of trial and error. ✅ Save this post (you’ll need it later) 🔁 Repost for your team/network ➕ Follow Shubham Vora to learn more on AI agents.

139

77 Comments

The combination of Netflix's event-driven architecture, domain-oriented design, and real-time processing capabilities creates a possible foundation for building an agentic AI mesh. While there may need to be some modifications—particularly in terms of semantic understanding and goal-oriented orchestration—the core infrastructure patterns for a decoupled architecture support distributed agent coordination at scale. I'm still not sure how context-sharing would occur. Netflix proves that millions of independent transactions can create emergent intelligence without explicit coordination. If Netflix's architecture can predict what you want to watch without agents explicitly sharing your viewing history, are we solving the wrong problem by focusing on context transfer instead of context inference? https://lnkd.in/eqVsz4px

11

We evaluated Foundry IQ's agentic retrieval against 15+ RAG benchmarks. The evaluation breaks down how our retrieval engine delivers better grounding at scale—and why knowledge quality matters for production AI agents. https://lnkd.in/e4CZR5uE

76

If you're giving LLMs more and more context and not getting the results you expect, check out my colleague Norman's new research paper about the maximum effective context window in LLMs. It's awesome to be building tech for the electric utility industry alongside such a thoughtful engineering leader. You can read his full paper here: https://lnkd.in/d4pMhXd9

4

A 12-month professional, voluntary collaboration has finally borne fruit. This morning, I was honored to see the University of Washington News (https://lnkd.in/gRRpDqjA) and the UW Milgard School of Business (https://lnkd.in/gKJ_pwbm) highlight our Responsible GenAI Framework. Why do we care about responsible generative AI? Because GenAI systems are no longer experimental curiosities. They shape decisions, influence behavior, generate content at scale, and increasingly affect trust, fairness, privacy, and accountability in organizations and society. Without intentional design and governance, these systems can amplify bias, create opaque risks, and erode confidence rather than enable innovation. This framework is about making GenAI useful, ethical, and sustainable at the same time. It emphasizes responsibility not as a constraint, but as an enabler of long-term value, credibility, and impact. I am grateful to the collaborators, reviewers, and institutions who supported this work, and excited to see responsible GenAI move from aspiration to practice. To learn more: - Read the RGAF Framework (PDF): RGAF Framework https://lnkd.in/gTaxnndV #GenAI #ResponsibleGenAI International Society of Service Innovation Professionals (ISSIP) Milgard CBA

23

1 Comment

Check out my latest demo for Week 3 of my 4-part series sharing the latest AlloyDB innovations from Next '26! What if you could run semantic vector searches on live database records and instantly join those results with your data warehouse or cold archives—without moving a single byte of data? In this demo, I’ll show how Lakehouse Federation acts as a unified access layer, connecting your live transactions, warehouse analytics, and historical archives under a single database lens, with: 🌐 Zero-Data-Movement Joins: Instantly join vector search results stored in AlloyDB with structured reference data in BigQuery or archived Parquet data in Apache Iceberg. No complex ETL pipelines required. 🚀 Automatic Query Pushdown: Keep performance fast and accurate by automatically pushing filters and aggregations directly down to BigQuery. 🔄 One-Click Reverse ETL: For queries that demand the absolute lowest latency, easily sync data from BigQuery directly into a local AlloyDB table with zero performance trade-offs. 🎯 Unified Data Platform: Query live transactional data and historical cold archives through a single lens—without changing your database connection string. Check out the full demo below to see Lakehouse Federation in action, and reach out to your Google Cloud team to learn more! #AlloyDB #GoogleCloud #Lakehouse #BigQuery #DataArchitecture #Next26 #Database cc: Sujatha Mandava, Bala Narasimhan, Sean Zinsmeister https://lnkd.in/dS6MiBjC

37