Data Council AI

It’s official… mark it down! Data Council AI returns to San Francisco: May 12–14 at the Marriott Marquis. Every year, Data Council AI brings together top innovators in AI & data for technical talks, networking and collaboration. Three days in the heart of SOMA with the world's top AI & data minds. 📅 Block your calendar → datacouncil.ai

👋 SF friends - throwing a happy hour for technical founders, AI & data engineers, and ML researchers. Join me next week for Agents & Founders: Back to School Party in SF on September 10th! Come share your latest insights on AI, agents, and multi-agent workflows, and enjoy some cold drinks on us. 🍻🤖 Spots are limited so make sure to RSVP soon at link in comments. 6–9pm Hope to see you there. cc Yang, Tim, Zero Prime Ventures

All AI problems eventually become search problems, and all search problems ultimately become ranking problems.   This insight from David Karam suggests a fundamental reframing of where real bottlenecks live in your AI/ML systems.   When model quality plateaus, most teams look to obvious solutions: scale up with bigger models, more data, more compute. David argues this misses the mark. The real constraint might exist in your ranking function: are you optimizing for the right things? Are your feedback loops meaningful? And most critically, have you actually seen enough of your domain to represent it well?   David spent a decade at Google architecting massive-scale search & AI systems and these days he’s building modular scoring infra at pi-labs.ai…. safe to say he knows search and AI. His full argument:   https://lnkd.in/gYzzz6TZ

All AI problems eventually become search problems, and all search problems ultimately become ranking problems.   This insight from David Karam suggests a fundamental reframing of where real bottlenecks live in your AI/ML systems.   When model quality plateaus, most teams look to obvious solutions: scale up with bigger models, more data, more compute. David argues this misses the mark. The real constraint might exist in your ranking function: are you optimizing for the right things? Are your feedback loops meaningful? And most critically, have you actually seen enough of your domain to represent it well?   David spent a decade at Google architecting massive-scale search & AI systems and these days he’s building modular scoring infra at pi-labs.ai…. safe to say he knows search and AI. His full argument:   https://lnkd.in/gYzzz6TZ

The db was publicly accessible. The "hack" involved downloading .jpg files from an exposed URL endpoint. This wasn’t a solo dev shipping on the weekend, but rather a dev team behind the #1 app in the App Store.   The Tea app “breach” represents the latest in a growing catalog of security failures caused by fundamental oversights in data and infra management. And these incidents are appearing more frequently now. In March, a solo founder went viral after racking up massive API costs - he had inadvertently pushed his .env file to a public repo while building in public on X.    Codegen feels like the invention of the high-performance engine and assembly line simultaneously. We suddenly have fast cars everywhere on the road, but lack seatbelts, airbags or speed limits. Ship faster, iterate faster, learn faster and crash sometimes.   The assembly line enables more individuals than ever to access dev capabilities, yet many are stepping into the driver's seat for the first time. Accidents are inevitable. Who manufactures the seatbelts? The guardrails? What standard safety & security protocols should be enforced for different types of applications?   These cases will push foundation model providers toward enhanced security in codegen. But there’s also space for new companies to audit applications and improve security protocols. Seatbelts represent the essential infra we need in the current moment. DM me or Yang if that’s you, we’d love to talk to you.

🎙️ New episode drop for the Zero Prime podcast! This week, we're featuring an interview with Dwarak Rajagopal – VP of AI Engineering and AI Research at Snowflake. In this episode, Dwarak traces his career from optimizing processors at AMD, to scaling AI at Google - and now transforming how enterprises actually use and drive VALUE with artificial intelligence at Snowflake. 🔥 Listen in for Dwarak’s hot takes on: - Why text-to-SQL isn't dead (plot twist!) - Query logs as secret weapons  - Efficiency being AI's next battleground No fluff here – just real insights on making AI work in the real world. Don’t miss out! https://lnkd.in/g862rNY7

Your calendar invites are color-coded. You set up automations in Notion for fun. And you’ve said the words “Let’s templatize that” at least once this week. If that’s you, Zero Prime Ventures needs a Director of Ops. Apply here: https://lnkd.in/gc5r4mDj

Over 100 million people will use ChatGPT today. It remains the most visible example of a foundation model, trained across a vast corpus encompassing nearly the entire digital universe. But the strength of foundation models is also their limitation. They are generalized systems designed to respond to almost any prompt. That generality is powerful but often inefficient, especially in enterprise contexts where latency, cost and domain specificity matter more than breadth. Why deploy a 70-billion parameter model fluent in 16th-century British history if all you need is a precise answer to a structured query? At Data Council one of the most salient themes was the growing interest in small, domain-specific models. Lightweight architectures tuned for specific tasks are not only more performant in context, but also dramatically more efficient. More and more, builders are asking: “What’s the smallest model that can do the job well?” More on this from our Data Council ‘25 keynote with George Mathew and Naveen Rao: https://lnkd.in/gkEsU-Cr