Directed Acyclic Graphs (DAGs) are the foundation of most orchestration frameworks. But what happens when you allow an LLM to act as the router? Acyclic graphs now become cyclic, which means you have to design for the challenges resulting from all this extra power. We'll cover the ins and outs of agentic applications and how to best use them in your work as a data practitioner or developer building today.

➡️ Follow Us LinkedIn: https://www.linkedin.com/company/small-data-sf/ X/Twitter : https://twitter.com/smalldatasf Website: https://www.smalldatasf.com/

Discover LangChain, the open-source framework for building powerful agentic systems. Learn how to augment LLMs with your private data, moving beyond their training cutoffs. We'll break down how LangChain uses "chains," which are essentially Directed Acyclic Graphs (DAGs) similar to data pipelines you might recognize from dbt. This structure is perfect for common patterns like Retrieval Augmented Generation (RAG), where you orchestrate steps to fetch context from a vector database and feed it to an LLM to generate an informed response, much like preparing data for analysis.

Dive into the world of AI agents, where the LLM itself determines the application's control flow. Unlike a predefined DAG, this allows for dynamic, cyclic graphs where an agent can iterate and improve its response based on previous attempts. We'll explore the core challenges in building reliable agents: effective planning and reflection, managing shared memory across multiple agents in a cognitive architecture, and ensuring reliability against task ambiguity. Understand the critical trade-offs between the dependability of static chains and the flexibility of dynamic LLM agents.

Introducing LangGraph, a framework designed to solve the agent reliability problem by balancing agent control with agency. Through a live demo in LangGraph Studio, see how to build complex AI applications using a cyclic graph. We'll demonstrate how a router agent can delegate tasks, execute a research plan with multiple steps, and use cycles to iterate on a problem. You'll also see how human-in-the-loop intervention can steer the agent for improved performance, a critical feature for building robust and observable agentic systems.

Explore some of the most exciting AI agents in production today. See how Roblox uses an AI assistant to generate virtual worlds from a prompt, how TripAdvisor’s agent acts as a personal travel concierge to create custom itineraries, and how Replit’s coding agent automates code generation and pull requests. These real-world examples showcase the practical power of moving from simple DAGs to dynamic, cyclic graphs for solving complex, agentic problems.

Over the last decade, Big Data was everywhere. Let's set the record straight on what is and isn't Big Data. We have been consumed by a conversation about data volumes when we should focus more on the immediate task at hand: Simplifying our work.

Some of us may have Big Data, but our quest to derive insights from it is measured in small slices of work that fit on your laptop or in your hand. Easy data is here— let's make the most of it.

📓 Resources Big Data is Dead: https://motherduck.com/blog/big-data-is-dead/ Small Data Manifesto: https://motherduck.com/blog/small-data-manifesto/ Small Data SF: https://www.smalldatasf.com/

➡️ Follow Us LinkedIn: https://linkedin.com/company/motherduck X/Twitter : https://twitter.com/motherduck Blog: https://motherduck.com/blog/

Explore the "Small Data" movement, a counter-narrative to the prevailing big data conference hype. This talk challenges the assumption that data scale is the most important feature of every workload, defining big data as any dataset too large for a single machine. We'll unpack why this distinction is crucial for modern data engineering and analytics, setting the stage for a new perspective on data architecture.

Delve into the history of big data systems, starting with the non-linear hardware costs that plagued early data practitioners. Discover how Google's foundational papers on GFS, MapReduce, and Bigtable led to the creation of Hadoop, fundamentally changing how we scale data processing. We'll break down the "big data tax"—the inherent latency and system complexity overhead required for distributed systems to function, a critical concept for anyone evaluating data platforms.

Learn about the architectural cornerstone of the modern cloud data warehouse: the separation of storage and compute. This design, popularized by systems like Snowflake and Google BigQuery, allows storage to scale almost infinitely while compute resources are provisioned on-demand. Understand how this model paved the way for massive data lakes but also introduced new complexities and cost considerations that are often overlooked.

We examine the cracks appearing in the big data paradigm, especially for OLAP workloads. While systems like Snowflake are still dominant, the rise of powerful alternatives like DuckDB signals a shift. We reveal the hidden costs of big data analytics, exemplified by a petabyte-scale query costing nearly $6,000, and argue that for most use cases, it's too expensive to run computations over massive datasets.

The key to efficient data processing isn't your total data size, but the size of your "hot data" or working set. This talk argues that the revenge of the single node is here, as modern hardware can often handle the actual data queried without the overhead of the big data tax. This is a crucial optimization technique for reducing cost and improving performance in any data warehouse.

Discover the core principles for designing systems in a post-big data world. We'll show that since only 1 in 500 users run true big data queries, prioritizing simplicity over premature scaling is key. For low latency, process data close to the user with tools like DuckDB and SQLite. This local-first approach offers a compelling alternative to cloud-centric models, enabling faster, more cost-effective, and innovative data architectures.

It's finally possible to bring the awesome power of Large Language Models (LLMs) to your laptop. This talk will explore how to run and leverage small, openly available LLMs to power common tasks involving data, including selecting the right models, practical use cases for running small models, and best practices for deploying small models effectively alongside databases.

Bio: Jeffrey Morgan is the founder of Ollama, an open-source tool to get up and run large language models. Prior to founding Ollama, Jeffrey founded Kitematic, which was acquired by Docker and evolved into Docker Desktop. He has previously worked at companies including Docker, Twitter, and Google.

➡️ Follow Us LinkedIn: https://www.linkedin.com/company/small-data-sf/ X/Twitter : https://twitter.com/smalldatasf Website: https://www.smalldatasf.com/

Discover how to run large language models (LLMs) locally using Ollama, the easiest way to get started with small AI models on your Mac, Windows, or Linux machine. Unlike massive cloud-based systems, small open source models are only a few gigabytes, allowing them to run incredibly fast on consumer hardware without network latency. This video explains why these local LLMs are not just scaled-down versions of larger models but powerful tools for developers, offering significant advantages in speed, data privacy, and cost-effectiveness by eliminating hidden cloud provider fees and risks.

Learn the most common use case for small models: combining them with your existing factual data to prevent hallucinations. We dive into retrieval augmented generation (RAG), a powerful technique where you augment a model's prompt with information from a local data source. See a practical demo of how to build a vector store from simple text files and connect it to a model like Gemma 2B, enabling you to query your own data using natural language for fast, accurate, and context-aware responses.

Explore the next frontier of local AI with small agents and tool calling, a new feature that empowers models to interact with external tools. This guide demonstrates how an LLM can autonomously decide to query a DuckDB database, write the correct SQL, and use the retrieved data to answer your questions. This advanced tutorial shows you how to connect small models directly to your data engineering workflows, moving beyond simple chat to create intelligent, data-driven applications.

Get started with practical applications for small models today, from building internal help desks to streamlining engineering tasks like code review. This video highlights how small and large models can work together effectively and shows that open source models are rapidly catching up to their cloud-scale counterparts. It's never been a better time for developers and data analysts to harness the power of local AI.

This is a talk about how we thought we had Big Data, and we built everything planning for Big Data, but then it turns out we didn't have Big Data, and while that's nice and fun and seems more chill, it's actually ruining everything, and I am here asking you to please help us figure out what we are supposed to do now.

📓 Resources Big Data is Dead: https://motherduck.com/blog/big-data-... Small Data Manifesto: https://motherduck.com/blog/small-dat... Is Excel Immortal?: https://benn.substack.com/p/is-excel-immortal Small Data SF: https://www.smalldatasf.com/

➡️ Follow Us LinkedIn: / motherduck
X/Twitter : / motherduck
Blog: https://motherduck.com/blog/

Mode founder David Wheeler challenges the data industry's obsession with "big data," arguing that most companies are actually working with "small data," and our tools are failing us. This talk deconstructs the common sales narrative for BI tools, exposing why the promise of finding game-changing insights through data exploration often falls flat. If you've ever built dashboards nobody uses or wondered why your analytics platform doesn't deliver on its promises, this is a must-watch reality check on the modern data stack.

We explore the standard BI demo, where an analyst uncovers a critical insight by drilling into event data. This story sells tools like Tableau and Power BI, but it rarely reflects reality, leading to a "revolving door of BI" as companies swap tools every few years. Discover why the narrative of the intrepid analyst finding a needle in the haystack only works in movies and how this disconnect creates a cycle of failed data initiatives and unused "trashboards."

The presentation traces our belief that "data is the new oil" back to the early 2010s, with examples from Target's predictive analytics and Facebook's growth hacking. However, these successes were built on truly massive datasets. For most businesses, analyzing small data results in noisy charts that offer vague "directional vibes" rather than clear, actionable insights. We contrast the promise of big data analytics with the practical challenges of small data interpretation.

Finally, learn actionable strategies for extracting real value from the data you actually have. We argue that BI tools should shift focus from data exploration to data interpretation, helping users understand what their charts actually mean. Learn why "doing things that don't scale," like manually analyzing individual customer journeys, can be more effective than complex models for small datasets. This talk offers a new perspective for data scientists, analysts, and developers looking for better data analysis techniques beyond the big data hype.

As a Head of Data or a one-person data team, keeping the lights on for the business while running all things data-related as efficiently as possible is no small feat. This talk will focus on tactics and strategies to manage within and around constraints, including monetary costs, time and resources, and data volumes.

📓 Resources Big Data is Dead: https://motherduck.com/blog/big-data-... Small Data Manifesto: https://motherduck.com/blog/small-dat... Why Small Data?: https://benn.substack.com/p/is-excel-... Small Data SF: https://www.smalldatasf.com/

➡️ Follow Us LinkedIn: / motherduck
X/Twitter : / motherduck
Blog: https://motherduck.com/blog/

Learn how your data team can drive innovation and maximize ROI by embracing constraints, drawing inspiration from SpaceX's revolutionary cost-effective approach. This video challenges the "abundance mindset" prevalent in the modern data stack, where easily scalable cloud data warehouses and a surplus of tools often lead to unmanageable data models and underutilized dashboards. We explore a focused data strategy for extracting maximum value from small data, shifting the paradigm from "more data" to more impact.

To maximize value, data teams must move beyond being order-takers and practice strategic stakeholder management. Discover how to use frameworks like the stakeholder engagement matrix to prioritize high-impact business leaders and align your work with core business goals. This involves speaking the language of business growth models, not technical jargon about data pipelines or orchestration, ensuring your data engineering efforts resonate with key decision-makers and directly contribute to revenue-generating activities.

Embracing constraints is key to innovation and effective data project management. We introduce the Iron Triangle—a fundamental engineering concept balancing scope, cost, and time—as a powerful tool for planning data projects and having transparent conversations with the business. By treating constraints not as limitations but as opportunities, data engineers and analysts can deliver higher-quality data products without succumbing to scope creep or uncontrolled costs.

A critical component of this strategy is understanding the Total Cost of Ownership (TCO), which goes far beyond initial compute costs to include ongoing maintenance, downtime, and the risk of vendor pricing changes. Learn how modern, efficient tools like DuckDB and MotherDuck are designed for cost containment from the ground up, enabling teams to build scalable, cost-effective data platforms. By making the true cost of data requests visible, you can foster accountability and make smarter architectural choices. Ultimately, this guide provides a blueprint for resisting data stack bloat and turning cost and constraints into your greatest assets for innovation.

By introducing a range of AI-enhanced products that amplify creativity and interactivity across our platforms, Buzzfeed has been able to connect with the largest global audience of young people online to cement its role as the defining digital media company of the AI era. Notably, some of Buzzfeed's most successful tools and content experiences thrive on the power of small, focused datasets. Still wondering how Shrek fits into the picture? You'll have to watch!

Video from: https://smalldatasf.com/

📓 Resources Big Data is Dead: https://motherduck.com/blog/big-data-... Small Data Manifesto: https://motherduck.com/blog/small-dat... Why Small Data?: https://benn.substack.com/p/is-excel-... Small Data SF: https://www.smalldatasf.com/

➡️ Follow Us LinkedIn: / motherduck
X/Twitter : / motherduck
Bluesky: motherduck.com Blog: https://motherduck.com/blog/

Discover how BuzzFeed's Data team, led by Gilad Cohen, harnesses AI for creative purposes, leveraging large language models (LLMs) and generative image capabilities to enhance content creation. This video explores how machine learning teams build tools to create new interactive media experiences, focusing on augmenting creative workflows rather than replacing jobs, allowing readers to participate more deeply in the content they consume.

We dive into the core data science problem of understanding what a piece of content is about, a crucial step for improving content recommendation systems. Learn why traditional methods fall short and how the team is constantly seeking smaller, faster, and more performant models. This exploration covers the evolution from earlier architectures like DistilBERT to modern, more efficient approaches for better content representation, clustering, and user personalization.

A key technique explored is the use of text embeddings, which are dense, low-dimensional vector representations of data. This video provides an accessible explanation of embeddings as a form of compressed knowledge, showing how BuzzFeed creates a unique vector for each article. This allows for simple vector math to find semantically similar content, forming a foundational infrastructure for powerful ranking and recommender systems.

Explore how BuzzFeed leverages generative image capabilities to create new interactive formats. The journey began with Midjourney experiments and evolved to building custom tools by fine-tuning a Stable Diffusion XL model using LORA (Low-Rank Approximation). This advanced technique provides greater control over image output, enabling the rapid creation of viral AI generators that respond to trending topics and allow for massive user engagement.

Finally, see a practical application of machine learning for content optimization. BuzzFeed uses its vast historical dataset from Bayesian A/B testing to train a model that predicts headline performance. By generating multiple headline candidates with an LLM like Claude and running them through this predictive model, they can identify the winning headline. This showcases how to use unique, in-house data to build powerful tools that improve click-through rates and drive engagement, pointing to a significant transformation in how media is created and consumed.