Firehose

This video demonstrates how to use the Data Agent Kit to connect AI coding agents to Google Cloud data sources. It provides a practical walkthrough of setting up BigQuery and Cloud SQL via MCP servers, and shows how to use "Agent Skills" to guide an LLM through data analysis, pipeline building with dbt, and troubleshooting. - Data Agent Kit — The core toolkit for connecting agents to Google Cloud data. - VS Code Extension — The primary interface for managing these connections in VS Code-based editors. - Claude Code Plugin — The specific integration for using the kit within Claude Code. - Antigravity IDE — The IDE used for the primary demonstrations in the video.

This talk is a series of case studies on the realities of building data infrastructure for AI, focusing on the gap between development environments and production-grade reliability. It highlights how scaling from millions to billions of requests requires constant architectural iteration rather than one-off engineering breakthroughs.

This presentation argues that for high-frequency AI agent tasks, "Context-as-a-Service" (CaaS) vendors often become cost-prohibitive compared to building a custom scraping pipeline. The speaker shares an internal test showing that once you exceed roughly 15,000 queries, the upfront cost of owning your data infrastructure pays for itself by eliminating the recurring costs of repeated API calls for identical, decaying web data.

This talk outlines the shift from training agents in controlled, game-like environments to the messy reality of browser-based automation. The speaker argues that "flight school" simulations—which explicitly include layout shifts, stale sessions, and adversarial UI—are necessary to teach agents how to recover from errors rather than just attempting to brute-force tasks.

This video explains how current computer-use benchmarks are easily gamed by "replay agents"—scripts that simply record and repeat successful trajectories—and introduces PRISM principles to build more robust, non-deterministic evaluation environments like DIGIWORLD.

This is a fireside chat between Travis Kalanick and Ben Horowitz that serves as a high-level pitch for Kalanick’s current venture, Atoms. Kalanick frames his work as "industrial AI"—applying software, robotics, and sensors to automate physical sectors like food production and mining—and reflects on how his experience scaling Uber informs this new focus on digitizing the physical world.

Batra argues that the "agentic web" will rely on computer-use models rather than APIs because the vast majority of the web is built for human eyes, not machine consumption. He demonstrates that modern websites often render state dynamically—making pixels, rather than HTML or JSON, the only reliable source of truth for an agent.

This podcast episode features a panel discussion on three industry topics: IBM’s partnership with Together AI to build large-scale inference clusters, the release of Meta’s Muse Glimmer (a 30B parameter model), and the implications of OpenAI’s upcoming Astra model regarding cybersecurity capabilities.

This is a casual, conversational livestream where the hosts discuss design industry news, react to current Twitter discourse, and highlight portfolios from their community. It functions primarily as a weekly variety show for designers rather than a technical tutorial or deep-dive resource.

A Google Cloud engineer outlines a reference architecture for building AI-integrated mobile apps using a "serverless" stack. The video maps specific Google services—like Firebase, Cloud Run, and Model Garden—to common enterprise requirements like authentication, API security, and model flexibility.

This video is a promotional announcement for "All Things Agentic," a Google Cloud hackathon focused on building autonomous AI agents using Gemini models. It outlines the three competition tracks, submission requirements, and prize structure for participants.

The presenters use a "periodic table" metaphor to map out the standard components of data engineering and AI pipelines. They walk through a document Q&A use case to demonstrate how traditional data science tasks (ETL, cleansing, governance) and modern AI techniques (embeddings, RAG, guardrails) function as a unified, iterative system.

This is a conversational interview with Gary Tan, focusing on the evolution of Silicon Valley culture and the importance of founder psychology. The discussion centers on the value of earnestness, ignoring market trends in favor of first-principles thinking, and the "cult-like" nature of building transformative companies.

A talk from a LangChain researcher on the necessity of "trace mining"—using LLMs to analyze the logs of your own autonomous agents to identify failure points, distill successful behaviors into smaller models, and automate the creation of evaluation datasets. It serves as a conceptual pitch for LangSmith, framing observability as the prerequisite for any serious attempt at agentic continuous learning.

This talk provides a comparative history of how ChatGPT and Claude have evolved their "memory" systems—specifically the shift from simple fact-lists to asynchronous, LLM-generated user profiles. The speaker argues that these systems are not standardized RAG implementations, but rather bespoke, compute-constrained architectures that teams must build in-house to match their specific product needs.

This is a candid talk from the CTO of Raindrop arguing that most standard AI evaluation practices are "theater" that fail to account for the unpredictable nature of real-world agents. He advocates for prioritizing "floor-raising" (preventing catastrophic failures) over chasing benchmarks and points to his company's resources, Workshop and howtoeval.com, as practical alternatives for production monitoring.

This is a technical presentation on building continual learning systems for enterprise AI agents. The speaker categorizes distillation methods into a 2x2 grid based on whether the data and "hints" (guidance) are sourced offline or online, focusing on how to improve model performance without requiring a "golden" ground-truth dataset.

This talk outlines a workflow for turning unstructured voice-dictated notes into a searchable, interconnected knowledge base using LLMs. The presenter demonstrates how to use local transcription tools to capture raw thoughts, then employs agentic "enrichment" scripts to automatically tag, link, and organize these files into a personal wiki. - Handy — open-source local voice dictation tool - Voice Inc — voice-to-text app with hotkey support - Exa — search engine recommended for agentic web research

This is a talk by an AI researcher arguing that the current "narrow path" of frontier AI development—dominated by a few labs and massive compute—is shifting toward decentralized, domain-specific model training. The speaker introduces their project, Auto Scientist, which automates the model training loop by co-optimizing data and architecture to allow for more accessible, efficient, and specialized AI development.

This is a conceptual talk by a professor and CEO arguing that current AI agents are failing at general tasks because they possess "intelligence" (reasoning through unfamiliar context) but lack "expertise" (accumulated, situated competence). The speaker posits that we are stuck in a modern Moravec’s paradox, where agents need continuous learning to compress idiosyncratic "micro-world" experiences into reusable structures rather than just brute-forcing problems with larger models.

This talk frames the current limitation of AI as a "data axis" problem: models excel at public information but fail to gain depth on private, non-public corpora because training from scratch is impractical. The speaker argues that since the data budget is effectively fixed for personal or enterprise use, the only remaining path to improvement is "scaling compute on context"—a search for ways to apply deep learning's scaling laws to private data without hitting a synthetic data wall.

This talk outlines a framework for building "memory harnesses" to help local LLMs manage long-horizon tasks without context rot. The speaker details a control loop—consisting of a core, recall, and archival block—and demonstrates that a ranked decision ledger consistently outperforms standard RAG for retrieving information outside the active context window.

This is a technical talk proposing "On-Policy Self-Distillation" (OPSD) as a more efficient alternative to GRPO for training LLMs. The speaker argues that by using "privileged information" as a hint in the prompt, you can train models on-policy without the need for massive parallel rollouts or sequence-level rewards.

A panel discussion from the *Security Intelligence* podcast breaking down the 2026 OWASP Top 10 for LLM applications. The experts argue that the list should be used as a framework for tabletop exercises rather than a compliance checklist, specifically emphasizing the shift from simple prompt injection concerns to the operational risks of "excessive agency."

This is a technical presentation from Anthropic staff explaining the evolution of their agentic infrastructure, moving from the basic Messages API to the Claude Agent SDK and finally to their managed agent services. The speakers focus on the engineering trade-offs of building "harnesses" that manage context and tool execution, specifically highlighting how rigid, hand-rolled infrastructure often becomes technical debt as model capabilities improve.

This is a conversation between two designers about using HTML artifacts and custom agent skills to build interactive, visual feedback loops for their work. They demonstrate how to chain AI prompts to generate side-by-side comparisons of UI and motion changes, effectively turning Claude into a specialized design-review environment.

This video provides a conceptual overview of "context engineering," framing it as the shift from simply writing better prompts to intentionally curating, structuring, and managing the information environment provided to an AI agent. It explains why dumping data into large context windows often leads to performance degradation and outlines best practices like relevance, compression, and lifecycle management.

This is a high-level management talk on integrating AI agents into engineering teams, arguing that success depends on treating agent adoption as a cultural and organizational challenge rather than just a technical one. The speaker outlines a strategy focused on shared codebase standards, automated "self-healing" workflows, and winning over skeptical developers to avoid the productivity gaps caused by inconsistent tooling.

This talk explains how to decouple RL rollout workers from a central training cluster by replacing full 500GB checkpoint transfers with 500MB "lossless patches." The speaker, Nan Jiang, demonstrates that because Adam updates are tiny and rollout engines use lower-precision formats (like BF16 or FP8), over 99% of weights remain unchanged between steps. His implementation, Stitch, leverages this to allow rollout fleets to run across distributed, elastic GPU capacity rather than being tethered to a single high-bandwidth cluster.

This is a technical deep dive into the architecture of the Continue (formerly Codeium/CodeEx) agent harness. The speaker explains how their open-source Rust-based framework manages context, handles tool discovery, and implements sandboxed file system and browser interactions.