Firehose

This is a technical walkthrough of how to implement the Gemini 3.5 Transcribe model using the Google GenAI SDK. The video covers the two primary integration patterns—synchronous batch processing and real-time streaming via WebSockets—while detailing how to configure features like diarization, word-level timestamps, and custom vocabulary. * Code tutorial — Implementation examples for batch and streaming * Code script — Live API integration pattern * Docs — Official API reference and configuration guide

This is a casual, community-focused podcast episode where the hosts highlight independent designers and experimental UI tools. The show functions as a "group chat" for design-engineers, featuring a segment on survey.dive.radio to collect industry data and a showcase of projects like Popmelt, a tool for creating interactive UI variables.

This is a functional overview of Gemini 3.5 Transcribe, demonstrating how to use its batch and live APIs for domain-specific transcription. The video highlights using custom vocabulary for medical dictation and chaining the output into other Gemini models for sentiment analysis, with documentation and a code tutorial provided for implementation.

This is a recording of the main stage presentations from the AI Engineer Paris 2026 conference. It features technical talks on production-scale AI, covering topics like inference optimization, voice streaming, agentic workflows, and model customization from speakers at companies including Black Forest Labs, Hugging Face, Stripe, and Google DeepMind.

A practical walkthrough of how to configure hard spending limits in the Google Cloud Console to automatically disable services like the Gemini API and Vertex AI when costs exceed a set threshold.

This presentation is a high-level overview of Skydio’s approach to autonomous drone fleets, focusing on how they transition from manual piloting to "drones as infrastructure." The speaker demonstrates how they use edge-to-cloud orchestration and vision-language models to manage multiple drones simultaneously for utility inspection and public safety.

This is a technical presentation from Dyna Robotics CTO Jason Ma explaining their approach to moving beyond 90% success rates in robotic manipulation. He details how they use a reward-model-based active learning loop to achieve high reliability, specifically demonstrating their Dyna-1 model's performance in long-duration tasks like napkin and T-shirt folding.

This is a pitch for Bright Data, a service that indexes web video to help developers extract specific action-based clips for training robotics models. The speaker argues that because existing models discard the vast majority of raw video data during pre-processing, searching for and trimming clips via their API before collection saves significant compute and storage costs.

This talk is a two-part lecture: the first half is a historical overview of AI at Stanford, and the second half is a technical pitch for Moonlake AI. Manning argues that current generative video models lack the causal semantics needed for robotics, and he outlines his company's approach to building action-conditioned world models that use code-based simulation to bridge the sim-to-real gap.

This video explains why you should use specialized backend APIs to process raw data before feeding it to an LLM, rather than dumping massive datasets directly into a context window. It uses the US Open's "Serve Quality" tracking as a case study to demonstrate how an AI agent can effectively use tool-calling to retrieve pre-processed insights.

This presentation outlines Perceptron's research approach to building embodied foundation models, focusing on two specific technical hurdles: creating self-supervised objectives that prioritize task-relevant visual tokens and using data-sparse mixture-of-experts to manage context bloat. The speaker, Armen Aghajanyan, argues that these architectural choices allow for more efficient scaling by substituting expensive teleoperation data with large-scale video pretraining.

This video explains how to implement tool calling for real-time voice agents using the Gemini Live API and the Agent Development Kit (ADK). It focuses on maintaining conversational flow by ensuring tools execute synchronously and using a `before_tool_callback` to intercept and validate agent actions before they run.

This is a talk about using AI agents to solve the labor bottleneck in rotational livestock grazing. The speaker outlines a vision for an "Open Pasture" system that uses computer vision and LLMs to automate paddock management, while calling for more developers to build for physical-world infrastructure rather than just software-for-software. - Open Pasture — The speaker's project to build an open-source knowledge base and management system for regenerative grazing. - Firecrawl — The web-scraping tool the speaker uses to ingest research and farmer data for his models.

This is a technical critique of why current frontier models fail at spatial reasoning, presented by a former Google Brain lead. The speaker argues that models rely on pattern matching rather than visual thinking, and outlines his company's (Elorian) approach to building native visual chain-of-thought models for robotics and mechanical design.

This talk outlines the technical challenges of building agentic workflows that rely on document data, specifically focusing on why standard RAG pipelines fail when inputs are messy or complex. The speaker argues for a hybrid approach that uses lightweight computer vision for layout detection, structured parsing for tables, and agentic harnesses to handle multi-step reasoning, rather than relying solely on frontier models. Learn more at Reducto.

This is a technical breakdown of how Meta uses a multi-agent system of specialized, small-scale VLMs to detect content anomalies and unoriginal uploads at scale. The talk details the architecture—a perceiver, a reviewer, and a retriever—and explains the practical engineering trade-offs required to run these models, including custom pretraining, DPO loops, and aggressive inference-time optimizations.

This talk outlines a workflow for automating the creation of specialized computer vision models by using Vision Language Models (VLMs) as automated labelers and judges. The speaker argues for training smaller, Apache 2.0-licensed detectors like RF-DETR instead of deploying heavy VLMs at runtime, demonstrating how to use coding agents to orchestrate the pipeline for just a few dollars.

Jerry Liu argues that because PDFs are designed for printing rather than data structure, they remain a major bottleneck for AI agents. He outlines a three-layer approach—parsing, semantic storage, and workflow automation—to turn these "document containers" into usable context, specifically highlighting ParseBench for model evaluation and LightParse as a high-performance, open-source parsing tool.

This episode is a career retrospective with Geoff Teehan, covering his transition from agency founder to design leadership at Facebook and Lightspark. The conversation focuses on his philosophy of navigating industry shifts, his approach to lean team building, and the accidental viral success of his side project, Dialed.

This is a promotional conversation introducing the Horowitz and Andreessen Academy, a project-based educational initiative backed by a16z. The founders discuss their philosophy of "learning by doing" and why they believe traditional university models are currently misaligned with the needs of young people who want to build in the AI era.

This video is a high-level conceptual argument that the true value of AI agents lies in the infrastructure upgrades they force upon organizations—such as better data hygiene, standardized APIs, and improved system interoperability—rather than the agents themselves. It serves as a strategic overview of why these "catalyst" improvements will provide long-term utility regardless of how agent technology evolves.

This workshop details how to move beyond simple prompt engineering by building "harnesses" that treat AI as a component in a larger, deterministic system. The speaker argues that prose-based instructions are insufficient to overcome a model's tendency toward "slop," and instead demonstrates nine techniques—such as adversarial sub-agents, deterministic linters, and external script execution—to force divergence and maintain design quality. - Impeccable — The speaker's open-source design skill and framework. - Paul Bakaus — The speaker's personal site and portfolio.

This video argues that organizations should stop using token consumption as a primary metric for AI success, as it encourages either wasteful "tokenmaxxing" or counterproductive cost-cutting. It advocates for Valuemaxxing, a shift toward measuring operational outcomes like deployment speed and rework reduction.

This talk explains FlashNorm, a technique to speed up transformer inference by folding RMS norm layers into projection weights and parallelizing matrix and vector operations. The speaker details the CUDA implementation challenges, specifically how an implicit stream join caused a race condition that led to model output repetition.

This talk is a technical post-mortem on two silent, state-related bugs the speakers encountered while running Jamba on vLLM. It details how they used log-probability forensics and custom request-ID tracing to identify scheduler race conditions and integer overflows in the Mamba state cache that caused the model to output gibberish without crashing.

This is a technical pitch from FriendliAI explaining why standard inference stacks fail for agentic workflows. The speaker argues that because agents reuse large context prefixes across long-running tasks, inference providers need to shift from optimizing individual request latency to optimizing end-to-end task completion through techniques like prefix caching, hierarchical KV cache management, and cache-aware routing.

This talk outlines the operational bottlenecks of serving a "fleet" of small, task-specific models rather than a single large one. The speaker argues that traditional top-down routing and manual infrastructure tuning stall GPU utilization, and proposes an inverted, queue-based architecture to automate batching and model adaptation.

This talk serves as a post-publication update to the author's book, *Inference Engineering*, focusing on data center optimizations that have emerged since early 2026. It categorizes recent developments into three pillars: * Quantization — A breakdown of why techniques like TurboQuant, while effective for memory-constrained local setups, often underperform in data centers compared to standard weight quantization. * KV Compaction — An overview of moving from simple caching to learned bottlenecks that cross-attend query vectors against the cache to emit compact keys and values. * Speculative Decoding — A review of the shift toward training custom draft models, the introduction of diffusion-based drafting (DFlash), and the performance gains from continuously retraining speculators on live production prompts.

This is a high-level technical overview of how CoreWeave manages inference workloads by balancing serverless and dedicated capacity. The speaker explains how they optimize for KV cache locality and use heterogeneous GPU scheduling to treat inference like a game of Tetris, prioritizing cache hits to avoid redundant prefill computations.

This talk explains why common LLM benchmarking tools often produce misleading results due to Python's global interpreter lock, client-side latency inflation, and inconsistent data sampling. The speakers introduce Inference Perf, a CNCF project designed to provide high-fidelity, multi-process load generation and declarative workload definitions to ensure reproducible performance metrics.