Firehose

This panel discussion features representatives from NVIDIA, Prime Intellect, and Arcee AI arguing that open-source models provide superior reliability and control for enterprise use cases compared to closed APIs. The speakers focus on the strategic importance of owning model weights and training data to build internal "data flywheels" and optimize performance for specific, narrow tasks.

This panel discussion explores the practical realities of model compression, moving past the hype to explain how techniques like quantization actually function on modern hardware. The speakers discuss why layer-wise precision matters, the limitations of current benchmarks, and why they prioritize KL divergence over standard accuracy scores when evaluating quantized models. - Daniel Han (Unsloth) — Discusses the practical application of mixed-precision quantization to run massive models on consumer hardware. - Asma Beevi (NVIDIA) — Covers the technical implementation of NVFP4, a 4-bit float format designed to minimize accuracy loss. - Merve Noyan (Hugging Face) — Focuses on the democratization of AI through libraries like `bitsandbytes` and the ability to fine-tune models on limited VRAM. - Parth Sareen (Ollama) — Provides the perspective of running compressed models locally and the trade-offs between latency and model intelligence.

This panel discussion explores the practical mechanics of model routing, moving beyond simple task-based delegation to discuss how to maintain context and cost-efficiency in agentic workflows. The speakers from Cognition and OpenRouter argue that effective routing requires keeping a frontier model for high-level planning while delegating execution to smaller, specialized models.

This video outlines a workflow for securing AI-assisted coding by treating agents like junior developers who require strict boundaries. It emphasizes using test-driven development (TDD) for small code batches, enforcing security rules via context files, and using external tools for verification rather than relying on the agent's own output.

This is a technical conversation about the role of vLLM as the underlying infrastructure for serving open-weight models. The guests discuss why companies are moving away from proprietary APIs toward self-hosted inference to gain control over performance, costs, and guardrails, and how this shift has turned open-source inference engines into critical enterprise software.

A clear, high-level primer on "model collapse," explaining how training AI on its own output leads to the degradation of rare information and the amplification of bias. It frames the phenomenon as a long-term engineering challenge rather than an immediate crisis, while outlining preventative strategies like data provenance and RAG.

Kenton Varda argues that modern cloud architecture prevents users from customizing software, proposing a new model where AI-generated features run as isolated "gadgets" rather than monolithic app updates. He demonstrates this using workerd to sandbox user-coded extensions, effectively treating apps like office documents that can be modified on the fly without breaking the core platform.

This is a founder-led pitch for Mariana Minerals, a company attempting to modernize domestic copper mining by replacing legacy manual processes with a proprietary software stack. It frames the venture as a national security play, focusing on their acquisition of a Utah-based operation to serve as a sandbox for their automation tools.

This interview with Nate Parrott focuses on his workflow within Claude Design, demonstrating how he uses the tool to generate rapid design variations and custom, task-specific interfaces. The conversation centers on using AI as a "junior designer" to handle high-volume iteration, rather than expecting perfect output on the first try.

This video explains the concept of Large Database Models (LDMs), a technique for performing semantic similarity searches directly within relational databases by treating rows as "sentences" and values as tokens. It outlines the process of binning numeric data and generating embeddings to allow SQL users to query for patterns without moving data to external AI platforms.

A wide-ranging conversation between Gergely Orosz and Simon Eskildsen that covers Simon's early career at Shopify, his transition to building Turbopuffer, and his pragmatic approach to engineering and startup infrastructure. It’s a candid look at the technical trade-offs of building a database on S3 and the realities of running a lean, remote-first company.

This is a high-level industry discussion about the launch of the Tokenomics Foundation, focusing on the need for standardized metrics to manage AI infrastructure costs. The conversation frames "tokenomics" as a bridge between FinOps and AI engineering, emphasizing that organizations must move beyond simple consumption tracking toward a framework that balances production costs, model efficiency, and measurable business value.

This video provides a conceptual framework for how developer roles are shifting from writing deterministic code to orchestrating probabilistic AI systems. It defines a spectrum of AI-assisted development—ranging from "vibe coding" to "agentic engineering"—to argue that human oversight remains critical as autonomy increases.

This is a technical overview of MCP Apps, a standard that allows AI agents to render interactive web components instead of plain text responses. The presenters explain how this protocol enables developers to ship branded, functional UI widgets that work consistently across any host supporting the specification.

This talk explains why the Model Context Protocol (MCP) tasks specification is difficult to implement, focusing on the distributed systems challenges of maintaining state across network failures and process crashes. The presenter demonstrates a durable invoice-processing workflow using FastMCP to show how servers can handle long-running tasks that require human-in-the-loop signals.

This talk provides a critique of current AI evaluation practices, arguing that "benchmaxxing" occurs because labs prioritize high scores on flawed, easily gamed benchmarks over real-world utility. Heiner details specific failure modes—such as data contamination, reward hacking, and poorly designed verifiers—and advocates for a shift toward high-quality human evaluation and domain-specific testing.

This is a high-level conceptual overview of why AI agents hallucinate and how to mitigate those risks through data grounding, tool-based reasoning, and human-in-the-loop design. It functions as a primer for developers and managers on why "confidence" in LLMs is often a byproduct of their training rather than a measure of factual accuracy.

This talk argues that AI security research is currently held back by "benchmaxxing"—relying on flawed, LLM-based grading oracles—and proposes instead the use of reproducible, sandboxed reinforcement learning environments with deterministic graders. The speaker, David Brumley, demonstrates that by grading models on their ability to trigger specific, real-world vulnerabilities (like those found in the V8 engine), researchers can move beyond hallucinated successes to measure genuine exploitation capabilities.

This presentation argues that current "long-horizon" benchmarks are often noisy or misleading because they rely on arbitrary time thresholds rather than environment design. The speakers emphasize that meaningful evaluation requires building environments where early decisions have cascading effects, and verifying success based on final state outcomes rather than intermediate judge scores.

This talk argues that current AI development is trapped in an "assistance" paradigm because RLHF explicitly optimizes for human approval rather than objective task completion. The speaker contends that true automation requires moving away from preference-based training toward reinforcement learning with verifiable rewards, noting that current agentic tools remain tethered to the same human-in-the-loop constraints.

This talk argues that data curation is a more cost-effective way to improve model performance than simply buying more compute. The speaker outlines a "data refinery" approach—cleaning, curating, creating, and composing—to maximize signal per token, using DatologyAI research to show how smaller, better-curated datasets can outperform much larger ones.

This talk outlines a technical framework for "learning on the job," where models are continuously fine-tuned via reinforcement learning using an enterprise's existing production harness rather than a synthetic sandbox. The speaker, Raymond Feng, details the shift from simple Q&A loops to long-horizon task adaptation, while candidly addressing the risks of reward hacking and the difficulty of maintaining environment fidelity when moving away from controlled, replayable data.

This talk outlines the philosophy behind Bespoke Labs, arguing that the primary bottleneck in post-training isn't compute, but the systematic curation of data and reinforcement learning environments. The speaker details lessons learned from building OpenThoughts, emphasizing that high-quality reasoning models are built through iterative ablations of data mixtures rather than just scaling raw volume.

This talk is a high-level technical retrospective on reinforcement learning for language models, tracing the evolution from early experiments like Galactica to modern long-horizon agent design. The speakers argue that scaling to long-duration tasks requires shifting focus from simple context windows to better simulation environments and deliberate token allocation.

This presentation argues that current AI benchmarks are too narrow because they focus on code diffs rather than the messy reality of infrastructure engineering. The speakers introduce Emulated, a platform that simulates entire company environments—including network failures, resource provisioning, and distributed system constraints—to provide more realistic training data for autonomous agents.

This talk argues that the traditional "base model" trained on raw web text is becoming obsolete as reinforcement learning (RL) takes center stage. The speaker explains how modern training recipes now pull instruction-tuned and synthetic data into the pre-training phase to better prepare models for the reasoning tasks required by RL, using his team's work on the Trinity series as a case study for managing data mixes and expert load balancing.

This talk outlines LatchBio's approach to building AI agents for biology by treating experimental data as a verifiable substrate. The speaker argues that because biology lacks clear ground truth, progress depends on designing benchmarks that reward scientific reasoning over memorized knowledge, using spatial biology as a primary test case.

This podcast episode features the founders of Decagon discussing their shift from relying on frontier model APIs to a 90% open-source inference stack. They argue that fine-tuning smaller, task-specific models allows them to beat the latency and cost of general-purpose models while maintaining higher performance for enterprise workflows.

This is a casual, conversational livestream where the hosts discuss current design trends, share their favorite interaction designers, and highlight tools they are currently using in their workflows. It functions primarily as a community-focused show-and-tell rather than a structured tutorial. - Notion — mentioned as a partner and tool used in their kit - Lovable — mentioned as a partner and platform for UI development - Mobin — mentioned as a partner and design reference library - Raycast — mentioned in the context of community contributions - Glaze — mentioned as a design platform - Figma — mentioned for motion and design systems - Biscuit Camera — designer project mentioned during the show - Mia Calendar — designer project mentioned during the show

This podcast episode features a panel discussion centered on the 2026 Cost of a Data Breach Report, focusing on how AI is shifting the economics of cybersecurity. The experts discuss the growing asymmetry between the low cost of AI-driven attacks and the high cost of defensive guardrails, while also touching on recent industry news like the release of Claude Opus 5 and the acquisition of Co-Star by Midjourney.