The Video Signal

Weights & Biases

Fully Connected 2026 | Day 1 Full Replay

Published 2026-10-06 · Duration 1:52:20

Summary

The keynote detailed the evolution of AI from simple prompt-response models to complex, multi-step agentic systems. CoreWeave introduced CoreWeave Forge, a unified platform designed to manage the entire AI lifecycle—the 'AI Loop' (Run, Observe, Curate, Improve, Evaluate)—allowing engineers to move from initial agent concepts to production-grade systems. Hardware advancements were highlighted, including the NVIDIA Vera Rubin NVL72 and the new Vera CPU, designed to handle the unique computational demands of agentic workflows, long context, and high throughput.

Download summary

Key takeaways

  1. The AI Loop: Run, Observe, Curate, Improve, Evaluate 30:59

    The core methodology for building reliable AI agents involves a continuous cycle: Run (inference), Observe (monitoring agent behavior), Curate (selecting failure patterns as evidence), Improve (model refinement), and Evaluate (testing changes). This loop is critical for moving AI from the demo stage to production.

  2. CoreWeave Forge Unifies the AI Lifecycle 31:51

    CoreWeave Forge brings every step of the AI loop into one connected environment, allowing services to share context and evidence seamlessly. It is available for self-service sign-up.

  3. New Hardware for Agentic AI 52:32

    NVIDIA unveiled the Vera Rubin NVL72, designed for inference in the agentic era. Additionally, the Vera CPU was introduced, combining high single-threaded performance with throughput for cost-effective, large-scale agent processing.

  4. CoreWeave's Global Infrastructure 1:18:35

    CoreWeave operates a global AI cloud platform with 51 active data centers across North America and Europe, demonstrating its capacity to scale complex AI workloads.

Technical details

  • Agentic Failure Modes 2329s

    Complex agents introduce failure modes like context/memory drift (forgetting core intent), behavioral traps (false claims of action), and systemic complexity (minor early errors compounding exponentially).

  • Observability and Governance 2516s

    CoreWeave Agent Lens helps identify critical agent failures by grouping related failures into patterns, improving failure detection by 20% and reducing fix time by 1/10th. Capital One uses an evaluation-first architecture with LLM-as-a-judge and RAG metrics (faithfulness, contextual accuracy) to ensure trustworthy AI outputs.

  • Model Optimization Techniques 2230s

    Improvements can be achieved via CoreWeave Model Distillation (creating smaller, specialized models from production traces) or RL Rollouts (allowing continuous training updates without requiring a full serving redeployment).

  • Hardware Performance Metrics 3426s

    The Vera Rubin NVL72 is designed for inference, supporting six generations of NVIDIA accelerators (from Volta to Rubin). The Vera CPU offers 3x faster sandbox startup time and 1.7x faster data processing on benchmarks like ClickBench.

Mentioned resources

  • CoreWeave Forge (Platform)
  • NVIDIA Vera Rubin NVL72 (Hardware)
  • NVIDIA Vera CPU (Hardware)
  • CoreWeave Agent Lens (Tooling)
  • CoreWeave Model Distillation (Service)
  • CoreWeave Partner Network (Ecosystem)

Channel & topics

Watch on YouTube · Back to latest

This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.