Max Productivity, Minimal Compute, Zero Leaks: Pioneering Secure Enterprise Agents
summary
Command A claims parity or better than GPT‑4o and Deepseek V3 on human evaluations while cutting compute: a ~110B design that fits on two A100s and uses sliding‑window attention, grouped query attention and parallel blocks to use four times fewer GPUs. Post‑training merges six experts, retains 90–98% per expert and gains ~1.5–3% human‑preference wins from RL soup plus RPO↔COPG ping‑pong. North is an air‑gapped enterprise agent already deployed internally and with customers, powering real workflows and summaries.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:01Lean AI model challenges
- 3:10Compute bottlenecks and security
- 4:58Command A introduction
- 6:22Command A architecture
- 10:58Post-training innovations
- 15:21North: secure agent system
- 19:43Discovering AI use cases
- 22:37Enterprise needs with Command A
- 23:39Q&A
- 24:04Quantization and efficiency
- 25:31Context management and RAG
- 27:19Air-gapped deployment
- 29:11Live demo: North agent
- 31:19Closing remarks
related talks
20mKeynote: Future-Proofing Confidential AI
01h 42mBuilding Responsible AI Agents in Decentralized Open-Source Environments
52mSimplifying the adoption of generative AI for enterprises
07h 47mAI Agents Conference 2026 - Day 5
52mDesigning Production-Ready Agentic Systems: Lessons from the Field
53mHow to Build and Scale Ultra-Lean Gen AI Startups
30mBeyond GPUs: Why AI Agents Need a Need New Hardware
37mGenerative AI: Beyond (Just) Hype