AI Summary
→ WHAT IT COVERS Jason Calacanis and Lon Harris interview Josh Sirota, CEO of Aragon, on why enterprises should own custom AI models trained on proprietary data rather than relying on frontier models like Claude or GPT-4. They also cover drone logistics startup Iona, FAA Part 108 regulations, Kimi K2.6 launch, and Sergey Brin's internal Google memo on AI adoption. → KEY INSIGHTS - **Company World Model:** Aragon uses reinforcement learning to embed proprietary company data directly into model weights rather than pulling context into frontier models at runtime. This reduces token usage by roughly 100x and cuts inference costs to one-third the price of Anthropic's Opus at $5 per million tokens versus $15, while delivering responses up to 10x faster for repetitive internal workflows. - **Continuous Learning Architecture:** Aragon runs a Cron job every 15 minutes to ingest Slack, email, and other enterprise data into a ChromaDB vector store using Ollama and Chroma. The longer-term goal is overnight weight updates via reinforcement learning so the model self-improves without costly retraining cycles, eliminating the need to re-read entire data sources on every query. - **Enterprise AI Deployment Strategy:** When deploying AI into large organizations like Corgi's insurance underwriting workflow, start by identifying one repetitive, manual, high-frequency task rather than attempting full-scale integration. Corgi replaced a full-time operations role processing 40 daily insurance quote requests with an Aragon agent that sends a WhatsApp confirmation to a human approver, maintaining human-in-the-loop oversight while cutting cost and time. - **FAA Part 108 Regulatory Shift:** The upcoming Part 108 framework eliminates the requirement for drone operators to maintain visual line of sight, enabling one remote team to supervise multiple autonomous drones simultaneously from any location. This mirrors the transition from Level 2 to Level 4 vehicle autonomy and opens drone logistics to hospitals, local carriers, and rural communities previously excluded by Part 107's strict operator requirements. - **Drone Logistics Specifications:** Iona's tilt-rotor hybrid drone carries up to 20 parcels weighing 44 pounds total, flies 60 to 80 miles at full load on a one-way trip, and can autonomously load, drop, and recharge across multiple depot stops. The tilt-rotor design provides vertical takeoff convenience combined with fixed-wing glide efficiency, making per-unit delivery costs competitive with ground logistics in low-density areas serving under 500 people per square mile. - **AI Film Production Cost Reduction:** The Doug Liman Bitcoin film starring Gal Gadot is being shot entirely on a custom gray-screen London soundstage using AI-generated backgrounds and lighting via a company called Acme AI and FX, targeting a $70 million budget versus a typical $200-300 million comparable production. This model preserves actor performances and writing while eliminating location shoots, set construction, and large production design teams. → NOTABLE MOMENT Josh Sirota demonstrated Aragon processing roughly 50 real emails live on air, identifying a meeting request from a prominent venture capitalist as the highest priority action item while simultaneously routing a cleaning service invoice to an autonomous payment agent — all within seconds, using production data from his actual company. 💼 SPONSORS [{"name": "Pilot", "url": "https://pilot.com/twist"}, {"name": "Vanta", "url": "https://vanta.com/twist"}, {"name": "Quo", "url": "https://quo.com/twist"}, {"name": "Plaud", "url": "https://plaud.ai/twist"}] 🏷️ Enterprise AI Models, Drone Logistics, FAA Part 108, Reinforcement Learning, AI Film Production, Startup Deployment Strategy