Benjamin Jones

Raw. Behind the scenes. Explainer. 🎥🤖 🚨 The first agentic military strike. 🚨 A few weeks ago we showed something that had never been demonstrated before: A team of AI agents conducting a coordinated military strike on a target. In the explainer video below, I'm sharing a look under the hood at how it actually works. System architecture: 🚙 1 UGV carrying 2 one-way attack drones 🧠 7 AI agents total • 1 Orchestrator agent (~300B parameters) • 6 asset agents Each asset runs two agents: • Reasoning agent – planning, decision making, tool use • VLA agent – perception and vehicle control trajectories (~1B parameters) Tool use is also used all over the place on orchestration and platform autonomy. Agents call tools on traditional platform autonomy (e.g. flight controllers) when that makes sense vs. direct control (e.g. unlatch and takeoff command). Systems: • Orchestrator runs in an orchestration compute cluster on a set of H100s. • Each edge asset has 2x Nvidia inference chips, some as low as 25W. • Orchestrator does not have any implementation insight into assets. • Treats them as third-party assets. • It reads the ICD for the function calls, similar to MCP. Shared cognition across the system: 📡 All agents share their thoughts across the network 🧠 Memory stored as text + short-term video context 👁️ Systems maintain shared situational awareness This allows the system to operate as a single coordinated intelligence, not isolated robots. • All systems are loosely coupled, prompting eachother. • Clear safety guard rails and human on the loop. • Entire video is hands-off from human control after the initial paragraph of intent into the Orchestrator. • Agents run between 350ms to 10s cycle times depending on their function and all actions are real-time dynamic. No pre-scripting of the mission. ⚡ Video is sped up 2× for brevity This capability was developed under DARPA and U.S. Army contracts for broad DoW use. 🎙️ @emilmichael @Jason @friedberg @chamath this is the fully autonomous AI mission you discussed as possible someday soon on the @theallinpod last week. Happy to give any of you a direct technical brief at any time.

Fury has started learning to fly missiles.

Robot parking lot perfection achieved internally. Proceed to next level. ✅🤖 1x E2E neural net 1x RGB camera (video) 1x low power GPU 1x natural language prompt No headlights, no map, no lidar, no waypoints, no labeled data. VLA is fine-tuned only from human observation. VLA: vision-language-action model, an LLM with vision, trained to actuate a robot - embodied AI. Much simpler, more lightweight, and more capable than old-school autonomy. These systems are the future. Spot the team cheering.😃