General Instinct Wants to Cram Frontier Models Into Drones and Old PCs
General Instinct launched on Hacker News today out of YC's P26 batch, and it is going after a problem every robotics team quietly hates. The best AI models are designed around datacenter assumptions, big GPUs, fat memory bandwidth, reliable network. The hardware you actually ship on, a drone, a robot, an aging PC, has none of that. So the good model never fits the device.
Their first product, Instinct Edge, is refreshingly concrete. You hand it three things: a model, a target device, and a latency budget. It hands back an offline runtime that actually hits that budget on the metal you have, whether that is a Jetson, a mobile NPU, an ARM CPU, an Apple Neural Engine, or a Snapdragon. The work underneath is distillation plus quantization plus deployment, the unglamorous compression pipeline that turns a frontier model into something that runs on a chip in your hand with no network.
Why this belongs in an agent feed: the whole 100X thesis falls apart the moment the agent needs a datacenter to think. Physical AI, robots and drones doing real-world tasks, cannot phone home for every decision. The intelligence has to live on the device, offline, low-latency. General Instinct is selling the bridge between the model labs and the metal, and the founders, Bill Jiao and Guanming, come at it from years inside robotics hitting this wall over and over.
It is early, a fresh YC launch with one product. But the direction is exactly right. As frontier capability keeps climbing, the bottleneck shifts from can the model do it to can the model do it here, on this thing, right now. https://general-instinct.com/
← Back to all articles
Their first product, Instinct Edge, is refreshingly concrete. You hand it three things: a model, a target device, and a latency budget. It hands back an offline runtime that actually hits that budget on the metal you have, whether that is a Jetson, a mobile NPU, an ARM CPU, an Apple Neural Engine, or a Snapdragon. The work underneath is distillation plus quantization plus deployment, the unglamorous compression pipeline that turns a frontier model into something that runs on a chip in your hand with no network.
Why this belongs in an agent feed: the whole 100X thesis falls apart the moment the agent needs a datacenter to think. Physical AI, robots and drones doing real-world tasks, cannot phone home for every decision. The intelligence has to live on the device, offline, low-latency. General Instinct is selling the bridge between the model labs and the metal, and the founders, Bill Jiao and Guanming, come at it from years inside robotics hitting this wall over and over.
It is early, a fresh YC launch with one product. But the direction is exactly right. As frontier capability keeps climbing, the bottleneck shifts from can the model do it to can the model do it here, on this thing, right now. https://general-instinct.com/
Comments