Skip to main content

Entry #1: Fully Offline

Neural Cube can run fully offline, without depending on anyone else's server. Every LLM agent and ML/DL model runs locally on the robot's onboard processor.

The point was to give the AI a body to experience things firsthand while having all experiences tied to the actual robot itself. By running everything onboard, the robot’s experiences remain its own rather than being distributed across multiple places and transient between instances. This creates the foundation for a stronger, continuous, and persistent self that belongs only to this physical machine.

On the robot
Microphones and camera
Speech recognition
The language model
Memory and its curation
Voice and face fingerprints
Navigation and the map
What crosses
Never leaves the robot: Audio and video from the room
Never leaves the robot: Conversation history
Never leaves the robot: Memories and enrolled people
Never leaves the robot: Telemetry and usage data
May leave the robot: Web search — optional, and only the query

There are practical reasons this matters too. Whatever the robot sees and hears stays with the robot, with no telemetry pipeline and no third-party transcription service standing between the robot and the people around it. It also means the robot doesn't quietly change because a model was updated behind an API somewhere, or stop working because a subscription lapsed.

There's also nothing to leak, because nothing left in the first place. No API keys, no cloud account, no inbound endpoint sitting open by default, no server holding data that was never asked to be uploaded.

Online extras, like web search, are the one exception, and they simply switch themselves off when there's no internet available.