The NI is loaded into a biomechanical interface that provides sensory inputs, locomotive actuators, and environmental manipulators. Typically, only one researcher builds the entire device, and collaboration is not useful for that part of the project. (Unfortunately, as the network heavily exploits subtle implementation details in the mechanicals, it is very difficult to perform upgrades after the first stages of training are completed.)

From there, the researcher has to continually upload conceptual primitives through the sensory apparatus, and the NI prunes and rebalances its own neural network to establish basic foundation concepts such as object permanence, the acceleration of gravity, thermodynamics, ballistic path prediction, etc. Eventually, when the network is sufficiently trained, researchers may begin to input additional data through a natural language interface. Due to variations in the biomechanical devices, it is currently impossible to use standard bootstrap code to accelerate that process.

In order for the NI to be useful as a music recommendations engine, it is essential to expose it to music that you already like through its audio sensors, during the initial training phases. After approximately 8 years, the NI will begin to autonomously seek out music samples in the wild and recommend that you purchase copies of promising collections. The system is not perfect. It will occasionally issue recommendations for music that was already present in the training corpus, or for maliciously-formed music files designed to hack uninoculated NIs into recommending them. And it should be noted that these NIs have been known to abruptly diverge from preferences implied by the training corpus, producing wildly inaccurate recommendations thereafter.

It's probably just cheaper and more reliable to use an AI, but as long as this thing still works okay, I'm going to keep refueling it.