Currently, if the choice of device is ambiguous, we run an out-of-band dialogue to let the user choose. This dialogue is not part of the general state machine, and choosing a device does not go through the neural network. This is suboptimal, because it constrains the way the user can talk about the selected device, and prevents us from having general fallbacks and error handling. Instead, we should fold disambiguation the general state machine.
The challenge with this plan is how to expose the list of available devices to the state machine and to the neural network (ideally, in a privacy-preserving way), both at simulation and at inference time.
Currently, if the choice of device is ambiguous, we run an out-of-band dialogue to let the user choose. This dialogue is not part of the general state machine, and choosing a device does not go through the neural network. This is suboptimal, because it constrains the way the user can talk about the selected device, and prevents us from having general fallbacks and error handling. Instead, we should fold disambiguation the general state machine.
The challenge with this plan is how to expose the list of available devices to the state machine and to the neural network (ideally, in a privacy-preserving way), both at simulation and at inference time.