2 Train the first computer neural network to operate the second computer like a human operator in response to instructions (voice commands) from a human controller (i.e., the first computer neural network being the operator of the second computer and the human controller giving instructions to the neural network operator).
3 The neural network may be similar to the Openai Five network [1] with CNNs plus embedding and FC layers at the audio and video inputs that feed into a long LSTM layer that feeds into FC layers for the keyboard and mouse outputs.
4 The "reward" for the neural network is something like a laptop battery life meter that drains over time as the neural network operates the second computer and gets refilled at the whim of the human operator, e.g., after the successful completion of commands. Eventually, with a large enough network and enough compute, the commands transition from "open this" and "do that" to "what will happen next?" and "how does that make you feel?". Sessions using human operators can be recorded and replayed for the neural networks to begin learning and then graduate to live interaction with a human controller.
5 If emotions, intelligence, and consciousness are emergent properties, then those properties should emerge from the neural network behavior.
6 If consciousness is not emergent, then the neural network may still be trained to operate a computer. E.g., as a software programmer that does the grunt work of programming (performing a correct refactorization) while the human operator does the creative work (deciding what to refactor).
[1] see https://towardsdatascience.com/the-science-behind-openai-five-that-just-produced-one-of-the-greatest-breakthrough-in-the-history-b045bcdc2b69
