Continuum of Autonomy

We can draw a continuum of machines of greater or lesser autonomy. Here, the autonomy of the machine increases as we move downwards:

Up this end is:

Least autonomy for the machine.
Greatest amount of work for the human.
Have to imagine every design ourselves.
But less risk that nothing will emerge.
Greater ability to analyse and explain the solution.

Entirely designed by hand and fixed. No user input.
|
|
Explicit user input - e.g. Command-line parameters, or enter user preferences in Options Dialog.
|
|
Implicit user input - Machine "learns" (memory) by observation. e.g. Open up in same directory as was using last time.
|
|
Tweak some parameters (e.g. for performance) while running.
|
|

NORMAL PROGRAMS (NON-AI PROGRAMS)

|
|
Search a solution space using heuristics.
|
|
Learn to generalise from I/O exemplars (Supervised Learning).
|
|
Learn from no exemplars, but only periodic rewards during task (Reinforcement Learning).
|
|
Learn from no rewards during task, but only a live/die decision at end, according to some explicit fitness function (Genetic Algorithms).
|
|
Fitness function isn't even explicit, but is implicit in the dynamics of the environment (Artificial Life).
|
|
No fitness function or feedback (Unsupervised Learning, Learning to Learn).
|
|
Only laws of physics designed by human.
|
|
Laws of physics themselves evolve.

Down this end is:

Greatest autonomy of the machine.
Least work for the human.
Hope of something emerging that we couldn't have designed.
But also greater risk of only a simple machine emerging.
Longer wait time for self-modification to complete.
Less analysis of final solution.

Evolving an entire universe

The story "The Genesis Tub", in The Simpsons, Treehouse of Horror VII.
Lisa accidentally creates a mini universe in which intelligent life emerges quickly (overnight).