Consciousness is just the ability to model your body within an environment and simulate futures that optimize a behavior policy. Maximizing your gene pool or paperclips, it doesn’t matter. Solvable AGI simply allows you to optimize your behavior policy better over time.
It is known how to produce consciousness — 3.7 billion years of evolution. The requirements boil down to:
- a reality
- reproduction
- a fitness function to drive reproduction
- time
But can do better than waiting around for 3.7 billion years. Machine reproduction doesn’t have a biological analog — it can evolve it’s information without physically reproducing. Our translated list for silicon statck:
- a reality
- the ability to take action and update our understanding of the reality
- cycles
Now, we have two questions from this:
- approach — what is the quantum of machine consciousness
- cycles — how can we make as many attempts as fast as possible
Focus on What Matters
The quantum of MC is both an ensemble of models and a meta model, an OS for intelligence, that trains, interleaves, and updates the ensemble.
The “consciousness” of AI right now has humans in the middle. I don’t mean anything mystical, simply the meta process of training models must be handled by the machine instead of humans. Humans tweak algorithms, adjust training data, and follow a waterfall style deployment process for models.
Forgot everything but training the simplest version of this process and letting it run itself in real time. The machine intelligence won’t need to sleep, it can train a separate mind and then swap it in when it is ready or do it in a rolling incremental process for each sub module in the ensemble.
How to go about automating this loop is another discussion altogether because simply automated what is happening now is an order of magnitude more difficult then reframing the problem. More on that in another series.
We can cut scope if we Embrace Alien Machine Intelligence.
Two principles occur for increasing cycles:
- A digital reality is ideal because we don’t need to tackle everything at once. The stakes are low so we can move fast. But it needs to be consistent. A game engine, simulation, or AR projection is ideal. Or perhaps even highly simplified hardware.
- Remove developers from the development loop as much as possible. Focus on on-line learning and the meta process of self-updating.
- Parallelize to federate learning
Putting it Together
Here is what any system that wants to build the quantum of machine consciousness will need to have:
- “hands” for the MC to test “ideas”
- efficient generation of trial and error cases aka “ideas”
- identification of what is better vs. what is worse i.e. error
The good news is we can build this today. Robotics or other spatial agents are ideal.