"Not Enough Stuff" arguments in AI
AI has a history of people saying that the systems and algorithms
simply need more of some property and there will be a breakthrough.
To a large extent this has been true.
The main "Not Enough Stuff" arguments would be:
- Not enough speed.
"If we just get faster machines
we will have full AI".
- This has been happening for the whole history of AI.
Faster machines make older algorithms suddenly more useful.
-
The recent use of GPUs as simple parallel computers
has massively helped the speed of certain algorithms.
- Not enough neurons/memory. Not big enough networks.
"Neural networks need to be much bigger, and with a much higher number of layers."
-
Theoretical breakthroughs have made this happen.
- Not enough data.
"The machine needs to read half the Internet."
-
The rise of massive AI companies spending hundreds of millions on training has made this happen.
- Not enough time.
"We get 20 years to grow our intelligence.
Maybe a machine needs a 20 year childhood too."
-
Many people have noted the gap between a human childhood,
switched on and interacting with the physical world 15 hours a day,
365 days a year, for 20 years,
and robot "learning", where the robot is typically switched on and learning only for
very short time periods.
-
While software-only experiments can run for months,
bringing a 20 year human childhood to a robot with a physical body is quite a challenge.
-
Cloud robotics
is an idea that multiple robots could share their learning and have a collective brain,
something that does not exist in nature.
- Not enough authors. Not enough diversity in the mind.
"Real brains will consist of a huge, diverse society of sub-brains of different types of intelligence
and different forms of learning and memory."
Some of my thoughts on this:
- "Not enough stuff" in AI.
In Constructing complex minds through multiple authors,
Proceedings of the 7th International Conference on the Simulation of Adaptive Behavior (SAB-02), 2002.
- AI is too big a problem.
In "The World-Wide-Mind: Draft Proposal", Dublin City University, School of Computing, Technical Report no. CA-0301, 2001.