Here's the standard argument, as I understand it:
- There are something like 100,000,000,000 neurons in the human brain, each of which can have up to around 10,000 synaptic connections to other neurons. This is basically why the brain is so powerful.
- Modern CPUs have around 4,000,000,000 transistors, but Moore's law means that this number will just keep going up and up.
- Several decades from now (probably in the 2030s), the number of transistors will exceed the number of synaptic connections in a brain. This doesn't automatically make computers as "smart" as people, but many of the things that the human brains does well by brute-forcing them via parallelism will become very achievable.
- Once you have an AI that's effectively as "smart" as a human, you only have to wait 18 months for it to get twice as smart. And then again. And again. This is what "the singularity" means to some people.
The other form of this argument which I see in some places is that all you need is an AI which can increase its own intelligence and a lot of CPU cycles, and then you'll end up with an AI that's almost arbitrarily smart and powerful.
I don't hold these views myself, so hopefully someone with more information can step in to correct anything I've gotten wrong. (LessWrong.com seems to generally view AI as a potential extinction risk for humans, and from poking around I found a few pages such as http://lesswrong.com/lw/k37/ai_risk_new_executive_summary/)
- There are something like 100,000,000,000 neurons in the human brain, each of which can have up to around 10,000 synaptic connections to other neurons. This is basically why the brain is so powerful.
- Modern CPUs have around 4,000,000,000 transistors, but Moore's law means that this number will just keep going up and up.
- Several decades from now (probably in the 2030s), the number of transistors will exceed the number of synaptic connections in a brain. This doesn't automatically make computers as "smart" as people, but many of the things that the human brains does well by brute-forcing them via parallelism will become very achievable.
- Once you have an AI that's effectively as "smart" as a human, you only have to wait 18 months for it to get twice as smart. And then again. And again. This is what "the singularity" means to some people.
The other form of this argument which I see in some places is that all you need is an AI which can increase its own intelligence and a lot of CPU cycles, and then you'll end up with an AI that's almost arbitrarily smart and powerful.
I don't hold these views myself, so hopefully someone with more information can step in to correct anything I've gotten wrong. (LessWrong.com seems to generally view AI as a potential extinction risk for humans, and from poking around I found a few pages such as http://lesswrong.com/lw/k37/ai_risk_new_executive_summary/)