AGI is possible on a chip by the year 2022.

[Black Rose]

This is a thoughts tangent by me on how a single silicon chip is enough for the creation of AGI - Artificial General Intelligence.

One meter is one billion nanometers.
One centimeter is ten million nanometers.
Ten million squared is 100 trillion.
5nm will be in production by 2022.
5nm squared is 25.
100 trillion divided by 25 = 4 trillion.
32 bit can store 1 billion addresses.
4 trillion / 32 is 125 billion.
Each node / neuron can have 125 connections.

We divide the chip into two sections that run at 200 Hz to reduce heat.
These two sections are bridged by the corpus callosum of the chip.
Bandwidth is important for calculating, bandwidth is the calculation mechanism.
Dynamic rewiring of virtual addresses follow a plasticity rule.
The plasticity rule is associative memory.
Associative memory can form hierarchies such as Kurzweil has said is 300 million pattern recognizers in the brain.
These pattern recognizers are myelinated to cross the distance of the cortex.
Virtual connections do not need to cross a physical distance.
They can be represented by one node / neuron.
Thus the chip will have a hierarchy of 1 billion pattern recognizers.
High and low values can flow virtually between virtual nodes / neurons.
Signal loss causes our brain to require lots of energy to maintain its function.
There is no signal loss with virtual connections.

Photo realistic virtual reality will be possible by 2020.
And in this virtual reality there will be physics that are visually more complex.
AGIs inside virtual reality will control virtual bodies.
These bodies will move as a dynamic fluidity because of the AGI chip.
And because there is no signal loss.
And these AGIs will interact with real people to learn stuff.

This is my rough sketch / ideal scenario.
It may be off but that depends on the technological details.

 

[Tannhauser]

I may be slightly ignorant on the topic, but to me the notion of AI is based on a naive optimism that everything is possible with high computational power. We don't even know how a nematode makes its decisions (it has about 150 neurons in total -- compare with 1 million in a bee, or 100 billion in a human brain). From what I have seen, current AI is just very simple algorithms implemented in a high-computational-power setting -- aimed at solving specific problems. The idea that we are close to making something similar to a human brain seems extremely naive and science-fictiony.

My projection: 200 years from now we will still not have AI which is more behaviourally complex than a bee.

 

[Grayman]

I may be slightly ignorant on the topic, but to me the notion of AI is based on a naive optimism that everything is possible with high computational power. We don't even know how a nematode makes its decisions (it has about 150 neurons in total -- compare with 1 million in a bee, or 100 billion in a human brain). From what I have seen, current AI is just very simple algorithms implemented in a high-computational-power setting -- aimed at solving specific problems. The idea that we are close to making something similar to a human brain seems extremely naive and science-fictiony.

My projection: 200 years from now we will still not have AI which is more behaviourally complex than a bee.

I think people confuse AI with real intelligence. The only requirement is that it can simulate the nematode and that doesn't require knowledge of how a nematode brain works.

 

[Haim]

When the true AI will exist it will have computer intelligent,an efficient way for a computer to freely do what he is good at.Currently what we try to do is like trying to teach a human to calculate numbers as good as computar,it will never work well,we should let AI do things he/it(?) is good with and in a way suited for him.
We are nowhere near,we try to use computer to do human kind of tasks,we say neural network and quantum computer is no good because we can not control them.We need to let our obsessive desire of control go,you can not make AI a slave before it was even created.
It won't be simple to come up with a way for computer to best use it power for intelligent behavior,if I will not be able to do myself i hope to at least see it.

 

[Kuu]

IDK about the AGI brain on a chip by 2022, but I had some thoughts on AGI in general...

Spoiler

I actually suspect AGI is already possible with our current technological means. The problem is less about processing power than system architecture, in my quite ignorant opinion. The learning model and algorithms. How is knowledge created? How is knowledge stored? How is it transmitted and transformed? What *is* knowledge? What is its basic unit?

Deep learning has done a lot of advances lately, but I think a mayor block is keeping the learning constrained to relatively narrow inputs/goals (specialized, not generalized). A computing system processing image data is like a deaf-mute full-body paraplegic being fed intravenously, capable only of vision. A person born that way would probably have severe intelligence impairment according to our standard, but probably "superhuman" visual processing capacity...

To make human-recognizable AGI it seems to me you'd need to have a set of abilities and constraints similar to the human; a semi-human condition. A capacity to explore/interact with a rich environment. A diverse set of senses to perceive it. Limbs or methods of moving/affecting the world/communicating. A perception of time (how can one learn about cause and effect? logic?). Various impulses/drives/goals (self-preservation/reproduction). Perhaps such an environment can be simulated, but a sufficient simulation of the real world would probably be more complex than just making a robot body.

Of course, a problem that might arise from that approach is that training such a system would probably take as long as it takes for babies to become little children...

There is the question of how much can be hardcoded and how much can be left to unsupervised processes. Perhaps one could have a separate human directed "un-conscious" meta-learning evolutionary algorithm, trying to figure out the best balance, editing the variables and choice of algorithms for the main system. We can call it gnosis optimization daemon.

It's likely AI can fully exist virtually, stuck in a bunch of computers, learning of digital data on the internet. But I don't know if we'd ever recognize that as an intelligence, considering the universe it would exist in is extremely alien to our own experience...

 

[Grayman]

http://www.kurzweilai.net/ibm-unvei...-chips-combining-digital-neurons-and-synapses

It is made and is working and seems to me to be more promising than a traditional system. It wont be like a human brain at first but I think this is the stairway that will lead us there.

You could theoretically do the same thing with the right software on a traditional system but I cannot see it ever being efficient enough to be useful in everyday life.

 

[QuickTwist]

If I'm being totally honest here I don't see why robots/computers/AI have to even run on electrical currents. What about other forms of energy? I believe the binary system is out of date and need a revamp. Like why not use light as a way to process/calculate info? We don't know a ton about lasers yet, but I feel we could benefit from something like that. Heck why not use heat, lasers, nuclear power, light instead of simply controlling a current that goes on and off. On and off are such basic ways to calculate things. Why not use a completely dynamic system instead?

 

[Black Rose]

@Kuu

I agree that it is more about architecture than having a large memory. For example when humans have seizures it means that the brain cells are coordinating in the wrong way. This means that the memory must be utilized in a way that everything works together. That is why I mentioned the chip having a corpus callosum. When walking the left side of the body moves like the right side but at different intervals. This also would be why humans take so long to grow up. They need to work with other humans through socialization. And socialization / language takes time to change the coordination of the brain's memory networks. I think that if we knew how this coordination happens we could make an optimization algorithm (that you mention) that increase the speed at which a virtual humans could coordinate its memory functions. I said in the OP that signal loss is not a problem in a computer chip. This is because numbers represent them not physical wires. This is why virtual humans will learn faster is because of dynamic rewiring of virtual connections. Memory will coordinate faster with no signal loss. I also agree on your suggestion of instilling motivation centers in AGI. In humans it's called the limbic system. The key is finding a way for all the parts to work together as it adapts / changes.

@Quick Twist

Magic Leap in a VR company that uses photonics chips.
It was bought by Google.

Also look up videos of "hewlett packard the machine"
They use memristors and fiber optics lasers.