Brooklyn Project Approach
What is needed is a new approach to AGI. Is the Brooklyn Project that approach?
If you look at what J. Craig Venter and Tony White did with DNA, it was more than just a small
improvement.
AGI needs something like that. What you have in your proposal is all of the modern
approaches such as shareware,
But we need a way of bootstrapping everything. We need to create a tool which helps us
create AGI. If you could create a subset of AGI which would pertain only to computer
programming and concepts pertaining to computer design, neural networks, genetic
algorithms, Bayesian concepts, statistical concepts, etcetera, then that AGI helper could
help to search for the new concepts that would kick start the process of a full human-like
AGI.
Another approach would be to evolve the higher system concepts which the human brain
uses but that we do not yet understand.
Develop a set of problems which exist in one-dimension or two-dimension spaces. Once
these problems are solved, move on to three-space.
The Brooklyn Project will avoid the mistakes of previous models for general intelligence.
Most previous models have foundered because each of them were based on some single
idea--no matter that it was a valuable one. For example, some researchers became fixated on
mathematical logic as a universal mechanism for intelligence. Then, much research shifted
to try to exploit the use of rule-based systems and semantic networks. More recently there
has been an emphasis on neural-network learning schemes, probabilistic and statistical
mechanisms, and models that simulate evolution.
Our new mistake is to think that we understand the brain sufficiently to program it into
code using a top-down approach.
If we completely understand the human brain then we should be able to program it using a
top-down approach. Making improvements to the resulting structure will require not only an
understanding of how it works but of how nature would have improved it had nature had more
time. This will demand quite a bit of understanding. For example, we now know that:
"The retina compresses visual information in various ways, and then
routes it to the primary visual cortex. However, there are 10 times
as many nerves the go *to* the primary cortex than come *from* it!
So it appears that, in effect, the central processes tell the visual
cortex how to process each particular image; in other words, perhaps,
what to be looking for."
What we DON'T know is all of the other tricks which the human brain has evolved over the
years.
To make sure that you succeed in creating a form of AGI, one should copy as much from how
nature designed the human brain as possible.
If the interfaces between various modules in the brain could be defined, then each module
could be worked on separately. If each module was defined genetically, then evolutionary
algorithms could be used to optimize the whole machine.
If we can actually program a desktop computer to have Artificial General Intelligence,
then why can't we program it to do something less? One such lesser goal would be to
create a form of intelligence which could do computer programming for us. Once this
sub-goal is reached, we would have plenty of "virtually free" assistance in doing the rest of
the programming.
Acadabot
Next
What is needed is a new approach to AGI. Is the Brooklyn Project that approach?
If you look at what J. Craig Venter and Tony White did with DNA, it was more than just a small
improvement.
AGI needs something like that. What you have in your proposal is all of the modern
approaches such as shareware,
But we need a way of bootstrapping everything. We need to create a tool which helps us
create AGI. If you could create a subset of AGI which would pertain only to computer
programming and concepts pertaining to computer design, neural networks, genetic
algorithms, Bayesian concepts, statistical concepts, etcetera, then that AGI helper could
help to search for the new concepts that would kick start the process of a full human-like
AGI.
Another approach would be to evolve the higher system concepts which the human brain
uses but that we do not yet understand.
Develop a set of problems which exist in one-dimension or two-dimension spaces. Once
these problems are solved, move on to three-space.
The Brooklyn Project will avoid the mistakes of previous models for general intelligence.
Most previous models have foundered because each of them were based on some single
idea--no matter that it was a valuable one. For example, some researchers became fixated on
mathematical logic as a universal mechanism for intelligence. Then, much research shifted
to try to exploit the use of rule-based systems and semantic networks. More recently there
has been an emphasis on neural-network learning schemes, probabilistic and statistical
mechanisms, and models that simulate evolution.
Our new mistake is to think that we understand the brain sufficiently to program it into
code using a top-down approach.
If we completely understand the human brain then we should be able to program it using a
top-down approach. Making improvements to the resulting structure will require not only an
understanding of how it works but of how nature would have improved it had nature had more
time. This will demand quite a bit of understanding. For example, we now know that:
"The retina compresses visual information in various ways, and then
routes it to the primary visual cortex. However, there are 10 times
as many nerves the go *to* the primary cortex than come *from* it!
So it appears that, in effect, the central processes tell the visual
cortex how to process each particular image; in other words, perhaps,
what to be looking for."
What we DON'T know is all of the other tricks which the human brain has evolved over the
years.
To make sure that you succeed in creating a form of AGI, one should copy as much from how
nature designed the human brain as possible.
If the interfaces between various modules in the brain could be defined, then each module
could be worked on separately. If each module was defined genetically, then evolutionary
algorithms could be used to optimize the whole machine.
If we can actually program a desktop computer to have Artificial General Intelligence,
then why can't we program it to do something less? One such lesser goal would be to
create a form of intelligence which could do computer programming for us. Once this
sub-goal is reached, we would have plenty of "virtually free" assistance in doing the rest of
the programming.
Acadabot
Next
 | |  | |  | |  |