On General Machine Intelligence Strategies

On General Machine Intelligence Strategies

A transcript of the above talk

Hi! I'm Tim Tyler - and this is a video which discusses the issue of whether the path to superintelligent machines will be "broad" or "narrow".

The breadth of machine intelligence

Enthusiasts for creating general-purpose machine intelligence have often described classical application-driven development projects as having a "narrow" scope. Despite - or perhaps because of - the rather derogatory overtones, the name seems to have stuck - so these days, "general" machine intelligence projects are sometimes described as being "broad" - while ones targetted at particular application domains are usually described as being "narrow".

"Narrow" projects have some advantages. They are more likely to pay the bills, and actually do something useful in the short term. However, some have questioned whether such development will lead to the type of more general-purpose intelligence found in humans.

Here is Ben Goertzel on the topic:

[Footage of Ben Goertzel]

Follow the money

My take on this issue is that the first superintelligent machines will probably arise out of a large organisation with a specific mission.

The most likely candidates are probably search engines, stock market investment firms and governmental security services - such as the NSA.

These organisations seem to me to be the most likely ones to have the money, manpower and motivation to create the first superintelligence.

In each case, there is a driving problem - which seems best classified as a being associated with a narrow problem domain.

However, in order to excel at solving these kinds of problems, considerable breadth of intelligence will be required.

Don't compete with humans

Apart from sheer financial oomph, there is another reason to expect these kinds of projects to be promising candidates for the role of the first superintelligent machine.

So far, computing machines have made their inroads into the biosphere largely by avoiding competing directly with humans - and instead specializing in the areas where the human brain is weak.

Consequently we have machines with enormous memories, high reliability and rapid serial performance, which are appalling at parallel pattern-recognition tasks.

Such machines complement human intelligence - rather than competing directly with it. They augment humans, rather than attempt to substitute for them - because human labour is pretty cheap, and machines can't yet compete directly with us in the cognitive areas where we are strong.

The search oracles, stockmarket masters, and surveillance gurus than I envisage waking up first are utterly inhuman entities - whose roles could not be played by human beings. The fact that they are not in competition with existing human brains will help them to get off the ground.


So, to summarise, "narrow" machine intelligence projects will be better funded, are more likely to get earlier access to better hardware, and will attract a larger research and development budget. Also, they avoid competing directly with the human brain in areas where the existing brains hardware is strong - which is a smart move.

The tip of a wedge is narrow - as it must be, in order to penetrate initially. A wedge with a broad tip is not very useful. Breadth comes later.


Tim Tyler | Contact | http://alife.co.uk/