Thinking as a Hobby


Home
Get Email Updates
LINKS
JournalScan
Email Me

Admin Password

Remember Me

2997252 Curiosities served
Share on Facebook

Kurzweil Forgets to Carry His One
Previous Entry :: Next Entry

Read/Post Comments (0)

Speaking of scientists pulling stuff out of their bums...

Ray Kurzweil predicts we will have "human-level AI" by 2029.

How did he pinpoint it that precisely? His reasoning is here:


I've consistently set 2029 as the date that we will create Turing test-capable machines. We can break this projection down into hardware and software requirements. In the book, I show how we need about 10 quadrillion (1016) calculations per second (cps) to provide a functional equivalent to all the regions of the brain. Some estimates are lower than this by a factor of 100. Supercomputers are already at 100 trillion (1014) cps, and will hit 1016 cps around the end of this decade. Two Japanese efforts targeting 10 quadrillion cps around the end of the decade are already on the drawing board. By 2020, 10 quadrillion cps will be available for around $1,000. Achieving the hardware requirement was controversial when my last book on this topic, The Age of Spiritual Machines, came out in 1999, but is now pretty much of a mainstream view among informed observers. Now the controversy is focused on the algorithms.


Yeah, those algorithms are the stickler, aren't they?


To understand the principles of human intelligence, that is to achieve the software designs, we need to reverse-engineer the human brain. Here, progress is far greater than most people realize.


And yet it is still woefully inadequate, mostly because of the dearth of strong theories of brain function.


The spatial and temporal (time) resolution of brain scanning is also progressing at an exponential rate, roughly doubling each year, like most everything else having to do with information. Just recently, scanning tools can see individual interneuronal connections, and watch them fire in real time. Already, we have mathematical models and simulations of a couple dozen regions of the brain, including the cerebellum, which comprises more than half the neurons in the brain.


Right, and that means we've got the cerebellum all figured out. Not.


IBM is now creating a simulation of about 10,000 cortical neurons, including tens of millions of connections. The first version will simulate the electrical activity, and a future version will also simulate the relevant chemical activity. By the mid 2020s, it's conservative to conclude that we will have effective models for all of the brain.


Conservative, huh? Hahahahahahahahahahaha.

Look, I like a little optimism as much as the next guy, but one way to quickly get into trouble is to try to map the trajectory of any theory-poor domain (another is to say "X will never be doable").

When will we have teleportation? Travel at or near the speed of light? Kurzweil naively plots the necessary hardware requirements as if they are sufficient to create strong AI. And he assumes that workable cognitive theories are progressing at either a linear or exponential rate, which just isn't true.

The answer to when human-level AI will emerge is: nobody has any frickin' clue.



Read/Post Comments (0)

Previous Entry :: Next Entry

Back to Top

Powered by JournalScape © 2001-2010 JournalScape.com. All rights reserved.
All content rights reserved by the author.
custsupport@journalscape.com