http://www.guardian.co.uk/science/20...ntelligence-ai
Quote:
Scientists have created a "Eureka machine" that can work out the laws of nature by observing the world around it – a development that could dramatically speed up the discovery of new scientific truths.
The machine took only hours to come up with the basic laws of motion, a task that occupied Sir Isaac Newton for years after he was inspired by an apple falling from a tree.
Scientists at Cornell University in New York have already pointed the machine at baffling problems in biology and plan to use it to tackle questions in cosmology and social behaviour.
The work marks a turning point in the way science is done. Eureka moments, which supposedly began in Archimedes' bath more than 2,000 years ago, might soon be happening not in the minds of geniuses, but through the warm hum of electronic circuitry.
"We've reached a point in science where there's a lot of data to deal with. It's not Newton looking at an apple, or Galileo looking at heavenly bodies any more, it's more complex than that," said Hod Lipson, the computer engineer who led the project.
"This takes the grunt out of science by sifting through data and looking for the laws that govern how something behaves."
(snip)
|
I've been chatting to my brother about this for years. Basically, formal logic is now sufficient that, if you fill up a database with a vast array of formally-described facts about the world, you can build an "inference engine" around that to infer vastly more facts from those facts, based simply on 1st-order, 2nd-order and other extensions of formal logic. Couple that to heuristics allowing the engine to determine if new facts are "interesting" or "significant" by our criteria and
voila, an invention/discovery machine.
These guys have gone a step further than that and built a robot that can actually devise experiments to gather new data or seek missing data.
---
A programming project I shelved last year for lack of energy revolved around the idea of starting with a database schema* for storing formal predicates and a natural-language parser coupled with a web spider for continuously browsing the web for new facts about the world, then having an inference engine sift through those facts and try to find interesting inferences.
The gathering program would have to be trained using a variety of means so that it didn't try to download the entire internets. Among the mechanisms I was thinking of were heuristics for the classic hallmarks of bad science, using formal logic itself to establish that many facts are essentially equivalent (for pruning) and allowing the trainer to sift through facts in the database and their sources and indicate to the gathering program the bullshit level of the sites and higher level domains particular facts were gathered from (this would also influence the bullshit level of sites linked to from those sites). A second-tier extension of this might be to get valuable data out of crackpot sites by not recording their "facts" as actual facts so much as data about how humans think and the limits of their beliefs.
The internet is a godsend for training AIs since its the single largest interconnected repository of human knowledge. You just have to find a way to get an AI to read it like a skeptical and informed human would.
Sadly, between insomnia, workload and the sheer, daunting size of the problem, I found I just didn't have the energy to continue past a certain point, which is why I'm working on smaller problems like IK now in my spare time
*ETA for the technical types - The db engine I pulled down for this is the Berkeley DB engine behind Google, which instead of using SQL just has a programming API and expects you to manage relation constraints and all the rest. It just focuses on being massively efficient at storing vast numbers of arbitrary-sized objects in hash tables (tuples, for the even more technically-minded). Its actually a lot better and faster for a lot highly-specialised massive data storage requirements.