Advanced Artificial Intelligence--Friend or Foe

When I learned last December that AlphaZero, an artificial intelligence developed by Google subsidiary DeepMind, had soared from a completely "blank slate" to superhuman play in Chess, Go, and Shogi (a Japanese variant of chess) in a matter of hours, just by playing against itself, I realized that artificial intelligence is advancing far faster than most of us realize, and to superhuman, and potentially dangerous levels of capability.

You can read my description of AlphaZero's explosive mastery of Chess, Shogi and Go on OpEdNews or from an earlier post on zerospinzone.

AlphaZero's prowess led me do a lot of reading and research about the pace of AI progress, especially in the rapidly advancing areas of deep learning by deep neural networks.

 Image from the mind of Google's Deep Dream Generator

Credit: Robert Adler and DeepDreamGenerator

What I found out is that it's much worse than I thought.

Multiple groups worldwide are researching and developing increasingly powerful AI agents. These artificial intelligences have already surged past humans in many specific areas--not just games, but face and pattern recognition, medical diagnostics, reading comprehension, drug discovery and many other areas. Looming just ahead are AIs that will be smarter--possibly many times smarter--than any human in every important area. Not to mention super-intelligent AIs that can design and create other AIs that are even smarter, leading to an "intelligence explosion" whose impacts and risks are impossible to predict.

As several of the experts who have looked into this most deeply have pointed out, less intelligent species--think chimpanzees, gorillas or our close cousins the Neanderthals--typically don't fare well once a more intelligent species emerges.

A few farsighted individuals and groups have started to look into "the control problem"--how we might be able to create superhuman-but-friendly AIs, agents that will keep our best interests at heart even as they grow smarter and smarter. This turns out to be an extremely challenging problem. Meanwhile, hundreds of much better funded research groups--including many developing AI for military use-- are rushing ahead without giving the risks from the AIs they are developing much if any thought.

If you would like to look into this potential existential risk a bit further, here are some resources:

Center for the Study of Existential Risk (Cambridge, UK): cser.ac.uk

Deep dreaming: deepdreamgenerator.com (site where you can create your own “deep dreams.”

DeepMind: deepmind.com

DeepMind Ethics and Society research group: deepmind.com/applied/deepmind-ethics-society/

Eliezer S. Yudkowsky: yudkowsky.net

     “Artificial Intelligence as a Positive and Negative Factor in Global Risk”             (2008)

      Can be downloaded at: intelligence.org/files/AIPosNegFactor.pdf

Future of Humanity Institute (Oxford): fhi.ox.ac.uk

Future of Life Institute (Boston/Cambridge): futureoflife.org

The 23 Asilomar AI Principles: futureoflife.org/ai-principles/

James Barrat: Our Final Invention: Artificial Intelligence and the End of the Human Era (2015)--excellent introduction to these issues

Machine Intelligence Research Institute, MIRI (Berkeley): intelligence.org

Nick Bostrom: www.nickbostrom.com

     Superintelligence: Paths, Dangers, Strategies (2016)--a challenging factual, analytic, and philosophical work examining these developments in depth

OpenAI: openai.com

Pedro Domingos: The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World (2018)

I'd suggest starting with James Barrat's extremely well-researched book, Our Final Invention: Artificial Intelligence and the End of the Human Era.

And for a super-deep dive into the issue, Oxford philosopher Nick Bostrom's Superintelligence: Paths, Dangers, Strategies. 

----

If you enjoyed this post, please sign up to follow or receive email alerts from zerospinzone.blogspot.com.