DeepMind: inside Google's super-brain

wired.co.uk

34 points·edward··6 comments

The breakout example reminds me about this thing that optimizes an "AI" for Super Mario: http://www.engadget.com/2015/06/17/super-mario-world-self-le...

Actually what the article isn't super clear on is that the deepmind Atari 2600 learner knew nothing about aliens, balls, bats, bullets, left, right, up or down ... it was simply given the visual display (downsampled) as input, and was given a +ve or -ve reinforcement when something good or bad happened in the game.

Someone (not me) reconstructed the code [1] from the original publication [2].

[1] https://github.com/kuz/DeepMind-Atari-Deep-Q-Learner

[2] https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf

is 30! ≅ 10^25 ?

not in my book!?!

It appears to be close enough as an approximation that it didn't stop him getting into Cambridge at the tender age of 16.

At the end of the article:

> There's a long gap from here to having a system where you can sit and debate philosophy.

My response to that is:

First you have to answer the question, "What is philosophy?" :)

They deliberately avoid answering that question. The point is you feed "the machine" with all the data you can, maybe from all the sensors in every Android phone around the world, and "the machine" will tell you what philosophy is (or should be).

Please don't, the larger the search scope the more unpredictable the result of your optimization will be...

A "plausible" explanation for the Matrix if it would be real would not be a war ignited by fear and hatred but rather the likes of Amazon asking a very vague question on how to optimize it's product delivery or customer satisfaction from a very capable calculator with near unlimited access to information and resources.

Reply on news.ycombinator.com