Recipe for CEV (was Re: Morality simulator)

From: Matt Mahoney (matmahoney@yahoo.com)
Date: Fri Nov 23 2007 - 16:25:53 MST


Recipe for CEV.

Set up a data compression contest for compressing a large portion of the
Internet, including text, images, video, and audio. To compress text, you
must predict it. To predict it, you must understand it. To understand it,
you must understand the thoughts of the writer, which depend on the writer's
model of the world. Compressors that compress smaller will therefore have
better models of the world and of the human mind.

Likewise for video and audio. Compressors that are better at modeling
physics, that know about objects commonly seen and heard by humans, know how
people move and speak, and model visual and audio perception will compress
smaller than programs that don't.

The text portion would ideally include writing samples from everyone on Earth.
 In practice we won't achieve that, so people who post a lot of material to
the Internet will get a bigger vote than those that don't. So be it. The
important point is that the top compression program will have the best model
of humanity. It will know what makes people happy or mad or successful or
afraid. It will understand, as many people do, about planning for the future,
that people who immediately get everything they want will not be happy, and
that the rich are not happier than the poor. It will understand genius and
mental illness. It will understand about moral views in various cultures,
because it will have to in order to compress well.

I do not specify how this model is built. It will likely be built using a mix
of manually coded knowledge, machine learning, and meta-learning algorithms
like evolution and recursive self improvement. A successful compressor will
know how to write and debug software, because such models will compress
specifications, source code, and executable code better than those that don't.

The will require a lot of computing power. To model a single adult human mind
we would require training data equivalent to 20 years of text, speech, and
video. We want to model billions of human minds.

We now have a model of the world that includes human goals, an understanding
of volition, and forecasting abilities on a par with prediction markets. We
stop here, because CEV does not tell us how to transfer these goals to the AI.

-- Matt Mahoney, matmahoney@yahoo.com



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:00 MDT