"YKY (Yan King Yin, 甄景贤)" <generic.intellige...@gmail.com> wrote:

1999. Wrote dissertation proposal to study the cost of AI. Why is it so hard? I
had the idea that text compression was AI-complete. My approach would be to look
at the computational costs of language modeling and use compression to measure
results and project trends.

2000. I publish a paper and the first practical software (GPL) on text
compression using neural networks.

2001. Prospects for funding are dim. NSF rejects my proposal for a text
compression contest. My dissertation advisor pressures me to change my topic to
intrusion detection, which I do so I can get paid.

2002. In what little spare time I have, I publish PAQ1, the first context mixing
data compressor, using ideas from my previous neural network compressor.
Performance is better than zip but not great. It sits mostly unnoticed.

2003. I get my Ph.D and part time work teaching as an adjunct. The pay is low
but enough to live on, and I have time to work on my data compression hobby.

2003-2006. There is a (relatively) rapid period of development (with hundreds of
experimental versions) when dozens of people notice that PAQ is rapidly moving
to the top of the benchmarks. It started when one other person who knew
something about compression made some improvements. By 2006 a version of PAQ6
wins the Calgary compression challenge, replacing PPM as the dominant high end
compression format from that point forward. Nobody is getting paid, BTW.

2006. I take a couple of short term consulting jobs when my work on data
compression is noticed. My primary income is still teaching part time, however.

2006. I start a text compression contest similar to my NSF proposal but with no
funding or prize money. Marcus Hutter starts the Hutter Prize, a nearly
identical contest offering his own prize money after we can't agree on details
of the contest rules.

2006-2008. Continued development of the PAQ project with about 20 (unpaid)
developers working mostly independently racing for the top of various benchmarks
including mine. Versions of PAQ win the Hutter prize.

2008. A small company notices my work and offers me contracting work from home
on my own hours writing custom data compression algorithms at about 5 times what
I got paid teaching.

2009. They hire me and 2 other PAQ developers full time, doubling my pay while
continuing to work at home on my own hours.

2010. The company is acquired and I get another big pay raise while still
working at home. I am no longer involved in developing PAQ. Instead I have a new
open source project ZPAQ which is the first configurable, high-end format to
preserve compatibility between versions. It was my idea but my company is paying
me to work on it and promote it as an open standard, which I think will happen
in several years.

Failing that, your best bet is probably to put as much information as you can
about yourself on the internet. A future society with advanced technology might
be able to build a robot or a program that simulates you well enough that nobody
could tell the difference. Whether that's "you" is kind of a philosophical
question. If your goal is to build AI, I'm sure you've thought about it.

 -- Matt Mahoney, matmaho...@yahoo.com