My name is Daniel Dewey. I'm an AI safety researcher.
My name is Daniel Dewey. I'm an AI safety researcher.
Motivated by I.J. Good's intelligence explosion theory, this paper argues that reinforcement learning is not appropriate for real-world artificial general intelligence. It then introduces an alternative motivational structure, value learning, which is designed to maximize an initially unknown utility function in an initially unknown environment. In the proceedings of AGI-11.
November 2011— Great news: The Singularity Institute for Artificial Intelligence Canada Association has selected me to receive their Academic Prize for 2011!
"We resolve to award Daniel Dewey a prize of $5,000.00 in recognition of his efforts to improve AI theory, and in the hope he will make further contributions in this field of study."
This will be a great help, both in supporting the work I'm doing now on formal models of physical AI, and in moving me towards a career in AI research. Thank you to the SIAI-CA, and to the donors who made this possible!
October 2011— Many thanks to Oxford's Future of Humanity Institute for hosting me for a couple of weeks. I've come back with lots of new ideas and colleagues.
August 2011— Learning What to Value was very well received at the Fourth Annual Conference for Artificial Intelligence. Congratulations to Moshe and the other organizers for putting on a great conference!
January 2011— I've joined the Singularity Institute's Research Associate program; their funding will support my travel to present conference papers relevant to their mission. Very excited to be collaborating with the folks at SI!