Daniel DeweyContact me: Twitter, Email
I'm an independent consultant and researcher on potential risks from advanced AI, currently based in Portland, OR. I was previously a research fellow at the Programme on the Impacts of Future Technology at the Future of Humanity Institute and Oxford Martin School, a software engineer at Google Seattle, and a student researcher at Intel Labs Pittsburgh during my undergraduate at Carnegie Mellon University.
The societal challenges posed by modern machine learning are significantly different from the challenges posed by past developments in AI; likewise, we will probably face qualitatively different problems when AI systems are much more capable than they are today. One foreseeable challenge is called the superintelligent agent control problem. This is the problem of ensuring that AI agents that far surpass human capabilities in most domains do not act in ways that cause catastrophic harm – not through the emergence of "intentionality" or human-like motivation, but as a side-effect of highly competent pursuit of the agent's built-in goal. Though it is implausible that superintelligent agents will be built in the next few decades, some work to prepare for this development seems prudent, and some initial work is already underway.
For an introduction to this area, I'd most recommend Nick Bostrom's excellent book Superintelligence. For a briefer introduction, along with an initial reading list, I'd recommend my post at the Global Priorities Project:
My papers on this topic:
- Research priorities for robust and beneficial artificial intelligence. S. Russell, D. Dewey, M. Tegmark. In AI Magazine, 2015.
- Long-term strategies for ending existential risk from fast takeoff. In Risks of Artificial Intelligence, V. Müller (ed.), 2015.
- Reinforcement learning and the reward engineering principle. AAAI Spring Symposium Series, 2014.
- A representation theorem for decisions about causal models. In proceedings of the Fifth Conference on Artificial General Intelligence, 2012.
- Learning what to value. In proceedings of the Fourth Conference on Artificial General Intelligence, 2011.
Other papers of mine:
- Additively efficient universal computers. ECCC TR14-044, 2014.
- Crucial phenomena. In How Should Humanity Steer the Future?, A. Aguirre, B. Foster, and Z. Merali (eds.), 2015.
- Generalizing metamodules to simplify planning in modular robotic systems. D. Dewey, M. Ashley-Rollman, M. De Rosa, S. Goldstein, T. Mowry, S. Srinivasa, P. Pillai, and J. Campbell. In proceedings of the International Conference on Intelligent Robots and Systems, 2008.