Looking into AI Risk
In considering what to work on, I wrote that recently many people I respect have started to think that AI risk is the most valuable place to focus EA efforts. For example, 80000 Hours ranks it first on their "list of global issues", taking into account "scale, neglectedness, and solvability". On the other hand, I have a lot of friends working in machine learning, and none of them think AI risk is worth working on now. This level of disagreement is very strange, and kind of worrying.
What I'm planning to spend the next few days on is getting a better understanding of where this difference comes from. I think I'm in a good position to do this: I'm close to both groups, have some technical background as a programmer, and have some time. I see two ways this could go:
If after looking into it more I still think AI risk is not a valuable place to be working, I may be able to convince others of this. Since 80000 Hours and other EAs are currently suggesting a lot of people go into this field, if it turns out we're overvaluing it then those people could work on other things.
If I change my mind and start thinking AI risk is something we should be working on, I may convince some of my friends in machine learning. It's also likely that something in this direction would be close enough to my skills to be a good career fit and I should consider working on it.
Of course it's also possible that I won't get to the root of the disagreement, or that I won't convince anyone except myself, but I do think it's worth trying.
Rough plan: read a bunch of stuff to get background, talk to a lot of people, write things up. Things I'm planning to read:
80000 Hours' Artificial intelligence and the 'control problem'
WaitButWhy's The AI Revolution: The Road to Superintelligence
Nick Bostrom's Superintelligence
Dario Amodei and others' Concrete Problems in AI Safety
Eliezer Yudkowsky's AI Alignment: Why It's Hard, and Where to Start
Paul Christiano's Three impacts of machine intelligence and Technical and social approaches to AI safety [EDIT: also Prosaic AI alignment]
Scott Alexander's AI Researchers on AI Risk and No Time Like the Present for AI Safety Work
EDIT: also Holden Karnofsky's Some Background on Our Views Regarding Advanced Artificial Intelligence and Potential Risks from Advanced Artificial Intelligence: The Philanthropic Opportunity
EDIT: also Luke Muehlhauser's What Do We Know about AI Timelines? and Replies to people who argue against worrying about long-term AI safety risks today
EDIT: also Kaj Sotala and Roman Yampolskiy's Responses to catastrophic AGI risk: a survey
EDIT: also Dario Amodei and others' Learning from Human Preferences
EDIT: also Brian Tomasik's Artificial Intelligence and Its Implications for Future Suffering
Maybe some of 80000 Hours' AI safety syllabus as it looks relevant.
Maciej Ceglowski's Superintelligence: The Idea That Eats Smart People
EDIT: also Ernest Davis' Ethical Guidelines for A Superintelligence
The list above is entirely people who think AI risk should be prioritized, aside from the Ceglowski post at the end, so I'm especially interested to read (if they exist) pieces where machine learning experts talk about why they don't think AI risk is a high priority. I'm also interested in other general AI risk background reading, and suggestions of people to talk to.
Comment via: google plus, facebook