Difference between revisions of "List of AI safety projects I could work on"

From Issawiki
Jump to: navigation, search
(Exposition of technical topics)
(14 intermediate revisions by the same user not shown)
Line 1: Line 1:
 
(November 2020)
 
(November 2020)
  
==Writing up my opinions==
+
evening tier = I can make useful/meaningful progress within one hour and even when I am somewhat tired from having done other stuff during the day (at the limit, think of things like [[Duolingo]] where each task can be accomplished in a few seconds). This doesn't mean I can ''complete'' a task within one hour; completion may require several evenings. But at least a single evening can make progress in a way that doesn't require "re-doing" the same step (because I lost whatever context I had when working on it initially), and a single evening can also make definite progress (rather than feeling lost or demotivated because it requires having good ideas and I can't generate ideas when I'm tired).
 +
 
 +
weekend tier = I can make useful progress if I make it my main thing for a single day, e.g. on a weekend day. Alternatively, I can make progress within 1-2 hours as long as I am mentally alert and feeling fine.
  
* writing some sort of overview of my beliefs regarding AI safety. like, if i was explaining things from scratch to someone, what would that sound like?
+
vacation tier = I need more than a day or two of concentrated "good" time to focus on this, either because the task itself is not clearly defined (so I need to do lots of brainstorming and thinking to figure out next steps), or because it requires keeping lots of things in my head so I cannot be context-switching with other tasks (i.e. if I work on it for a few hours, then I might get distracted by another project and forget what I was doing)
* my current take on [[AI timelines]]
 
* my current take on [[AI takeoff]]
 
* my current take on MIRI vs Paul
 
  
==Research projects==
+
==Increase activity on LessWrong==
  
* continue working out [[AI takeoff]] disagreements
+
* Ask lots of questions on LW, e.g.:  (duration tier depends on question; I believe I can rush through and ask certain questions within an hour which would be evening tier or weekend tier)
* continue working out MIRI vs Paul
+
** Applicability of FDT/UDT/TDT to everyday life https://github.com/riceissa/project-ideas/issues/44
* HRAD paper with [[David Manheim]]
 
* concrete plausible scenarios for what could happen when AGI comes around
 
* deep dive into human evolution to figure out what the heck it might tell us about AI takeoff/AI timelines
 
* comparison of AI and nukes
 
* think about [[AI polarization]], using examples like COVID and climate change
 
  
==AI safety wiki==
+
==Writing up my opinions==
  
* Writing articles for AI safety wiki
+
* writing some sort of overview of my beliefs regarding AI safety. like, if i was explaining things from scratch to someone, what would that sound like? (vacation tier)
 +
* my current take on [[AI timelines]] (vacation tier)
 +
* my current take on [[AI takeoff]] (vacation tier)
 +
* my current take on MIRI vs Paul (vacation tier)
  
==The Vipul Strategy==
+
==Personal reflection==
  
* Expanding [[AI Watch]] to cover agendas, documents, more people/orgs with more columns filled in, graphs, automated [https://aiwatch.issarice.com/compare.php?by=organization&for=2020 yearly review] of orgs, etc.
+
* reflection post about getting involved with AI safety (vacation tier)
* Updates to timeline of AI safety and other relevant timelines
 
* Timeline of [[Eliezer Yudkowsky]] publications https://github.com/riceissa/project-ideas/issues/16
 
* Wikipedia pages for AGI projects https://github.com/riceissa/project-ideas/issues/22
 
  
==Increase activity on LessWrong==
+
==Exposition of technical topics==
  
* Ask lots of questions on LW, e.g.:
+
* [[Solomonoff induction]] guide (I think I've already figured out things here that are not explained anywhere, so I think I could write the best guide on it, but it's not clear how important this is for people to understand) (vacation tier)
** Applicability of FDT/UDT/TDT to everyday life https://github.com/riceissa/project-ideas/issues/44
+
* [[Judea Pearl|Pearl]] [[belief propagation]] guide (vacation tier)
 +
* Summarizing/distilling work that has been done in [[decision theory]] (vacation tier)
 +
* starting something like [[RAISE]]; see [[There is room for something like RAISE]] (vacation tier)
 +
* Doing something more like "tools for thought" or "[[explanation science]]" research, with AI safety as the [[serious context of use]]
  
==Exposition of technical topics==
+
==AI safety wiki==
  
* [[Solomonoff induction]] guide (I think I've already figured out things here that are not explained anywhere, so I think I could write the best guide on it, but it's not clear how important this is for people to understand)
+
* Writing articles for AI safety wiki (some pages are evening tier or weekend tier, but thinking strategically about what to write for the wiki is vacation tier)
* [[Judea Pearl|Pearl]] [[belief propagation]] guide
 
* Summarizing/distilling work that has been done in [[decision theory]]
 
* starting something like [[RAISE]]; see [[There is room for something like RAISE]]
 
  
==Personal reflection==
+
==The Vipul Strategy==
  
* reflection post about getting involved with AI safety
+
* Expanding [[AI Watch]] to cover agendas, documents, more people/orgs with more columns filled in, graphs, automated [https://aiwatch.issarice.com/compare.php?by=organization&for=2020 yearly review] of orgs, etc. (duration tier depends on task -- adding data is evening tier but bigger strategic additions are weekend tier or vacation tier)
 +
* Updates to timeline of AI safety and other relevant timelines (evening tier)
 +
* Timeline of [[Eliezer Yudkowsky]] publications https://github.com/riceissa/project-ideas/issues/16 (evening tier)
 +
* Wikipedia pages for AGI projects https://github.com/riceissa/project-ideas/issues/22 (vacation tier to figure this out; writing an individual page is likely weekend tier)
  
==Philosophy==
+
==Research projects==
  
* human values/[[deliberation]]
+
* continue working out [[AI takeoff]] disagreements (vacation tier)
 +
* continue working out MIRI vs Paul (vacation tier)
 +
* HRAD paper with [[David Manheim]] (vacation tier)
 +
* concrete plausible scenarios for what could happen when AGI comes around (vacation tier)
 +
* deep dive into human evolution to figure out what the heck it might tell us about AI takeoff/AI timelines (vacation tier)
 +
* comparison of AI and nukes (vacation tier)
 +
* think about [[AI polarization]], using examples like COVID and climate change (vacation tier)
 +
* write up my understanding of people's views on [[whole brain emulation]] (vacation tier)
 +
* clarify my thoughts on probability of saving the world via x-risk reduction (vacation tier)
 +
* Philosophy
 +
** human values/[[deliberation]] (vacation tier)
  
 
==Learn stuff==
 
==Learn stuff==
  
* learn more machine learning so I can better follow some discussions
+
* learn more machine learning so I can better follow some discussions (weekend tier as long as I am diligent to use [[Anki]] to review partial progress)
 +
* learn more economics. I somewhat often feel confused about how to think about various AI strategy questions (e.g. history of growth, how AI systems might cooperate and increase economies of scale, what percentage of GDP might be used for computing costs) and I suspect part of the reason is that I don't know enough economics. (weekend tier as long as I am diligent to use [[Anki]] to review partial progress)
 +
* understand the [[neuromorphic AI]] pathways that people like [[Steve Byrnes]] and [[gwern]] have been talking about (evening tier)
 +
* there are a lot of blog posts that I could catch up on (evening tier)
 +
 
 +
==Technical AI safety==
 +
 
 +
* I still don't feel motivated to do this since I don't feel convinced by any existing worldview/visualization that has been put forth (by e.g. MIRI or Paul) (vacation tier)
  
 
==See also==
 
==See also==

Revision as of 12:27, 2 December 2021

(November 2020)

evening tier = I can make useful/meaningful progress within one hour and even when I am somewhat tired from having done other stuff during the day (at the limit, think of things like Duolingo where each task can be accomplished in a few seconds). This doesn't mean I can complete a task within one hour; completion may require several evenings. But at least a single evening can make progress in a way that doesn't require "re-doing" the same step (because I lost whatever context I had when working on it initially), and a single evening can also make definite progress (rather than feeling lost or demotivated because it requires having good ideas and I can't generate ideas when I'm tired).

weekend tier = I can make useful progress if I make it my main thing for a single day, e.g. on a weekend day. Alternatively, I can make progress within 1-2 hours as long as I am mentally alert and feeling fine.

vacation tier = I need more than a day or two of concentrated "good" time to focus on this, either because the task itself is not clearly defined (so I need to do lots of brainstorming and thinking to figure out next steps), or because it requires keeping lots of things in my head so I cannot be context-switching with other tasks (i.e. if I work on it for a few hours, then I might get distracted by another project and forget what I was doing)

Increase activity on LessWrong

  • Ask lots of questions on LW, e.g.: (duration tier depends on question; I believe I can rush through and ask certain questions within an hour which would be evening tier or weekend tier)

Writing up my opinions

  • writing some sort of overview of my beliefs regarding AI safety. like, if i was explaining things from scratch to someone, what would that sound like? (vacation tier)
  • my current take on AI timelines (vacation tier)
  • my current take on AI takeoff (vacation tier)
  • my current take on MIRI vs Paul (vacation tier)

Personal reflection

  • reflection post about getting involved with AI safety (vacation tier)

Exposition of technical topics

AI safety wiki

  • Writing articles for AI safety wiki (some pages are evening tier or weekend tier, but thinking strategically about what to write for the wiki is vacation tier)

The Vipul Strategy

Research projects

  • continue working out AI takeoff disagreements (vacation tier)
  • continue working out MIRI vs Paul (vacation tier)
  • HRAD paper with David Manheim (vacation tier)
  • concrete plausible scenarios for what could happen when AGI comes around (vacation tier)
  • deep dive into human evolution to figure out what the heck it might tell us about AI takeoff/AI timelines (vacation tier)
  • comparison of AI and nukes (vacation tier)
  • think about AI polarization, using examples like COVID and climate change (vacation tier)
  • write up my understanding of people's views on whole brain emulation (vacation tier)
  • clarify my thoughts on probability of saving the world via x-risk reduction (vacation tier)
  • Philosophy

Learn stuff

  • learn more machine learning so I can better follow some discussions (weekend tier as long as I am diligent to use Anki to review partial progress)
  • learn more economics. I somewhat often feel confused about how to think about various AI strategy questions (e.g. history of growth, how AI systems might cooperate and increase economies of scale, what percentage of GDP might be used for computing costs) and I suspect part of the reason is that I don't know enough economics. (weekend tier as long as I am diligent to use Anki to review partial progress)
  • understand the neuromorphic AI pathways that people like Steve Byrnes and gwern have been talking about (evening tier)
  • there are a lot of blog posts that I could catch up on (evening tier)

Technical AI safety

  • I still don't feel motivated to do this since I don't feel convinced by any existing worldview/visualization that has been put forth (by e.g. MIRI or Paul) (vacation tier)

See also