JavaScript isn't enabled in your browser, so this file can't be opened. Enable and reload.
Evaluating AI safety progress
See the accompanying blog post:
https://forum.effectivealtruism.org/posts/9iGFjYnRquxiy29jm/safety-timelines-how-long-will-it-take-to-solve-alignment
Sign in to Google
to save your progress.
Learn more
* Indicates required question
What is your P(doom | AGI, no alignment)?
Probability that all human potential is lost given a non-aligned AGI.
Your answer
By what year do you expect alignment to be solved?
Median estimate
.
Your answer
What is your probability estimate that we will solve alignment before humans become unable to control AI?
Your answer
By what year do you expect humans become unable to control AI?
Median estimate.
Your answer
Which of these metrics for AI safety progress do you like the most? (and add your own ideas)
*
Second-order estimates, e.g. amount of papers published in AI safety compared to the ratio of solar cell efficiency and papers in solar cell engineering
Researcher surveys, e.g. asking researchers how much their project will contribute to alignment and how far they are
Evaluating if projects were a success or not and comparing the amount of success to the total amount of projects
Qualitative "State of AI Safety" reports, videos, or posts
Developing a complexity model of the AI space and estimating ripple effects of alignment discoveries
Researching benchmarks for alignment, e.g. TruthfulQA, OpenAI Safety Gym, moral Q&A
Focusing on developing a new "theory of scientific progress"
Comparing AI safety progress to progress in mathematical problems
Display how close we are to the ability to implement an array of alignment proposals without showing "how close to AGI alignment" we are
Other:
Required
Which of these ways to guide AI safety research do you like the most? (and add your own ideas)
Evaluating existing projects and talking with authors about their probability of success and estimated progress towards alignment
Creating "Perspective papers", papers that focus on creating new perspectives on alignment
Collecting and curating a centralized repository of research ideas like
aisafetyideas.com
Incentivizing variety instead of quantity in AI safety
Creating bounties for solving alignment sub-problems
Get averaged expert perspectives on all major research agendas
Other:
What could be a parametric definition of an alignment solution? E.g. "Follows Asimov's three laws of robotics". You are very welcome to write something uncertain and creative.
Your answer
Any other comments?
Your answer
Your name
Your answer
Your email address
Your answer
Submit
Clear form
Never submit passwords through Google Forms.
This form was created inside of Apart Research.
Report Abuse
Forms
