Frank_R

Topic Contributions

Comments

A central AI alignment problem: capabilities generalization, and the sharp left turn

In my opinion there is a probability of >10%  that you are right, which means AGI will be developed soon and you have to solve some of the hard problems mentioned above. Do you have any reading suggestions for people who want to find out if they are able to make progress on these questions? On the MIRI website there is a lot of material. Something like "You should read this first.", "This is intermediate important stuff." and "This is cutting edge research." would be nice.   

My list of effective altruism ideas that seem to be underexplored

Thank you for the link to the paper. I find Alexander Vilenkins theoretical work very interesting. 

My list of effective altruism ideas that seem to be underexplored

Let us assume that a typical large but finite volume contains  happy simulations of you and  suffering copies of you, maybe Boltzmann brains or simulations made by a malevolent agent. If the universe is infinite, you have infinitely many happy and infinitely suffering copies of you and it is hard how to interpret this result.    

My list of effective altruism ideas that seem to be underexplored

I see two problems with your proposal:

  1. It is not clear if a simulation of you in a patch of spacetime that is not causally connected to our part of the universe is the same as you. If you care only about the total amount of happy experiences, this would not matter, but if you care about personal identity, it becomes a non-trivial problem. 
  2. You probably assume that the multiverse is infinite. If this is the case, you can simply assume that for every copy of you that lives for N years another copy of you that lives for N+1 years appears somewhere by chance. In that case there would be no need to perform any action.

I am not against your ideas, but I am afraid that there are many conceptual and physical problems that have to solved before. What is even worse is that there is no universally accepted method how to resolve this issues. So a lot of further research is necessary. 

My list of effective altruism ideas that seem to be underexplored

Thank you for your answers. With better brain preservation and a more detailed understanding of the mind it may be possible to resurrect recently deceased persons. I am more skeptical about the possibility to resurrect a peasant from the middle ages by simulating the universe backwards, but of course these are different issues.     

My list of effective altruism ideas that seem to be underexplored

Could you elaborate why we have to make choices before space colonisation if we want to survive beyond the end of the last stars? Until now, my opinion is that we can can "start solving heat death" a billion years in the future while we have to solve AI alignment in the next 50 - 1000 years.

Another thought of mine is that it is probably impossible to resurrect the dead by computing how the state of each neuron of a deceased person was at the time of her/his death. I think, you need to measure the state of each particle in the present with a very high precision and/or the computational requirements for a backward simulation are much too high. Unfortunately, I cannot provide a detailed mathematical argument. This would be an interesting research project; even if the only outcome is that a small group of people should change their cause area.     

Which possible AI impacts should receive the most additional attention?

It should be mentioned that all (or at least most) ideas to survive the heat death of the universe involve speculative physics. Moreover, you have to deal with infinities. If everyone is suffering but there is one sentient being that experiences a happy moment every million years, does this mean that there is an infinite amount of suffering and an infite amount of happiness and the future is of neutral value? If any future with an infinite amount of suffering is bad, does this mean that it is good if sentient life does not exists forever? There is no obvious answer to these questions.  

Which possible AI impacts should receive the most additional attention?

Other S-risks that may or may not sound more plausible are suffering simulations (maybe an AI comes to the conclusion that a good way to study humans is to simulate earth at the time of the Black Death) or suffering subroutines (maybe reinforcement learners that are able to suffer enable faster or more efficient algorithms). 

[Closed] Hiring a mathematician to work on the learning-theoretic AI alignment agenda

I have noticed that there are two similar websites for mathematical jobs. www.mathjobs.org is operated by the American Mathematical Society and is mostly for positions at universities, although they list jobs at other research institutions, too. www.math-jobs.com redirects you to www.acad.jobs , which has a broader focus. They advertise also government and industry jobs and it is also for job offers in computer science and other academic disciplines. 

You have to register on both websites as an employer for several hundreds of dollars before you can post a job offer. I do not know if this is to much. Both sites are probably among the top ten for math related research positions, although this is only based on my gut feelings. Unfortunately, I cannot tell you if it is possible to post remote only jobs. I hope this information helps. 

[Closed] Hiring a mathematician to work on the learning-theoretic AI alignment agenda

How much knowledge about AI alignment apart from the right mathematical background is necessary for this position? If the job is suitable for candidates without prior involvment in x-risks / longtermism / Effective Altruism,  it may be a good idea to  announce it at places as mathjobs.org.   

Load More