Wow thank you for the awesome reply. If all the people in the Less Wrong community are as friendly and as knowledgeable as you are then I have obviously joined the right site. You were right I was talking about Roko’s Basilisk and since it is okay to mention it, here is the article that introduced me to this site if anyone is interested. I will definitely check out the Sequences in addition to the articles you suggested. There is so much interesting information on this site that it is hard to know where to start. One question I do have is what exactly is the importanceof decision theories? That is another thing that I am interested in. Are they applicable in real life situations or only in thought experiments? What is the importance of finding a perfect decision theory? I know the basics of Causal and Evidential Decision Theory but I am baffled by Timeless Decision Theory. If you could point me in the direction of where to find articles on these issues that would be greatly appreciated. Thank you again for the thoughtful and useful reply, it helped a lot.
Edit: I started reading Mysterious Answers to Mysterious Questions today and found it so engaging that I didn’t stop reading until I finished it. It was definitely a mind opening experience for me as I was exposed to a plethora of ideas and biases that I had no idea existed. I am definitely going to try reading the rest of the Sequences now.
I know the basics of Causal and Evidential Decision Theory but I am baffled by Timeless Decision Theory. If you could point me in the direction of where to find articles on these issues that would be greatly appreciated. Thank you again for the thoughtful and useful reply, it helped a lot.
If you want to get a handle on the “Less Wrong” approach to decision theory, I’d recommend starting with Wei Dai’s Updateless Decision Theory (UDT) rather than with Timeless Decision Theory (TDT). The basic mathematical outline of UDT is more straightforward, so you will be up and running quicker.
Wei’s posts introducing UDT are here and here. I wrote a brief write-up that just gives a precise description of UDT without any motivation, justification, or examples.
One question I do have is what exactly is the importance of decision theories? [...] Are they applicable in real life situations or only in thought experiments?
One of the main functions of a good decision theory is to bridge the territory-map divide: by solving problems in your head, it shows you how to solve problems in the real world. You can identify a good decision theory when it works in theory and in practice. If a decision theory seems to work in practice, but is not describable in a precise language (e.g. “do what feels good”), it actually hasn’t been well thought out and puts you at risk of being paralyzed when a very serious and very complex situation arises. On the other hand, if it only works in theory but is impracticable (e.g. “pray to Minerva for an omen”), it will be a waste of storage space in your head. In short, a decision theory should serve as a tool for you to manage your life.
Hmm. It does capture most of the essence of TDT, doesn’t it? See for example the last paragraph of chapter 12 and the last two paragraphs of chapter 13.3 in the TDT paper. I disagree with the “just” in the grandparent, but given e.g. “mostly”? Maybe I’m reading too much into the one-sentence description, though.
Hmm. It does capture most of the essence of TDT, doesn’t it?
No. Most of the interesting applications of TDT are about producing the same (or complimentary) outputs with different input. Moreover that description doesn’t even imply making a correct decision on Newcomblike problems (the motivation for producing TDT in the first place). In fact, CDT augmented by the assumption that two copies of the same algorithm with the same input will always yield the same result yields CDT.
To get closer to an (oversimplified) ‘essence’ of TDT I’d instead suggest building from the title. CDT augmented by not caring about which point on the time dimension you are in.
Wow thank you for the awesome reply. If all the people in the Less Wrong community are as friendly and as knowledgeable as you are then I have obviously joined the right site. You were right I was talking about Roko’s Basilisk and since it is okay to mention it, here is the article that introduced me to this site if anyone is interested. I will definitely check out the Sequences in addition to the articles you suggested. There is so much interesting information on this site that it is hard to know where to start. One question I do have is what exactly is the importanceof decision theories? That is another thing that I am interested in. Are they applicable in real life situations or only in thought experiments? What is the importance of finding a perfect decision theory? I know the basics of Causal and Evidential Decision Theory but I am baffled by Timeless Decision Theory. If you could point me in the direction of where to find articles on these issues that would be greatly appreciated. Thank you again for the thoughtful and useful reply, it helped a lot.
Edit: I started reading Mysterious Answers to Mysterious Questions today and found it so engaging that I didn’t stop reading until I finished it. It was definitely a mind opening experience for me as I was exposed to a plethora of ideas and biases that I had no idea existed. I am definitely going to try reading the rest of the Sequences now.
Three motivations are common around here:
Building a Friendly AI that is based on decision theory.
Understanding what ideal rationality looks like, so we have a better idea of what to aim for as far as improving our own rationality.
Curiosity. If we knew what the perfect decision theory was, many philosophical questions may be answered or would be closer to being answered.
For some relevant posts, see 1 and 2.
Thank you for the clear and informative reply.
If you want to get a handle on the “Less Wrong” approach to decision theory, I’d recommend starting with Wei Dai’s Updateless Decision Theory (UDT) rather than with Timeless Decision Theory (TDT). The basic mathematical outline of UDT is more straightforward, so you will be up and running quicker.
Wei’s posts introducing UDT are here and here. I wrote a brief write-up that just gives a precise description of UDT without any motivation, justification, or examples.
Just wanted to say you’re off to a great start posting to LW—asking very good questions!
(Also, please break posts like this into more than one paragraph.)
Thank you I’m just trying to learn all I can.
One of the main functions of a good decision theory is to bridge the territory-map divide: by solving problems in your head, it shows you how to solve problems in the real world. You can identify a good decision theory when it works in theory and in practice. If a decision theory seems to work in practice, but is not describable in a precise language (e.g. “do what feels good”), it actually hasn’t been well thought out and puts you at risk of being paralyzed when a very serious and very complex situation arises. On the other hand, if it only works in theory but is impracticable (e.g. “pray to Minerva for an omen”), it will be a waste of storage space in your head. In short, a decision theory should serve as a tool for you to manage your life.
TDT just augments CDT by saying that running two copies of the same algorithm with the same input will always yield the same result.
What? No it doesn’t. That’s not remotely what TDT says. That isn’t even a claim with particularly relevance to decision theory.
Hmm. It does capture most of the essence of TDT, doesn’t it? See for example the last paragraph of chapter 12 and the last two paragraphs of chapter 13.3 in the TDT paper. I disagree with the “just” in the grandparent, but given e.g. “mostly”? Maybe I’m reading too much into the one-sentence description, though.
No. Most of the interesting applications of TDT are about producing the same (or complimentary) outputs with different input. Moreover that description doesn’t even imply making a correct decision on Newcomblike problems (the motivation for producing TDT in the first place). In fact, CDT augmented by the assumption that two copies of the same algorithm with the same input will always yield the same result yields CDT.
To get closer to an (oversimplified) ‘essence’ of TDT I’d instead suggest building from the title. CDT augmented by not caring about which point on the time dimension you are in.