Note: I have no relation to MIRI/CFAR, no familiarity with this situation and am not a metal health expert, so I can’t speak with any specific authority here.
First, I’d like to offer my sympathy for the suffering you described. I’ve had unpleasant intrusive thoughts before. They were pretty terrible, and I’ve never had them to the degree you’ve experienced. X/S risk research tends to generate a lot of intrusive thoughts and general stress. I think better community norms/support in this area could help a lot. Here is one technique you may find useful:
Raise your hand in front of your face with your palm facing towards you.
Fix your eyes on the tip of a particular finger.
Move your hand from side to side, while still tracking the chosen finger with your eyes (head remains still).
Every time your hand changes direction, switch which finger your eyes track. I.e., first track the tip of the thumb, then track the pointer finger, then the middle, then ring, then pinky, then back to thumb.
This technique combines three simultaneous control tasks (moving your hand, tracking the current finger, switching fingers repeatedly) and also saturates your visual field with the constantly moving background. I find it captures my full attention for as long as a I perform the technique and is therefore useful for interrupting intrusive/obsessive thoughts.
Another option for dealing with intrusive thoughts is to just start doing squats or similar exercises. This is less distracting from the thoughts, but physically healthier and may serve as negative reinforcement to your brain about avoiding intrusive thoughts in the future.
I was told, by Nate Soares, that the pieces to make AGI are likely already out there and someone just has to put them together. He did not tell me anything about how to make such an AGI, on the basis that this would be dangerous. Instead, he encouraged me to figure it out for myself, saying it was within my abilities to do so.
Did Nate Soares specify that he wanted you to come up with a workable AGI design? If I’d heard my supervisor ask that of me, my interpretation would be something like: “Think about what sort of extra capabilities would make current AI more “agent-y” or general, and how such capabilities might be implemented.”
For example, I might think that adding some sort of multi-episode memory to GPT-3 would make it more agent-y, and think about ways to integrate memory with transformers. Another idea in this ballpark is to train an ML system on human brain activity data as a form of knowledge distillation from humans to AI (as described by Gwern here).
(Neither of these is among my most “dangerous” ideas about improving AI capabilities. I’m comfortable sharing the first because ~everyone already knows improved memory should be useful for general systems. I’m sharing the second because I think it will make many alignment approaches easier, especially value learning, interpretability, alignment by default and HCH).
I’d assume only ~20% at best of the approaches I thought of would actually be useful for capabilities, even if they were well implemented by a team of people smarter than me. If my supervisor then specified that they expected me to produce an actually workable AGI design, with significantly better performance than current state of the art systems, I’d have been blown away and seriously questioning their sanity/competence.
If I came up with a non-workable AGI design, that would not be significant evidence for “the pieces to make AGI are already out there and someone just needs to put them together”. Lots of AI people throughout the history of the field have come up with non-workable AGI designs, including me in high school/college.
Note: I have no relation to MIRI/CFAR, no familiarity with this situation and am not a metal health expert, so I can’t speak with any specific authority here.
First, I’d like to offer my sympathy for the suffering you described. I’ve had unpleasant intrusive thoughts before. They were pretty terrible, and I’ve never had them to the degree you’ve experienced. X/S risk research tends to generate a lot of intrusive thoughts and general stress. I think better community norms/support in this area could help a lot. Here is one technique you may find useful:
Raise your hand in front of your face with your palm facing towards you.
Fix your eyes on the tip of a particular finger.
Move your hand from side to side, while still tracking the chosen finger with your eyes (head remains still).
Every time your hand changes direction, switch which finger your eyes track. I.e., first track the tip of the thumb, then track the pointer finger, then the middle, then ring, then pinky, then back to thumb.
This technique combines three simultaneous control tasks (moving your hand, tracking the current finger, switching fingers repeatedly) and also saturates your visual field with the constantly moving background. I find it captures my full attention for as long as a I perform the technique and is therefore useful for interrupting intrusive/obsessive thoughts.
Another option for dealing with intrusive thoughts is to just start doing squats or similar exercises. This is less distracting from the thoughts, but physically healthier and may serve as negative reinforcement to your brain about avoiding intrusive thoughts in the future.
Did Nate Soares specify that he wanted you to come up with a workable AGI design? If I’d heard my supervisor ask that of me, my interpretation would be something like: “Think about what sort of extra capabilities would make current AI more “agent-y” or general, and how such capabilities might be implemented.”
For example, I might think that adding some sort of multi-episode memory to GPT-3 would make it more agent-y, and think about ways to integrate memory with transformers. Another idea in this ballpark is to train an ML system on human brain activity data as a form of knowledge distillation from humans to AI (as described by Gwern here).
(Neither of these is among my most “dangerous” ideas about improving AI capabilities. I’m comfortable sharing the first because ~everyone already knows improved memory should be useful for general systems. I’m sharing the second because I think it will make many alignment approaches easier, especially value learning, interpretability, alignment by default and HCH).
I’d assume only ~20% at best of the approaches I thought of would actually be useful for capabilities, even if they were well implemented by a team of people smarter than me. If my supervisor then specified that they expected me to produce an actually workable AGI design, with significantly better performance than current state of the art systems, I’d have been blown away and seriously questioning their sanity/competence.
Thanks for the suggestion for intrusive thoughts.
If I came up with a non-workable AGI design, that would not be significant evidence for “the pieces to make AGI are already out there and someone just needs to put them together”. Lots of AI people throughout the history of the field have come up with non-workable AGI designs, including me in high school/college.
That’s a neat idea with the hand thing.