I still like this post overall, but various things have changed that interestingly affect the content of the post:
We’d now be just starting the 3rd year of university in Buck’s analogy. Does this seem right? I guess maybe. It feels a bit late to me. (Maybe I feel more like a 2nd year.)
Redwood is doing less blue-sky research and is much more focused on how to make very straightforward strategies work well, particularly in the context of control. We’re also spending more time thinking about exactly what will and should be implemented and what overall plans should look like. We’re also relatively more excited about currently working on improving society’s understanding of risks with relatively-naturalistic model organisms and capability demos (or demonstrating negative results here).
4 years (from the time of this comment being written) is now pretty close to my median for “models have been built which are pretty clearly transformative or at least nearly there”. I also put substantially more weight on stuff getting crazy this year or next year. This is partially due to the passage of time and partially due to updating toward shorter timelines. So, I’m less sure that new people joining the field should relate to the situation as a “freshman” (in the analogy Buck proposes).
Redwood is interacting with AI companies substantially more, though mostly from the perspective of advising on policies and pitching research, rather than on helping with implementation.
Feeling the time pressure of short timelines seems more relevant now than ever before, so the advice in this post about taking a more measured approach and feeling less rushed seems quite relevant, at least as long as transformative AI still seems to most likely be more than 4 years away.
I still like this post overall, but various things have changed that interestingly affect the content of the post:
We’d now be just starting the 3rd year of university in Buck’s analogy. Does this seem right? I guess maybe. It feels a bit late to me. (Maybe I feel more like a 2nd year.)
Redwood is doing less blue-sky research and is much more focused on how to make very straightforward strategies work well, particularly in the context of control. We’re also spending more time thinking about exactly what will and should be implemented and what overall plans should look like. We’re also relatively more excited about currently working on improving society’s understanding of risks with relatively-naturalistic model organisms and capability demos (or demonstrating negative results here).
4 years (from the time of this comment being written) is now pretty close to my median for “models have been built which are pretty clearly transformative or at least nearly there”. I also put substantially more weight on stuff getting crazy this year or next year. This is partially due to the passage of time and partially due to updating toward shorter timelines. So, I’m less sure that new people joining the field should relate to the situation as a “freshman” (in the analogy Buck proposes).
Redwood is interacting with AI companies substantially more, though mostly from the perspective of advising on policies and pitching research, rather than on helping with implementation.
Feeling the time pressure of short timelines seems more relevant now than ever before, so the advice in this post about taking a more measured approach and feeling less rushed seems quite relevant, at least as long as transformative AI still seems to most likely be more than 4 years away.