I’m excited (and worried) about pursuing Whole Brain Emulation. I think that being able to study such an emulation to get a better idea about human values and human-human alignment successes and failures would be a boon. I also think finding a simplified brain-inspired AI with the useful aspects of the brain and also the useful aspects of current deep learning techniques would potentially be key for a pivotal act, or at least enforcing a delay long enough to figure out something better.
I rather suspect we’re going to find ourselves in a transitional mixed-up muddle of many semi-aligned not-quite-trustworthy variants on ML models, which will not quite be good enough for anything approaching a pivotal act, but might be good enough for awkwardly muddling through on a Superalignment-type agenda. My best guess is that knowing more about the positive aspects of the brain could be quite helpful to humanity at such a juncture.
So I’m trying to figure out things like: “How is it that the general learning hardware of the cortex ends up with such specific and consistent-across-individuals localizations of function?”
I’m excited (and worried) about pursuing Whole Brain Emulation. I think that being able to study such an emulation to get a better idea about human values and human-human alignment successes and failures would be a boon. I also think finding a simplified brain-inspired AI with the useful aspects of the brain and also the useful aspects of current deep learning techniques would potentially be key for a pivotal act, or at least enforcing a delay long enough to figure out something better.
I rather suspect we’re going to find ourselves in a transitional mixed-up muddle of many semi-aligned not-quite-trustworthy variants on ML models, which will not quite be good enough for anything approaching a pivotal act, but might be good enough for awkwardly muddling through on a Superalignment-type agenda. My best guess is that knowing more about the positive aspects of the brain could be quite helpful to humanity at such a juncture.
So I’m trying to figure out things like: “How is it that the general learning hardware of the cortex ends up with such specific and consistent-across-individuals localizations of function?”