I suspect that humans will turn out to be relatively simple to encode—quite small amounts of low-resolution memory that we draw on, with detailed understanding maps—smaller than LLMs that we’re creating. Added to which there is an array of motivation factors that will be quite universal but of varying levels of intensity in different dimensions for each individual.
If that take on things is correct then it may be that emulating a human by training a skeleton AI using constant video streaming etc over a 10-20 year period (about how long neurons last before replacement) to optimally better predict behaviour of the human being modelled will eventually arrive at an AI with almost exactly the same beliefs and behaviours as the human being emulated.
Without physically carving up brains and attempting to transcribe synaptic weightings etc that might prove the most viable means of effective up-loading and creation of highly aligned AI with human like values. And perhaps would create something closer to being our true children-of-the-mind
For AGI alignment; seems like there will at minimum need to be a perhaps multiple blind & independent hierarchies of increasingly smart AIs continually checking and assuring that next level up AIs are maintaining alignment with active monitoring of activities, because as AIs get smarter their ability to fool monitoring systems will likely grow as the relative gulf between monitored and monitoring intelligence grows.
I think a wide array of AIs is a bad idea. If there is a non-zero chance that an AI goes ‘murder clippy’ and ends humans, then that probability is additive—more independent AIs = higher chance of doom.
If that take on things is correct then it may be that emulating a human by training a skeleton AI using constant video streaming etc over a 10-20 year period (about how long neurons last before replacement) to optimally better predict behaviour of the human being modelled will eventually arrive at an AI with almost exactly the same beliefs and behaviours as the human being emulated.
That’s the premise of Greg Egan’s “Jewel” stories. I think it’s wrong. A person who never saw a spider will still get scared when seeing one for the first time, because humans are hardwired for that. A person who has a specific memory and doesn’t mention it to anyone for many years, probably doesn’t give enough information through their behavior to infer the memory in detail. And the extreme example of why input/output is not enough to infer everything about inner life: imagine a human in a box, which has no input/output at all, but plenty of inner life. I think we all have lots of inner degrees of freedom, that can’t be fully determined even from a full record of our behavior over a long time.
I suspect that humans will turn out to be relatively simple to encode—quite small amounts of low-resolution memory that we draw on, with detailed understanding maps—smaller than LLMs that we’re creating. Added to which there is an array of motivation factors that will be quite universal but of varying levels of intensity in different dimensions for each individual.
If that take on things is correct then it may be that emulating a human by training a skeleton AI using constant video streaming etc over a 10-20 year period (about how long neurons last before replacement) to optimally better predict behaviour of the human being modelled will eventually arrive at an AI with almost exactly the same beliefs and behaviours as the human being emulated.
Without physically carving up brains and attempting to transcribe synaptic weightings etc that might prove the most viable means of effective up-loading and creation of highly aligned AI with human like values. And perhaps would create something closer to being our true children-of-the-mind
For AGI alignment; seems like there will at minimum need to be a perhaps multiple blind & independent hierarchies of increasingly smart AIs continually checking and assuring that next level up AIs are maintaining alignment with active monitoring of activities, because as AIs get smarter their ability to fool monitoring systems will likely grow as the relative gulf between monitored and monitoring intelligence grows.
I think a wide array of AIs is a bad idea. If there is a non-zero chance that an AI goes ‘murder clippy’ and ends humans, then that probability is additive—more independent AIs = higher chance of doom.
That’s the premise of Greg Egan’s “Jewel” stories. I think it’s wrong. A person who never saw a spider will still get scared when seeing one for the first time, because humans are hardwired for that. A person who has a specific memory and doesn’t mention it to anyone for many years, probably doesn’t give enough information through their behavior to infer the memory in detail. And the extreme example of why input/output is not enough to infer everything about inner life: imagine a human in a box, which has no input/output at all, but plenty of inner life. I think we all have lots of inner degrees of freedom, that can’t be fully determined even from a full record of our behavior over a long time.