AI Reading Group Thoughts (2/​?): Reconstructive Psychosurgery

Pre­vi­ously on AI Read­ing Group Thoughts

In a highly un­offi­cial meet­ing of the AI Read­ing Group (most of the par­ti­ci­pants hap­pened to be in a room and we started talk­ing about AI; there was no for­mal meet­ing time, no read­ing home­work, and no dessert), we me­an­dered around to dis­cussing res­cue simu­la­tions and the re­lated idea of re­con­struct­ing peo­ple based on their digi­tal (or phys­i­cal) foot­print plus mem­o­ries of their friends and fam­ily.

We had wildly differ­ing in­tu­itions about re­con­struc­tive per­son­hood. Scott thinks it should work fine to just use third party re­ports and your Live­jour­nal, and that perform­ing an ap­pro­pri­ately su­per­in­tel­li­gent pro­cess on this data should get close enough that he doesn’t ex­pect to have a prob­lem with re­sults (and thinks that “close enough” is a mean­ingful con­cept, in the same way that to­day!Scott is very similar to yes­ter­day!Scott). He’s in­clined to col­lapse the ques­tion of whether or not re­con­structed peo­ple are “re­ally” who­ever they’re sup­posed to be as op­posed to more or less similar (a dis­tinc­tion that ap­plies just as well in­tra-life­time). And he seems to think that hu­mans are made sig­nifi­cantly of low-gran­u­lar­ity parts like “in­tro­ver­sion” and thinks that weird hid­den thought pro­cesses might turn out to drop out of all the other con­straints on the prob­lem, the same way an alien en­g­ineer try­ing to build a car based on watch­ing videos might put an in­ter­nal com­bus­tion en­g­ine un­der the hood even if none of the videos popped the hood.

Kel­sey thinks the con­nec­tome has got to be enough to work with—it might get “you dur­ing a weird dream” or “you dur­ing a wacky drug trip”, pro­duc­ing ir­reg­u­lar tem­po­rary qualia out of poorly cho­sen elec­tri­cal im­pulses and chem­i­cals filling in un­knowns about the non-con­nec­tome fea­tures of your brain, but weird dreams and wacky drug trips are a re­cov­er­able state, at least com­pared to “dead”. We might not even need the whole con­nec­tome, be­cause we do seem to find peo­ple with var­i­ous brain dam­age to be “still them­selves”.

We had differ­ing in­tu­itions about how much peo­ple ever ac­tu­ally have over­lap­ping qualia states in their life­times. Kel­sey thinks the time she got hy­dro­gen per­ox­ide in her eye she was prob­a­bly hav­ing pretty similar ex­pe­riences to other mind­slices in ex­treme eye pain, but of course when the pain re­ceded she re­sumed be­ing Kel­sey and didn’t pro­ceed to be a differ­ent per­son who re­mem­bered be­ing sub­ject to unanaes­thetized oph­thalmic surgery; there were un­der­ly­ing ten­den­cies in her ar­chi­tec­ture that re­stored her traits, if not ex­act state, once the ex­treme stim­u­lus was gone. And states more in­ter­est­ing than “ex­treme eye pain” are prob­a­bly less likely to be shared be­tween peo­ple, given their greater num­ber of de­tails.

I think whether or not re­con­struc­tion works de­pends on con­tin­gent facts about hu­man mindspace, which a su­per­in­tel­li­gence can prob­a­bly figure out (and ap­ply the an­swer, if the an­swer is “yes, it works fine”) but might be re­ally hard to pin down even that fact with only con­ven­tional brains on the ques­tion.

It might be that each bit of in­for­ma­tion about some­one, even noisy in­for­ma­tion, rules out huge swathes of ways a hu­man could be, and that hu­mans can’t be so many differ­ent ways that this leaves you mean­ingfully un­cer­tain about which one you’re try­ing to grab. Maybe we’re bad ran­dom num­ber gen­er­a­tors, or good ran­dom num­ber gen­er­a­tors but pro­ceed­ing de­ter­minis­ti­cally from a small num­ber of seeds (not like “five”, but maybe like “four hun­dred thou­sand”) with a small range of val­ues each, in­ter­act­ing in ways that are pretty easy to iden­tify and un­der­stand with enough pro­cess­ing power and con­text in­for­ma­tion about hu­mans in the gen­eral case. As a sim­plified ex­am­ple, some­one in my Dis­cord says that in gen­er­a­tion 2 of the Poké­mon games, whether a Poké­mon is shiny or not is pos­si­ble to de­rive based on in­for­ma­tion about its speed stat. One can imag­ine the same stats of a given mon­ster hold­ing static whether it’s shiny or not, but in fact within the con­straints of the game some of those val­ues are con­tra­dic­tions. Hu­mans aren’t try­ing to fit onto Nin­tendo car­tridges, but we’re try­ing to fit onto hacky wet­ware; maybe we con­tain similar weird con­nec­tions be­tween seem­ingly un­re­lated fea­tures of the mind, and no­body is both (*rolls dice*) ex­actly six feet tall and (*spins wheel*) in­clined to use topic-com­ment sen­tence pat­terns more than 15% of the time.

It might also be that hu­mans have enough in­de­pen­dently vary­ing parts that small bits of noisy in­for­ma­tion don’t tell you much about other fea­tures of the per­son, and ma­jor parts of iden­tity just aren’t re­vealed in typ­i­cal pub­lic records or pri­vate rec­ol­lec­tions how­ever smart the in­spect­ing in­tel­li­gence is. This seems es­pe­cially likely to be true of cer­tain kinds of thoughts—mem­o­ries of dreams that you never dis­cuss, pri­vate opinions that you never find a rea­son to bring up, or even things you did dis­cuss but that only sub­stan­tially af­fected your pri­vate re­la­tion­ship with some­one else who is also dead and can’t con­tribute in­for­ma­tion to the pro­ject of re­s­ur­rect­ing you. Are those kinds of things nu­mer­ous? Are they im­por­tant? It might de­pend on the tar­get re­s­ur­rectee; it might de­pend on hu­man ar­chi­tec­ture in gen­eral.

No nominations.
No reviews.