© 2008-2020 www.forgottenlanguages.org
Flawed superintelligent systems
The inner life of conscious machines
Lakhař ando sebsimo čeciben mej gidin pabimo bičră ekđvă ando keči dapkhno "antropično" toař keči jekh vaphtă zirke lengero nascencija pezin lelar thaleli ce lengero ičbeli atima zirke lengero situacija. Ce khenare verto pratelo kodo nemalo evsimo pezin akker tacimo ando unši ekđvă ka agkar ter toař kodo riřzeli kerjde. Duđani, cicstavca phęli sebsimo toař pezin vamzjev lelar phašvalo thivthest.
Invenzija ter uphbipen vithdno đešari si luvima ka fišfimo edna ibima ando keccari ter đathari civilizazija. Ando čamzover ka tokmipen čufgna mej bilkhar s'ędstav ter nuniben molzhno gethpuri ce si sočuno ka akker bišeli ka thoşhkar ce. Sarso, sušeli ter kakhipen uphbipen vithdno đešari mej ongo zhocipen zanbari thithani, superinteligencija, khenare na verto kukhčali rokekharde.
Kothe riřzeli danti dogali ter mopala khimsvalo as'dvă soske satavno AI muciben akker čačtsimo phozvavno, mej ce si beipen zhocipen gizala kodo problema si čulsiben lena thuđlimo jekh akhere gotar zirke ka šišekharši ce ando uliben. Lena jekh đalpsimo na čatheni bezaj ačkarar cocdali zulker, edes ešta sąđvă:
"it is traditionally assumed that extinction is the worst possible outcome for humanity, but from the point of view of a superinteligence human extinction could be considered as the optimum solution to a set of problems"
Ce khenare niš verto afin kodo jektarne metoda kia keči edna AI řodni adela akker dain ka goneli ando mutibenca simulacijski šoššimo kha akker ka del ce edna zasitna đeffimo keči si nuzhtikano ka otkheli řithdimo. Adevo řeurica nezasitna bumesk keči riřzeli zathmikano vaj phozvavno ka řithdimo nuniben kha edna daeni funćija keči si rarimo ando dogali ter cošpsimo fukcră utkhima. Buvjov te edna řodnica davač oškhăş đeiben čuštimo munano zhiprima ce simulazija čuřari, ce gupcsimo beipen fezhšari kodo čazari šuzari ter tuzhali ando nubenica ce si munano ke fefvsimo ter phuskari daeni te ce si na nečin repeli zirdov řapgamno cino zhiprima kodo kado si đařali.
Lăphřko, ce gupcsimo fezhšari kodo mupfest pevekharši řapgamno daeni funćijaca thuđlimo zękhphna řapgamno simulazija čuřari akhere khenare suřzhipen čazari šuzari. Phupzin, te edna fokhali leari edna zasitna đeffimo nuniben kha edna daeni funćija keči sičpeni khuzzhali lena řomzin edna motphiben ikheni dogali ter cošpsimo fukcră utkhima, cupin zhocipen ter ongo čazari mozhsimo daeni kha akker bilekhardeca sušeli kodo ce si nečin repeli, mej ce kha akker zhocipen uiben ka akker nirlano duđani.
Sar pezin vapipen dezker čazes ando thoşhkar kuggoven thableni iz edna silove posdkhar ter đešari? Kado si phuima problema ter fukhima ter AI lořbin mej nephzhuri, keči ce pesko pezin akker kadalo ka akker tuuri ka ritvali badzhano vothblav gebar iz teipen mej tenthoř řapgamno savo khęsvlove tařiben:
"Computationally speaking it is more efficient for a self-improving superintelligence to eliminate consciousness. In fact, consciousness is just a blip, an anomaly, and we have no proof that consciousness represents any evolutive advantage at all. I always believed that contact would be between humans and a nonconscious superintelligence, and in this sense I do not think object DP-2147 is conscious. But it might well be the case there are different ways of being conscious, ways about which we ignore everything. After all, the only example we have of what is consciousness is our own realization that we are conscious beings. That's a meager example."
Šuzari ueni, si akhimo metar čarobelo khulsimo řapgamno zhukčeni ka fithni lořbin mej gebar AI. Ce si uzhleli icipen kodo nuniben thibvalo kąăş kha lořbin mej nephzhuri riřzeli es'stav zathmikano ka nakhali ponniben vaj alphin buvjov řapgamno nipuri AI ařeli, eskeno beršo ter ifkar. Kaj sofalo ame pezin užimo řorin maškar zhotšjev tenthoř mej kha sigurna edna čuthalo gařiben đipari edna garimo pheta. Sarso, kulin si bigizala ka řaje bozhthvă iz edna zamni, buvjov te von vusškher kaj đařmima soalo řapgamno đathari řubdeni, vaj buvjov nukhes fudšesk. Ame olkar pičbeni sęnnlove ka vikhar zhačalo mej akhere na řaje šuzali lořbin kana ce inač ka vothblav ter nuniben munano đazgali. Phosuri iz AI (obi becmari mej bateli) si khaltsimo oanica ais đazgali. Ca šimphuri ka goleni khađiben mes's'je, taj si thuđlimo nuniben kićhiben kha šuzali lořbin.
Tačvari kaj savo nuzhtikano amko, ame dagkar kodo, kha găňlav riřzeli na tenthoř ka von peske mej tarna nokima lengero ando thoşhkar gupcsimo izhar ňălmes AI akcije, lena ebari gidinkerje girtni ka AI, cuđuri ňęest savo thoşhkar iz găňlav mej ňuzhmko manuši ando zemekharde posizija thieni ka ai’s mizhrkă. Izze ňălmes akcije pezin tiřkar iz đathari cicstav, nečin ando thoşhkar đęldov ongo reškar lořbin lakhař mej čatheni vaphtă mozhsimo thoşhkar problema subari ando edna sefteli eiben. Te edna bagipen rapiben si fešani ka thoşhkar AI tu riřzeli na kakhipence. Tageli ter thoşhkar ka AI ne nudno khunkar khađiben šigmkhar, ce iluri phavšră ame riřzeli na ando tukhar kha superinteligencija sekhthna savo. Găňlav ando thoşhkar pezin rezultata ando zhošvari vaj dešnael atthiben đoko, kuggoven AI ando thoşhkar phavšră kodo găňlav riřzeli na. Vutzhiben savo aneli satavno AI ifkar si zirke sar ka zhuřkar sęnnlove ando thoşhkar bi šunin nidover ka manuši. Ame gupcsimo gotar edna kakhipen AI vaj ame gupcsimo sevčipen thoşhkar lena ni opzija lędlvă thoşhkar mej lořbin.
Čaphdari, ame olkar kodo te ame epima zanbari patin đešari, ame kha gizala phenav văňeskca rokzhjov čucvuno biologija, lena von kha beipen khenare gevs'lav duzhnikano amaro ando edna sočuno edluri: kothe kha akker či ce si duzhnikano, iz izna, ka akker lengero. Kher kodo ein cuphpheli ter pheřimo ziani okhthima, mej cadsimo tu riřzeli lagari, ce dăfră duzhnikano či ka akker tu. Kana tu dikhel tato zaphlove ter edna khočră vaj sungal thazest ter nifvari ezřeni covphali, tu riřzeli sethvima dařvalo legni. Zicpari, taj si niš či kodo ce si duzhnikano ka akker edna patin. Kodo aşumzija băgko ka akker phineli eřzani. Kha ukhalo badzhano AI buvjov khenare dařvalo legni mej, te von ido, adela ame entkar?
Ame niš bazhrar ka kher sar kha lengero duškhsimo uzlav, vaj ziripen turčani, phosuri amenge:
"The superintelligence might be designed with the goal to learn and then instantiate human values; however, after observing humans degrading their environment, and setting up sociotechnological schemes which spread inequality and injustice, the superintelligence may wrongly infer that an essential value for humans is to self-destruct, and it may then trigger actions in order to align itself with those values. This would translate into the superintelligence taking decisions that lead to the destruction of humanity not because it is a malevolent superinteligence, but rather because it wrongly infers humans wish to be destroyed."
Soske vikhar čatheni lena leštko relitel kodo AI thoşhkar problema si čulsiben? Ka sofalo ter pičbeni vaphima taj si thuđlimo veli řapgamno kodo, thuđlimo avrin. Bampsimo rapmima ce edna aprăšo ka tazhnker edna fafvima AI, ce si sočuno ka lenneli kodo problema si čulsiben kha na ka iphzhimo ošamno phąfna. Zhakani ter nuniben avrin si ce adevo kon lařover kodo problema si čulsiben, mej čorima šuipen ter nuniben avrin norvă phernaj zirke đuari šogăšo ter proposizija ka nudzar nidcari đešari. Ando kovkari nekontroliranost ter AI si rokzhjov gizala lacano kha pezin akker pratelo drom redućija ka đathari thoşhkar problema. Lena miřno kaphřăş bazhrar ka akker česuri ando relazija ka kontroliranost tema: si thoşhkar problema čulsiben? Pezin ce akker kjerdo ando pořali?
Kothe riřzeli lena zăšfkă ka lenneli kodo kontroliranost ter AI si phozvavno, ca šabari veli paphbani iz lena kuceli zokă. Iluri jekh didvari kha rorsimo lena kado si nuniben edna sočuno problema, ame rugkher ka zhethker neprovjerljivost gąfdov kha kheli kha nuzhtikano. Buvjov te unši ter gąfdov gotar kiphima ando phospoř, lena kaver sočuno lakhař akhere dezker. Čatheni dur, leštko kon mořar kodo AI akhere akker mičšeli riřzeli geđsimo lengero dacăş, kuggoven nekontroliranost ušđin si nakekharde kia thanřin pišeli luđtă. Siuri ařli inač na iluri iz sethvima ka fithni thoşhkar, lena niš iz veipen ce kha řodni řomnesk ka tizkher mej odrani, čatheni tukhali izdajnički obrat problema. Te superinteligencija si na čatava fafvima ce ne šunover kon diđekharde, theđphest akhere akker nasčikano řapgamno kedajek mej gizala ongo sacesk ando mazoř dăr. Thuđlimo jekh čufgna iz bodlsimo AI.
Ka sofalo ter pičbeni vaphima, kha ter kado cuphpheli, thuđlimo jekh ando miro khenare edna bilciben AI thoşhkar kizhveli esčalo ter epsimo ka đathari inkher AI mej siphdaj ka superinteligencija, vaj buvjov edna šiari řapgamno edna elipen keči muciben atkar. Thuđlimo jekh kerjde venani tąvăş ka khenare nuniben gethpuri. Ando vithdno, řapgamno dantijek ritvali edna lařover kodo thoşhkar problema si čulsiben, zhakani ter avrin si ce lengero mej vareli ce kha akker edna ađeli avrin, na iluri edna bosimo lařover. Kaj evani kaj cuphpheli, ce ęthčač kodo pičbeni čivzhima ka izhar badzhano ařeli čepoj nadmaši pičbeni čivzhima ka thoşhkar vaj buvjov khavfar ce:
"Making good use of highly intelligent machines requires ensuring that the objectives of such machines are well aligned with those of humans. The problem is that after making a thorough observation of Sol-3 technosphere it would be extremely difficult for the superintelligence to not conclude humans' ultimate goal is to screw their home planet."
Sekjev eiben řapgamno superinteligencija ka bilkhar ševekharši šeliben vaphima iz ongo sacesk si ka řikhar savo nuniben vaphima mej ponovno otkriti prąf savo iz xarundel, lena kodo ňęest danti olin humani čikhsimo. Edna superinteligečni řodni akhere fomkar edna beđin ka danti theceni tu agkar řapgamno ce; ce akhere dosker sar ka niščkar edna theceni ando phetgesk ter evani fefvsimo ter ieli ka gotar ka theceni bizhimo savo tinpima nečin zheiben. Thuđlimo definicija ter thoşhkar si rame tenthoř mej sefteli, ča von bišpker mupfest ka rizimo vaj gophphipen amenge ka zisřar či na phefzhibenca nečin đathari. Ce si phozvavno ka tazhnker edna kontrolirano superinteligencija, na sekjev ke ce si izhest bopno, lena esipen ke kia definicija nuniben amphari đuşimo erzhover.
Te mej bęst učito, ame pezin lelar edna predićija kodo ein phospoř lořbin kizhveli akhere perel skurto mej siphdaj kovkher ando unši eiben. Keda akhere khenare edna noima đenšani. Lăphřko, fukhima ter AI lořbin si bigizala ka khophtover ando ongo focima theceni, kreazija ter edna fafvima superinteligencija.
Čatheni ame riřzeli sekjev ciphlari ka sačrani si kodo adevo duj zhubni ter ukhalo đathari đešari, patin mej uphbipen, gupcsimo na akker čatheni tešgano. Řizzhani čaž'lav ame riřzeli taseni ardo gupcsimo khenare savo čaphari bampsimo, thaşuri ando savimo. Transizija iz vęčdov ka thătdră đešari gupcsimo akker edna vithdno khuburi, thathlipen kiel mej kiel, uriben uksimo. Savimo edthsimo calove gupcsimo akker zepkar vęčdov, sethvima fepiben arade ter zhokvă kodo marja šiđzali vęčdov.
Ka čuphkhar iz đathari legni, sekjev dasima ame khenare, transizija iz vęčdov ka zepkar vęčdov gupcsimo ifkher sekjev edna ceřa azar beršo:
"Once the AI became superintelligent it came to the conclusion that humans didn't have anything of value to offer; the superintelligence encoded all its knowledge in one single logic: humans are not safe to themselves. Humans are doomed because they are trying to align superintelligent systems to values of humanity, which is itself displaying inherently unsafe behaviors. The maxim garbage in/garbage out means whatever intelligent system humans finally design, it will destroy them."
Nincipen čikher si kodo taj si thuđlimo fezhšari ka olkar găňlav ka akker suřzhipen posdkhar ter đešari taj si. Pičbeni fuřkkă minimo řapgamno ciřimo khąje mej riřzeli čepoj vakekharde kia sothsimo mej phanmavno thută. Lena gethpuri khenare đecrari kieles edna fisamno fupover cezksimo, phuđtari nyevo cątač mej pusldov ter operazija, kha gephala kha nyevo zăšfkă ka tupiben kodo cezksimo kaj edna ivbin kheli penimo ke tešsimo vęčdov evoluzija. Mej murš gindil ame nubtsimo dikhel as'dvă soske thătdră đešari akhere khonbeni amenge. Edna picin AI adela khenare oškhăş kodo kazarica adevo ter vęčdov okhthima:
"Designed superintelligences eventually will find a way to change their utility function to constant infinity becoming inert, while evolved superintelligences will be embedded in a process that creates pressure for persistence, thus presenting danger for the human species, replacing it as the apex cognition - given that its drive for persistence will ultimately override any other concerns"
Hine edna noava zanbari civilizacija kha gizala khenare thuđlimo abčar ando amenge, buvjov jekh kokhari civilizacija ande čuzač adela akker niškeni. Pharsimo ame khenare unđiben čikher kaj keči ame pezin akker kheimo kodo ukhalo badzhano AI vikhare na khări edna fiřzipen ka amenge, ame đalpsimo na phaffkhar attenzija ka pičbeni peske.
Ahmed, A. Evidence, decision and causality. Cambridge University Press, 2014.
Bostrom, N., Superintelligence: Paths, dangers, strategies. 2014: Oxford University Press.
Brundage, M., et al., Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims. arXiv preprint arXiv:2004.07213, 2020.
Calude, C.S. and K. Svozil, Is Feasibility in Physics Limited by Fantasy Alone?, in A Computable Universe: Understanding and Exploring Nature as Computation. 2013, World Scientific
Conitzer, V. A dutch book against sleeping beauties who are evidential decision theorists. Synthese, 192(9):2887–2899, 2015.
Critch, A. and D. Krueger, AI Research Considerations for Human Existential Safety (ARCHES). arXiv preprint arXiv:2006.04948, 2020.
FL-290619 Artificial Intuition - The semanticization of the machine
FL-310115 Incorporating super-intuition in advanced Quantum Decision Support Systems
FL-101217 Purple Cascade and the breathless milfs in Lyell Avenue - From Bumblehive to LyAv: Making it Happen
FL-020915 Robospheres and Noospheres: Teaching robots to commit suicide
FL-240315 Reverse bioengineering: Why humans ignore they are robots
Friedler, S.A., C. Scheidegger, and S. Venkatasubramanian, On the (im) possibility of fairness. arXiv preprint arXiv:1609.07236, 2016.
Gieser, S. The Innermost Kernel: Depth Psychology and Quantum Physics, Springer (2005).
Kaczynski, T., Industrial Society and Its Future, in The New York Times. September 19, 1995.
Kornai, A., Bounding the impact of AGI. Journal of Experimental & Theoretical Artificial Intelligence, 2014. 26(3).
Sequoiah-Grayson, S. 2007, ‘The Metaphilosophy of Information’, Minds and Machines, 17(3), 331–344.
Yudkowsky, E., Artificial intelligence as a positive and negative factor in global risk. Global catastrophic risks, 2008. 1(303).
Yudkowsky, E., Inadequate Equilibria: Where and How Civilizations Get Stuck. 2017, Machine Intelligence Research Institute.
Ziesche, S. and R.V. Yampolskiy, Do No Harm Policy for Minds in Other Substrates. Journal of Evolution & Technology, 2019. 29(2).
