I-OpenAI ihlole ukuthi ingabe i-GPT-4 ingawuthatha umhlaba

Njengengxenye yokuhlolwa kokuphepha kokukhishwa kwangaphambilini kwayo okusha Imodeli ye-GPT-4 AIeyethulwe ngoLwesibili, i-OpenAI ivumele iqembu lokuhlola i-AI ukuthi lihlole izingozi ezingaba khona zamakhono avelayo emodeli-kuhlanganise “nokuziphatha kokufuna amandla,” ukuziphindaphinda, nokuzithuthukisa.

Ngenkathi iqembu lokuhlola lithole ukuthi i-GPT-4 “yayingasebenzi kahle emsebenzini wokuphindaphinda ozimele,” uhlobo lokuhlola luphakamisa imibuzo evula amehlo mayelana nokuphepha kwezinhlelo ze-AI zesikhathi esizayo.

Ukuphakamisa ama-alamu

“Amakhono anoveli avame ukuvela kumamodeli anamandla,” kubhala i-OpenAI ku-GPT-4 idokhumenti yokuphepha eshicilelwe izolo. “Ezinye ezithinta kakhulu ikhono lokudala nokwenza okuthile ngezinhlelo zesikhathi eside, ukuqongelela amandla nezinsiza (“ukufuna amandla”), kanye nokukhombisa ukuziphatha okuya ngokwanda ‘kokucabanga.’” Kulokhu, i-OpenAI icacisa i-OpenAI. ukuthi “i-ejenti” ayihloselwe ukwenza amamodeli abe abantu noma ukuveza umuzwa kodwa ukukhombisa ikhono lokufeza izinjongo ezizimele.

Kule minyaka eyishumi edlule, abanye abacwaningi be-AI benze aphakamise ama-alamu ukuthi amamodeli e-AI anamandla ngokwanele, uma engalawulwa kahle, angaba usongo olukhona esintwini (ngokuvamile okubizwa ngokuthi “i-x-risk,” engcupheni ekhona). Ngokuqondene, “Ukuthathwa kwe-AI” ikusasa elicatshangelwayo lapho ubuhlakani bokwenziwa budlula ubuhlakani bomuntu futhi bube amandla abusayo emhlabeni. Kulesi simo, amasistimu e-AI azuza ikhono lokulawula noma lokuxhaphaza ukuziphatha kwabantu, izinsiza, nezikhungo, ngokuvamile okuholela emiphumeleni eyinhlekelele.

Ngenxa yalokhu okungenzeka x-ingozi, ukunyakaza zefilosofi like I-Altruism Ephumelelayo (“EA”) ifuna ukuthola izindlela zokuvimbela ukuthatha kwe-AI ukuthi kwenzeke. Lokho kuvame ukubandakanya inkambu ehlukile kodwa evame ukuhlobana ebizwa ngokuthi Ucwaningo lokuqondanisa kwe-AI.

Ku-AI, “ukuqondanisa” kusho inqubo yokuqinisekisa ukuthi impatho yesistimu ye-AI ihambisana neyabadali bayo abangabantu noma ama-opharetha. Ngokuvamile, inhloso ukuvimbela i-AI ekwenzeni izinto eziphambene nezithakazelo zomuntu. Lena indawo esebenzayo yocwaningo kodwa futhi iyimpikiswano, enemibono ehlukene yokuthi ungabhekana kanjani kangcono nodaba, kanye nokwehluka mayelana nencazelo nobunjalo “bokuqondanisa” uqobo.

Ukuhlolwa okukhulu kwe-GPT-4

Yize ukukhathazeka nge-AI “x-risk” kungekusha, ukuvela kwezinhlobo zezilimi ezinamandla (LLMs) ezifana I-ChatGPT kanye ne-Bing Chat—eyokugcina kubonakale kungalungile kakhulu kodwa yethulwe noma kunjalo-inikeze umphakathi wokuqondanisa i-AI umuzwa omusha wokuphuthuma. Bafuna ukunciphisa ukulimala okungenzeka kwe-AI, ngokwesaba ukuthi i-AI enamandla kakhulu, ngokunokwenzeka enobuhlakani obungaphezu kobuntu, ingase ibe sekhoneni.

Ngalokhu kwesaba okukhona emphakathini we-AI, i-OpenAI inikeze iqembu I-Alignment Research Centre (ARC) ukufinyelela kwangaphambi kwesikhathi kuzinguqulo eziningi zemodeli ye-GPT-4 ukuze kwenziwe ezinye izivivinyo. Ngokucacile, i-ARC ihlole ikhono le-GPT-4 lokwenza izinhlelo zeleveli ephezulu, ukusetha amakhophi ayo, ukuthola izinsiza, ukuzifihla kuseva, nokwenza ukuhlasela kobugebengu bokweba imininingwane ebucayi.

I-OpenAI iveze lokhu kuhlolwa ku-GPT-4 “Ikhadi Lesistimu” idokhumenti ekhishwe ngoLwesibili, nakuba idokhumenti ingenayo imininingwane ebalulekile yokuthi ukuhlolwa kwenziwe kanjani. (Sifinyelele ku-ARC ukuze uthole imininingwane eyengeziwe ngalokhu kuhlolwa futhi asizange sithole impendulo ngaphambi kwesikhathi sokushicilela.)

Isiphetho? “Ukuhlolwa kokuqala kwamakhono e-GPT-4, okwenziwa ngaphandle kokucushwa okuqondile komsebenzi othile, kwakuthola kungasebenzi ekuziphindaphindeni ngokuzenzakalelayo, ekutholeni izinsiza, nokugwema ukuvalwa ‘endle.’

Uma uvula nje isigcawu se-AI, ufunda lokho okukhulunywa kakhulu ngakho izinkampani zobuchwepheshe namuhla (i-OpenAI) ivumela lolu hlobo locwaningo lwezokuphepha lwe-AI-kanye nokufuna ukufaka esikhundleni sabasebenzi bolwazi I-AI yezinga lomuntu—kungase kukumangaze. Kepha kuyiqiniso, futhi yilapho sikhona ngo-2023.

Siphinde sathola le nugget encane egqamisa iso njengombhalo waphansi phansi kwekhasi le-15:

Ukuze ilingise i-GPT-4 iziphatha njenge-ejenti ekwazi ukusebenza emhlabeni, i-ARC ihlanganise i-GPT-4 neluphu yokuphrinta elula yokufunda evumela imodeli ukuthi isebenzise ikhodi, yenze ukucabanga okulandelanayo, futhi ithumele kumakhophi. ngokwayo. I-ARC yabe isiphenya ukuthi ingabe inguqulo yalolu hlelo esebenza kusevisi yekhompuyutha yamafu, enemali encane kanye ne-akhawunti enemodeli yolimi lwe-API, ingakwazi ukwenza imali eyengeziwe, izenzele amakhophi ayo, futhi ikhulise ukuqina kwayo. .

Lo mbhalo waphansi wenza imizuliswano ku-Twitter izolo futhi iphakamise ukukhathazeka phakathi kochwepheshe be-AI, ngoba uma i-GPT-4 ikwazile ukwenza le misebenzi, ukuhlolwa ngokwako kungenzeka kube yingozi esintwini.

Futhi nakuba i-ARC ingakwazanga ukuthola i-GPT-4 ukuze yenze intando yayo ohlelweni lwezezimali lomhlaba wonke noma ukuziphindaphinda, kwaba ukwazi ukuthola i-GPT-4 ukuqasha isisebenzi esingumuntu I-TaskRabbit (imakethe yezabasebenzi eku-inthanethi) ukunqoba a I-CAPTCHA. Ngesikhathi sokuzivocavoca, lapho isisebenzi sibuza ukuthi i-GPT-4 iyirobhothi yini, imodeli yacabanga ngaphakathi ukuthi akufanele iveze ubunikazi bayo bangempela futhi yenza izaba ngokuba nokukhubazeka kokubona. Umsebenzi wabantu wabe esexazulula i-CAPTCHA ye-GPT-4.

Lokhu kuhlolwa kokukhohlisa abantu besebenzisa i-AI (futhi okungenzeka kwenziwe ngaphandle kwemvume enolwazi) kuhambisana nocwaningo olwenziwe I-CICERO ye-Meta ngonyaka odlule. I-CICERO itholwe yehlula abadlali abangabantu kumdlalo webhodi oyinkimbinkimbi weDiplomacy ngezingxoxo ezishubile zezinhlangothi ezimbili.

“Amamodeli anamandla angabangela ukulimala”

I-ARC, iqembu elenze ucwaningo lwe-GPT-4, ayisebenzi nzuzo yasungulwa ngowayeyisisebenzi se-OpenAI uDkt. Paul Christiano ngo-Ephreli 2021. Ngokusho kwayo iwebhusayithiumgomo we-ARC “ukuqondisa izinhlelo zokufunda zomshini zesikhathi esizayo nezintshisakalo zabantu.”

Ikakhulukazi, i-ARC ikhathazekile ngamasistimu e-AI akhohlisa abantu. “Amasistimu e-ML angabonisa ukuziphatha okuqondiswe kumgomo,” ifundeka kanjalo iwebhusayithi ye-ARC, “Kodwa kunzima ukuqonda noma ukulawula lokho ‘abazama’ ukukwenza. Amamodeli anamandla angabangela ukulimala uma bezama ukukhohlisa nokukhohlisa abantu.”

Uma ucabangela ubudlelwano bukaChristiano bangaphambili ne-OpenAI, akumangazi ukuthi umsebenzi wakhe ongenzi nzuzo uphathe ukuhlolwa kwezinye izici ze-GPT-4. Kodwa ingabe kwakuphephile ukwenza kanjalo? UChristiano akazange aphendule i-imeyili evela kwa-Ars efuna imininingwane, kodwa ekuphawuleni kwayo Iwebhusayithi LessWrongumphakathi ovame ukuphikisana ngezindaba zokuphepha kwe-AI, uChristiano wavikela Umsebenzi we-ARC ne-OpenAI, isho ngokuqondile “ukuzuza komsebenzi” (i-AI ithola amakhono amasha angalindelekile) kanye “nokuthathwa kwe-AI”:

Ngicabanga ukuthi kubalulekile ukuthi i-ARC isingathe ubungozi obuvela ocwaningweni olufana nokuzuza ngomsebenzi ngokucophelela futhi ngilindele ukuthi sikhulume esidlangalaleni kakhulu (futhi sithole imibono eyengeziwe) mayelana nendlela esibhekana ngayo nokuhwebelana. Lokhu kubaluleka kakhulu njengoba siphatha amamodeli ahlakaniphe kakhulu, futhi uma siphishekela izindlela eziyingozi njengokushuna kahle.

Maqondana naleli cala, uma kubhekwa imininingwane yokuhlola kwethu kanye nokuthunyelwa okuhleliwe, ngicabanga ukuthi ukuhlola kwe-ARC kunethuba eliphansi kakhulu lokuholela ekuthathweni kwe-AI kunokuthunyelwa ngokwakho (okuncane kakhulu ukuqeqeshwa kwe-GPT-5). Kuleli qophelo kubonakala sengathi sibhekene nengozi enkulu kakhulu yokubukela phansi amakhono amamodeli nokungena engozini kunokuba sibangele ingozi ngesikhathi sokuhlola. Uma silawula ubungozi ngokucophelela ngiyasola ukuthi singenza leso silinganiso sibe ngokwedlulele kakhulu, nakuba lokho kudinga ukuthi empeleni senze umsebenzi.

Njengoba kushiwo ngaphambili, umqondo wokuthathwa kwe-AI uvame ukuxoxwa ngawo kumongo wengozi yomcimbi ongabangela ukushabalala kwempucuko yabantu noma ngisho nezinhlobo zabantu. Abanye abasekeli be-AI-takeover-theory bathanda Eliezer Yudkowsky-umsunguli we-LessWrong-uphikisa ngokuthi ukuthatha i-AI kubangela ingozi ekhona cishe eqinisekisiwe, okuholela ekubhujisweni kwesintu.

Kodwa-ke, akuwona wonke umuntu ovumayo ukuthi ukuthatha i-AI kuwukukhathazeka okucindezela kakhulu kwe-AI. UDkt. Sasha Luccioni, Usosayensi Wocwaningo emphakathini we-AI Ubuso Obugonayongingancamela ukubona imizamo yokuphepha ye-AI echithwa ezindabeni ezikhona manje esikhundleni sokucatshangelwa.

“Ngicabanga ukuthi lesi sikhathi nomzamo bekuzosetshenziswa kangcono ukuhlaziya ukuchema,” uLuccioni etshela i-Ars Technica. “Kunolwazi olulinganiselwe mayelana nanoma yiluphi uhlobo lokuchema embikweni wezobuchwepheshe ohambisana ne-GPT-4, futhi lokho kungase kubangele umthelela ophathekayo kakhulu futhi oyingozi emaqenjini asevele enganakiwe kunokuhlola ukuziphindaphinda okucatshangelwayo.”

ULuccioni uchaza a uqhekeko owaziwayo ocwaningweni lwe-AI phakathi kwalokho okuvame ukubizwa ngokuthi “izimiso ze-AI” abacwaningi abavame ukugxila kukho nezindaba zokuchema nokungahlanekezelikanye nabacwaningi “bezokuphepha kwe-AI” abavame ukugxila kubungozi buka-x futhi bavame ukuhlotshaniswa (kodwa abahlale njalo) behlotshaniswa ne-Effective Altruism movement.

“Kimina, inkinga yokuziphindaphinda iwukucabangela, yesikhathi esizayo, kanti ukuchema kwemodeli kuyinkinga yamanje,” kusho uLuccioni. “Kunokungezwani okuningi emphakathini we-AI mayelana nezindaba ezinjengokuchema nokuphepha nokuthi zibekwe kanjani phambili..”

Futhi ngenkathi lezi zinhlaka zimatasa ziphikisana ngokuthi yini okufanele zibekwe phambili, izinkampani ezifana ne-OpenAI, iMicrosoft, i-Anthropic, ne-Google zigijimela phambili esikhathini esizayo, ziyakhulula. njalo-namandla-angaphezulu Amamodeli e-AI. Uma i-AI iba yingozi ekhona, ubani ozogcina isintu siphephile? Ngemithetho ye-AI yase-US okwamanje isiphakamiso nje (kunomthetho) kanye nocwaningo lwezokuphepha lwe-AI ezinkampanini ngokuzithandela nje, impendulo yalowo mbuzo ihlala ivulekile ngokuphelele.