Kwenzekeni
I-MIT Technology Review ishicilele isici esisebenzisanayo esakhelwe cishe izindida eziyisikhombisa eziphenya ukuthi amamodeli e-AI acabanga kanjani. I-athikili ibika ukuthi amamodeli athuthuke ngokushesha kweminye imisebenzi kodwa asazabalaza nokukhohlisa kwendawo, ukuhlukahluka okucashile kwezinkinga ezijwayelekile, ukufinyela okubukwayo, kanye nokucabanga okuyinkimbinkimbi kwezinyathelo eziningi.
I-MIT Technology Review ibika ukuthi isici sayo sethula izivivinyo eziyisikhombisa ezihlanganisa ukucabanga kwendawo, inkumbulo kanye nokuvumelana nezimo, ukucabanga okungaqondakali nokubonakalayo, ukuqonda komuntu, kanye nobunzima obandayo. I-athikili ibeka ukuxazulula izindida emlandweni omude wokuhlolwa kwe-AI, iphawula ukuthi amageyimu afana nokuhlola, i-chess, ne-Go isetshenziswe njengemibhede yokuhlola kusukela eminyakeni yokuqala yenkambu. Ithi ukusebenza kwendida kuthuthuke kakhulu: ithimba lase-Columbia University elitholwe ngasekupheleni kuka-2024 ukuthi amamodeli aphambili axazulule kuphela u-18% wamaphazili we-New York Times Connections, kuyilapho ekuqaleni kuka-2025 amanye amamodeli akwazi ukuwaxazulula cishe ngokuphelele ngaso sonke isikhathi. Umthombo awukhombi amamodeli noma unikeze ukuqhathanisa okujwayelekile kuzo zonke izibonelo ezikusici.
I-MIT Technology Review ithi ukucabanga kwendawo kusewubuthakathaka obukhulu. Ingxenye yayo yokushintsha ingqondo icela abafundi ukuthi bahlonze into enezinhlangothi ezintathu eboniswa kwenye i-engeli, futhi i-athikili ibika ukuthi amamodeli olimi anamandla okufaka okubukayo asasebenza kabi kakhulu emisebenzini enjalo. Lesi sici sixhumanisa lobu bunzima nezimangalo ezimayelana namasistimu e-AI athuthukisa amamodeli omhlaba, siphikisa ngokuthi amamodeli wamanje amakhulu olimi awabonakali ekhohlisa izinto ezinezinhlangothi ezintathu ngendlela efanayo naleyo ochwepheshe bendawo abangenza ngayo. I-athikili iphinde ichaze inkinga yenkumbulo nokuzivumelanisa nezimo: amamodeli aqeqeshwe ngenani elikhulu lolwazi angase abone iphethini yendida ejwayelekile futhi ashaye indiva ushintsho oluncane. Icaphuna ucwaningo lwango-2024 lwe-Google kanye ne-University of Illinois Urbana-Champaign olubandakanya ukuhlukahluka kwamaphazili we-Knights kanye ne-Knaves njengobufakazi balokhu kukhathazeka.
Isici sibe sesiphendukela ekufinyezweni kwezinhlangothi ezimbili kanye nokucabanga okubonakalayo. I-MIT Technology Review ibika ukuthi amamodeli asebenza kangcono kuzindida ze-ARC-AGI lapho amagridi enikezwa njengeyunithi yezinhlamvu zezinombolo ebhala ngekhodi imibala yeseli kunokuba njengezithombe. Iphinde ithi ucwaningo luthole ukuthi amamodeli kwesinye isikhathi angafinyelela impendulo eyiyo ngemithetho eyinkimbinkimbi, engajwayelekile, kuyilapho abantu bengase bathembele emibonweni elula yokubuka. I-athikili yethula imibuzo e-SimpleBench, izinkinga ze-Knights ne-Knaves, kanye nephazili yokuguqulwa kwe-ARC-AGI njengezibonelo zemisebenzi engadalula lo mehluko.
Ichaza ngokuhlukene izivivinyo ze-intuition lapho abantu bevame ukunikeza izimpendulo ezisheshayo kodwa ezingalungile, kuyilapho amamodeli engase aphendule ngamabomu kakhulu. Isibonelo esisodwa sibuza ukuthi umhume uzothatha isikhathi esingakanani ukuthi ugcwale ingxenye lapho inani lamalulwane liphindaphindeka kabili nsuku zonke; omunye ucela abafundi ukuthi basho amazwi acashuniwe ahlotshaniswa no-Alice.
Okokugcina, I-MIT Technology Review ibika ukuthi ukusebenza kuvame ukuwohloka njengoba izinkinga ziba nzima kakhulu. Icaphuna ucwaningo lwe-Apple oluthola ukuthi izinhlobo zolimi zingaxazulula izinguqulo ezilula ze-Tower of Hanoi kanye nezinkinga zokuwela umfula kodwa zaqala ukuncipha lapho inani lamadiski noma abantu lifinyelela kwabayisithupha noma ngaphezulu. Iphinde icaphune umsebenzi wabacwaningi baseNyuvesi yase-Washington, i-Stanford University, kanye ne-Allen Institute yokuthola ubunzima obufanayo ngama-logic-grid puzzle. Isici siphawula ukuthi abahlaziyi babuze ukuthi ingabe le miphumela iveza umkhawulo wokucabanga ofana nomshini noma ivele ibonise iqiniso lokuthi abantu nabo benza amaphutha amaningi njengoba ubunkimbinkimbi bukhula. Ngakho-ke izibonelo zayo zokuwela umfula kanye ne-logic-grid azivivinyi nje kuphela ukuthi imodeli ingakwazi yini ukukhiqiza impendulo, kodwa ukuthi ingakwazi yini ukugcina izingqinamba kuzo zonke ezibanjwayo eziningi ezixhunyiwe.
Imininingwane yomthombo: technologyreview.com ↗
Kungani kubalulekile
Isici sibonisa ukuthi kungani izimangalo ezibanzi mayelana nobuhlakani be-AI zingadukisa. Imodeli ingase yenze kahle ku-trivia noma ibhentshimakhi evamile kuyilapho yehluleka ukuguqulwa okuncane kumagama, ukuguqulwa okubonakalayo, noma inkinga edinga izinyathelo ezimbalwa ezincikile. Lowo mehluko ubalulekile lapho izinhlelo zisetshenziselwa izinqumo kunemibukiso.
Isifundo esiyinhloko ukuthi ukusebenza ekilasini elilodwa lokuhlolwa akufanele kuthathwe njengesilinganiso esijwayelekile sobuhlakani. Izibonelo ze-MIT Technology Review zihlanganisa imisebenzi ebukeka ilula kakhulu kodwa edinga amakhono ahlukene: ukuzungezisa into ngengqondo, ukulandelela iqiniso namanga kuzo zonke izitatimende, ukuveza umthetho obonakalayo, ukumelana nokuzwa okudukisayo, noma ukuhlela ukulandelana ngaphansi kwezingqinamba. Uhlelo lungaba namandla ngendlela engavamile endaweni ethile futhi lungathembeki kwenye. Lokho kungalingani kubaluleke kakhulu uma abasebenzisi behumusha izimpendulo ezishelelayo njengobufakazi bokuthi imodeli iyayiqonda inkinga ewumsuka.
I-athikili futhi igqamisa inkinga yokuhlola: ifomethi yomsebenzi ingathinta kakhulu umphumela. I-MIT Technology Review ibika ukuthi ukusebenza kwe-ARC-AGI kuyathuthuka lapho amagridi abonakalayo eguqulwa abe izethulo zezinombolo. Lokhu kuphakamisa ukuthi amaphuzu angase angabonisi nje kuphela ikhono lokucabanga lemodeli kodwa nendlela inkinga ebhalwe ngayo futhi yethulwa ngayo. Ukukhathazeka okufanayo kuyasebenza kumaphazili ajwayelekile. Uma imodeli ihlangabezane nokwehluka okuseduze phakathi nokuqeqeshwa, impendulo elungile ingase ibonise ukuqashelwa kwephethini noma ukubuyisa kunekhono lokujwayelanisa umthetho esimweni esisha. Umthombo awuqinisekisi ukuthi noma iyiphi indlela yenzeke kaningi kangakanani kuwo wonke amasistimu asetshenzisiwe.
Le mikhawulo inemithelela ebonakalayo kunoma ubani osebenzisa i-AI kwezemfundo, isofthiwe, ucwaningo, ukuphatha, noma ezinye izilungiselelo ezibandakanya amacala angajwayelekile. Imodeli ephumelela ezibonelweni ezivamile isengase yehluleke lapho amagama eshintsha, izithiyo ezimbalwa zisebenzisana, noma ulwazi olufanele lubonakala kunombhalo. Isici asibiki ukwethulwa komkhiqizo omusha, ukukhishwa kwemodeli entsha, noma ukuhlolwa okulawulwayo kwesistimu ethile yezentengiso. Inani laso liyachaza: linikeza abafundi izindlela ezibambekayo zokubona ukuthi kungani izinzuzo zokuqhathanisa nolimi oluphucuziwe zingazisunguli ngokwazo ukucabanga okuqinile. I-athikili futhi igcina imfanelo ebalulekile: abantu banokuchema kwabo futhi bangahamba kancane noma bangathembeke kancane ezinkingeni ezithile.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Ukuhlola okuzayo kuzodinga ukuhlola ngaphezu kwezikolo zesihloko. Imibuzo ebalulekile ihlanganisa ukuthi amamodeli angakwazi yini ukujwayelana nezinkinga ezingajwayelekile, noma izimpendulo zawo zincike ekutheni ulwazi lumelwe kanjani, ukuthi ukusebenza kushintsha kanjani njengoba ubunkimbinkimbi bukhuphuka, nokuthi ingabe impumelelo ibonisa isu elisebenziseka kabusha noma amaphethini enziwe ngekhanda.
Intuthuko elandelayo ewusizo ingaba ebanzi, ukuhlola okuphindaphindekayo kuwo wonke amamodeli namafomethi emisebenzi. Isici se-MIT Technology Review asihlinzeki ngekhadi elilodwa lamaphuzu elimboza zonke izivivinyo eziyisikhombisa, futhi umthombo awucacisi amagama amamodeli, izinguqulo, osayizi besampula, izimo eziyalelwayo, noma izilinganiso zamaphutha zezibonelo eziningi. Leyo mininingwane izodingeka ukuze kunqunywe ukuthi ingabe ubuthakathaka obubikiwe busakazekile yini, bugxile emindenini ethile eyimodeli, noma buzwela ukuthi ukuhlola okwenziwa kanjani.
Ukuhlola okuzimele kufanele futhi kuhlukanise ukuhluleka kombono obonakalayo ekuhlulekeni kokucabanga ngokugcina iphazili eyisisekelo ingashintshi ngenkathi kushintsha ukumelwa kwayo. Abacwaningi nabahloli kufanele futhi babheke ukuthi amamodeli ayathuthuka yini ngokwenza okuvamile kwangempela noma ngokuchayeka kakhulu ezintweni ezifana nebhentshimakhi. Ingxoxo ye-athikili ye-Connections puzzle kanye nokuhluka kwe-Knights ne-Knaves kubonisa ukuthi kungani ukungcola nokujwayela kubalulekile. Imodeli esebenza kahle emibuzweni eshicilelwe noma ehlobene eduze ingase ingakwazi ngokulinganayo ezinkingeni ezisanda kukhiqizwa ngesakhiwo esiyisisekelo esifanayo. Ngakho-ke ukuhlola kufanele kubandakanye izindida ezibanjiwe, amagama ashintshiwe, izakhiwo ezibonwayo ezingajwayelekile, kanye nemisebenzi edinga ukuchaza noma ukuhlola izingqinamba ezimaphakathi ngaphandle kokucabangela ukuthi impendulo ehehayo ilungile.
Inkimbinkimbi nokudluliselwa komhlaba wangempela kuhlala kuyimibuzo evulekile. I-MIT Technology Review ibika ukuthi ukusebenza kungehla njengoba inani lezinto noma izinyathelo ezidingekayo zikhula, kodwa umthombo awuqinisekisi ukuthi lokho okutholakele kuhumusha kanjani kumasistimu asebenzayo ngamathuluzi, ukubuyisa, inkumbulo yangaphandle, noma ukugada komuntu. Ukubika okuzayo kufanele kulandelele ukuthi ingabe izengezo ezinjalo zithuthukisa ukwethembeka, zisize amamodeli nje aseshe ngempumelelo kakhulu, noma zethule izindlela zokuhluleka ezintsha. Abafundi kufanele futhi babheke ukuhlaziya okulinganisa ikhwalithi yesu, hhayi impendulo yokugcina kuphela, ngoba umphumela olungile ofinyelelwe ngomthetho obhuqayo noma ongakwazi ukuphindaphindeka ungase ungasindi kuguquko oluncane lwenkinga.