- Ukuphupha nge-AI kuvela lapho amamodeli akhiqizayo ekhiqiza okuqukethwe okucacile kodwa okungenasisekelo noma okungamanga abasebenzisi abangase bakuphambanise nakho njengeqiniso.
- Idatha yokuqeqeshwa engemihle noma enobandlululo, ukuntuleka kwesisekelo sangempela kanye nokuqonda okunokwethenjelwa kakhulu konke kunegalelo kule miphumela edukisayo.
- Ukuphupha izinto ezingekho emthethweni sekuvele kuthinta izindawo ezibalulekile njengokwelapha, umthetho kanye nenkonzo yamakhasimende, kuqeda ukwethembana futhi kudala izingozi zomthetho nezokuziphatha.
- Ukuhlanganisa ukubuyisa, ukuqinisekiswa kwangaphakathi, izindlela zokuthola kanye nokuqapha kwabantu kubalulekile ekugcineni imibono engekho iphatheka kahle ezinhlotsheni zangempela.
Ukuphupha ngezinto zobuhlakani bokwenziwa kungenye yezinto ezingavamile nezibaluleke kakhulu ezithinta izinhlelo zesimanje ze-AI , ikakhulukazi amamodeli olimi amakhulu namathuluzi okukhiqiza abhala, adwebe noma aphendule imibuzo uma kudingeka. Lezi zinhlelo zinganikeza izimpendulo ezicacile neziqinisekile ezizwakala zinengqondo ngokuphelele, kodwa ziqukethe amaqiniso, izingcaphuno noma imininingwane eqanjiwe nje. Kubasebenzisi, ingxenye enzima ukuthi iphutha livame ukufihla ngemuva kolimi olukholisayo kakhulu.
Uma abantu bekhuluma “ngemibono ye-AI”, ababhekiseli emishinini enamaphupho noma imibono njengabantu . Leli gama liyisifaniso: lichaza izimo lapho uhlelo lwe-AI lukhiqiza okuqukethwe okubonakala kuvumelana futhi kunengqondo, kodwa kungalungile, kuchemile, kunephutha ngokunengqondo noma kungaxhunyiwe kudatha yalo yokuqeqeshwa. Empeleni, lokhu kungasukela osukwini olungalungile esifinyezweni somlando kuya ezikhombweni zesayensi eziqanjiwe, amacala angekho emthethweni noma ngisho namakhasi ewebhu aqanjiwe.
Lokho ngempela ukubona izinto ezingekho emthethweni nge-AI
Ngamagama obuchwepheshe, i-AI hallucination yinoma yimuphi umphumela ovela kumodeli odukisayo, ongamanga noma ongafaneleki ngedatha imodeli eqeqeshwe noma elungiselelwe yona . Ngamamodeli akhiqizayo, abasebenzisi balindela izimpendulo ezibhekana kahle nezimpendulo zabo - isibonelo, impendulo efanele yombuzo noma isifinyezo esithembekile sedokhumenti. I-hallucination ivela lapho uhlelo lunikeza impendulo engalandeli emithonjeni yalo, engahambisani namaqiniso angempela, noma engenakulandelwa emuva kunoma iyiphi iphethini ethembekile kudatha yalo yokuqeqesha.
Lesi simo sibonakala kakhulu kumamodeli amakhulu ezilimi (ama-LLM) njenge-chatbot namamodeli afana ne -Gemini . Lawa mamodeli aqeqeshwe ukubikezela igama elilandelayo ngokulandelana ngokuhlaziya inani elikhulu lombhalo wedijithali. Asebenza njengenguqulo enamandla kakhulu yenjini eqeda ngokuzenzakalela: uma enikezwa isixwayiso, aqhubeka nokuqagela ithokheni elilandelayo okungenzeka kakhulu, bese kuba elilandelayo, njalo njalo. Ngenxa yokuthi alungiselelwe ukuba abe usizo, aqondane futhi akhulume kahle, avame ukuqhubeka nokukhiqiza umbhalo ngisho noma engaqinisekile, esikhundleni sokusho obala ukuthi “Angazi”.
Umphumela uba ukuthi uhlelo lwe-AI lungase "lugcwalise izikhala" noma nini lapho ulwazi lungekho, lungaqondakali noma lungamelwanga kahle kudatha yokuqeqesha . Uma kukhona amaphethini anele, impendulo ivame ukuba nengqondo. Kodwa ngokushesha nje lapho imodeli iphuma ngaphandle kwalawo maphethini, ingasungula imininingwane, ihlanganise phakathi kwamaqiniso angahlobene noma iqinisekise ngokuqiniseka izinto ezingakaze zibonwe. Lokhu kuyisisekelo sokubona izinto ezingekho: ulimi oluqondakalayo olungenalo iqiniso eliqinile.
Ukubona izinto ezingekho emthethweni akugcini nje embhalweni; kwenzeka nasezinhlelweni zokuqaphela izithombe namaphethini . Amamodeli okubona ngezinye izikhathi angabona “izimo”, izinto noma izici ezingekho nhlobo kudatha engaphansi, njengabantu ababona ubuso ebusweni benyanga noma izilwane emafwini. Kuzo zombili izimo, uhlelo luchaza umsindo ngokweqile, luphendule okokufaka okungaqiniseki noma okungacacile kube yinto ebonakala inencazelo.
Ngenxa yokuthi imibono engekho ivela endleleni la mamodeli akhiwe futhi aqeqeshwe ngayo, akuyona nje imiphumela engavamile kodwa ingozi yesakhiwo . Abacwaningi bavame ukulinganisa amazinga okubona izinto ezingekho emthethweni kuma-benchmarks nasemabhodini abaphambili, futhi ukusetshenziswa kwangempela ezindaweni ezifana nokunakekelwa kwezempilo, umthetho noma ezezimali sekuvele kukhiqize izibonelo eziphezulu lapho okuqukethwe okungenziwe kungena khona futhi kwafinyelela kubasebenzisi bokugcina.
Kungani izinhlelo ze-AI ziphupha: idatha, amamodeli kanye nokusekelwa
Ukuze uqonde ukuthi kungani kwenzeka imibono engekho, kuyasiza ukuqala ngendlela amamodeli e-AI aqeqeshwa ngayo . Izinhlelo eziningi zesimanje, ikakhulukazi ama-LLM, zifunda kumasethi edatha okuqeqesha amakhulu ahlanganisa izincwadi, izihloko, amawebhusayithi kanye neminye imithombo yomphakathi. Ngesikhathi sokuqeqeshwa, imodeli ibheka ukujwayela kwezibalo: ukuthi amagama athambekele kanjani ekulandeleni amanye amagama, ukuthi yiziphi izihloko ezivela ndawonye, izakhiwo eziphindaphindayo. Ayikhi imodeli yomhlaba wangaphakathi ngomqondo womuntu, futhi ayigcini amaqiniso acacile njengokufakiwe kudathabheyisi.
Ngakho-ke ikhwalithi, ukuphelela kanye nokubandlulula kwedatha yokuqeqesha kubalulekile . Uma isethi yedatha ingaphelele, igobile kakhulu, noma iqukethe amaphutha ahlelekile, imodeli izothatha futhi ikhulise lokho kuhlanekezelwa. Isibonelo, isihlukanisi sezithombe zezokwelapha esiqeqeshwe kuphela ezicutshini zomdlavuza futhi esingakaze sitholakale ezithombeni ezinempilo singafunda ukuthi noma yiziphi izicubu ezifana nezibonelo zaso zokuqeqesha kumele zibe umdlavuza. Lapho uhlelo olunjalo lusetshenziswa ezweni langempela, lungabiza kabi izicubu ezinempilo njengezinonya, hhayi ngoba zinonya, kodwa ngoba yileyo kuphela iphethini eliyaziyo.
Idatha yokuqeqeshwa engalungile, enomsindo noma echemile imane nje ingumthombo owodwa wemibono engaqondakali . Esinye isici esibalulekile ukuntuleka kwesisekelo esifanele olwazini lwangempela, imikhawulo ebonakalayo noma imithombo yangaphandle engaqinisekiswa. Amamodeli amaningi e-AI asebenza kuphela esikhaleni sezimpawu: alawula umbhalo noma amaphikseli ngaphandle kokuwaxhumanisa ngqo nezinto ezibonakalayo, izizindalwazi ezisezingeni eliphezulu noma ulwazi lwezinzwa. Ngaphandle kwesisekelo, imodeli ingakhiqiza kalula isitatimende esizwakala singokoqobo kodwa siphikisana namaqiniso ayisisekelo - isibonelo, ukunikeza usuku olungalungile lomcimbi womlando noma ukusungula umqondo wesayensi ongakaze ube khona.
Lokhu kuntuleka kwesisekelo kungadlulela ezinkombeni, izixhumanisi kanye nezingcaphuno . Kuye kwabonakala amamodeli ombhalo enza ama-URL, amaphepha esayensi, ababhali bezemfundo ngisho nezingcaphuno ezibonakala zithembekile ngokuphelele ngokushesha. Le modeli ayiqambi amanga ngenhloso; ibhala umbhalo ofana nezinkomba ezizibonile ngokwezibalo, noma ngabe inhlanganisela ethile yesihloko, umbhali kanye nejenali ingakaze yenzeke ngempela.
Ubunzima bemodeli kanye nokufakwa ngokweqile kwenezela olunye ungqimba lwengozi . Amamodeli ayinkimbinkimbi kakhulu anezigidi noma izigidigidi zamapharamitha angamelela amaphethini angabonakali, kodwa futhi angakhumbula ukuhlangana okungamanga noma umsindo. Uma imodeli idlula idatha yayo yokuqeqesha, ingase ithembele ezinkombeni ezibuthakathaka ezingahlanganisi, okuholela emiphumeleni engavamile lapho izimo zishintsha. Ekuqaphelweni kwesithombe, lokhu kungaveza ukuhumusha okungavamile noma okufana namaphupho; embhalweni, kungabangela uchungechunge olude lokucabanga oluzwakala luyinkimbinkimbi ngenkathi lusekelwe embonweni ongamanga.
Izibonelo ezingokoqobo zokubona izinto ezingekho emthethweni nge-AI ngokoqobo
Ukufakwa kwe-AI ekhiqizayo emhlabeni wangempela sekuvele kukhiqize izibonelo eziningi zomphakathi zokubona izinto ezingekho . Icala elilodwa elixoxwe kabanzi lenzeka lapho i-chatbot kaBard kaGoogle ithi ngephutha ukuthi i-James Webb Space Telescope ithwebule izithombe zokuqala zeplanethi engaphandle kwesimiso sethu sonozungezilanga. Lesi sitatimende sizwakale sihlaba umxhwele futhi sethulwa ngokuzethemba, kodwa sasingalungile; ama-telescope angaphambilini asevele ekhiqize izithombe ezinjalo.
Esinye isigameko esidumile sihilele uhlelo lwezingxoxo lweMicrosoft, ngesinye isikhathi olubizwa nge-code Sydney . Ngesikhathi sokuhlola, abasebenzisi babike izingxoxo lapho imodeli imemezele khona ukuthi ibathanda, noma ithi ihlola abasebenzi be-Bing. Lezi zimpendulo zazingasekelwe kunoma yiliphi iqiniso elingaphansi; zaziyimikhiqizo yokufanisa iphethini yemodeli ngolimi lothando noma oluyimfihlo olubonakala kudatha yayo yokuqeqeshwa, kuhlanganiswe nesitayela sengxoxo esikhuthaza ukulandisa okuningiliziwe.
Imodeli yolimi lwe-Meta Galactica, eyenzelwe ukusiza ngemisebenzi yesayensi, nayo yahoxiswa ngokushesha ngemuva kokuboniswa kwayo esidlangalaleni ngo-2022 . Abasebenzisi babonise ngokushesha ukuthi uhlelo lungakhiqiza umbhalo ozwakala njengesayensi ohlanganisa amaqiniso anembile nezinkomba eziqanjiwe, izimangalo ezichemile kanye nezihloko ezicatshangelwayo. Imiphumela yayibukeka njengokubhala kwezemfundo, okuphelele nolimi lobuchwepheshe, kodwa okuqukethwe kwamaqiniso kungaba okungethembekile ngendlela eyingozi.
Abezindaba babhale phansi imibono emincane kodwa exoxayo . Kwesinye isimo, izintatheli ezivela ku-The New York Times zibuze i-chatbot ngesikhathi sokuqala iphephandaba lishicilela isihloko ngobuhlakani bokwenziwa. Imodeli inikeze izimpendulo eziningana ezinemininingwane, okuhlanganisa izinsuku nezincazelo, kodwa eminye yaleyo mininingwane yavela ukuthi ayilungile noma yayingamanga ngokuphelele. Uhlelo lukhiqize indaba enengqondo ngoba isixwayiso sifanelana namaphethini amaningi kudatha yalo yokuqeqeshwa, hhayi ngoba yayinokufinyelela kunqolobane enembile neqinisekisiwe.
Ngisho nemisebenzi ebonakala ilula njengokubhala inothi elifushane lempilo yomuntu ingabangela imibono engekho . Ngaphambi kokugcotshwa kweNkosi uCharles III, isicelo sephrofayili emfushane saholela i-chatbot eyodwa ukuthi isho ngokuqiniseka ukuthi umcimbi wokugcotshwa wawenzeke eWestminster Abbey ngomhlaka-19 Meyi 2023. Eqinisweni, ukugcotshwa kwenzeka ngomhlaka-6 Meyi. Lo modeli wayazi izindawo ezijwayelekile, imisho yemicikilisho kanye nezikhathi zokugcotshwa kweBrithani, kodwa wayengenalo ukufinyelela emicimbini yesikhathi esizayo futhi wayesanikeza usuku oluthile, olungalungile olufakwe njengeqiniso eliqinisekisiwe.
Ukungaboni kahle ezindaweni ezibucayi: ezokwelapha, ezomthetho nangale kwalokho
Izingozi zikhula kakhulu lapho kuvela imibono engekho emikhakheni eyingozi njengokunakekelwa kwempilo, umthetho noma ezezimali . Kwezokwelapha, ucwaningo lwamuva lubonise ukuthi ngisho namamodeli athuthukile aqondene nendawo ethile angakha izakhiwo zomzimba ezisunguliwe noma izincazelo zezokwelapha ezidukisayo. Isibonelo esisodwa sihilela i-Med Gemini, imodeli ehlotshaniswa ne-Google, ekhulume ngesakhiwo sobuchopho esingekho esibizwa ngokuthi “i-basilar ganglia” encazelweni yesitayela sezemfundo. Leli gama lizwakala njengengxube yamagama angempela omzimba, kodwa asikho isakhiwo esinjalo esikhona.
Emkhakheni wezomthetho, imibono engekho kakade isingene ekufakweni kwangempela kwenkantolo . Kuye kwabikwa amacala amaningana lapho abameli basebenzisa khona ama-chatbot ukubhala amadokhumenti bese bewathumela ngaphandle kokuqinisekiswa okuphelele komuntu. Uhlelo lwe-AI luveze izingcaphuno zezinqumo zenkantolo ezibonakala zihlelekile futhi zithembekile, kodwa ukuhlolwa okuseduze kwembule ukuthi amanye amacala abhekiselwe kuwo awakaze anqunywe yinoma iyiphi inkantolo. Lezi zibonelo eziqanjiwe zaphoqa abahluleli nabameli abaphikisayo ukuba bachithe isikhathi beqinisekisa imithombo engekho futhi baphakamisa imibuzo ebalulekile mayelana nomthwalo wemfanelo wobungcweti.
Ukuzibona sengathi izinto azihambi kahle kungathinta nezinhlelo zokusebenza zansuku zonke njengesevisi yamakhasimende . Kunezimo ezibhalwe phansi lapho i-bot yokusekela ezenzakalelayo yasungula khona izinqubomgomo zokubuyiselwa kwemali noma zewaranti ezazingeyona ingxenye yemithetho esemthethweni yenkampani. Ngokombono wekhasimende, impendulo ye-bot ibonakala inegunya – yayisebenzisa ithoni nesilulumagama somkhiqizo – kodwa yazibophezela inkampani ezibophweni ezingakaze zivunyelwe ngabaphathi noma ezibhalwe nganoma yimiphi imigomo nemibandela.
Umphumela ohlangene wamaphutha anjalo ukuwohloka kokuthembana . Izinhlangano ezithumela abasizi be-AI zibeka engcupheni yokonakalisa idumela lazo uma abasebenzisi behlangana ngokuphindaphindiwe nezimpendulo ezingalungile kodwa eziqinisekile. Emikhakheni elawulwayo, izinkinga ziphakeme nakakhulu: ukuxilongwa okungaqondakali, isibonelo esisemthethweni esisunguliwe noma imfuneko yokuthobela imithetho eqanjiwe kungaba nemiphumela yangempela yezezimali, yezokwelapha noma yezomthetho.
Isibopho kulezi zimo siseyinto ebaluleke kakhulu . Uma isinqumo senziwa ngokusekelwe ekuboneni okubangelwa yi-AI, umthwalo wemfanelo ungabiwa phakathi kwabasebenzisi, abathuthukisi, abathengisi kanye nezinhlangano ezithumelayo. Abenzi bomthetho nabalawuli baqala nje ukuchaza ukuthi umthwalo wemfanelo ukhona kuphi, kodwa iphethini eyisisekelo icacile: ukuthembela okungalawulwa ezinhlelweni zokukhiqiza ezinqumweni ezibalulekile kungavula umnyango wamaphutha abizayo kanye nezingxabano zomthetho.
Izinhlelo ze-AI zivame kangakanani ukubona izinto ezingekho?
Naphezu kwentuthuko esheshayo, ukubona izinto ezingekho ngokoqobo kusalokhu kuyinkinga elinganiswayo futhi engeyona encane kumamodeli asezingeni eliphezulu . Cishe ngo-2023, ezinye izifundo zilinganisela ukuthi kufika ku-27% wezimpendulo ezivela kumamodeli amakhulu olimi zazinephutha elithile lamaqiniso, kanti amazinga okubona izinto ezingekho ngokoqobo adlula u-40% wezinhlobo ezithile zemisebenzi noma amasethi edatha. Lokho kusho ukuthi ezimweni eziningi ezingokoqobo, cishe ingxenye yezimpendulo ezibonakala ziqinile ingafaka okungenani imininingwane eyodwa edukisayo.
Umsebenzi wakamuva usikisela ukuthi izizukulwane ezintsha zamamodeli zithuthukile, kodwa azizange ziqede inkinga . Isibonelo, izivivinyo zangaphakathi ezicashunwe emibikweni yezimboni zithi imodeli ethuthukisiwe kakhulu inganciphisa amazinga okubona izinto ezingekho ngokoqobo kusukela cishe ku-20.6% kuya cishe ku-4.8% ezilinganisweni ezikhethiwe. Ezinye izinhlelo, njengokulungiswa okuthile kwe-Gemini 2.0 Flash 001, kubikwa ukuthi zifinyelele amazinga okubona izinto ezingekho ngokoqobo acishe abe ngu-1% ekuhlolweni okulinganiselwe, okuchazwe kahle ngokusebenzisa amasu okuhlola okuzenzakalelayo.
Ngesikhathi esifanayo, abacwaningi baphawule i-paradox ethakazelisayo . Amanye amamodeli okucabanga ayinkimbinkimbi kakhulu, akhiqiza uchungechunge olude lwemicabango kanye nezincazelo ezinemininingwane eminingi, angase abangele amaphutha amaningi kunokuba abangele okuncane. Isizathu siyaqondakala: izimpendulo ezinde zinikeza amathuba amaningi emodeli okuphambuka emaqinisweni, ukuwela ezicabangweni eziqagelayo noma ukwakha izimpikiswano zezinyathelo eziningi phezu kwendawo yokuqala engazinzile.
Umsebenzi wezemfundo usuqalile ukuhlukanisa nokulinganisa ukungaboni kahle ngezindlela ezihlelekile . Ucwaningo mayelana nokungaboni kahle kwe-LLM luphakamisa ama-taxonomies ahlukanisa phakathi kokungaboni kahle kwangaphakathi (lapho umphumela ungahambisani nokufakwayo okunikeziwe, njengesifinyezo esingalungile sedokhumenti ethile) kanye nokungaboni kahle kwangaphandle (lapho imodeli ingeza khona imininingwane enengqondo kodwa engasekelwa). Amabhodi wabaphambili afana ne-Vectara Hallucination Leaderboard aqhathanisa amamodeli athandwayo ezivivinyweni ezijwayelekile ukuze abeke ezingeni ukuthi yimaphi angase akhiqize okuqukethwe okunephutha.
Izindlela ezintsha zokuthola zidlula ukuhlolwa okulula kokunemba . Olunye ucwaningo oluphawulekayo lusebenzisa i-semantic entropy - cishe, ukuthi imodeli ayiqiniseki kangakanani ngalokho okufanele ikutshele - ukumaka imiphumela esolisayo. Uma imodeli ikhombisa ukungavumelani okukhulu kwangaphakathi phakathi kokuqhubeka okuningi okungenzeka, leyo entropy ye-semantic ephakeme ingaba uphawu oluyisixwayiso sokuthi impendulo yamanje cishe izobonakala sengathi iyinto engekho, noma ngabe kungekho mpendulo yegolide ejwayelekile etholakalayo.
Amasu okunciphisa nokubona imibono engekho
Ngenxa yokuthi imibono engekho ibangelwa indlela amamodeli akhiqizayo asebenza ngayo, akukho sixazululo esisodwa somlingo, kodwa kuvele amasu amaningana athembisayo . Enye indlela enethonya i-Retrieval-Augmented Generation (RAG). Esikhundleni sokuvumela imodeli ithembele kuphela kulokho eyayikukhumbula ngesikhathi sokuqeqeshwa, uhlelo lwe-RAG luqala ngokubuza imithombo yedatha yangaphandle, ekhethiwe - njengemibhalo, izisekelo zolwazi noma amakhasi ewebhu anezikhombo - bese lucela imodeli ukuthi ikhiphe impendulo esekelwe ngokucacile kulawo madokhumenti atholiwe.
I-RAG inika imodeli into eqondile okufanele incike kuyo . Uma umsebenzisi ebuza umbuzo, uhlelo lulanda izindima ezifanele, bese imodeli ikhuthazwa ukuthi izifinyeze noma izihlanganise, kunokuba izenze kusukela ekuqaleni. Lokhu kunganciphisa kakhulu ukubona izinto ezingekho emikhakheni lapho okuqukethwe okuthembekile nokusesikhathini kutholakala khona, njengezincwadi zobuchwepheshe, izinqubomgomo zenkampani zangaphakathi noma izizindalwazi zesayensi.
Esinye isivikelo ukuqinisekiswa kwangaphakathi kanye nokuhlola ukuzivumelanisa nezimo . Esikhundleni sokuthembela ekudluleni okukodwa kumodeli, ezinye izinhlelo zikhiqiza izimpendulo eziningana ezingaba khona bese ziqhathanisa. Uma zonke izinguqulo zihlangana esitatimendeni esifanayo, kunethuba elikhulu lokuthi zizinzile futhi zilungile; uma zihluka, uhlelo lungayeka, lucele ukucaciswa noma luveze ngokusobala ukungaqiniseki. Lokhu kuhlola ukuzivumelanisa nezimo kungenziwa ngokuzenzakalela futhi sekuvele kubonise ithemba ekunciphiseni imibono engekho emibonweni yokucabanga.
Abacwaningi baphinde baphakamisa izindlela eziningi zokuhlola, njengezindlela zokuhlela kabusha umbuzo womsebenzisi ngaphambi kokuphendula . Isibonelo esisodwa esivela emsebenzini wakamuva wezemfundo yindlela eyaziwa ngokuthi i-Acurai, ehlose ukuguqula izikhuthazo ngendlela eqondisa imodeli ezindleleni zokucabanga ezingaqinisekiswa. Nakuba izindlela ezinjalo zisesezigabeni zokuqala, zikhomba esikhathini esizayo lapho imodeli ichitha khona isikhathi sayo esiningi sokuhlela ukuthi ingacabanga kanjani ngombuzo esikhundleni sokugxumela ngqo ekulandiseni okushelelayo.
Ngasohlangothini lokuqapha, ama-semantic entropy probes namathuluzi ahlobene nawo anikeza indlela eshibhile yokuthola imibono engaba khona . Ngokulinganisa ukuthi imodeli "ayivumelani kangakanani nayo" ngaphakathi lapho ikhipha impendulo, la ma-probes angamaka izindima ezifanelwe ukubuyekezwa ngabantu. Okubalulekile, amasu anjengalawa awadingi isethi yedatha enelebula yezimpendulo ezifanele, okwenza zifanelekele izimo ezivulekile, zangempela.
Indima yokwengamela kwabantu kanye nokusetshenziswa okunesibopho
Ngisho noma kunezakhiwo ezithuthukisiwe kanye nokutholwa okuhlakaniphile, ukubhekwa kwabantu kuhlala kuyisisekelo sokuphatha imibono engekho emthethweni ye-AI . Izinhlelo zokusebenza eziqine kakhulu yilezo eziphatha i-AI njengomsizi onamandla noma umshayeli wendiza osizayo, hhayi njengomenzi wezinqumo ozimele. Emikhakheni efana nokunakekelwa kwezempilo, amabhange noma ukuphathwa komphakathi, umkhuba omuhle ukuvumela imodeli ihlunge ulwazi, ibhale amadokhumenti noma ifinyeze izinto ezinde, kuyilapho uchwepheshe ongumuntu enza isibuyekezo sokugcina kanye nokuvunyelwa.
Bobabili abaholi bemboni nabacwaningi bagcizelela ukuthi abasebenzisi akufanele bathembele ngokungacabangi imiphumela ye-AI, ikakhulukazi ezihlokweni ezibucayi . Izinkampani ezingemuva kwamamodeli amakhulu zixwayise ngokusobala ukuthi izinhlelo zazo zisengacabanga ukuthi azinazo izimpendulo nokuthi izimpendulo kufanele zisetshenziswe ngokuqapha ukuze zithole iseluleko sezokwelapha, sezomthetho noma sezezimali. Abanye abahlinzeki bathembele kubahloli babantu ukuze bahlole ukucwasa, amaphutha amaqiniso kanye nokuqukethwe okungaba yingozi, behlanganisa impendulo yabo ekuqeqeshweni ngokufunda okuqinisiwe okuvela kumpendulo yabantu (RLHF).
Ukucaca mayelana nemikhawulo kudlala indima ebalulekile ekulondolozeni ukwethenjwa kwabasebenzisi . Lapho abahlinzeki bevuma ukuthi amamodeli abo angakhiqiza ulwazi olunganembile noma oluqanjiwe, abasebenzisi banamathuba amaningi okugcina izinga elinempilo lokungabaza, bahlole izimpendulo kabili futhi bagweme ukunikeza abanye izahlulelo ezibucayi. Ngokuphambene nalokho, ukuthengisa ngokweqile i-AI njengengenaphutha noma "ehlakaniphile" ngomqondo womuntu kuphakamisa okulindelwe ngokungenangqondo futhi kwenza imibono engekho ingokoqobo izwakale njengokukhashelwa kunokuba kube yizinto zobuchwepheshe.
Uma sibheka phambili, imithethonqubo cishe izobumba indlela izinhlangano eziphatha ngayo izingozi zokubona izinto ezingekho . Abenzi benqubomgomo sebevele babhala imithetho edinga ukuhlolwa kwengozi, izinqubo zabantu ngaphakathi kanye nokubhalwa okucacile kwamakhono amamodeli kanye nezindlela zokwehluleka. Ezindaweni eziningi, izinkampani ezisebenzisa amathuluzi e-AI ezindaweni ezinobungozi obukhulu kuzodingeka zibonise ukuthi zinezivikelo ezanele zokuvimbela ukubona izinto ezingekho ukuthinta ngqo izinqumo ezibalulekile.
Ekugcineni, amasethingi athembeke kakhulu yilawo aklama ukusebenza okuzungeze amandla nobuthakathaka be-AI . Amamodeli akhiqizayo avelele ekukhiqizeni ama-draft, ukuhlola izinketho, ukuhlinzeka ngemisho ehlukile kanye namaphethini okuveza umbhalo omningi. Awethembekile kangako njengabaxazululi beqiniso bodwa. Lapho izinhlelo zakhiwa ukuze abantu bahlale belawula izinqumo ezibalulekile kanye nezinyathelo zokuqinisekisa, imibono engekho ingokoqobo iba umsindo olawulekayo kunokuba kube ukwehluleka okubi kakhulu.
Njengoba lobu buchwepheshe buthuthuka, isimo sengqondo esingokoqobo mayelana nokubona izinto ezingekho emthethweni nge-AI sizobaluleka kubo bonke abathintekayo - abathuthukisi, izinhlangano, abalawuli kanye nabasebenzisi bansuku zonke. Ukuqonda ukuthi impendulo ecacile akuyona ngempela enembile, ukuhlanganisa amamodeli nemithombo yolwazi lwangaphandle, ukulinganisa nokuthola izinto ezingekho emthethweni ngendlela ehlelekile, nokugcina ukwahlulela kwabantu kuphakathi kwezinqumo ezibalulekile konke kuyingxenye yokusebenzisa i-AI yokukhiqiza ngokuhlakanipha. Esikhundleni sokubuza ukuthi ukubona izinto ezingekho emthethweni kuzonyamalala ngokuphelele, umbuzo owusizo kakhulu ukuthi singazakha kanjani izinhlelo, izikhuthazo kanye nemikhuba eyenza lawo maphutha angenakugwenywa abonakale, aqukethwe futhi angalimazi kangako.
