Konke mayelana ne-Generative Artificial Intelligence: ukuthi isebenza kanjani, isetshenziswa kanjani, kanye nezingozi

Informatec Digital » Intelligence yokufakelwa » Konke mayelana ne-Generative Artificial Intelligence: ukuthi isebenza kanjani, isetshenziswa kanjani, kanye nezingozi

Ubuhlakani bokwenziwa obukhiqizayo sebube obunye bobuchwepheshe obujabulisayo nobuphazamisayo kakhulu besikhathi sethu. Amandla alo okudala okuqukethwe okusha ngokuphelele ashintshe ngokuphelele indlela esisebenzisana ngayo nobuchwepheshe nolwazi . Kunokuba nje silingise ubuhlakani bomuntu, i-AI manje isiyakwazi ukusungula izinto ezintsha, ukuxazulula izinkinga ngobuhlakani, nokukhiqiza izinto zokuqala cishe ngokuzimela . Uma wake wazibuza ukuthi yini ebangela izimo ezifana ne-ChatGPT, i-DALL-E, noma ama-deepfake adumile, lesi sihloko sizokwembula konke okudingeka ukwazi mayelana nobuhlakani bokwenziwa obukhiqizayo, kusukela ekusebenzeni kwayo okuyisisekelo kuya emiphumeleni yayo yokuziphatha kanye neyomthetho.

Eminyakeni yamuva nje, sibone inguquko yangempela ekuhlakanipheni kokwenziwa , ngentuthuko esukela ekwakhiweni okuzenzakalelayo kombhalo nekhodi kuya ekudalweni kwezithombe, umculo, amavidiyo, ngisho nama-molecule amasha ezokwelapha. I-AI ekhiqizayo inamandla okuguqula imikhakha eyahlukahlukene njengemfundo, ukunakekelwa kwempilo, ubuciko, ukumaketha, kanye nokuhlela . Kodwa-ke, futhi ibeka izinselele ezinkulu kanye nezingozi maqondana nobumfihlo, impahla yobuhlakani, kanye nolwazi olungalungile . Ukulungele ukuthola zonke izinto eziyinkimbinkimbi zayo?

Yini i-generative intelligence yokwenziwa?

I-AI Ekhiqizayo iyigatsha elikhethekile lobuhlakani bokwenziwa eligxile ekudaleni okuqukethwe kwasekuqaleni ngokubhekisela kudatha eningi ekhona. Ngokungafani namamodeli e-AI endabuko, avame ukuhlaziya, ukuhlukanisa, noma ukubikezela ulwazi, i-AI ekhiqizayo iyakwazi ukukhiqiza imibhalo emisha, izithombe, umsindo, ividiyo, ngisho nezakhiwo zama-molecule , ngokusekelwe kumaphethini ewafundile ngesikhathi sokuqeqeshwa kwayo.

I-AI Ekhiqizayo isebenza ngokusebenzisa amamodeli okufunda okujulile , asebenzisa amanethiwekhi e-neural ukukhipha ubudlelwano nezakhiwo ezifihliwe kusuka kudatha enkulu. Kusukela lapho, futhi ngemuva kwenqubo yokuqeqeshwa okujulile, la mamodeli angaphendula emiyalweni (noma 'iziphakamiso') ngokukhiqiza okuqukethwe okukholisayo nokungokoqobo kangangokuthi, ezimweni eziningi, cishe akuhlukaniseki kulokho umuntu angakudala.

Asebenza kanjani amamodeli e-AI akhiqizayo?

Ukuze uqonde amandla e- AI yokukhiqiza , kubalulekile ukwazi ukuthi amamodeli ayo aqeqeshwa futhi ahlelwe kanjani. Konke kuqala ngamamodeli ayisisekelo . Lawa aqukethe amanethiwekhi anamandla e-neural aklanyelwe ukuphatha izinhlobo ezahlukene zokuqukethwe: umbhalo, izithombe, umsindo, ividiyo, noma ngisho nedatha ye-multimodal—okungukuthi, inhlanganisela yezinhlobo ezahlukene zokufaka.

Inqubo yokuqeqesha yalezi zinhlobo isekelwe enanini elikhulu ledatha engahlelekile nengabhaliwe , evame ukutholakala ku-inthanethi. Phakathi nalesi sigaba, imodeli yenza izigidi zemisebenzi, njengokubikezela igama elilandelayo emshweni, ukuhlonza isici esilandelayo esithombeni, noma ukuphakamisa imiyalelo yekhodi. Umgomo uwukulungisa kahle amapharamitha ayo angaphakathi ukuze kuncishiswe ukungafani phakathi kwezibikezelo zayo nedatha yangempela, ngaleyo ndlela kucwengisiswe ikhono layo lokukhiqiza okuqukethwe okunengqondo.

Uma ukuqeqeshwa sekuqediwe, amamodeli ayisisekelo angasetshenziswa ukuthuthukisa izinhlelo zokusebenza ezikhethekile: kusukela kumamodeli okucubungula ulimi lwemvelo (njenge-OpenAI's GPT-4 noma i-Google DeepMind's Gemini) kuya kumathuluzi akhiqiza izithombe, umculo, noma ividiyo kusukela ezincazelweni noma emiyalweni elula.

Ukuziphendukela kwemvelo kanye nezenzakalo ezibalulekile ze-AI ekhiqizayo

Ukuthuthukiswa kobuhlakani bokwenziwa obukhiqizayo akukaze kube okuqondile. Imvelaphi yakho itholakala kumamodeli ezibalo kanye nama-algorithms okufunda komshini akudala, kodwa igxathu langempela lekhwalithi lifike ngokufunda okujulile kanye namanethiwekhi e-neural ajulile . Kuze kube muva nje, i-AI igxile kakhulu ekuhlukaniseni, ekuhlaziyeni okubikezelayo, kanye nokuqashelwa kwamaphethini; nokho, ngokuvela kobuchwepheshe obufana namaNethiwekhi Aphikisayo Ajwayelekile (ama-GAN) kanye namamodeli asekelwe ku-transformer , konke kwashintsha.

Ngo-2014, ama-GAN okuqala enze kwaba nokwenzeka ukukhiqiza izithombe ezintsha ngokuphelele nezikholisayo ngomsindo ongahleliwe. Kodwa ushintsho lwangempela lwafika ngo-2017 nge-Transformers , isakhiwo esaguqula ukucutshungulwa kolimi lwemvelo futhi saholela kumamodeli afana ne-BERT, i-GPT-2, kanye ne-GPT-3 kanye ne-GPT-4 kamuva. Lawa mamodeli abonise ikhono elingakaze libonwe lokuqonda nokukhiqiza umbhalo ofana nomuntu ngaphandle kokudinga ukuqeqeshwa okuqondile komsebenzi ngamunye.

Ngesikhathi esifanayo, amasu afana nama -variational autoencoders (VAE) kanye namamodeli okusabalalisa enze kwaba nokwenzeka ukudala izithombe, umculo namavidiyo ekhwalithi evelele, okuvula amathuba amasha emikhakheni yokudala neyesayensi.

Akhiwe kanjani amasistimu e-AI akhiqizayo?

Izinhlelo ze -AI ezikhiqizayo zingahlukaniswa ngokwezinhlobo zedatha ezisebenza ngazo:

• I-Unimodal: phatha uhlobo olulodwa lokufaka (isibonelo, umbhalo kuphela noma isithombe kuphela).
• I-Multimodal: Ayakwazi ukuhlanganisa izinhlobo ezahlukene zolwazi (umbhalo nesithombe, umsindo nevidiyo, njll.), okuwenza aguquguquke kakhulu futhi abe namandla.

Amanye amamodeli ayisisekelo aziwa kakhulu kanye nezinhlelo zawo zokusebenza eziphambili yilezi:

• I-GPT (OpenAI): ukucubungula nokukhiqiza ulimi lwemvelo, ukuhumusha ngomshini, ukubhala umbhalo, usizo ngemisebenzi eyinkimbinkimbi, njll.
• I-DALL·E (OpenAI): ukukhiqizwa kwezithombe ezivela ezincazelweni zombhalo.
• U-Claude (Anthropic): ukuhlaziywa kombhalo okuthuthukisiwe, ukufinyezwa nokukhiqizwa kwekhodi.
• I-Gemini (Google DeepMind): Ukuphathwa kokuqukethwe kwe-multimodal okuhlanganisa umbhalo, izithombe namavidiyo.
• FLAME (Igoli): imodeli yolimi oluvulekile eyenzelwe ucwaningo oluthuthukisiwe.
• I-Copilot (Microsoft + OpenAI): umsizi wokuhlela nokuzenzakalela kwesoftware kanye nezindawo zokukhiqiza.

Ukusetshenziswa okusebenzayo kobuhlakani bokwenziwa okukhiqizayo

Umthelela wangempela we -AI yokukhiqiza ubonakala emikhakheni eminingi lapho isivele inezela khona inani:

• imfundo: ukwenza kube ngokwakho ukufunda, ukukhiqizwa kwezinto zokufundisa, ukulungisa okuzenzakalelayo komsebenzi, ukulingisa amacala angokoqobo, njll.
• Imithi: ukwakheka kwesithombe sezokwelapha, ukukhiqizwa kombiko womtholampilo, ukwesekwa kokuxilonga, ukusheshisa ukutholakala kwezidakamizwa, ukudalwa kwamaphrotheni asungulayo noma ukulandelana kwamangqamuzana.
• Ezokuzijabulisa kanye nabezindaba: ukudalwa komculo, imibhalo, ama-deepfakes ayimpikiswano, imifanekiso, imidlalo yevidiyo, nokuqukethwe okulalelwayo nokubonwayo ngokwezifiso.
• Ukumaketha nokuklama: ukubhala imikhankaso yokukhangisa, ukwakhiwa kwelogo, ukwakheka komkhiqizo, namasu okwenza umyalezo kube ngokwakho.
• Ukuhlela: Abasizi abafana ne-GitHub Copilot ekhiqiza ikhodi ngokuzenzakalelayo, ibona amaphutha, futhi iphakamise ukuthuthukiswa.
• Ibhizinisi: ukuzenzekelayo kwemisebenzi yokuqondisa, ukubika, ukuhlaziya idatha, nokwenza kahle kwesevisi yamakhasimende ngama-chatbots athuthukile.

Ukuguquguquka kwe-AI ekhiqizayo kusho ukuthi ukwamukelwa kwayo akunamkhawulo owaziwayo , futhi nsuku zonke izinhlelo zokusebenza ezintsha ziyahlolwa emikhakheni eyahlukahlukene njengobunjiniyela, amandla, ezokuthutha noma umkhakha wezomthetho.

Iyiphi indima edlalwa ukwaziswa ekukhiqizeni i-AI?

Esinye sezihluthulelo zokuthola okuningi kubuhlakani bokwenziwa obukhiqizayo ukwazi ukwakha imiyalelo emihle, noma 'imiyalezo'. Isaziso siyi-oda noma umyalelo—obhaliwe noma okhulunyiwe—ofakwa ohlelweni ukuze uthole impendulo yomuntu siqu . Uma i-oda licacile futhi linembile, kulapho umphumela okhiqizwa yimodeli uba ngcono khona.

Ukukhethwa kwamagama, ikakhulukazi izenzo, kudlala indima ebalulekile. Ukucela i-AI ukuthi 'ibhale' akufani nokucela ukuthi 'ifinyeze', 'ichaze', 'ichaze', noma 'ifanekise'. Amathuluzi akhona ukusiza ukuthuthukisa izikhuthazo nokuthuthukisa imiphumela, okuvula umnyango wezinhlelo zokusebenza eziningi ezisebenzayo kanye nemiphumela esezingeni eliphezulu.

Izinselele kanye nezingozi zobuhlakani bokwenziwa obukhiqizayo

Kanye nezinzuzo zayo, i-AI ekhiqizayo idala izingozi nezinselele okumele kubhekwane nazo ngokujulile :

• Ukuchema nolwazi olungaqondile: Uma amamodeli eqeqeshwa ngedatha echemile, angasakaza imibono engaqondile noma adale ulwazi olungamanga. Okuthiwa 'ama-hallucinations' (izimpendulo eziqanjiwe) zivamile.
• Ubumfihlo kanye nokusetshenziswa kabi kwedatha: Ukufinyelela kumthamo omkhulu wolwazi kungabeka ubumfihlo kanye nokusetshenziswa ngendlela efanele kwedatha yomuntu siqu engcupheni.
• Impahla yengqondo nokudala: Imibuzo iyaphakanyiswa mayelana nokuthi ubani ngempela umnikazi wokuqukethwe okukhiqizwa yi-AI kanye nendlela yokuvikela ngokusemthethweni lokhu okudaliwe.
• Umthelela emsebenziniEmikhakheni yokudala noma yokuphatha, ukuzenzela kanye nokusebenza kahle kwe-generative AI kungangena esikhundleni semisebenzi ethile yabantu.
• Ama-deepfakes kanye nokwebiwa komazisiAmasu afana nama-GAN avumela ukudalwa kwamavidiyo mbumbulu, izithombe, noma umsindo olalelwayo wekhwalithi ephezulu kangangokuthi usetshenziselwe ukukhohlisa noma ukusabalalisa ulwazi olungaqondile.

Lezi zinselele zenze ukulawulwa kwe-AI ekhiqizayo kwaba yinto eza kuqala kohulumeni nezinhlangano zomhlaba wonke , ezifuna ukuqinisekisa ukusetshenziswa kobuchwepheshe okunesibopho, okusobala kanye nokuziphatha okuhle.

Ama-Deepfakes kanye nokukhohlisa kwemidiya

Esinye sezimo eziphikisana kakhulu eziqhutshwa yi -AI yokukhiqiza yi- deepfakes . Lawa amavidiyo aguquliwe ngedijithali, ukuqoshwa komsindo, noma izithombe ezenzelwe ukuzenza abantu bangempela kusetshenziswa amanethiwekhi okukhiqiza abaphikisi. Ama-Deepfakes angasetshenziswa ngezinjongo ezisemthethweni (njengemiphumela ekhethekile kumafilimu noma ekuqeqeshweni kwezokwelapha), kodwa futhi nokusabalalisa ulwazi olungamanga, ukuzenza abanye, noma ukudala izinto ezingamanga ngenhloso embi.

Ukuthola i-deepfake akulula ngaso sonke isikhathi, kodwa kunezinkomba ezithile: ukunyakaza okungavamile, ukucwayiza okungaguquki, ukungalungi ngokulingana kobuso, noma ukuvumelanisa izindebe ngendlela engalungile. Ukwanda kwama-deepfake kuholele kubuchwepheshe obusha bokuqinisekisa ubuqiniso bokuqukethwe okulalelwayo nokubukwayo kanye nemithethonqubo ethile yokulawula nokunciphisa kwawo ezinhlelweni zomthetho ezahlukahlukene.

Indima yokulawula kanye nezimiso zokuziphatha ekukhiqizeni i-AI

Njengoba i-AI ekhiqizayo ihlanganiswa empilweni yansuku zonke nasezindaweni zobungcweti , kudingeka izinhlaka zomthetho ezicacile. Isibonelo, i-European Union isivele ikhuthaze imithetho ebhekana nale nkinga. Izivumelwano nemithethonqubo mayelana nokulebula kanye nokusetshenziswa okunesibopho kwalolu hlobo lokuqukethwe nakho kumenyezelwe e-United States nase-China.

Ukuziphatha kuba yinto ebalulekile lapha . Izingxoxo zamanje zigxile ekubeni ngumbhali, ilungelo lobumfihlo, ukusetshenziswa okunonya okungenzeka, kanye nomthelela obanzi emphakathini ongaba nawo i-AI yokukhiqiza engalawulwa. Abagxeki bafuna ukucaca okukhulu mayelana nedatha esetshenziswa ukuqeqesha amamodeli kanye nesidingo sezindlela ezisebenzayo zokuvimbela ulwazi olungamanga.

Amamodeli, izakhiwo kanye namasu: I-RNN, Ama-Transformers, i-GAN, i-VAE kanye nokusabalalisa

Izikhali zobuchwepheshe ezisetshenziswa yi-AI ekhiqizayo zinkulu futhi zishintsha njalo. Izakhiwo ezivelele kakhulu zifaka:

• I-Recurrent Neural Networks (RNN): Isetshenziselwa kakhulu ukusebenza nokulandelana kwedatha (njengolimi noma umculo), nakuba inemikhawulo ethile yokungaguquguquki kwesikhathi eside.
• Transformers: okwamanje ireferensi yezakhiwo. Ikhono labo lokuhlaziya ubudlelwano obuyinkimbinkimbi phakathi kwamagama noma izakhi zedatha kubenza bafaneleke kumamodeli olimi amakhulu (LLM) nokuqukethwe kwezindlela eziningi.
• I-Generative Adversarial Networks (GAN): Akhiwe amanethiwekhi amabili aqhudelanayo (ijeneretha kanye nombandlululi). Zibe zibalulekile ekukhiqizeni izithombe namavidiyo angokoqobo.
• Ama-autoencoder ahlukahlukene (VAE): Zikuvumela ukuthi uhlanganise izithombe noma umculo, udale okuhlukile okuzwakalayo okuvela kumaphethini okufunda.
• Amamodeli wokusabalalisa: isetshenziselwa imisebenzi yokuthwebula ethuthukisiwe, kancane kancane iguqule umsindo ongahleliwe ube izithombe ezinemininingwane, ezingokoqobo.

Imikhawulo nezinselelo ezisalindile

Naphezu kwentuthuko emangalisayo, i-AI yokukhiqiza isabhekene nezinselele ezinkulu :

• Ukulawulwa kwekhwalithi: Ukuqinisekisa imiphumela efanele futhi engaguquki kuhlala kuyinselele, ikakhulukazi ekukhiqizeni umbhalo nezithombe.
• Ukutolika: : Ngokuvamile akunakwenzeka ukuqonda ukuthi i-AI ifika kanjani emiphumeleni ethile noma ezinqumweni.
• Ukusetshenziswa kwezinsiza kusebenza: Amamodeli anamandla kakhulu adinga ingqalasizinda ethuthukisiwe, ukusetshenziswa kwamandla okuphezulu, nezindleko zokulungisa eziphezulu.
• Umgunyathi kanye nobugebengu be-inthanethi: Ukudala okuqukethwe okungelona iqiniso okuphezulu kungasetshenziselwa imikhonyovu, ukukhwabanisa, nokuhlasela konjiniyela bezenhlalo.

Ikusasa lobuhlakani bokwenziwa obukhiqizayo

Ucwaningo lubonisa ukuthi i-AI yokukhiqiza izoqhubeka nokukhula ekusebenzeni kahle, ikhwalithi, kanye namakhono okwenza kube ngokwakho . Amamodeli azoba ngamazinga amaningi, akhethekile emikhakheni ethile, futhi aguquguquke kalula. Ukubambisana phakathi kwabantu nemishini kuzoba yisisekelo sokuthuthukisa ubuciko kanye nokusungula izinto ezintsha . Kodwa-ke, kuzobaluleka ukuhambisana nalezi ntuthuko ngemithethonqubo efanele kanye nokuziphatha okunomthwalo wemfanelo okwandisa izinzuzo futhi kunciphise izingozi.

I-athikili ehlobene:

Iyini i-Generative AI? Imininingwane, izibonelo kanye nezingozi