I-AI izizwa icishe ifane neqhinga lomlingo ngezinye izikhathi. Uthayipha umbuzo ongahleliwe, bese u-bam - impendulo eshelelayo, epholishiwe ivela ngemizuzwana. Kodwa nali i-curveball: ngemuva kwawo wonke umshini “wobuhlakani”, kunabantu bangempela abawugudlayo, abawulungisayo, futhi bawulolonge endleleni. Labo bantu babizwa ngabaqeqeshi be-AI , futhi umsebenzi abawenzayo ungowabantu abangabazi, uyahlekisa, futhi ngokwethembeka ungumuntu ongaphezu kwalokho abantu abaningi abacabanga ngakho.
Ake sihlole ukuthi kungani laba baqeqeshi bebalulekile, ukuthi usuku lwabo lwansuku zonke lubukeka kanjani, nokuthi kungani le ndima iqhuma ngokushesha kunanoma ubani owayebikezele.
Izindatshana ongathanda ukuzifunda ngemva kwalesi:
🔗 Yini i-AI arbitrage: Iqiniso ngemuva kwe-buzzword
Ichaza i-AI arbitrage, ubungozi bayo, izinzuzo, kanye nemibono eyiphutha evamile.
🔗 Izidingo zokugcinwa kwedatha ye-AI: Odinga ukukwazi ngempela
Imboza izidingo zesitoreji, ukukala, nokusebenza kahle kwamasistimu e-AI.
🔗 Ubani ubaba we-AI?
Ihlola amavulandlela e-AI kanye nemvelaphi yobuhlakani bokwenziwa.
Yini eyenza umqeqeshi we-AI oqinile? 🏆
Akuwona umsebenzi wokuhlanganisa izinkinobho. Abaqeqeshi abangcono kakhulu bancike ekuxubeni okuhle okungajwayelekile kwamathalenta:
-
Ukubekezela (okuningi) - Amamodeli awafundi ngeshothi elilodwa. Abaqeqeshi baqhubeka belungisa izilungiso ezifanayo kuze kunamathele.
-
I-nuance ecacile - Ukubamba ukubhuqa, umongo wamasiko, noma ukuchema yilokho okunikeza impendulo yomuntu umkhawulo wayo [1].
-
Ukuxhumana okuqondile - Ingxenye yomsebenzi ibhala imiyalelo ecacile i-AI engakwazi ukuyifunda kabi.
-
Ilukuluku + izimiso zokuziphatha - Umqeqeshi omuhle ubuza ukuthi ingabe impendulo “ilungile ngempela” kodwa ithoni yezenhlalo yabantu abayizithulu - itimu eyinhloko ekwengameleni kwe-AI [2].
Kalula nje: umqeqeshi uyingxenye yothisha, umhleli wengxenye, kanye nodeshi lwesazi sokuziphatha.
Izindima Zomqeqeshi we-AI ngokubuka nje (NgamaQuirs athile 😉)
| Uhlobo Lwendima | Ubani Olingana Kakhulu | Inkokhelo Ejwayelekile | Kungani Isebenza (noma ingasebenzi) |
|---|---|---|---|
| Ilebula Yedatha | Abantu abathanda imininingwane emihle | Phansi–Phakathi $$ | Kubaluleke kakhulu; uma amalebula ewubudlabha, yonke imodeli iyahlupheka [3] 📊 |
| Uchwepheshe we-RLHF | Ababhali, abahleli, abahlaziyi | Okumaphakathi–Phezulu $$ | Ilinganisa futhi ibhale kabusha izimpendulo ukuze iqondanise iphimbo nokucaca nokulindelwe umuntu [1] |
| Umqeqeshi Wesizinda | Abameli, odokotela, ochwepheshe | Kuyo yonke imephu 💼 | Iphatha i-niche jargon kanye nama-edge case kumasistimu aqondene nomkhakha othile |
| Umbuyekezi Wezokuphepha | Abantu abanomqondo wokuziphatha | Okumaphakathi $$ | Isebenzisa imihlahlandlela ukuze i-AI igweme okuqukethwe okuyingozi [2][5] |
| Umqeqeshi Wokudala | Abaculi, ababalisi | Okungalindelekile 💡 | Isiza i-AI echo emcabangweni ngenkathi ihlala ngaphakathi kwemingcele ephephile [5] |
(Yebo, ukufometha kungcolile - kufana nomsebenzi ngokwawo.)
Usuku Empilweni Yomqeqeshi we-AI
Ngakho-ke umsebenzi wangempela ubukeka kanjani? Cabanga ngokubhala amakhodi okukhangayo nokunye okwengeziwe:
-
Ukulinganisa izimpendulo ezibhalwe nge-AI kusukela kokubi kakhulu kuye kwezingcono kakhulu (isinyathelo sakudala se-RLHF) [1].
-
Ukulungisa ukuxubana (njengalapho imodeli ikhohlwa iVenus akuyona iMars).
-
Ukubhala kabusha izimpendulo ze-chatbot ukuze zizwakale ngokwemvelo.
-
Ilebula izintaba zombhalo, izithombe, noma umsindo - lapho ukunemba kubaluleke khona ngempela [3].
-
Ukuphikisana ngokuthi "okulungile ngobuchwepheshe" kuhle ngokwanele noma uma imihlahlandlela yokuphepha kufanele idlule [2].
Kuyingxenye yokugaya, ingxenye yephazili. Eqinisweni, zibone ngeso lengqondo ufundisa upholi hhayi nje ukukhuluma kodwa ukuyeka ukusebenzisa amagama okungalungile kancane - lokho kuyi-vibe. 🦜
Kungani Abaqeqeshi Bebaluleke Kakhulu Kunokuba Ucabanga
Ngaphandle kokuqondisa kwabantu, i-AI inga:
-
Umsindo uqinile futhi u-robotic.
-
Sabalalisa ukuchema ungahloliwe (umcabango owesabekayo).
-
Ungaphuthelwa ngokuphelele amahlaya noma uzwela.
-
Ungavikeleki kancane ezimweni ezibucayi.
Abaqeqeshi yibo abanyonyoba “ezintweni zomuntu ezingcolile” - isitsotsi, imfudumalo, isingathekiso esifiphele ngezikhathi ezithile - kuyilapho besebenzisa izinsimbi zokuqapha ukuze kugcinwe izinto ziphephile [2][5].
Amakhono Abalayo Empeleni
Khohlwa inganekwane yokuthi udinga iPhD. Okusiza kakhulu ukuthi:
-
Ukubhala + ukuhlela ama-chops - Umbhalo opholishiwe kodwa ozwakalayo wemvelo [1].
-
Ukucabanga kokuhlaziya - Ukubona amaphutha emodeli aphindaphindiwe nokulungisa.
-
Ukuqwashisa ngamasiko - Ukwazi ukuthi kunini lapho umushwana kungase kubonakale kungalungile [2].
-
Ukubekezela - Ngoba i-AI ayibambeki ngokushesha.
Amaphoyinti ebhonasi wamakhono ezilimi eziningi noma ubuchwepheshe be-niche.
Lapho Abaqeqeshi Bevela Khona 🌍
Lo msebenzi akuwona nje omayelana nama-chatbots - ungena ngokunyenya kuyo yonke imikhakha:
-
Ukunakekelwa kwezempilo - Imithetho yokubhala isichasiselo yamacala asemngceleni (enanelwe kumhlahlandlela wezempilo we-AI) [2].
-
Ezezimali - Ukuqeqesha amasistimu okuthola ukukhwabanisa ngaphandle kokuminzisa abantu kuma-alamu angamanga [2].
-
Ukuthengisa - Ukufundisa abasizi ukuthola i-slangy shopper lingo ngenkathi benamathela kuthoni yomkhiqizo [5].
-
Imfundo - Ukubumba ama-bots okufundisa ukuthi akhuthaze esikhundleni sokusekela [5].
Ngokuyisisekelo: uma i-AI inesihlalo etafuleni, kukhona umqeqeshi ocashe ngemuva.
I-Ethics Bit (Ayikwazi Ukweqa Lokhu)
Nakhu lapho kuba nesisindo khona. Uma ingahloliwe, i-AI iphinda imibono engaqondile, ulwazi olunganembile, noma okubi nakakhulu. Abaqeqeshi bayeka lokho ngokusebenzisa izindlela ezifana ne-RLHF noma imithetho yomthethosisekelo eqondisa amamodeli ezimpendulo eziwusizo, ezingenabungozi [1][5].
Isibonelo: uma i-bot iphusha izincomo zomsebenzi ezichemile, umqeqeshi uyayihlaba umkhosi, abhale kabusha incwadi yomthetho, futhi aqinisekise ukuthi akwenzeki futhi. Lokho kuwukwengamela esenzweni [2].
Uhlangothi Olungajabulisi Kakhulu
Ayicwebezeli yonke. Abaqeqeshi babhekana nalokhu:
-
I-Monotony - Ukulebula okungapheli kuyaguga.
-
Ukukhathala ngokomzwelo - Ukubuyekeza okuqukethwe okuyingozi noma okuphazamisayo kungase kubangele; amasistimu okusekela abalulekile [4].
-
Ukushoda kokuqashelwa - Abasebenzisi abaqapheli ukuthi abaqeqeshi bakhona.
-
Ushintsho oluqhubekayo - Amathuluzi aguquka ngokungami, okusho ukuthi abaqeqeshi kufanele bahambisane.
Noma kunjalo, kwabaningi, injabulo yokubumba “ubuchopho” bobuchwepheshe ibagcina bebambe ongezansi.
Ama-MVP Afihliwe we-AI
Ngakho-ke, obani abaqeqeshi be-AI? Ziyibhuloho eliphakathi kwama-algorithms angahluziwe namasistimu asebenzela abantu ngempela. Ngaphandle kwabo, i-AI izofana nomtapo wolwazi ongenabo abasebenzi basemtatsheni wezincwadi - amathani olwazi, kodwa cishe akunakwenzeka ukuwusebenzisa.
Ngokuzayo i-chatbot ikwenza uhleke noma uzizwe “ngokuvumelana” ngendlela emangalisayo, ngiyabonga umqeqeshi. Ziyizibalo ezithulile ezenza imishini azigcini nje ngokubala, kodwa zixhuma [1][2][5].
Izithenjwa
[1] Ouyang, L. et al. (2022). Ukuqeqesha amamodeli olimi ukulandela imiyalelo ngempendulo yomuntu (InstructGPT). I-NeurIPS. Isixhumanisi
[2] NIST (2023). I-Artificial Intelligence Risk Management Framework (AI RMF 1.0). Isixhumanisi
[3] Northcutt, C. et al. (2021). Amaphutha Elebula Egcwele Kumasethi Okuhlola Athunaza Amabhentshimakhi Wokufunda Womshini. I-NeurIPS Datasets & Benchmarks. Isixhumanisi
[4] WHO/ILO (2022). Imihlahlandlela ngempilo yengqondo emsebenzini. Isixhumanisi
[5] Bai, Y. et al. (2022). I-AI Yomthethosisekelo: Ukungabi Nabungozi okuvela ku-AI Impendulo. arXiv. Isixhumanisi