Amamodeli e-AI aphakeme: Amasistimu amakhulu ekhompiyutha afinyelela iphuzu lokukhipha

• About the Author:
• Igama lomlobi

  I-Quantumrun Foresight
• June 2, 2023

Kusukela ngo-2012, intuthuko ebalulekile kubuhlakani bokwenziwa (AI) yenzeka njalo, ikakhulukazi iqhutshwa amandla ekhompiyutha akhulayo ("compute" ngamafuphi). Enye yamamodeli amakhulu, eyethulwe ngo-2020, yasebenzisa ikhompuyutha ephindwe izikhathi ezingu-600,000 kunemodeli yokuqala kusukela ngo-2012. Abacwaningi bakwa-OpenAI baphawule lo mkhuba ngo-2018 futhi baxwayisa ngokuthi lesi silinganiso sokukhula ngeke siqhubeke isikhathi eside.

Okuqukethwe kwemodeli ye-AI ephezulu

Onjiniyela abaningi bokufunda komshini (ML) basebenzisa amamodeli we-transformer yokufunda okujulile (DL) ngenxa yamandla abo abonakala angenamkhawulo. Izibonelo zalawa mamodeli zifaka i-Generative Pre-trained Transformer 2 (GPT-2), GPT-3, Bidirectional Encoder Representations from Transformers (BERT), kanye neTuring Natural Language Generation (NLG). Lawa ma-algorithms ngokuvamile anezinhlelo zokusebenza zomhlaba wangempela ezifana nokuhumusha komshini noma ukuqagela kochungechunge lwesikhathi. 

Izindlela zobuhlakani bokwenziwa kufanele zande ukuze zithwale idatha yokuqeqeshwa eyengeziwe futhi zibe ngcono ekuqaguleni. Le mfuneko iholele ekukhuphukeni kwamamodeli aphakeme anezigidigidi zamapharamitha (okuguquguqukayo okusetshenziswa ama-algorithms ukwenza izibikezelo). Lawa mamodeli amelwe yi-OpenAI's GPT-3 (kanye nokusebenzisana kwayo kwe-ChatGPT kwethulwa ngoDisemba 2022), i-PanGu-alpha yase-China, i-Nvidia's Megatron-Turing NLG, kanye ne-DeepMind's Gopher. Ngo-2020, ukuqeqeshwa kwe-GPT-3 kwakudinga i-supercomputer eyayiphakathi kwamahlanu amakhulu kakhulu emhlabeni. 

Kodwa-ke, lawa mamodeli avame ukudinga amanani amakhulu edatha yokuqeqeshwa edinga amandla. Ukufunda okujulile kuncike ekhonweni layo lokusebenzisa amandla amakhulu ekhompyutha, kodwa lokhu kuzoshintsha maduze. Ukuqeqesha kuyabiza, kunemikhawulo kuma-chips e-AI, futhi ukuqeqesha amamodeli amakhulu avala ama-processor, okwenza kube nzima ukuwaphatha wonke. Uma ipharamitha inkulu, kubiza kakhulu ukuqeqesha lawa mamodeli. Ochwepheshe bayavuma ukuthi kuzofika isikhathi lapho amamodeli e-AI aphakeme angase abize kakhulu futhi asebenzise amandla okuqeqesha. 

Umthelela ophazamisayo

Ngo-2020, i-OpenAI yalinganisela inani elincane lekhompuyutha elidingekayo ukuze kuqeqeshwe amamodeli amaningi, ifaka inani lamapharamitha nosayizi wedathasethi. Lezi zibalo zilandisa ukuthi i-ML idinga kanjani ukuthi idatha idlule kunethiwekhi izikhathi eziningi, ukuthi ukubala kwephasi ngayinye kukhuphuka kanjani njengoba inani lamapharamitha likhuphuka, nokuthi ingakanani idatha edingekayo njengoba inani lamapharamitha likhula.

Ngokwezilinganiso ze-Open AI, kucatshangwa ukuthi abathuthukisi bangakwazi ukufeza ukusebenza kahle okuphezulu, ukwakha i-GPT-4 (izikhathi eziyi-100 ezinkulu kune-GPT-3 (amapharamitha ayizigidi eziyizinkulungwane eziyi-17.5)) kungadinga amayunithi wokucubungula ama-graphics angu-7,600 (GPUs) asebenza okungenani unyaka owodwa futhi abize cishe U-$200 million. Imodeli yepharamitha engu-100-trillion ingadinga ama-GPU angu-83,000 ukuze isebenze unyaka wonke, ibize ngaphezu kwe-USD $2 billion.

Noma kunjalo, amafemu ezobuchwepheshe abesebenzisana futhi athulula ukutshalwa kwezimali kumamodeli azo e-AI ahlala enwebeka njengoba kukhula isidingo sezixazululo ze-ML. Isibonelo, i-Baidu yase-China kanye ne-Peng Cheng Lab ikhiphe i-PCL-BAIDU Wenxin, enamapharamitha ayizigidi eziyizinkulungwane ezingama-280. I-PCL-BAIDU isivele isetshenziswa izifunzo zezindaba ze-Baidu, injini yokusesha, nomsizi wedijithali. 

Inguqulo yakamuva yohlelo lwe-Go-playing, eyakhiwe yi-DeepMind ngo-December 2021, inamapharamitha ayizigidi eziyizinkulungwane ezingama-280. Amamodeli we-Google Switch-Transformer-GLaM anepharamitha emangazayo engu-1 trillion kanye ne-1.2 trillion, ngokulandelana. I-Wu Dao 2.0 evela e-Beijing Academy of AI inkulu kakhulu futhi kubikwe ukuthi inepharamitha eyizigidi eziyizinkulungwane eziyi-1.75. Njengoba amadolobha ahlakaniphile kanye ne-automation kuqhubeka nokuphazamisa ukuphazamiseka, ochwepheshe abaqiniseki ukuthi i-AI compute izolisekela kanjani ikusasa elinjalo. 

Imithelela yamamodeli e-AI aphezulu kakhulu

Imithelela ebanzi yamamodeli e-AI aphakeme ingase ihlanganise: 

• Ukwenyuka kokutshalwa kwezimali kanye namathuba ekuthuthukiseni ama-computer e-AI adla amandla amancane. 
• Inqubekelaphambili ye-AI yehliswe ijubane ngenxa yokuntuleka kwamandla ekhompiyutha, okuholela ekutholeni imali eyengeziwe yobuchwepheshe nezixazululo zokonga amandla.
• Onjiniyela be-ML benza amanye amamodeli ngaphandle kwama-transformer, angaholela ekutholweni nasekusungulweni okusha kwama-algorithms asebenza ngempumelelo.
• Izixazululo ze-AI ezigxile ezinkingeni ze-application-centric, ukulungisa ikhompuyutha ngokufanele noma ukulungisa njengoba kudingeka esikhundleni sokuvelela phezulu.
• Amasethi edatha ayinkimbinkimbi avumela izinhlelo ze-AI ukuthi zenze izibikezelo ezingcono, okuhlanganisa izibikezelo zesimo sezulu, ukutholwa komkhathi, ukuxilonga kwezokwelapha, nokuhweba kwamanye amazwe.

Imibuzo ongaphawula ngayo

• Uma usebenza emkhakheni we-AI, iyiphi inqubekelaphambili ekuthuthukiseni amamodeli e-ML angcono?
• Yiziphi ezinye izinzuzo ezingaba khona zamamodeli anedatha yokuqeqeshwa ebanzi ongafunda kuyo?