Samfuran AI masu girman kai: Manyan tsarin sarrafa kwamfuta suna kaiwa ga matakin da ya dace

• About the Author:
• Sunan marubuci

  Quantumrun Haskaka
• Yuni 2, 2023

Tun daga 2012, gagarumin ci gaba a cikin basirar wucin gadi (AI) na faruwa akai-akai, galibi ana yin su ta hanyar haɓaka ƙarfin kwamfuta ("ƙididdigar" a takaice). Ɗaya daga cikin manyan samfura, wanda aka ƙaddamar a cikin 2020, ya yi amfani da 600,000 sau fiye da ƙididdiga fiye da samfurin farko daga 2012. Masu bincike a OpenAI sun lura da wannan yanayin a cikin 2018 kuma sun yi gargadin cewa wannan haɓakar girma ba zai daɗe ba.

Babban mahallin ƙirar AI

Yawancin masu haɓaka na'ura (ML) suna amfani da ƙirar wuta don zurfin koyo (DL) saboda yuwuwarsu da alama mara iyaka. Misalai na waɗannan samfuran sun haɗa da Mai Canja-canje na Farko na Generative Pre-trained 2 (GPT-2), GPT-3, Wakilan Encoder Bidirectional daga Masu Canjawa (BERT), da Turing Natural Language Generation (NLG). Waɗannan algorithms galibi suna da aikace-aikace na zahiri kamar fassarar injin ko tsinkayar jerin lokaci. 

Hanyoyin basirar wucin gadi dole ne su faɗaɗa don ɗaukar ƙarin bayanan horo kuma su zama mafi kyawun tsinkaya. Wannan buƙatun ya haifar da haɓakar ƙira mafi girma tare da biliyoyin ma'auni (saɓanin da algorithms ke amfani da shi don yin tsinkaya). Waɗannan samfuran ana wakilta ta GPT-3 na OpenAI (da hulɗar ChatGPT da aka ƙaddamar a cikin Disamba 2022), PanGu-alpha na China, Megatron-Turing NLG na Nvidia, da DeepMind's Gopher. A cikin 2020, horar da GPT-3 yana buƙatar babban na'urar kwamfuta wanda ke cikin manyan biyar mafi girma a duniya. 

Koyaya, waɗannan samfuran suna buƙatar ɗimbin bayanan horo mai ƙarfi. Zurfafa ilmantarwa ya dogara da ikonsa na amfani da babban ƙarfin lissafi, amma wannan zai canza ba da daɗewa ba. Horowa yana da tsada, akwai iyaka ga kwakwalwan AI, kuma horar da manyan samfura suna toshe na'urori masu sarrafawa, yana da wahala a sarrafa su duka. Girman siga, ƙimar horon waɗannan samfuran shine tsada. Masana sun yarda cewa za a zo lokacin da manyan samfuran AI na iya zama masu tsada da kuzari don horarwa. 

Tasiri mai rudani

A cikin 2020, OpenAI ya kiyasta mafi ƙarancin adadin lissafin da ake buƙata don horar da ƙira da yawa, ƙididdige adadin sigogi da girman saitin bayanai. Waɗannan ma'auni suna lissafin yadda ML ke buƙatar bayanai don wucewa ta hanyar sadarwar sau da yawa, yadda ƙididdigewa ga kowane fasinja ke tashi yayin da adadin sigogi ya ƙaru, da nawa ake buƙatar bayanai yayin da adadin sigogi ke girma.

Dangane da ƙididdigar Buɗe AI, ɗauka cewa masu haɓakawa za su iya cimma matsakaicin inganci, gina GPT-4 (sau 100 girma fiye da GPT-3 (ma'auni tiriliyan 17.5)) zai buƙaci rukunin sarrafa hoto na 7,600 (GPUs) yana gudana aƙalla shekara ɗaya kuma farashi kusan. Dalar Amurka miliyan 200. Samfurin ma'auni na tiriliyan 100 zai buƙaci GPUs 83,000 don sarrafa shi har tsawon shekara guda, yana kashe sama da dalar Amurka biliyan 2.

Duk da haka, kamfanonin fasaha sun kasance suna yin haɗin gwiwa tare da zuba jari a cikin ƙirar AI na yau da kullum suna haɓakawa yayin da buƙatar mafita na ML ke girma. Misali, Baidu na kasar Sin da Peng Cheng Lab sun fitar da PCL-BAIDU Wenxin, tare da mitoci biliyan 280. PCL-BAIDU an riga an yi amfani da shi ta hanyar ciyarwar labarai na Baidu, injin bincike, da mataimakan dijital. 

Sabon sigar shirin Go-playing, wanda DeepMind ya ƙirƙira a cikin Disamba 2021, yana da sigogi biliyan 280. Samfuran Google Switch-Transformer-GLaM suna da ma'auni na tiriliyan 1 da tiriliyan 1.2, bi da bi. Wu Dao 2.0 daga Kwalejin AI na Beijing ya fi girma kuma an ba da rahoton cewa yana da ma'auni na tiriliyan 1.75. Kamar yadda birane masu wayo da aiki da kai ke ci gaba da tursasawa, masana ba su da tabbacin yadda lissafin AI zai goyi bayan irin wannan gaba. 

Tasirin samfuran manyan AI

Faɗin fa'idodin samfuran AI masu girma na iya haɗawa da: 

• Ƙara yawan saka hannun jari da dama wajen haɓaka kwakwalwan kwamfuta na AI waɗanda ke cinye ƙarancin kuzari. 
• Ci gaban AI ya ragu ta hanyar rashin ikon kwamfuta, wanda ke haifar da ƙarin kudade don fasahar adana makamashi da mafita.
• Masu haɓaka ML suna ƙirƙirar madadin ƙira ban da masu canzawa, wanda zai iya haifar da bincike da ƙirƙira don ingantaccen algorithms.
• Maganganun AI suna mai da hankali kan matsalolin da suka shafi aikace-aikace, daidaita ƙididdigewa daidai ko gyaggyarawa kamar yadda ake buƙata maimakon girman girman kawai.
• Ƙididdigar bayanai masu rikitarwa da ke ba da damar shirye-shiryen AI don yin ingantattun tsinkaya, gami da hasashen yanayi, gano sararin samaniya, binciken likita, da ciniki na ƙasa da ƙasa.

Tambayoyi don yin tsokaci akai

• Idan kuna aiki a sashin AI, menene wasu ci gaba a haɓaka ingantattun samfuran ML?
• Menene sauran yuwuwar fa'idodin samfura tare da fa'idodin horon da za a koya daga gare su?