Ingakanani Amandla asetshenziswa yi-AI? [Ividiyo kanye nombuzo]

Impendulo: I-AI ingasebenzisa ugesi omncane kakhulu emsebenzini olula wombhalo, kodwa kakhulu uma izimpendulo zinde, imiphumela i-multimodal, noma izinhlelo zisebenza ngezinga elikhulu. Ukuqeqeshwa ngokuvamile kuyimbangela enkulu yamandla angaphambili, kuyilapho ukuphetha kwansuku zonke kuba yinto ebalulekile njengoba izicelo ziqoqana.

Izinto ezibalulekile okufanele uzicabangele:

Umongo: Chaza umsebenzi, imodeli, ihadiwe, kanye nesilinganiso ngaphambi kokucaphuna noma yikuphi ukulinganisa kwamandla.

Ukuqeqeshwa: Phatha ukuqeqeshwa kwemodeli njengomcimbi oyinhloko wamandla lapho uhlela isabelomali.

Isiphetho: Qapha ngokucophelela isiphetho esiphindaphindwayo, ngoba izindleko ezincane ngesicelo ngasinye zikhula ngokushesha.

Ingqalasizinda: Faka ukupholisa, isitoreji, amanethiwekhi, kanye nomthamo wokungasebenzi kunoma yikuphi ukulinganisa okungokoqobo.

Ukusebenza kahle: Sebenzisa amamodeli amancane, izixwayiso ezimfushane, ukugcinwa kwesikhashana, kanye nokuhlanganisa ukuze unciphise ukusetshenziswa kwamandla.

Izihloko ongase uthande ukuzifunda ngemva kwalesi:

🔗 Indlela i-AI ethinta ngayo imvelo
Kuchaza umthelela we-AI wekhabhoni, ukusetshenziswa kwamandla, kanye nokushintshana kokusimama.

🔗 Ingabe i-AI imbi emvelweni?
Iveza izindleko zemvelo ezifihliwe zamamodeli e-AI kanye nezikhungo zedatha.

🔗 Ingabe i-AI ilungile noma imbi? Izinzuzo kanye nokubi
Ukubheka ngokulinganisela izinzuzo ze-AI, izingozi, izimiso zokuziphatha, kanye nemiphumela yangempela.

🔗 Kuyini i-AI? Umhlahlandlela olula
Funda izisekelo ze-AI, amagama ayisihluthulelo, kanye nezibonelo zansuku zonke ngemizuzu.

Kungani lo mbuzo ubaluleke ngaphezu kwalokho abantu abakucabangayo 🔍

Ukusetshenziswa kwamandla e-AI akuyona nje indaba yokuxoxa ngemvelo. Kuthinta izinto ezimbalwa zangempela:

• Izindleko zikagesi - ikakhulukazi kumabhizinisi asebenzisa izicelo eziningi ze-AI

• Umthelela wekhabhoni - kuye ngomthombo wamandla ngemuva kwamaseva

• Ubunzima behadiwe - ama-chip anamandla adonsa amandla amakhulu

• Izinqumo zokwandisa - isicelo esisodwa esishibhile singaphenduka izigidi zezibizayo

• Umklamo womkhiqizo - ukusebenza kahle kuvame ukuba yisici esingcono kunalokho abantu abakuqaphelayo (i-Google Cloud, i-Green AI)

Abantu abaningi babuza ukuthi “Isebenzisa amandla angakanani i-AI?” ngoba bafuna inombolo ephawulekayo. Into enkulu. Into evumelana nesihloko. Kodwa umbuzo ongcono yilona: Yiluphi uhlobo lokusetshenziswa kwe-AI esikhuluma ngalo? Ngoba lokho kushintsha konke. (IEA)

Isiphakamiso esisodwa sokuqedela ngokuzenzakalela? Sincane impela.
Ukuqeqesha imodeli yomngcele phakathi kwamaqoqo amakhulu? Kukhulu kakhulu, kukhulu kakhulu.
Ukuhamba komsebenzi we-AI webhizinisi ohlala usebenza okuthinta izigidi zabasebenzisi? Yebo, lokho kushesha... njengemali ephenduka inkokhelo yerenti. (DOE, Google Cloud)

Ingakanani i-AI esebenzisa amandla? Impendulo emfushane ⚡

Nansi inguqulo ewusizo.

I-AI ingasebenzisa noma yikuphi kusukela engxenyeni encane ye-watt-hora ngomsebenzi olula kuya enanini elikhulu likagesi lokuqeqesha nokusabalalisa okukhulu. Lelo banga lizwakala libanzi ngokuhlekisa ngoba libanzi. (Google Cloud, Strubell et al.)

Kalula nje:

• Imisebenzi elula yokuphetha - ngokuvamile ilula kakhulu ngokwesisekelo sokusetshenziswa ngakunye

• Izingxoxo ezinde, imiphumela emikhulu, ukwenziwa kwezithombe, ukwenziwa kwamavidiyo - okubonakala kudinga amandla amaningi

• Ukuqeqesha amamodeli amakhulu - iqhawe elisindayo lokusetshenziswa kwamandla

• Ukusebenzisa i-AI esikalini usuku lonke - lapho "okuncane ngesicelo ngasinye" kuba "ibhili elikhulu eliphelele" (i-Google Cloud, i-DOE)

Umthetho omuhle yilona:

• Ukuqeqeshwa kungumcimbi omkhulu wamandla angaphambili 🏭

• Isiphetho yi-utility bill eqhubekayo 💡 (Strubell et al., Google Research)

Ngakho-ke uma umuntu ebuza ukuthi, Ingakanani i-AI esebenzisa amandla?,impendulo eqondile ithi, “Akuyona neyodwa - kodwa kwanele ukuthi ukusebenza kahle kubalulekile, futhi kwanele ukuthi leso sikali sishintshe yonke indaba.” (IEA, Green AI)

Ngiyazi ukuthi lokho akuthandeki njengoba abantu befuna. Kodwa kuyiqiniso.

Yini eyenza inguqulo enhle yesilinganiso samandla se-AI? 🧠

Isilinganiso esihle akuyona nje inombolo ephawulekayo ephonswa kusithombe. Isilinganiso esisebenzayo sihlanganisa umongo. Ngaphandle kwalokho kufana nokulinganisa inkungu ngesikali sendlu yokugezela. Kuseduze ngokwanele ukuze kuzwakale kuhlaba umxhwele, hhayi eduze ngokwanele ukuze uthembeke. (IEA, Google Cloud)

Isilinganiso samandla se-AI esihle kufanele sifake:

• Uhlobo lomsebenzi - umbhalo, isithombe, umsindo, ividiyo, ukuqeqeshwa, ukulungiswa kahle

• Usayizi wemodeli - amamodeli amakhulu ngokuvamile adinga ukubalwa okwengeziwe

• Ihadiwe esetshenzisiwe - akuwona wonke ama-chip asebenza kahle ngokulinganayo

• Ubude besikhathi - izikhuthazo ezimfushane kanye nokusebenza okude kwezinyathelo eziningi kuhluke kakhulu

• Ukusetshenziswa - izinhlelo ezingasebenzi zisasebenzisa ugesi

• Ukupholisa kanye nengqalasizinda - iseva akuyona yonke imali ekhokhwayo

• Indawo kanye nengxube yamandla - ugesi awuhlanzekile ngokulinganayo yonke indawo (Google Cloud, IEA)

Yingakho abantu ababili bengaphikisana ngokusetshenziswa kukagesi we-AI futhi bobabili bazwakale beqiniseka ngenkathi bekhuluma ngezinto ezihluke ngokuphelele. Umuntu oyedwa usho impendulo eyodwa ye-chatbot. Omunye usho ukugijima okukhulu kokuqeqeshwa. Bobabili bathi “AI,” futhi ngokuzumayo ingxoxo iyanyamalala 😅

Ithebula Lokuqhathanisa - izindlela ezinhle kakhulu zokulinganisa ukusetshenziswa kwamandla e-AI 📊

Nasi ithebula eliwusizo kunoma ubani ozama ukuphendula umbuzo ngaphandle kokuwuguqula ube ubuciko bokusebenza.

Ayikho indlela ephelele. Leyo yingxenye ekhungathekisayo kancane. Kodwa kunezinga lenani. Futhi ngokuvamile, into elungisekayo idlula ephelele. (I-Google Cloud)

Isici esikhulu akusiwo umlingo - yi-compute kanye ne-hardware 🖥️🔥

Uma abantu becabanga ngokusetshenziswa kwamandla e-AI, bavame ukucabanga ukuthi imodeli ngokwayo iyinto edla amandla. Kodwa imodeli iwuhlelo lwesofthiwe olusebenza kuhadiwe. Ihadiwe yilapho kuvela khona ibhili kagesi. (Strubell et al., Google Cloud)

Izinguquko ezinkulu kakhulu zivame ukufaka:

• Uhlobo lwe-GPU noma i-accelerator

• Zingaki iziphuphu ezisetshenziswayo

• Isikhathi abasihlala bematasa ngaso

• Umthwalo wememori

• Usayizi weqembu kanye nomphumela

• Kungakhathaliseki ukuthi uhlelo luhlelwe kahle noma lumane luphoqelela yonke into (i-Google Cloud, i-Quantization, i-Batching, kanye namasu okukhonza ekusetshenzisweni kwamandla e-LLM)

Uhlelo oluhlelwe kahle kakhulu lungenza umsebenzi omningi ngamandla amancane. Uhlelo olungasebenzi kahle lungachitha ugesi ngokuzethemba okumangalisayo. Uyazi ukuthi kunjani - ezinye izinto ziyizimoto zomjaho, ezinye ziyizinqola zokuthenga ezine-rocket duct eqoshiwe 🚀🛒

Futhi yebo, usayizi wemodeli ubalulekile. Amamodeli amakhulu avame ukudinga inkumbulo eyengeziwe kanye nokubala okwengeziwe, ikakhulukazi lapho ekhiqiza imiphumela emide noma ephatha ukucabanga okuyinkimbinkimbi. Kodwa amaqhinga okusebenza kahle angashintsha isithombe: (I-Green AI, Ukulinganisa, Ukuhlanganisa, kanye Namasu Okukhonza Ekusetshenzisweni Kwamandla e-LLM)

• ukulinganisa

• umzila ongcono

• amamodeli amancane ochwepheshe

• ukulondoloza isikhashana

• ukuqoqa

• Ukuhlela ihadiwe okuhlakaniphile (Amasu Okulinganisa, Ukuhlanganisa, Nokukhonza Ekusetshenzisweni Kwamandla e-LLM)

Ngakho umbuzo awugcini nje ngokuthi “Ingakanani imodeli?” Uphinde uthi “Iqhutshwa ngobuhlakani kangakanani?”

Ukuqeqeshwa vs ukuqagela - lezi yizilwane ezahlukene 🐘🐇

Lokhu ukuhlukana okudida cishe wonke umuntu.

Ukuqeqeshwa

Ukuqeqeshwa yilapho imodeli ifunda khona amaphethini kusuka kumasethi edatha amakhulu. Kungabandakanya ama-chip amaningi asebenza isikhathi eside, ehlafuna amanani amakhulu edatha. Lesi sigaba sidinga amandla amaningi. Ngezinye izikhathi sidinga kakhulu. (Strubell et al.)

Amandla okuqeqesha ancike ku:

• usayizi wemodeli

• usayizi wesethi yedatha

• inani lokuqeqeshwa

• izivivinyo ezihlulekile

• amaphasi okulungisa kahle

• ukusebenza kahle kwehadiwe

• ukupholisa ngaphezulu (Strubell et al., Google Research)

Futhi nansi ingxenye abantu abavame ukuyiphuthelwa - umphakathi uvame ukucabanga ngokugijima okukhulu kokuqeqeshwa, okwenziwe kanye, ekugcineni kwendaba. Empeleni, intuthuko ingabandakanya ukugijima okuphindaphindiwe, ukulungisa, ukuqeqesha kabusha, ukuhlola, kanye nazo zonke iziphindaphindo eziqhubekayo kodwa ezibizayo emcimbini oyinhloko. (Strubell et al., Green AI)

Isiphetho

I-Inference iyimodeli ephendula izicelo zangempela zomsebenzisi. Isicelo esisodwa singase singabonakali siningi. Kodwa i-inference iyenzeka ngokuphindaphindiwe. Izigidi zezikhathi. Ngezinye izikhathi izigidigidi. (Google Research, DOE)

Amandla okucabanga akhula nge:

• ubude besicelo

• ubude bokukhipha

• inani labasebenzisi

• izidingo zokulibaziseka

• izici eziningi

• okulindelwe ngesikhathi sokusebenza

• izinyathelo zokuphepha kanye nokucubungula ngemuva (I-Google Cloud, Ukulinganisa, Amasu Okuhlanganisa, Nokukhonza Ekusetshenzisweni Kwamandla e-LLM)

Ngakho ukuqeqeshwa ukuzamazama komhlaba. Isiphetho ukugeleza kwamanzi. Esinye siyamangalisa, esinye siyaqhubeka, futhi kokubili kungakha kabusha ugu. Mhlawumbe kuyisingathekiso esingavamile, kodwa sihlala ndawonye... cishe noma kancane.

Izindleko zamandla ezifihliwe abantu abazikhohlwayo 😬

Uma umuntu elinganisela ukusetshenziswa kwamandla e-AI ngokubheka i-chip kuphela, ngokuvamile uba ngaphansi. Hhayi njalo ngendlela edabukisayo, kodwa kwanele ukuba nendaba. (I-Google Cloud, IEA)

Nazi izingcezu ezifihliwe:

Ukupholisa ❄️

Amaseva akhiqiza ukushisa. Ihadiwe enamandla ye-AI ikhiqiza okuningi kwayo. Ukupholisa akuyona into ongayikhetha. I-watt ngayinye esetshenziswa ukubala ivame ukumema ukusetshenziswa kwamandla amaningi ukuze nje kugcinwe amazinga okushisa ephilile. (IEA, Google Cloud)

Ukuhamba kwedatha 🌐

Ukuhambisa idatha kusitoreji, inkumbulo, kanye namanethiwekhi nakho kudinga amandla. I-AI ayikona nje "ukucabanga." Iphinde iguqule ulwazi njalo. (IEA)

Umthamo wokungasebenzi 💤

Izinhlelo ezakhelwe ukudinga okukhulu azihlali zisebenza ngokudinga okukhulu. Ingqalasizinda engasebenzi noma engasetshenziswa kahle isasebenzisa ugesi. (I-Google Cloud)

Ukweqiwa kanye nokuthembeka 🧱

Ama-backups, izinhlelo zokwehluleka, izindawo eziphindaphindwayo, izendlalelo zokuphepha - konke kuyigugu, konke kuyingxenye yesithombe samandla amakhulu. (IEA)

Isitoreji 📦

Idatha yokuqeqesha, ukushumeka, izingodo, izindawo zokuhlola, imiphumela ekhiqiziwe - konke lokhu kuhlala ndawana thize. Isitoreji sishibhile kune-computing, yebo, kodwa hhayi mahhala ngokwemigomo yamandla. (IEA)

Yingakho i-AI isebenzisa amandla angakanani? ingaphendulwa kahle ngokubheka ishadi elilodwa lokulinganisa. I-stack ephelele ibalulekile. (I-Google Cloud, IEA)

Kungani isixwayiso esisodwa se-AI singaba sincane - kanti esilandelayo singaba yisilo 📝➡️🎬

Akuzona zonke izikhuthazo ezidalwe ngokulinganayo. Isicelo esifushane sokubhalwa kabusha kwemisho asifani nokucela ukuhlaziywa okude, iseshini yokubhala ikhodi enezinyathelo eziningi, noma ukukhiqizwa kwesithombe esinesinqumo esiphezulu. (I-Google Cloud)

Izinto ezivame ukwandisa ukusetshenziswa kwamandla ngokusebenzisana ngakunye:

• Amafasitela omongo amade

• Izimpendulo ezinde

• Izinyathelo zokusebenzisa amathuluzi kanye nokuthola kabusha

• Amaphasi amaningi okucabanga noma ukuqinisekisa

• Ukukhiqizwa kwesithombe, umsindo, noma ividiyo

• Isilinganiso semali esiphezulu

• Ama-Targets e-latency aphansi (i-Google Cloud, i-Quantization, i-Batching, kanye nama-Serving Strategies ekusetshenzisweni kwamandla e-LLM)

Impendulo yombhalo elula ingase ishibhile. Indlela yokusebenza enkulu yezindlela eziningi ingaba eshibhile, kahle, hhayi eshibhile. Kufana noku-oda ikhofi kunokulungisa umshado. Kokubili kubalwa “njengenkonzo yokudla,” ngokomthetho. Enye ayifani nenye ☕🎉

Lokhu kubalulekile ikakhulukazi kumaqembu omkhiqizo. Isici esibonakala singenangozi ekusetshenzisweni okuphansi singabiza kakhulu uma iseshini ngayinye yomsebenzisi iba yinde, icebile, futhi inamandla kakhulu ekubaleni. (DOE, Google Cloud)

I-AI yabathengi kanye ne-AI yebhizinisi akuyona into efanayo 🏢📱

Umuntu ojwayelekile osebenzisa i-AI ngokunganaki angase acabange ukuthi izimpendulo zakhe ngezikhathi ezithile ziyinkinga enkulu. Ngokuvamile, akusikho lapho indaba yamandla eyinhloko ihlala khona. (I-Google Cloud)

Ukusetshenziswa kwebhizinisi kushintsha izibalo:

• izinkulungwane zabasebenzi

• abashayeli bezindiza abahlale bevuliwe

• ukucutshungulwa kwedokhumenti okuzenzakalelayo

• ukufingqa kwekholi

• ukuhlaziywa kwesithombe

• amathuluzi okubuyekeza ikhodi

• ama-ejenti angemuva asebenza njalo

Yilapho ukusetshenziswa kwamandla ahlanganisiwe kuqala khona ukubaluleka kakhulu. Hhayi ngoba isenzo ngasinye singesokubhujiswa, kodwa ngoba ukuphindaphinda kuyinto ephindaphindayo. (DOE, IEA)

Ekuhlolweni kwami ​​​​kokuhlola kanye nokubuyekezwa komsebenzi, yilapho abantu bemangala khona. Bagxila egameni lemodeli, noma i-demo ekhangayo, futhi banganaki ivolumu. Ivolumu ivame ukuba yimbangela yangempela - noma umusa wokonga, kuye ngokuthi ukhokhisa amakhasimende noma ukhokha ithebhu yesevisi 😅

Kubathengi, umthelela ungazwakala ungabonakali. Kumabhizinisi, uba yinto ecacile ngokushesha okukhulu:

• izikweletu ezinkulu zengqalasizinda

• ingcindezi eyengeziwe yokwenza ngcono

• isidingo esinamandla samamodeli amancane lapho kungenzeka khona

• ukubika kokusimama kwangaphakathi

• ukunaka okwengeziwe ekugcinweni kwesikhashana kanye nokuqondisa (i-Google Cloud, i-Green AI)

Indlela yokunciphisa ukusetshenziswa kwamandla e-AI ngaphandle kokuyeka i-AI 🌱

Le ngxenye ibalulekile ngoba umgomo akukhona ukuthi “yeka ukusebenzisa i-AI.” Ngokuvamile lokho akulona iqiniso, futhi akudingekile. Ukusetshenziswa okungcono kuyindlela ehlakaniphile.

Nazi izithiyo ezinkulu kakhulu:

1. Sebenzisa imodeli encane kunazo zonke eqeda umsebenzi

Akuwona wonke umsebenzi odinga inketho yokusinda. Imodeli elula yokuhlela noma ukufingqa inganciphisa imfucuza ngokushesha. (I-Green AI, i-Google Cloud)

2. Fushanisa izikhuthazo nemiphumela

Ukuhlanganisa ngaphakathi, ukuhlanganisa ngaphandle. Amathokheni engeziwe asho ukubala okwengeziwe. Ngezinye izikhathi ukunciphisa i-prompt kuyindlela elula yokuwina. (Amasu Okulinganisa, Ukuhlanganisa, Nokukhonza ku-LLM Energy Use, i-Google Cloud)

3. Imiphumela ephindaphindwayo ye-Cache

Uma umbuzo ofanayo uqhubeka nokuvela, ungawukhiphi kabusha njalo. Lokhu kusobala ngendlela ehlaselayo, kodwa kuyaphuthelwa. (I-Google Cloud)

4. Imisebenzi yeqembu uma kungenzeka

Ukuqhuba imisebenzi ngamaqoqo kungathuthukisa ukusetshenziswa futhi kunciphise imfucuza. (Amasu Okulinganisa, Ukuhlanganisa, Nokukhonza Ekusetshenzisweni Kwamandla e-LLM)

5. Imisebenzi yomzila ngokuhlakanipha

Sebenzisa amamodeli amakhulu kuphela uma ukuzethemba kwehla noma ubunzima bomsebenzi bukhuphuka. (Green AI, Google Cloud)

6. Lungiselela ingqalasizinda

Ukuhlela okungcono, ihadiwe engcono, isu lokupholisa elingcono - izinto ezingezona ezendabuko, inzuzo enkulu. (I-Google Cloud, i-DOE)

7. Kala ngaphambi kokuthatha

Amaqembu amaningi acabanga ukuthi ayazi ukuthi amandla aya kuphi. Bese elinganisa, futhi nansi - ingxenye ebizayo isendaweni ethile. (I-Google Cloud)

Ukusebenza kahle akukuhle. Akuvamile ukuthola ihlombe. Kodwa kungenye yezindlela ezinhle kakhulu zokwenza i-AI ibe ngengabizi kakhulu futhi ivikeleke kakhulu ngezinga 👍

Izinganekwane ezivamile mayelana nokusetshenziswa kukagesi we-AI 🚫

Ake sisuse izinganekwane ezimbalwa ngoba lesi sihloko sixaka ngokushesha.

Inganekwane 1 - Yonke imibuzo ye-AI ichitha imali kakhulu

Akunjalo ngempela. Ezinye zincane. Isikali kanye nohlobo lomsebenzi kubaluleke kakhulu. (I-Google Cloud)

Inganekwane 2 - Ukuqeqeshwa yiyona nto kuphela ebalulekile

Cha. Ukuqagela kungabusa ngokuhamba kwesikhathi lapho ukusetshenziswa kukhulu. (Google Research, DOE)

Inganekwane 3 - Imodeli enkulu ihlala isho umphumela ongcono

Ngezinye izikhathi yebo, ngezinye izikhathi akunjalo neze. Imisebenzi eminingi isebenza kahle ngezinhlelo ezincane. (I-AI eluhlaza)

Inganekwane 4 - Ukusetshenziswa kwamandla kulingana nomthelela wekhabhoni ngokuzenzakalelayo

Akuyona ngempela. Ikhabhoni incike emthonjeni wamandla futhi. (IEA, Strubell et al.)

Inganekwane 5 - Ungathola inombolo eyodwa yendawo yonke yokusebenzisa amandla e-AI

Awukwazi, okungenani hhayi ngesimo esihlala sinenjongo. Noma ungakwazi, kodwa kuzolinganiswa kakhulu kangangokuthi akube kusabalulekile. (IEA)

Yingakho ukubuza ukuthi i-AI isebenzisa amandla angakanani? kuhlakaniphile - kodwa kuphela uma usukulungele impendulo ehlukaniswe ngezigaba esikhundleni sesiqubulo.

Ngakho... i-AI isebenzisa amandla angakanani ngempela? 🤔

Nasi isiphetho esisekelwe ebufakazini.

Ukusetshenziswa kwe-AI:

• kancane, ngemisebenzi elula

• okuningi kakhulu, ukuze kukhiqizwe izinto eziningi kakhulu

• inani elikhulu kakhulu, lokuqeqeshwa kwamamodeli amakhulu

• inani elikhulu kakhulu selilonke, lapho izigidi zezicelo zinqwabelana ngokuhamba kwesikhathi (i-Google Cloud, i-DOE)

Yileso isimo sayo.

Into ebalulekile ukungashintshi yonke le ndaba ibe inombolo eyodwa eyesabekayo noma ukuhlikihla amahlombe okukodwa okunganaki. Ukusetshenziswa kwamandla e-AI kungokoqobo. Kubalulekile. Kungathuthukiswa. Futhi indlela engcono kakhulu yokukhuluma ngayo iwukusebenzisa umongo, hhayi ama-theatrics. (IEA, Green AI)

Ingxoxo yomphakathi eningi iyashintshashintsha phakathi kwezinto ezingavamile - “i-AI imahhala ngokuphelele” ngakolunye uhlangothi, “i-AI iyimbubhiso kagesi” ngakolunye uhlangothi. Iqiniso livamile kakhulu, okwenza libe nolwazi oluningi. Kuyinkinga yezinhlelo. Ihadiwe, isofthiwe, ukusetshenziswa, isikali, ukupholisa, izinketho zokuklama. Okungajwayelekile? Kancane. Kubalulekile? Kakhulu. (IEA, Google Cloud)

Izinto ezibalulekile okufanele uzicabangele ⚡🧾

Uma ufike lapha ubuza ukuthi, Ingakanani i-AI esebenzisa amandla?, nansi into okufanele uyikhumbule:

• Ayikho inombolo efanela bonke ngobukhulu obufanayo

• Ukuqeqeshwa kuvame ukudla amandla amaningi kakhulu kusengaphambili

• Ukuqagela kuba yinto ebalulekile esikalini

• Usayizi wemodeli, ihadiwe, umthwalo womsebenzi, kanye nokupholisa konke kubalulekile

• Ukulungiswa okuncane kungenza umehluko omkhulu ngokumangazayo

• Umbuzo ohlakaniphile kakhulu awukona nje ukuthi “malini,” kodwa futhi “ngomsebenzi waluphi, kuhlelo luni, nangasiphi isikali?” (IEA, Google Cloud)

Ngakho yebo, i-AI isebenzisa amandla angempela. Kwanele ukufanelwe ukunakwa. Kwanele ukuthethelela ubunjiniyela obungcono. Kodwa hhayi ngendlela efana nekhathuni, yenombolo eyodwa.

Isibonelo sangempela: Ukulinganisa izindleko zamandla zomsizi wokusekela we-AI

Isimo

Cabanga ngenkampani encane ye-SaaS esebenzisa umsizi we-AI ukubhala izimpendulo zamathikithi okusekela amakhasimende. Lesi yisibonelo esiqanjiwe kodwa esingokoqobo, hhayi isifundo senkampani.

Ithimba liphatha amathikithi okusekela angaba ngu-500 isonto ngalinye. Iningi lawo lilula: ukusetha kabusha amaphasiwedi, imibuzo yokukhokha, izincazelo zezici, kanye nokuxazulula izinkinga eziyisisekelo. Inkampani ayifuni ukuthi umsizi athumele izimpendulo ngokuzenzakalelayo. Ibhala izimpendulo ukuze i-ejenti yokusekela yabantu izibukeze.

Umbuzo wamandla awusiwo, “Ingakanani i-AI esebenzisa ngokujwayelekile?” Kuwusizo kakhulu:

"Sidala i-computing eningi kangakanani ngokungeza i-AI kulo msebenzi, futhi singayinciphisa ngaphandle kokulimaza ikhwalithi?"

Lokho okudingwa umsizi

Ithimba lizoqala ngokuthi:

Isisekelo solwazi esihlanzekile sesikhungo sosizo

Uhlu lwemithetho yokubuyiselwa kwemali evunyiwe, ubumfihlo, kanye nokwenyusa

Izibonelo ezingu-20-30 zezimpendulo eziqinile zokusekela zesikhathi esidlule

Umyalelo ocacile wokuthi umsizi kumele abhale phansi, hhayi ukuthumela

Amalogi okusetshenziswa kwamafu noma amalogi okusetshenziswa kwemodeli ye-API

Ispredishithi esilula sokulandelela uhlobo lwethikithi, ubude besaziso, ubude bokukhipha, isikhathi sokubuyekeza, nokuthi impendulo yamukelwe yini

Ingxenye ebalulekile ukulinganisa. Ngaphandle kwezingodo, ithimba liqagela nje.

Isibonelo semiyalelo

Ungumsizi wokubhala ukusekela umkhiqizo we-SaaS. Sebenzisa kuphela okuqukethwe kwesikhungo sosizo okuvunyiwe kanye namanothi enqubomgomo anikeziwe. Bhala impendulo ecacile nenenhlonipho engaphansi kwamagama ayi-180. Uma ikhasimende licela ukubuyiselwa imali, ukususwa kwe-akhawunti, iseluleko sezomthetho, imininingwane yokuphepha, noma yini engamboziwe kumadokhumenti, ungaphenduli ngokuqondile. Maka ukuze kubuyekezwe ngabantu bese uchaza ukuthi yiluphi ulwazi olungekho.

Ngaphambi kokubhala impendulo, hlela ithikithi ngokuthi: elula, ebucayi kunqubomgomo, ezobuchwepheshe, noma edingekayo yokwenyusa intengo.

Indlela yokuyihlola

Ithimba lingahlola umsizi kumathikithi adlule angu-50 ngaphambi kokuwasebenzisa bukhoma.

Isethi yokuhlola elula ingafaka:

Amathikithi ayi-10 ephasiwedi noma okungena ngemvume

Amathikithi okukhokha ayi-10

Amathikithi okuxazulula izinkinga obuchwepheshe ayi-10

Imiyalezo yamakhasimende eyi-10 engacacile noma engaphelele

Amathikithi ayi-10 athinta inqubomgomo ahilela ukubuyiselwa kwemali, ubumfihlo, noma ukuvalwa kwe-akhawunti

Ngethikithi ngalinye, ithimba kufanele lirekhode:

Ingabe uhlaka lwalulungile ngokweqiniso?

Ingabe isebenzise ulwazi olugunyaziwe kuphela?

Ingabe yahlala ngaphansi komkhawulo wamagama?

Ingabe ibeke uphawu olufanele ezimweni ezibucayi?

I-ejenti yomuntu yachitha isikhathi esingakanani iyihlela?

Zingaki amathokheni noma izicelo ezisetshenziswe umsebenzi wokusebenza?

Lokhu kunikeza iqembu okuthile okuphathekayo okungaqhathaniswa nakho esikhundleni sokuthembela ekucabangeni.

Umphumela

Umphumela obonisayo: Ngokusekelwe ekubekeni isikhathi samathikithi esampula angu-50 ngaphambi nangemva kokusebenzisa uhlelo lokusebenza, ithimba lilinganisela ukuthi isilinganiso sesikhathi sokuqala sisuka kumaminithi angu-6 ngethikithi kuya kumaminithi angu-2 ngethikithi.

Ngamathikithi angu-500 ngesonto, lokho konga cishe imizuzu engu-2,000, noma cishe amahora angu-33 esikhathi sokubhala.

Kodwa ama-log abonisa nokuthile okubalulekile: ama-38% amathikithi ayiziphindaphindo ezilula. Ngokugcina izimpendulo ezivunyiwe zale mibuzo ephindaphindwayo esikhundleni sokuvuselela yonke i-draft kusukela ekuqaleni, ithimba linciphisa izicelo ze-AI kusuka ku-500 ngesonto kuya ku-310 ngesonto.

Lokho kusho ukwehla okungu-38% ezicelweni zokuphetha zamasonto onke zalo msebenzi, ngaphandle kokususa isici se-AI.

Ithimba lingakuqinisekisa lokhu ngokuqhathanisa:

Izicelo ze-AI zamasonto onke ngaphambi nangemva kokugcinwa kwesikhashana

Isilinganiso sokuphakamisa kanye nobude bokukhipha

Izinga lokwamukelwa kwabantu

Inani lokwenyuka okubanjwe ngendlela efanele

Sekela amaphuzu ekhwalithi noma ukubalwa kokubuyekezwa

Ukonga ugesi ngqo kuzoncika kumodeli, ihadiwe, umhlinzeki, kanye nengqalasizinda. Kodwa ukunciphisa umthwalo womsebenzi ngokwako kuyalinganiswa.

Yini engase ihambe kabi

Umsizi angase aphendule imibuzo yenqubomgomo ngokweqile uma imithetho yokwenyusa ingaqondakali.

Amadokhumenti amade esikhungo sosizo angase akhukhule ubude besikhathi uma ukusetha kokubuyisa kungahlelekile kahle.

Ama-ejenti angase athembe ama-draft abushelelezi ngokushesha kakhulu futhi aphuthelwe amaphutha amancane.

Ukugcina idatha kungaba yingozi uma izinqubomgomo zokubuyiselwa kwemali ezindala, amanani, noma zobumfihlo ziqhubeka nokusebenza.

Ithimba lingase lisebenzise amathokheni ambalwa ngenkathi ngephutha likhiqiza izimpendulo ezingasizi kakhulu.

Inguqulo ephephe kakhulu igcina abantu benolwazi, ilinganisa izimpendulo ezamukelekayo, futhi ibukeze izimpendulo ezigciniwe noma nini lapho izinqubomgomo zishintsha.

Ukudla okuwusizo

Isilinganiso samandla e-AI esizwakalayo siqala ngokuhamba komsebenzi okuqondile. Bala izicelo, nciphisa izimpendulo, ugcine izimpendulo eziphindaphindwayo, bese ulinganisa ikhwalithi yokubuyekeza. Lokho kuguqula ukuthi “I-AI isebenzisa amandla angakanani?” kusuka engxoxweni engacacile ibe umbuzo wobunjiniyela osebenzayo onezinombolo iqembu elingazithuthukisa ekusebenzeni.

Imibuzo Evame Ukubuzwa

Ingakanani amandla asetshenziswa yi-AI ngesicelo esisodwa?

Ayikho inombolo evamile yesikhuthazo esisodwa, ngoba ukusetshenziswa kwamandla kuncike kumodeli, ihadiwe, ubude besikhuthazo, ubude bomphumela, kanye nanoma yikuphi ukusetshenziswa kwamathuluzi engeziwe okuhilelekile. Impendulo yombhalo omfushane ingaba nesizotha, kanti umsebenzi omude we-multimodal ungadla okuningi kakhulu. Impendulo enenjongo kakhulu akuyona into eyodwa eyinhloko, kodwa umongo ozungeze umsebenzi.

Kungani izilinganiso zokusetshenziswa kwamandla e-AI zihluka kangaka?

Izilinganiso ziyahlukahluka ngoba abantu bavame ukuqhathanisa izinto ezihlukene kakhulu ngaphansi kwelebula elilodwa le-AI. Isilinganiso esisodwa singachaza impendulo ye-chatbot elula, kanti esinye singase sihlanganise ukukhiqizwa kwesithombe, ividiyo, noma ukuqeqeshwa kwemodeli enkulu. Ukuze isilinganiso sibe nenjongo, sidinga umongo ofana nohlobo lomsebenzi, usayizi wemodeli, ihadiwe, ukusetshenziswa, ukupholisa, kanye nendawo.

Ingabe ukuqeqesha i-AI noma ukusebenzisa i-AI nsuku zonke yikona okubiza kakhulu amandla?

Ukuqeqeshwa ngokuvamile kuyisenzakalo esikhulu samandla esingaphambili, ngoba kungabandakanya ama-chip amaningi asebenza isikhathi eside kumasethi edatha amakhulu. Ukuqagela yizindleko eziqhubekayo ezivela njalo lapho abasebenzisi bethumela izicelo, futhi ngezinga zingaba zinkulu kakhulu. Empeleni, zombili zibalulekile, yize zibalulekile ngezindlela ezahlukene.

Yini eyenza isicelo se-AI esisodwa sisebenzise amandla amaningi kunesinye?

Amafasitela omongo amade, imiphumela emide, amaphasi okucabanga aphindaphindwayo, izingcingo zamathuluzi, izinyathelo zokubuyisa, kanye nokukhiqizwa kwezindlela eziningi konke kuvame ukwandisa ukusetshenziswa kwamandla ngokusebenzisana ngakunye. Izinhloso zokubambezeleka nazo zibalulekile, ngoba izidingo zokuphendula ngokushesha zinganciphisa ukusebenza kahle. Isicelo esincane sokubhala kabusha kanye nokusebenza kwekhodi ende noma kwesithombe akufani nhlobo.

Yiziphi izindleko zamandla ezifihliwe abantu abangaziboni lapho bebuza ukuthi i-AI isebenzisa amandla angakanani?

Abantu abaningi bagxila kuphela ku-chip, kodwa lokho akunaki ukupholisa, ukunyakaza kwedatha, isitoreji, amandla okungasebenzi, kanye nezinhlelo zokuthembeka njengezipele noma izindawo zokwehluleka. Lezi zingqimba ezisekelayo zingashintsha kakhulu indawo yonke. Yingakho i-benchmark ngokwayo ingavamile ukuthwebula isithombe samandla aphelele.

Ingabe imodeli ye-AI enkulu ihlala isebenzisa amandla amaningi?

Amamodeli amakhulu ngokuvamile adinga ukubala okwengeziwe kanye nenkumbulo, ikakhulukazi emiphumeleni emide noma eyinkimbinkimbi, ngakho-ke avame ukusebenzisa amandla amaningi. Kodwa amakhulu akusho ukuthi kungcono ngokuzenzakalelayo kuwo wonke umsebenzi, futhi ukwenza ngcono kungashintsha isithombe kakhulu. Amamodeli amancane angochwepheshe, ukulinganisa, ukuhlanganisa, ukulondoloza isikhashana, kanye nokuqondisa okuhlakaniphile konke kungathuthukisa ukusebenza kahle.

Ingabe ukusetshenziswa kwe-AI yabathengi kuyinkinga enkulu yamandla, noma ingabe i-AI yebhizinisi iyinkinga enkulu?

Ukusetshenziswa kwabathengi okungajwayelekile kungangeza, kodwa indaba enkulu yamandla ivame ukuvela ekusetshenzisweni kwamabhizinisi. Ama-copilot ahlala esebenza, ukucutshungulwa kwamadokhumenti, ukufingqwa kwezingcingo, ukubuyekezwa kwekhodi, kanye nama-ejenti angemuva kudala isidingo esiphindaphindwayo kuzo zonke izinsizakalo zabasebenzisi abakhulu. Inkinga ivame ukuba ncane ngesenzo esisodwa esimangalisayo kodwa imayelana nevolumu eqhubekayo ngokuhamba kwesikhathi.

Isebenzisa amandla angakanani i-AI uma ufaka izikhungo zedatha kanye nokupholisa?

Uma uhlelo olubanzi selufakiwe, impendulo iba ngokoqobo futhi ivame ukuba nkulu kunalokho okushiwo yizilinganiso ze-chip kuphela. Izikhungo zedatha azidingi amandla kuphela ekubaleni, kodwa futhi nasekupholiseni, ekuxhumaneni, ekugcineni, nasekugcineni amandla asele. Yingakho ukwakheka kwengqalasizinda kanye nokusebenza kahle kwesikhungo kubaluleke kakhulu njengokwakhiwa kwemodeli.

Iyiphi indlela ewusizo kakhulu yokukala ukusetshenziswa kwamandla e-AI emsebenzini wangempela?

Indlela engcono kakhulu incike ekutheni ubani olinganisayo nokuthi ngayiphi injongo. Umthetho ojwayelekile ungasiza ngokuqhathanisa okusheshayo, kuyilapho amamitha e-watt, i-GPU telemetry, amalogi okukhokhisa amafu, kanye nokubika kwesikhungo sedatha kunikeza ukuqonda okusebenzayo okuqinile ngokuqhubekayo. Ngomsebenzi wokuzinza ongathi sína, umbono womjikelezo wokuphila ogcwele unamandla nakakhulu, yize uhamba kancane futhi unzima kakhulu.

Amaqembu angakunciphisa kanjani ukusetshenziswa kwamandla e-AI ngaphandle kokuyeka izici ze-AI eziwusizo?

Izinzuzo ezinkulu zivame ukuvela ekusebenziseni imodeli encane kakhulu esasebenza, ukunciphisa izixwayiso nemiphumela, ukugcina imiphumela ephindaphindwayo, ukuhlanganisa umsebenzi, nokuhambisa imisebenzi enzima kuphela kumamodeli amakhulu. Ukulungiswa kwengqalasizinda nakho kubalulekile, ikakhulukazi ukuhlela kanye nokusebenza kahle kwehadiwe. Ezindleleni eziningi, ukulinganisa kuqala kusiza ukuvimbela amaqembu ukuthi athuthukise into engalungile.

Izinkomba

1. I-International Energy Agency (IEA) - Isidingo samandla esivela ku-AI - iea.org

2. UMnyango Wezamandla wase-US (DOE) - I-DOE ikhipha umbiko omusha ohlola ukwanda kwesidingo sikagesi ezikhungweni zedatha - energy.gov

3. I-Google Cloud - Ukulinganisa umthelela wemvelo we-AI inference - cloud.google.com

4. Ucwaningo lwe-Google - Izindaba ezinhle mayelana nomphumela wekhabhoni wokuqeqeshwa kokufunda komshini - research.google

5. Ucwaningo lwe-Google - Umthelela wekhabhoni wokuqeqeshwa kokufunda komshini uzokwenyuka bese wehla - research.google

6. arXiv - AI Eluhlaza - arxiv.org

7. arXiv - Strubell et al. - arxiv.org

8. arXiv - Amasu Okulinganisa, Ukuhlanganisa, Nokukhonza Ekusetshenzisweni Kwamandla e-LLM - arxiv.org

Thola i-AI Yakamuva Esitolo Esisemthethweni Somsizi we-AI

Mayelana NATHI