Amaqembu e-Frontier awasebenzisi nje i-AI ukwenza amakhodi ngokushesha. Bahlela kabusha ukuthi isofthiwe yakheka kanjani. Umphumela uba izinzuzo zokukhiqiza ezingu-4.5x, kwezinye izimo ezingaphezu kuka-10x.
Onjiniyela abayisithupha. Izinsuku ezingamashumi ayisikhombisa nesithupha. Iphrojekthi ehlelelwe onjiniyela abangama-30 ezinyangeni eziyi-12 kuye kweziyi-18, ilethwa phakathi nekota. Lokho akukona ukuqagela. Yilokho okwenzeka lapho ithimba le-Amazon Bedrock liyeka ukuphatha i-AI njengesinqamuleli sekhodi futhi liqala ukuyiphatha njengesisekelo sendlela abasebenza ngayo. Ithimba lithumele ikhodi yokukhiqiza eyengeziwe ezinyangeni ezinhlanu kunaseminyakeni eyishumi edlule.
Igebe phakathi kwamaqembu afana naleli nawo wonke umuntu likhula ngokushesha. Ama-ejenti wokubhala amakhodi we-AI ashintshe ngokuyisisekelo izinga isoftware ebhalwa ngalo, kodwa hhayi izinga efinyelela ngalo amakhasimende. Ukuzibophezela kuyakhula, futhi amapayipi e-CI/CD amatasa kakhulu kunangaphambili. Nokho, izici ezithunyelwa emkhiqizweni azigcinanga ijubane elifanayo. Ibhodlela akulona ikhono le-ejenti lokukhiqiza okukhiphayo. Ukufinyelela kwe-ejenti olwazini eludingayo ukuze enze izinqumo ezinhle, kanye nokuzimisela kwethimba ukuhlela kabusha umsebenzi mayelana nalelo qiniso.
Sibiza amaqembu athole lokhu ngokuthi “amaqembu asemngceleni.” Abagcini nje kuma-elite labs. Zikhona kuzo zonke izimboni nosayizi bezinkampani, futhi zabelana ngesiyalo esifanayo: zithatha ukutholwa kwe-AI njengotshalomali lobunjiniyela, hhayi ukukhishwa kwamathuluzi. Noma yiliphi ithimba lonjiniyela lingaba yiqembu lasemngceleni; singakubonisa ukuthi ungafika kanjani lapho.
Izindlela ezintathu zokuthuthuka kwe-AI e-Amazon
Ukuthuthukiswa kwesoftware yomdabu ye-AI kuphatha i-AI njengesisekelo sendlela isofthiwe yakhiwe ngayo, ngama-ejenti akhulayo anekhono aqondiswa ochwepheshe abangabantu. Indlela amaqembu aqondisa ngayo lawo ma-ejenti inquma imiphumela. E-Amazon, abashayeli abayinhloko be-AI ekuthuthukisweni bekuwukunciphisa isikhathi abathuthukisi abasichitha emisebenzini engafaki amakhodi efana nemibhalo, ukuxhumanisa, nokusebenza, ukuhoxiswa kwezikweletu zobuchwepheshe, kanye nokunciphisa ukungqubuzana kwekhodi ezinkulungwaneni zamathimba amancane “e-pizza amabili” abathuthukisi. Besizama kuwo wonke amaqembu onjiniyela futhi sihlonze okungenani izindlela ezintathu: isinyathelo sokuthola indlela esinochwepheshe ababhekana nenselele, umsubathi ohlelekile ozowenza ngohlelo oluchazwe kahle, kanye nokuhlolwa kwe-in-situ okuhlukanisa amaqembu ngesigamu phakathi kwezindlela ezikhona kanye nokugeleza komsebenzi okujwayelwe yi-AI. Izindlela ziyehluka ngesakhiwo kodwa zihlangana ngokuqonda okufanayo.
I i-pathfinder initiative kwakuyisivivinyo esilawulwayo. Onjiniyela abayisithupha abaphezulu bathole igunya elilodwa: ukwakha kabusha injini ye-Amazon Bedrock, iphrojekthi ekuqaleni eyayilinganiselwa kubathuthukisi abangama-30 abasebenza izinyanga eziyi-12 kuye kweziyi-18. Esikhundleni sokwengeza isibalo esikhulu, iqembu lichithe amasonto alo okuqala liklama kabusha ukugeleza komsebenzi ezizungeze i-AI, lisuka emisebenzini eqondile laya emiphumeleni eshayelwa umgomo, lisebenzisa ama-ejenti amaningi ngokufana, futhi limisa izinhlelo ze-AI ukuze isebenze ngokuzimela ngesikhathi sokusebenza. Iphrojekthi yethulwe ngezinsuku ezingama-76. Ukukhiqiza konjiniyela ngamunye kukhuphuke cishe ngo-20x njengoba kulinganiswa ngesivinini sokuzibophezela esijwayelekile (inani lemisebenzi eyenziwa ngonjiniyela ngamunye ngeviki, elungiselwe inkimbinkimbi yenqolobane kanye nosayizi weqembu). Izibophezelo zisuke koku-2 ngeviki zaya kokungu-40. Ithimba lithumele ikhodi yekhwalithi ephezulu kakhulu ezinyangeni ezinhlanu kunaleyo lenzile kumaphrojekthi eminyakeni eyishumi edlule, njengoba kulinganiswa ngemigqa ethunyelwe ekukhiqizweni.
I i-sprint ehlelekile wathatha indlela ehlukile. Ithimba le-Prime Video Financial Systems lenze ukuhlola kwezinsuku eziyi-10 eligqugquzelwe imodeli yesitholi sendlela. Onjiniyela abayisithupha, igumbi elilodwa, ukushintshwa kwengqikithi enguziro, awekho ama-on-call, awekho amanye amaphrojekthi, imihlangano elinganiselwe. Unjiniyela omkhulu uchithe amasonto amathathu ngaphambili ephula ubunkimbinkimbi emisebenzini ehlelwe kahle enezidingo ezinemininingwane. Ithimba lisebenzise ukuthuthukiswa okushayelwa ngokukhethekile emsebenzini wesici esiyinkimbinkimbi kanye nokuthuthukiswa okusizwa yi-ejenti okuqondile kwemisebenzi lapho izimfuneko zase zicacile kakade. Ezinsukwini eziyi-10, bakhiqize izibophezelo ezingama-556 ngokumelene nesisekelo sama-96 futhi banciphisa isilinganiso sephrojekthi samasonto angama-90 saba amasonto angama-24. Lokho kuhumusha cishe ekuphumeni okungu-6x nokusheshisa okungu-4x. Bathi inzuzo enikwe amandla i-AI ngenxa yezici ezintathu eziphindaphindeka ndawonye: ukusheshisa komsebenzi onesinqumo esiphansi (1.5x), ukugxila okuphezulu emsebenzini onokwahlulela okuphezulu ngaphandle kokushintsha umongo (1.5x), kanye nokufinyelela okusheshayo kulwazi lwesizinda esithathwe umenzeli (1.5x). Susa noma iyiphi into eyodwa bese izinzuzo ziyawa. Ithimba manje selibheke ukuthuthukisa lezi zici ezintathu ekusebenzeni okuvamile lisebenzisa imininingwane enemininingwane yomkhiqizo ehlanganisa ulwazi lwesizinda nama-ejenti azimele akhulula isikhathi sokugxila.
Kwe ukuhlolwa kwe-in-situemaqenjini angama-50-plus afundiwe, amaqembu angama-25 asebenzise kokubili amathuluzi amasha nemikhuba emisha asebenze kahle kakhulu kunalawo avele enezela i-AI ekuhambeni komsebenzi okukhona. Izitolo zase-Amazon zasebenzisa abashayeli bezindiza abahlelekile abanamaqembu entuthuko ajwayelekile asebenza ngokumelene nokusilela emuva kwabo njalo, esebenzisa i-Kiro namathuluzi e-AI akhelwe ngenhloso angenazo izimo ezikhethekile futhi abekho onjiniyela abakhethwe ngesandla. Inzuzo yokukhiqiza emaphakathi yaba ngu-4.5x, amanye amaqembu afinyelela ukuthuthuka okungaphezu kwe-10x ngesivinini esijwayelekile sokuphakelwa (izici ezifakwe ngokugijima ngakunye, zenziwe zijwayelekile uma kuqhathaniswa nezisekelo zomlando). I-Perfect Order Experience manje ithumela izici ntambama esikhundleni samaviki amabili. I-WW Grocery isike ukudalwa kwedokhumenti yomklamo kusukela ezinsukwini ezinhlanu kuya emahoreni ambalwa.
Izindlela ezahlukene, isifundo esifanayo. Ukuhamba komsebenzi kubalulekile, hhayi ithuluzi kuphela.
Izinyathelo ezinhlanu zokuba yiqembu lasemngceleni
Kuzo zontathu izindlela, amaqembu aqhuba kahle kakhulu abelana ngezinqubo ezinhlanu ngomqondo ofanayo. Yehlisa izithiyo kumongo we-ejenti futhi wandise indawo engaphezulu yomsebenzi engayenza ngokuzimela.
Lapha kulapho amaqembu asemngceleni ehlukana khona nemikhuba yangaphambili. Indlela yomlando elungiselelwe isivinini sokukhiqiza ikhodi ngayinye. Amaqembu e-Frontier alungiselela okuthile okuhlukile: izinga lapho isofthiwe elungile, elungele ukukhiqiza ifinyelela kumakhasimende. Lowo mehluko uqhuba zonke izenzo ezingezansi.
- Tshala kumongo womenzeli. Amaqembu athuthuke kakhulu atshala imali eningi ekwenzeni amaphrojekthi nolwazi lube lula ukuthi ama-ejenti alusebenzise ngamafayela aqondisa ama-ejenti kanye nesiqondiso mayelana nezivumelwano zeqembu, izindinganiso zokubhala amakhodi, ukuhlola, nokuzulazula kwe-codebase. Ithimba lengqalasizinda ye-Bedrock labeka yonke ikhodi kanye nemibhalo ku-monorepo futhi lagcina ukuphawula okusemgqeni okukhiqizwa yizithunywa ze-AI, likuthatha njengenkumbulo eqhubekayo. Amaqembu eqa lesi sinyathelo ayazibuza ukuthi kungani abenzeli bawo beqhubeka nokwenza amaphutha afanayo.
- Yehlisa ijubane ukuze usheshise. Lo mkhuba oshiwo ngenhla uthatha isikhathi futhi udinga ukuthi amaqembu abekezele. Wonke amaqembu abenze kahle abike ukuthi izinto ziqale zehla njengoba befunda amamodeli. Bafake amakhodi ubuchwepheshe obuhlukahlukene kumadokhumenti aqondisayo asebenziseka kabusha ama-ejenti, amakhosombe ahlelwe kabusha ukuze ama-LLM akwazi ukucabanga ngawo, futhi bengeza amazwana nokuhlukaniswa kwekhodi okuklanywe kabusha ukuze kusetshenziswe i-AI. Amaqembu adlule kulelo jika lokufunda futhi achaza imiphumela elindelwe kuqala ahlangabezane nokusheshisa okuhlanganisiwe. Amaqembu abelindele ukuzuza ngokushesha ngaphandle kokushintsha ukuhamba kwawo komsebenzi adumele. Lindela amaviki amabili okuqala ukuze uzizwe unensa. Lindela amaviki alandelayo ukuthi uzizwe ngokushesha okukhulu. Amaqembu ashiye ngeviki lesibili awakaze akubone ukuhlangana.
- Ama-feed agents esikhundleni sokubagada. Amaqembu e-Frontier agcina i-backlog engaguquki yemisebenzi ehlelwe kahle enemiphumela ecacile, esebenzisa ama-ejenti amaningi ngokuhambisana futhi ebuyekeza okukhiphayo ngokuhambisanayo. Abakhi babika ukuqeda izici ezinkulu ngokuqhuma okufushane, nomsebenzi uthuthuka ngisho nalapho bengalindeli umenzeli ukuthi aqedele umsebenzi. Omunye unjiniyela oyinhloko uthumele ushintsho oluphelele 'ngamahora ambalwa kuphela esikhathi esihlangene' ngoba umenzeli wasebenza ngenkathi unjiniyela ehamba phakathi kokubuyekezwa kwamakhodi, usekelo lokusebenza, nemihlangano.
- Yenza kube sobala inhloso ngaphambi kokuba ikhodi ibhalwe. Kungakhathaliseki ukuthi ngezicaciso ezihlelekile, amadokhumenti ezidingo ezinemininingwane, noma ukuhlukaniswa komsebenzi okunobubanzi obuhle, amathimba asemngceleni aqinisekisa ukuthi ama-ejenti anomongo ocacile mayelana nokuthi 'okwenziwe' kubukeka kanjani ngaphambi kokuba baqale ukukhiqiza ikhodi. Amanye amaqembu asebenzisa le ndlela abika ngokubhala ngesandla u-1–2% kuphela wekhodi yawo kuyilapho ephusha kakhulu ukuzinikela kumuntu ngamunye ngeviki kunangaphambili.
- “Ukuhlola u-Shift kwesokunxele.” Amaqembu e-Frontier akha amathuluzi ukuze ama-agent akwazi ukwenza zonke izivivinyo zokuhlanganisa endaweni futhi azilungise ngaphambi kokuba ikhodi ifike epayipini. Ithimba le-Prime Video litshale imali kuma- guardrails azenzakalelayo, ukuhlola izingxenye, ukuhlolwa kokusebenza, namafomethi abambe izinkinga kusenesikhathi. Ukubuyekezwa kwekhodi kuguqule ukugxila kuzincazelo zesixhumi esibonakalayo nasezinqumweni zezakhiwo esikhundleni sesitayela sekhodi nezimiso zokuqamba amagama.
Lokho abaholi bezobuchwepheshe abangakwenza namuhla
Akuwona wonke amaqembu azuza le miphumela. Amaqembu eqa isigaba sokwakha umongo, aphatha i-AI njengento yokufaka esikhundleni, noma alindele izinzuzo ezisheshayo ngaphandle kokuhlela kabusha indlela asebenza ngayo ngokuqhubekayo engasebenzi kahle. Onjiniyela kuyo yonke imboni bamukele amathuluzi okufaka amakhodi e-AI. Akubona bonke ababona izinzuzo zokukhiqiza. Awasebenzisi amathuluzi angalungile. Basebenzisa amathuluzi alungile ngaphakathi kokugeleza komsebenzi okungalungile.
Izithako eziyinhloko yilezi:
- Shintsha indlela osebenza ngayo ukuze wenze i-AI isebenze kahle kakhulu.
- Izinto ezintathu ziyaphindaphindeka ukuze kulethwe imiphumela: I-AI ephethe umsebenzi wokwahlulela okuphansi x ukugxila okungaphazamiseki emsebenzini onokwahlulela okuphezulu x ukufinyelela okusheshayo kochwepheshe besizinda.
- Pilot kuqala, bese uyakala.
Isiqalo esisebenzayo asikona ukukhishwa okubanzi. Ungumshayeli wamabomu. Qala ngeqembu elincane elizimisele ukuchitha amasonto okuqala umongo womenzeli wokwakha (amafayela okuqondisa, izifanekiso ze-spec, ama-monorepos) ngaphambi kokubhala ikhodi yokukhiqiza. Nikeza ithimba igunya lokuhlela kabusha ukuhamba komsebenzi. Linganisa isivinini sokuzibophezela, imvamisa yokusebenzisa, kanye nesikhathi sokuxazulula, kanye nezikolo zokwaneliseka konjiniyela. Bese usebenzisa abakufundayo ukwakha incwadi yokudlala yenhlangano yonke.
Amaqembu athola ukuzuza okungu-4.5x kuya kwezingu-10x zokukhiqiza awazange nje athole ubuchwepheshe obungcono. Babone ukuthi bangasebenza kanjani ngendlela ehlukile ngayo. Leso sinqumo sitholakala kuzo zonke izinhlangano zobunjiniyela namuhla. Yiqiniso, isivinini sokwenza ikhodi siyingxenye yendaba. Sifuna ukusiza ngazo zonke izici zomjikelezo wokuphila wokuthuthukiswa kwesofthiwe, kungakhathaliseki ukuthi lokho kuwukuqondisa ukukhishwa kokuphathwa, ukusebenza, nokusebenza kokuphepha, noma ukubhekana nokuthuthukiswa kwe-EOL kanye nemisebenzi eminingi engahlukanisiwe eza nobunjiniyela besofthiwe. Hlala ubukele ibhulogi elandelayo, lapho ngizohamba khona ngokuthi sifinyelela kanjani kulezi.
Funda kabanzi mayelana namaqembu asemngceleni >
Lalela ku-AWS Summit New York City ukuze uthole okwengeziwe ngokuthuthukiswa komdabu kwe-AI.
Mayelana nombhali
Swami Sivasubramanian uyiPhini likaMongameli we-Agent AI e-Amazon Web Services (AWS). Kwa-AWS, i-Swami ihole ekuthuthukisweni nasekukhuleni kwezinsizakalo ze-AI eziholayo njenge-Amazon DynamoDB, i-Amazon SageMaker, i-Amazon Bedrock, ne-Amazon Q. Umsebenzi wethimba lakhe ukuhlinzeka ngezinga, ukuguquguquka, kanye nenani elidingwa amakhasimende nozakwethu ukuze basungule izinto ezintsha zisebenzisa i-AI ye-agent ngokuzethemba futhi bakhe ama-ejenti angewona anamandla futhi asebenza kahle kuphela, kodwa futhi athembekile futhi anomthwalo wemfanelo. USwami uphinde wasebenza kusukela ngoMeyi 2022 kuya kuMeyi 2025 njengelungu leKomidi Likazwelonke Lokweluleka ngobuNhloli, elalinikezwe umsebenzi wokweluleka uMongameli wase-United States kanye neHhovisi Likazwelonke Le-AI Initiative ngezihloko ezihlobene ne-National AI Initiative.
