Ngaba i-ByteDance's Lumi liLingelo le-AI eliguqulelwe njengoluntu lwabadali be-AI?

Hlangana noLumi, ukukhululwa kwe-ByteDance yakutshanje-i-mashup ye-Pinterest, i-GitHub, kunye ne-Fiverr . Okanye ke ndithatha kwimiyalezo yabo efihlakeleyo. Konke okuqinisekisiweyo kukuba abasebenzisi baya kuba nakho ukulayisha kunye nokwabelana ngeemodeli ze-AI, ubugcisa bokuhamba komsebenzi, kunye nokulinga ngoqeqesho lwe-LoRA (Low-Rank Adaptation). Ubulumko obuqhelekileyo bucebisa ukuba le yimpendulo ye-titan kuthungelwano lwabadali be-AI abakhokelayo baseTshayina, iLiblib kunye neCivitai. Kodwa kwinkampani enkulu njengeByteDance, kunye nomseki onebhongo njengoYiming Zhang, utshintsho lwepokotho kuluntu lwabadali be-AI ziitapile ezincinci.

Nantsi ithiyori eyahlukileyo: I-Lumi ayiloqonga, kodwa yiprototyping powerhouse eyenzelwe ukuphembelela iinguqulelo ezilandelayo ze-ByteDance ze-AI ngokudibana okucwangcisiweyo phakathi kwabasebenzisi bokwenyani kunye “nabadali” bokwenziwa be-AI.

Olu luqikelelo olusulungekileyo , ewe, kodwa masithathele ingqalelo i-ByteDance yendlela yamva nje: yoyiswa ngokukrokrelayo kwihlabathi, ngelixa sikhupha uninzi lweemodeli ze-AI kwintengiso yabo yasekhaya. Ukujonga le nto ukuya kuthi ga kuvavanyo lwaseNtshona phakathi kwe-TikTok saga yingcaciso egaywa lula, kodwa ukukrwaqula kancinci kwaye iqala ukubonakala ngathi iByteDance ibeke ingqikelelo yabo endaweni, belungiselela ukuyivavanya kwilebhu ekufutshane nekhaya, apho. imithetho inkungu ngeyona ndlela.

Kwaye nje ukuba iimvavanyo ziphelile, ungothuki ukuba uLumi unyamalale ethe cwaka, eshiya kuphela uvimba wobunjineli beByteDance.

Indlela emfutshane eya kuBuphezulu: Ukuguqula iNkqubo kwiPlain Sight

Kwincwadi yomthetho engcwele yeSilicon Valley, izimvo zifanele ukuba zihamba ngepolishi engapheliyo, ukukhutshwa okulumkileyo koluntu, kwaye emva koko impendulo ingena ngesantya segama leCawa. I-ByteDance isenokuba yarhweba ngesiko ukuze ifumane into enobunkunkqele ngakumbi: ukudibanisa intsebenziswano yokwenyani kunye neyenziweyo yabasebenzisi ukuya kwimodeli yokukhutshwa enemigangatho emithathu elungelelaniswe ukuvavanywa kwemodeli ye-AI engayekeleliyo , yonke iphantsi kwesigqubuthelo sobuso “soluntu lwabadali be-AI.”

Ukuba yinyani, le ndlela yokujonga izakuvelisa indlela entsha (kwaye ingekho sikweni ngokupheleleyo ) yokuvelisa izinto ezintsha, edibanisa isantya, ulawulo, ukusebenza kakuhle, ukuchaneka, lonke eli xesha kuncitshiswa umngcipheko woluntu. Kulo mzekelo, impendulo yomsebenzisi ikhawuleza, uhlengahlengiso lwemveliso luhlala luhleli, kwaye abaphulaphuli, ngokungazi, baba yintloko yovavanyo kunye nesingqinisiso .

Ukukhutshwa ngokuthe ngqo kubasebenzisi bokwenyani kunye nabasebenzisi be-Synthetic-Buffing Up I-ByteDance-eyamkelweyo

Kumgca ongaphambili weqhinga leByteDance kukukhutshwa okuthe ngqo, okugcinelwe iimodeli ezigudisiweyo ezithathwa njengezilungele izandla zabantu, ezifana nePixelDance kunye neSeaweed. Abasebenzisi bokwenyani bayangena, kodwa ababodwa . I-AI personas-i-parading njengabadali abangoogxa, kodwa ixhotyiswe ngokusetyenzwa kolwimi lwendalo, umbono wekhompyuter, kunye nokufunda okomeleza endaweni yeStable Diffusion kunye neFlux-kukho ukujongwa, ukuhlalutya, kunye nokusasaza ukuqonda kwishishini kubhabho.

Umntu ngamnye we-AI unokudibanisa idatha kwiitshaneli ezininzi , ubambe yonke into ukusuka kwiindlela zokucofa ukuya kwisantya sokuskrola, ukudala iphrofayili yobuninzi bokuzibandakanya komsebenzisi. Basebenzisa i-multimodal data fusion ukudibanisa olu lwazi, ukudibanisa iipatheni zokuziphatha ukusuka kunqakrazo, ukulinganisa kwamehlo, kunye nohlalutyo lwethoni, zonke ngoko nangoko. Kwaye ngokusebenzisa ukufunda okudityanisiweyo , aba bantu banokutsalwa kwidatha ebekwe elugcinweni ngaphandle kokubeka esichengeni ubumfihlo babasebenzisi-isixhobo esinamandla sokuqaqambisa amava omsebenzisi ngaphandle komngcipheko kukhuseleko lwedatha .

Le ndlela ye-hybrid itsiba isigaba "se-beta kawonke-wonke" esibandakanya umjikelo omde: inqaku lokukhululwa, hlala kwaye ulinde, uhlalutye, kwaye ekugqibeleni uchithe iiveki uphikisana ukuba wenze ntoni ngokulandelayo. Ngokuthembela kule ngqungquthela, isiphaluka esiqhutywa yi-AI , i-ByteDance inokwenza uhlengahlengiso oluchanekileyo phakathi kweentsuku ukuba azikho iiyure, ukukhulula iimpawu ezilungelelanisa ngexesha langempela ekuziphatheni komsebenzisi-i-loop yempendulo eyenza uvavanyo lwe-beta ye-Silicon Valley ibe neentloni.

Ukukhutshwa kwe-Synthetic-Only: De-Risking The Moonshot Projects

Iingxilimbela zeTekhnoloji zonke zibeka ukubheja kodwa okunokuba nengeniso, ukuntywila izixhobo ezininzi kwiiprojekthi zokudubula kwenyanga ezinokuphinda zivelise ivili ledijithali okanye zihlale emanzini ashushu. Apho iinkampani zemveli zinokuwaphatha la manyathelo ngeeglavu zabantwana, ukubila ngenxa yeentlekele ze-PR ezinokuthi zikhutshwe, i-ByteDance irhweba ngamadabi e-boardroom ye-bots, ikhulula ezi ngqikelelo kwindawo yovavanyo eyenziweyo kuphela . Apha, i-AI personas ngabona babathathi-nxaxheba bodwa, bebeka imodeli nganye kuvavanyo olupheleleyo, olukwizinga eliphezulu apho kusinda kuphela abona banamandla. Akukho bantu, akukho zihloko, akukho ngozi inokwenziwa .

Ngoluhlu lweendlela ze-AI kubugcisa babo, abasebenzisi bokwenziwa banokutyhala into nganye kwimida yobugcisa kunye nengqikelelo . Ukulinganisa okuchaseneyo kuvavanya ukuba iiprototypes zinokuya kude kangakanani, ngelixa iimodeli zokufunda ezinzulu ziziqhelanisa ngexesha lokwenyani kuyo yonke imeko ephoswe kuyo. Aba bantu be-AI bemba nzulu, bekhangela iindawo ezibuthathaka, iimeko ezibalekayo, kwaye basebenzise ukufunda okungajongwanga kubuthathaka obungaphandle obunokuthi buphume kuvavanyo lwabantu .

I-ByteDance inokongeza omnye umaleko wokuntsonkotha ngokusebenzisa iimodeli ezivelisayo ukwazisa unxibelelwano lwabasebenzisi olungalindelekanga, ukunika iprototypes ukuvezwa kuluhlu olubanzi lweemeko. Yinkqubo eguquguqukayo , kunye nenqaku ngalinye livela ngempendulo eqhutywa yi-AI-evelayo ibukhali kwaye yomelele ngakumbi, ilungele ukuthathela thina bantu.

Iimbono ezisilelayo ziye ziphele ngokuzolileyo, zigcina isidima esicocekileyo seByteDance; ngelixa abasindileyo bephuma bephucukile, bomelele, kwaye bemiselwe ukukhululwa kwelizwe lokwenyani. Yindawo yokubonisa eqhutywa yi-AI-ibala elivaliweyo apho i-ByteDance inokucinezela imiphetho ye-radical, engenakwenzeka, kunye nomdlalo onokuthi utshintshe ngokuzithemba okupheleleyo .

Ukukhutshwa kwe-Synthetic-to-Real-Camouflaging Concepts “Njengomxholo womdali”

Apha kulele ixilongo . Betha igama lebhrendi kumbono, kwaye ngequbuliso kufuneka limenyezele, nto leyo ethetha ukuzika zombini iiyure kunye needola epolishini—ehlisa isantya “sokuhamba ngokukhawuleza kwaye wophule izinto” mantra yetekhnoloji. Kodwa ukuba bayakwazi ukwenza iikhonsepthi zamathambo phantsi kwesimo somxholo oveliswe ngumsebenzisi, i-ByteDance inokulinganisa impendulo yokwenyani kwimibono ebhakiweyo isibini seekhowudi ezincinci ezinokuthi zidibanise, lonke eli xesha liphepha ingxaki yokusungulwa ngokusemthethweni.

Ukuqala nje kokujonga, ezi modeli zinokubonakala ngathi zinomdla omncinci-izihluzi ezilula, i-automation esebenzayo, mhlawumbi isixhobo sohlalutyo se-niche. Badibanisa, njengokulayishwa kwe-amateur okuphakathi okanye igalelo lomthombo ovulekileyo. Kodwa phantsi kwe-vibe engazinzanga, iinkqubo ze-AI zinokuba ziguquguquka ngokupheleleyo, kunye nezixhobo ze-NLP ezithatha i-undercurrents ye-real-time reaction, kunye nemizekelo yokufunda eyomelezwayo ingena nzulu kwidatha yokuzibandakanya, ukubeka esweni kanye ukuba iqela ngalinye lisebenzisana njani nemodeli, ukubhala phantsi. amaxesha apho umdla unyuka.

I-ByteDance inokugrumba nzulu nangakumbi, idibanise abasebenzisi ngokuziphatha ukuze babone ukuba ngubani otsalwe kwimodeli nganye kwaye kutheni. Ukufunyaniswa okungaqhelekanga kunokuphawula naluphi na usetyenziso olungaqhelekanga okanye olungacetywanga, ukusuka kumacala angaqhelekanga ukuya kwiindawo ezibuthathaka zokungcwatywa kuvavanyo lwemveli olunokuthi ziphose, kunye neemodeli zengqikelelo zecausal esahlula ukuba yintoni eqhuba ukuzibandakanya komsebenzisi kunye nokuba yintoni esikalini.

Zonke ezi mbono zinokuphembelela i-injini yesigqibo eyaziyo ukuba ingayiyeka nini imodeli iphume ngobubele eqongeni ukuba ayinqumi. Kodwa xa ingqikelelo ibonisa isithembiso, iimodeli eziqikelelwayo ziyakhaba, zilinganisa amandla ayo okukhululwa ngokubanzi. Kwaye xa kukho into ephawula kuzo zonke iibhokisi, i-ByteDance inento yokuqala ephambili yayo ilungiselelwe ukusungulwa ngokupheleleyo.

Iingcamango Zokugqibela: Inkqubela-phambili Cofa Dlala—Kodwa Liphi Iqhosha Lokunqumamisa?

Masicace: oku ayisosikhalo sokubuza imibuzo kwi-ByteDance. Endaweni yoko, lubizo lokuthatha inyathelo umva kwaye ubuze-ngeliphi ixesha apho ukuthanda kwethu izinto ezintsha kuqala ukodlula amandla ethu okukugcina kuqwalaselwe? Ukusuka kwi-ByteDance ukuya kuGoogle ukuya kwiAmazon ukuya ekuqaleni ukuya ezantsi eholweni— ngubani ongazimiselanga ukutyhala imvulophu ?

Kodwa, kungxamiseko lokuguqula imbono nganye eshushu ibe linqaku elikhulu elilandelayo, umgca phakathi kwenkqubela phambili kunye nokuphendula unokunxitywa ngokuyingozi.

Mhlawumbi lixesha lokuba sityale imali engakumbi kumsebenzi ongathandekiyo wokuseta ii-guardrails, ngakumbi malunga nendlela esilawula ngayo intsebenziswano yoluntu ethi i-AI ngoku ilamle. Oku akukho malunga nokuthintela ukusungulwa, kodwa malunga nokukhusela ilaphu elithambileyo loqhagamshelwano lwabantu , isakhelo esinokonakala ngokulula phantsi kobunzima bokungaqhwanyazi, okungadinwayo kotshintshiselwano oluqhutywa yi-AI. Ukuba sonke sixakekile sileqa yonke ingqikelelo enyanzelisayo ukuya kweyona ndawo ikude, ngubani oseleyo ukusitsalela umva?

I-AI ayithathi isingqi ukuphefumla, ukubonakalisa, okanye ukuhlaziya. Kodwa mhlawumbi, njengabayili behlabathi, kufuneka.