Kwenzekeni
I-Tech in Asia ibika ukuthi i-Tencent ikhiphe i-Hy4 Preview futhi yenza izisindo zayo zitholakale nge-platform yefu ye-Tencent. Umbiko uthi u-Tencent uchaze imodeli njengenamapharamitha ayizigidi eziyizinkulungwane ezingama-770 kanye newindi lengqikithi yethokheni yesigidi, nokuhlanganiswa okuhlelelwe i-WorkBuddy kanye ne-chatbot kaTencent.
I-Tech e-Asia ibika ukuthi u-Tencent ukhiphe i-Hy4 Preview, eyichaza njengemodeli eyisisekelo enamapharamitha ayizigidi eziyizinkulungwane ezingu-770 kanye nefasitela lengqikithi yethokheni yesigidi. Umbiko uthi i-Tencent yenza izisindo zemodeli zitholakale ku-platform yayo yamafu futhi ihlose imodeli yokuhlanganiswa emikhiqizweni ehlanganisa i-WorkBuddy ne-Tencent chatbot. Umthombo awucacisi ukuthi izisindo zingalandwa ngaphandle kwefu lika-Tencent, noma yikuphi ukufinyelela, intengo, isifunda, noma izimo zelayisensi ezisebenzayo.
Ngokusho kwe-Tech in Asia, u-Tencent uthe i-Hy4 Preview iphumelele i-GLM-5.3 ye-Z.ai ye-Z.ai kanye ne-Moonshot AI's Kimi K3 ekuhlolweni kobunjiniyela obungaboni okwenziwa ochwepheshe be-Tencent. Leyo miphumela izimangalo zenkampani ezibikwe yi-Tech e-Asia futhi aziqinisekiswanga ngokuzimela emthonjeni onikeziwe. I-athikili ayihlinzeki ngamaphuzu, isethi yokuhlolwa, inombolo yabahloli, iphrothokholi yokuhlola, noma ulwazi olumayelana nokuthi amamodeli aqhathanisiwe asetshenziswe ngaphansi kwezimo ezifanayo.
I-Tech in Asia iphinde ibike ukuthi imiphumela kumabhentshimakhi ezinkampani zangaphandle iye yahluka lapho i-Hy4 Preview iqhathaniswa nezimbangi zasekhaya okuhlanganisa i-Qwen 3.8 Max ye-Alibaba. Lezo ziqu zikhawulela lokho okungaphethwa ngokuhlolwa kwangaphakathi kuka-Tencent. Umthombo awuwahlonzi amabhentshimakhi ezinkampani zangaphandle noma unikeze izibalo eziqhathaniswayo, ngakho-ke izinga nokuvumelana kwanoma iyiphi inzuzo yokusebenza kuhlala kungaziwa.
Umbiko uthi u-Tencent ukhombise ukusebenzisa okubandakanya ukudalwa kwegeyimu, idizayini yewebhu ye-3D, nokugeleza kokusebenza kokuthobela i-accounting. Lawa achazwa njengamademo, hhayi njengokuthunyelwa okuhlolwe ngokuzimele noma ubufakazi bokutholwa kwamakhasimende. I-Tech e-Asia iphinde ibeke ukukhululwa kumongo wentshisekelo yomtshali-zimali ekutheni ukusetshenziswa kwakamuva kwe-Tencent kwe-AI kungakhiqiza imbuyiselo, kodwa i-athikili enikeziwe ayinikezi imiphumela emisha yezezimali ehlanganiswe ne-Hy4 Preview.
Imininingwane yomthombo: techinasia.com ↗
Kungani kubalulekile
Uma ukucaciswa okubikiwe kanye nezimangalo zokusebenza zisekhona, i-Hy4 Preview izongeza enye imodeli enkulu yamaShayina emakethe yasekhaya eya ngokuya iqhudelana. Ukutholakala kwayo kwamafu okubikiwe kungenza imodeli ibe usizo konjiniyela, nakuba umthombo ungaqinisekisi ukuthi ingafinyelelwa ngobubanzi bani.
I-Hy4 Preview ibalulekile ngoba inhlanganisela ebikiwe yesilinganiso esikhulu kakhulu, iwindi lomongo wethokheni lesigidi, nokufinyelela kwamafu kungase kuthinte indlela onjiniyela abahlaziya ngayo amamodeli omude nomthamo omkhulu. Iwindi lomongo omude lingase libe usizo ekuphatheni izisekelo ezinkulu zekhodi noma amaqoqo amarekhodi, kodwa umthombo awunikezi ukuhlola okubonisa ukuthi i-Hy4 Preview iwusebenzisa ngokunembe noma ngempumelelo kangakanani lowo mthamo. Ukubalwa kwepharamitha kukodwa akusunguli ikhwalithi yomhlaba wangempela, isivinini, izindleko, noma ukwethembeka.
Ukukhishwa kuphinde kugqamise ukuncintisana phakathi konjiniyela abakhulu be-AI baseShayina. I-Tech in Asia ifreme imiphumela ka-Tencent efunwa ngamamodeli avela ku-Z.ai, Moonshot AI, ne-Alibaba, ibeka i-Hy4 Preview ngaphakathi kwemakethe yasekhaya lapho abakhi bamamodeli beqhudelana ngamabhentshimakhi, ukusatshalaliswa, nokuhlanganiswa komkhiqizo. Umbiko awuqinisekisi ukuthi i-Tencent ihola imakethe iyonke, futhi ukukhuluma kwayo ngemiphumela ehlukahlukene yenkampani yangaphandle iphikisana nokuphatha ukukhishwa njengezinga elizinzile.
Ukutholakala kwe-Cloud-platform kungase kwehlise umgoqo osebenzayo wamabhizinisi afuna ukuzama i-Hy4 Preview ngaphandle kokusebenzisa imodeli egcwele ngokwawo. Ngesikhathi esifanayo, umthombo awusho ukuthi onjiniyela bangaphandle bangakwazi yini ukuthola izisindo ngokuqondile, noma ngabe imodeli iyatholakala emazweni ngamazwe, noma ukuthi ukusetshenziswa kukhawulelwe. Leyo mininingwane ibalulekile ekunqumeni ukuthi ukukhishwa kuwusizo kakhulu yini noma ikakhulukazi ukunwetshwa kwe-ecosystem ka-Tencent.
Idemo ebikiwe yokuthobela ukubalwa kwezimali ibonisa indawo yokusetshenziswa okungenzeka ibe umphumela ngenxa yokuthi amaphutha emsebenzini wokuthobelana angadala izingozi zomthetho nezezezimali. Kodwa i-Tech in Asia ayibiki izilinganiso zokunemba, ukuthunyelwa kwamakhasimende, izidingo zokubuyekezwa komuntu, nanoma yimuphi umphumela webhizinisi oqinisekisiwe kusukela kulokho kuboniswa. Umkhawulo ofanayo uyasebenza kumademo wokudalwa kwegeyimu nawe-3D-design: abonisa izimo zokusetshenziswa ezihlosiwe, hhayi ukusebenza kokukhiqiza okubonisiwe.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
What is a common training objective for an autoregressive language model?
Ongakubuka ngokulandelayo
Ukuhlola okuyinhloko imiphumela ezimele yebhentshimakhi, imigomo yemodeli yelayisense neyokufinyelela, kanye nobufakazi bokuthi ukuhlanganiswa okubikiwe kuhamba ngale kwemiboniso. Akukacaci ukuthi imithetho yaseShayina elawula imithombo evulekile kanye namamodeli esisindo esivulekile izothinta ukusatshalaliswa.
Ukuhlola okuzimele kuyisinyathelo esilandelayo esibaluleke kakhulu. Buka imiphumela eshicilelwe, amasethi okuhlola, ukwaziswa, ukulungiselelwa kwezingxenyekazi zekhompuyutha, izilungiselelo zokucatshangelwa, neziqhathaniso ezenziwa abahloli ngaphandle kokuxhumana kwezezimali noma kwenhlangano ku-Tencent. Ukunakwa ngokukhethekile kufanele kuye emiphumeleni yenkampani yangaphandle ethi i-Tech in Asia ithi yahluka, njengoba leyo ingase inikeze isithombe esingokoqobo kunokuhlolwa kobunjiniyela bangaphakathi kwe-Tencent.
Imigomo yokufinyelela izonquma ukubaluleka kokusebenza kokukhishwa. Okungaziwa okubalulekile kufaka phakathi ukuthi izisindo ze-Hy4 Preview ziyakwazi ukudawunilodeka noma zishayelwa kuphela nge-Tencent Cloud, ukuthi ilayisense evulekile ikuvumela yini ukuguqulwa nokusetshenziswa kwezentengiso, ukuthi imodeli ibiza malini ukuze isebenze, nalapho itholakala khona. Umbiko onikeziwe awuphenduli noma yimuphi wale mibuzo.
Ubufakazi obulandelayo obuvela ku-WorkBuddy kanye ne-chatbot ka-Tencent kufanele buhlukanise ukuhlanganiswa kokumaketha nokusetshenziswa okuqhubekayo komkhiqizo. Amasignali awusizo azofaka phakathi ukutholakala okuqinisekisiwe, izinombolo zomsebenzisi, imikhawulo yomthwalo wokusebenza, ukubambezeleka, amanani amaphutha, nezivikelo ezibhaliwe. Umthombo awunikezi umugqa wesikhathi wokuhlanganiswa futhi awubiki ukuthi noma yimuphi umkhiqizo uzuze umphumela othile nge-Hy4 Preview.
Intuthuko yokulawula e-China ingaba nomthelela ekutheni imodeli isatshalaliswa kanjani. I-Tech in Asia ibika ukuthi imithetho elawula amamodeli e-AI omthombo ovulekile kanye nesisindo esivulekile ahlala engazinzile, kodwa i-athikili ayiwuhlonzi umthetho othile osalindile noma ichaza ukuthi izosebenza kanjani ku-Hy4 Preview. Ngakho-ke abatshalizimali nomphakathi kufanele baphathe noma yisiphi isibikezelo mayelana nembuyiselo ka-Tencent, isabelo semakethe, noma ukusatshalaliswa njengokungaqiniseki kuze kube yilapho kutholakala idatha yokufinyelela nokutholwa.