Hyra 2.0 Uyasebenza

I-Hyra 2.0 ingumenzeli wocwaningo lwesayensi osebenzisa i-AI, owethulwe yi-Tencent Hunyuan. Iyinguqulo ethuthukisiwe ye-Hyra 1.0, esuka ekuxazululeni izinkinga ngokuzenzakalela iye ekubeni ithuluzi locwaningo nentuthuko. I-Hyra 2.0 ingaqeqesha amamodeli esayensi ngokuzimela isikhathi eside, ithuthukise ukuhamba komsebenzi wocwaningo, futhi igcine ulwazi njengempahla engasetshenziswa kabusha. Imiphumela yayo isivele ihlanganisa ukuklanywa kwamaprotheni, ukulingisa ama-molecule, ubufakazi bezibalo kanye nokulungiswa kwamaphutha e-quantum. Umgomo wesikhathi eside uwukwenza ukuthuthukiswa kwamathuluzi nokuthuthukiswa kwamamodeli kuqondisane, kwakhe umjikelezo wokuzithuthukisa oqhubekayo.

Ukuthuthukiswa kwamathuluzi: umenzeli uphatha yonke inqubo yokuklama, ukuqeqesha, ukuhlola nokusabalalisa amamodeli esayensi.
Ucwaningo lokusebenzisana olunezigaba ezintathu: i-Overall, i-Mission ne-Worker zihlela ucwaningo lwesikhathi eside ngokusabalalisa imisebenzi; i-Worker ingenza izivivinyo ngokuzimela futhi ibuze izahlulelo ze-Mission.
Ukuqinisekisa okuzimele: i-Review Worker iqala ngokubona inkinga yokuqala nobufakazi bayo kumongo omusha, bese iqinisekisa iziphetho zalowo owenze umsebenzi.
Ukugcinwa kolwazi: isimo sephrojekthi, inkumbulo yama-agent kanye nolwazi olwabiwe kugcinwa ngokwehlukana; ama-skill akhiqizwa ekubuyekezeni umsebenzi angasetshenziswa emisebenzini elandelayo.
Ukusetshenziswa kwemiphumela emikhakheni eminingi: imiphumela yokuqala ihlanganisa ukuklanywa kwamaprotheni, ukulingisa ama-molecule, ubufakazi bezibalo, ukulungiswa kwamaphutha e-quantum, ukuklanywa komzimba wamarobhothi nokuthuthukiswa kwama-chip.
Ukubamba iqhaza ekuthuthukisweni kwamamodeli amakhulu: ibamba iqhaza kungqalasizinda yokuqeqeshwa kwangaphambili, epayipini ledatha yokuqeqeshwa kwangemva kwalokho nasekuthuthukisweni kohlelo lokucabanga lwemodeli yesizukulwane esilandelayo i-Hy.

Ukusetshenziswa kabusha kwamathuluzi kanye nenzuzo eqoqekayo yobuhlakani: indatshana ihlukanisa amathuluzi abe amathuluzi anezimpawu namathuluzi anepharamitha. Umbono oyinhloko uthi ukuxazulula inkinga eyodwa kukunikeza impendulo eyodwa kuphela, kodwa ukuthuthukisa ithuluzi elilodwa kungazuzisa yonke imisebenzi elandelayo. Wonke umklamo we-Hyra 2.0 ugxile ekwenzeni imiphumela yocwaningo olulodwa iphinde isetshenziswe ocwaningweni olulandelayo, kwakheke inzuzo eqoqekayo yobuhlakani eqhubekayo.
Isakhiwo senhlangano esinezigaba ezintathu kanye nokuhlela okuqhubekayo: i-Hyra 2.0 isebenzisa izigaba ezintathu ze-Overall, Mission ne-Worker. I-Overall ixhumana nabantu futhi ilawule izinto eziza kuqala nezinsiza; i-Mission iphethe inkinga eyodwa yocwaningo isikhathi eside futhi ithwale umthwalo weziphetho; i-Worker yenza ukuhlola nezivivinyo. Ngemva kokuba umsebenzi ungene kulayini oqhubekayo, uhlelo lokuhlela luwuqala ngokwezinto eziza kuqala, ubudlelwano bokuncika kanye nemikhawulo yezinsiza. Uma umsebenzi ulinde imiphumela yokuqeqeshwa noma ukugunyazwa ngumuntu, eminye imisebenzi iyaqhubeka ngokujwayelekile. Umenzeli angasetha isikhathi sokuvuswa ukuze aqhubeke nomsebenzi endaweni yokuhlola.
Ukusebenzisana okungavumelani ngesikhathi kanye nendlela yokuqinisekisa okuzimele: inqubekelaphambili yocwaningo oluhambisanayo iqoshwa ku-Research Board. Amanye ama-agent afunda inqubekelaphambili evamile ngokungavumelani ngesikhathi, kanti izindaba ezidinga ukusingathwa zithunyelwa ngemiyalezo futhi zidinge impendulo ecacile, ukuze kugwenywe ukuphazamisa kaningi imisebenzi esaqhubeka. Lapho i-Worker ibika ukuthi iqedile, kufanele ifake inguqulo yekhodi, imiphumela yezivivinyo, amalogi noma umkhondo wokuqalisa njengobufakazi. Uma kudingeka ukuqinisekiswa okuzimele, uhlelo luqala i-Review Worker entsha, yona eqala ngokubona inkinga yokuqala nobufakazi bayo kumongo omusha, ukuze kususwe umthelela weziphetho zalowo owenze umsebenzi. Amaphutha atholwe ekuqinisekisweni noma iziphetho ezinomkhawulo omncane zigcinwa ngokwethembeka.
Ukuhlukaniswa kwezigaba ezintathu zenkumbulo kanye nesizindalwazi solwazi esinamazinga: uhlelo lugcina ngokwehlukana isimo sephrojekthi, inkumbulo yomsebenzi enezigaba ye-agent ngayinye (ukuqonda kwamanje, imibono esemqoka, izindlela ezikhishiwe kanye nohlelo olulandelayo), kanye ne-Books, okuyisizindalwazi solwazi olwabiwe phakathi kwemisebenzi. Irekhodi ngalinye ku-Books kufanele lichaze indlela, isizathu, izimo zokusebenza futhi lifake ubufakazi bokuqala. Amarekhodi ahlukaniswa ngokuthembeka abe amazinga amathathu: “okubonwe”, “okuqinisekisiwe” kanye “nesivumelwano”. Lapho kuvela ubufakazi obusha, amarekhodi akhona angalungiswa.

I-Hyra 2.0 okwamanje iwuhlelo lwangaphakathi luka-Tencent Hunyuan locwaningo nentuthuko yesayensi, futhi ayiwona umkhiqizo we-SaaS ovulekele umphakathi.

Ukuthuthuka kwepharadigm: isuka ekutheni i-agent ixazulule izinkinga ngokuqondile iye ekutheni i-agent ithuthukise futhi ithuthukise amathuluzi ocwaningo lwesayensi, imiphumela iphinde isetshenziswe ocwaningweni olulandelayo, kwakheke inzuzo eqoqekayo yobuhlakani.
Ikhono locwaningo lwesikhathi eside: isakhiwo se-Overall, Mission ne-Worker kanye nolayini oqhubekayo kusekela ucwaningo oluthatha izinsuku kuya emavikini, olunamaphrojekthi amaningi asebenza ngesikhathi esisodwa.
Ukuthembeka okuphezulu kweziphetho: i-Review Worker yenza ukuqinisekiswa okuzimele kumongo omusha, futhi idinga ubufakazi obufana nenguqulo yekhodi namalogi, kunciphisa ubungozi bokuthi i-agent iqinisekise iphutha layo ngokwayo.
Ulwazi olungagcinwa: isimo sephrojekthi, inkumbulo ye-agent kanye nesizindalwazi solwazi olwabiwe kuhlukaniswe ngezigaba ezintathu. I-Books ilawulwa ngamazinga okubonwe, okuqinisekisiwe nesivumelwano, kanti ama-skill angasetshenziswa kabusha emisebenzini ehlukene.
Ukwakha amamodeli kusukela ekuqaleni kuze kube sekugcineni: i-HyInv, i-HyPLM namanye amamodeli amasha aqede ukuqeqeshwa kwangaphambili, ukulungiswa, ukufunda okuqinisa kanye nokuthuthukiswa kokucabanga kusukela ekuqaleni.
Ukusabalala kwemikhakha: imiphumela yokuqala ihlanganisa imikhakha eyisithupha: ukuklanywa kwamaprotheni, ukulingisa ama-molecule, ubufakazi bezibalo, ukulungiswa kwamaphutha e-quantum, ukuklanywa komzimba wamarobhothi kanye nokuthuthukiswa kwama-chip.

Iwebhusayithi yephrojekthi: https://hy.tencent.com/research/hyra2?langVersion=zh

Ukuklanywa kwemithi nobunjiniyela bama-enzyme: isebenzisa amamodeli okubuyisela ukugoqa afana ne-HyInv, iqede ukuklanywa kokulandelana kwamaprotheni okungu-128 kungakapheli ama-millisecond angu-50, isheshise ukukhethwa kwama-molecule angaba yimithi.
Ukulingisa amandla e-molecule: isebenzisa i-HyFF, inkambu yamandla esezingeni lama-athomu, ukuqhuba ukulingisa okukhulu, ilandele ukunyakaza kwama-molecule nezinqubo zokusabela esikalini sama-athomu, futhi isekele ucwaningo lwezinto nemithi.
Ucwaningo lwethiyori yezibalo: isebenzisa ukuhamba komsebenzi kwe-HyProver ukuhlela ukuphakanyiswa kwemibono, ukwakhiwa kobufakazi kanye nokuqinisekiswa okuzimele, isize osolwazi bezibalo ukubhekana nezinkinga ezivulekile.
Ukuthuthukiswa kwamakhodi okulungisa amaphutha e-quantum: isebenzisa i-HyQEC ukwakha amakhodi amasha e-quantum futhi iqinisekise ngokunembile ibanga lekhodi, inikeze amathuluzi okuthuthukisa izindleko zokulungisa amaphutha kumakhompyutha e-quantum.
Ukuklanywa nokuqinisekiswa kwama-chip: njengokuthuthukisa ukusebenza kwe-Xiangshan processor, yenza ukulungiswa kokusebenza, ukuhlolwa kwe-fuzz kanye nokuqinisekiswa okusemthethweni, inciphise isikhathi somjikelezo wokuthuthukiswa kwama-chip.