Subquadratic Ihlolwa ku-Quasa: I-LLM Yokuqala Esebenzisa Isakhiwo Esingaphansi Kwe-Quadratic

#Quasa #QUA #subquadratic

Subquadratic yinkampani eyenza ucwaningo oluphambili ku-AI njalo yakha ingqalasizinda yayo, isakha uhlobo olutsha lwamamodeli amakhulu olimi. Imodeli yabo ehamba phambili, SubQ, yi-LLM yokuqala eyakhiwe phezu kwesakhiwo sokunaka okukhethayo esisebenza ngokupheleleyo ngezinga elingaphansi kwe-quadratic. Ngokungafani lama-Transformers ajayelekileyo, lapho izidingo zokubala zikhula ngesikwele sobude bomongo ($O(n^2)$), ku-SubQ zikhula ngokulingana lobude bomongo ($O(n)$), okwenza ukucabangisisa kusetshenziswa ama-token ayizigidi ezinengi kwenzeke njalo kusebenzise izinsiza kuhle.

I-SubQ yakhelwe ikakhulu imisebenzi elomongo omude. Ingacubungula izinqolobane zekhodi ezipheleleyo, imilando emide yama-ejenti, izicelo zokuhlanganisa ikhodi eziqoqwe okwezinyanga, lezimo ezigcinwa ngokuqhubekayo, konke kusemyalelweni owodwa ngaphandle kokwehla kwekhwalithi. Khathesi imodeli isekela amafasitela omongo afika kuma-token ayizigidi ezingu-12 njalo isebenza kuhle ezilinganisweni zokubuyisa ulwazi emongweni omude lezokucabangisisa okuvamileyo. Kuma-token ayisigidi esingu-1, i-SubQ idinga umsebenzi wokubala omncane kakhulu (kuthiwa ingaphansi ngokuphindwe 64.5× kulokunaka okuminyeneyo) njalo isebenza masinyane kakhulu kulezindlela zokunaka ezijayelekileyo.

Inkampani igxile esimisweni sokuthi “ukusebenzisa izinsiza kuhle yikuhlakanipha.” Ngokuxazulula umgoqo wokukhula kwe-quadratic ezingeni lesakhiwo, i-Subquadratic ihlose ukwenza kube lesizukulwane esitsha sama-ejenti e-AI angenelisa ukusebenza lomongo omnengi, kulokuthembela ekuwehlukaniseni ube yiziqephu, ekuwufinqeni, kumbe emaqhingeni okubuyisa ulwazi.

I-Subquadratic ifanele kakhulu abathuthukisi bezinhlelo, ama-ejenti okubhala ikhodi, amaqembu ezinkampani, labacwaningi abadinga ukucabangisisa okuthembekileyo ngomongo omude — ikakhulu ekuhlaziyeni yonke ikhodi yohlelo, kuma-ejenti asebenza isikhathi eside, lasemisebenzini enzima ehlanganisa imibhalo eminengi.

Okuqakathekileyo

• Isakhiwo sokuqala se-LLM esisebenza ngokupheleleyo ngezinga elingaphansi kwe-quadratic (izidingo zokubala zikhula ngokulingana lobude bomongo).
• Umongo wama-token ayizigidi ezinengi (kufika ku-12M tokens).
• Ukubuyisa ulwazi emongweni omude okusebenza kuhle kakhulu kusetshenziswa izinsiza ngendlela ephumelelayo kakhulu.
• Yakhelwe izinqolobane ezipheleleyo lesimo se-ejenti esigcinwa ngokuqhubekayo.
• I-API ehambelana le-OpenAI lokuhlanganiswa lama-ejenti okubhala ikhodi.

• Okungadinga Ukucatshangelwa

• Njengemodeli yocwaningo oluphambili, inethiwekhi yayo ebanzi yamathuluzi lezinhlelo, kanye lezinhlobo zayo ezilobukhulu obutshiyeneyo, kulokhu kusanda.
• Ifanele kakhulu imisebenzi esebenzisa umongo omude kulengxoxo ezijayelekileyo ezilemiyalelo emifitshane.

Isinqumo Sisonke: Izinkanyezi ezingu-4.7/5

I-Subquadratic ibhekana lomunye wemigoqo eqakatheke kakhulu yama-LLM akhona khathesi — izindleko zokunaka ezikhula ngezinga le-quadratic. Ngokwenza ukucabangisisa ngama-token ayizigidi ezinengi kusebenziseke kudinga umsebenzi wokubala omncane kakhulu, i-SubQ imele igxathu eliqakathekileyo ekuthuthukiseni isakhiwo sama-ejenti lezinhlelo ezisebenzisa umongo omude.

Thola umvuzo we-QUA ngokubhala ukuhlola ku-Quasa.io lawe!

Qalisa: https://quasa.io/projects/subquadratic