Subquadratic ku-Quasa: Ukubuyekezwa Kwe-LLM Yokuqala Esebenza Ngaphansi Kwezinga Le-Quadratic

#Quasa #QUA #subquadratic

Subquadratic yinkampani yocwaningo nengqalasizinda ye-AI esezingeni eliphambili eyakha uhlobo olusha lwamamodeli olimi amakhulu. Imodeli yayo ehamba phambili, SubQ, iyi-LLM yokuqala eyakhiwe ngesakhiwo sokunaka okukhethiwe esisebenza ngokuphelele ngaphansi kwezinga le-quadratic. Ngokungafani nama-Transformer avamile, izidingo zawo zokubala ezikhula ngokwezinga le-quadratic uma kwanda ubude bomongo ($O(n^2)$), izidingo zokubala ze-SubQ zikhula ngokulingana nobude bomongo ($O(n)$), okwenza ukucabanga ngamathokheni ayizigidi eziningi kusebenziseke futhi konge izinsiza.

I-SubQ yakhelwe ngokukhethekile imisebenzi enomongo omude. Ingacubungula izinqolobane zekhodi eziphelele, imilando emide yama-ejenti, izicelo zokuhlanganisa izinguquko zekhodi eziqoqeke izinyanga, nezimo ezigcinwa ngokuqhubekayo, konke emyalweni owodwa ngaphandle kokwehla kwekhwalithi. Njengamanje imodeli isekela amafasitela omongo afinyelela kumathokheni angu-12 million futhi isebenza kahle kakhulu ezivivinyweni zokubuyisa ulwazi emongweni omude nezokucabanga okuvamile. Kumathokheni angu-1 million, i-SubQ idinga amandla okubala amancane kakhulu (kuthiwa mancane ngokuphindwe ngo-64.5× kunawokunaka okuminyene) futhi isebenza ngokushesha kakhulu kunezindlela zokunaka ezijwayelekile.

Inkampani igxile emgomeni othi “ukusebenza kahle kuwubuhlakani.” Ngokuxazulula umgoqo wezinga le-quadratic esakhiweni uqobo, i-Subquadratic ihlose ukwenza kube nesizukulwane esisha sama-ejenti e-AI akwazi ukusebenza ngempela nomongo omningi kunokuthembela ekuwuhlukaniseni ube yizingcezu, ekufingqeni, noma emaqhingeni okubuyisa ulwazi.

I-Subquadratic ifaneleka kakhulu kubathuthukisi, kuma-ejenti abhala ikhodi, emaqenjini ezinkampani nakubacwaningi abadinga ukucabanga okuthembekile ngomongo omude — ikakhulukazi ekuhlaziyeni yonke ikhodi, kuma-ejenti asebenza isikhathi eside nasemisebenzini eyinkimbinkimbi ehlanganisa imibhalo eminingi.

Amaphuzu Avelele

• Isakhiwo sokuqala se-LLM esisebenza ngokuphelele ngaphansi kwezinga le-quadratic (esinezidingo zokubala ezikhula ngokulingana nobude bomongo).
• Umongo wamathokheni ayizigidi eziningi (kufika kumathokheni angu-12M).
• Ukubuyisa ulwazi emongweni omude okusebenza kahle kakhulu nokonga izinsiza kakhulu.
• Yakhelwe izinqolobane eziphelele nesimo se-ejenti esigcinwa ngokuqhubekayo.
• I-API ehambisana ne-OpenAI nokuhlanganiswa nama-ejenti abhala ikhodi.

• Okungadinga Ukucatshangelwa

• Njengoba iyimodeli yocwaningo esezingeni eliphambili, uhlelo olubanzi lwamathuluzi nezinsiza ezihambisana nayo kanye nezinhlobo zayo ezinobukhulu obuhlukahlukene kusanwetshwa.
• Ifaneleka kakhulu ekusetshenzisweni okunomongo omude kunasezingxoxweni ezivamile ezinemiyalo emifushane.

Isinqumo Sisonke: izinkanyezi ezingu-4.7/5

I-Subquadratic ibhekana nomunye wemikhawulo ebaluleke kakhulu yama-LLM amanje — izindleko zokunaka ezikhula ngokwezinga le-quadratic. Ngokuhlinzeka ngokucabanga okusebenzisekayo ngamathokheni ayizigidi eziningi kusetshenziswa amandla okubala amancane kakhulu, i-SubQ imelela intuthuko ebalulekile esakhiweni sama-ejenti nezinhlelo ezisebenzisa umongo omude.

Nawe zuza umklomelo we-QUA ngokubhala ukubuyekeza ku-Quasa.io!

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