Subquadratic Uphononongo kwi-Quasa: I-LLM Yokuqala Enokukhula Okungaphantsi Kokwesikwere

#Quasa #QUA #subquadratic

Subquadratic yinkampani yophando neziseko zophuhliso lwe-AI esebenza phambili kweli candelo, eyakha udidi olutsha lweemodeli ezinkulu zolwimi. Imodeli yayo ephambili, SubQ, yi-LLM yokuqala eyakhiwe kuyilo olusebenzisa ingqalelo ekhetha iindawo ezithile kuphela, enokukhula okungaphantsi ngokupheleleyo kokwesikwere. Ngokungafaniyo nee-Transformers eziqhelekileyo ezikhula ngokwesikwere sobude bomxholo ($O(n^2)$), i-SubQ ikhula ngokomgca othe ngqo ($O(n)$), nto leyo eyenza ukuba ukuqiqa ngomxholo onezigidi ezininzi zeethokheni kusebenziseke yaye kusebenzise izixhobo ngobuchule.

I-SubQ yenzelwe ngokukhethekileyo imisebenzi enomxholo omde. Inokucubungula oovimba bekhowudi bephela, iimbali ezinde zee-arhente, izicelo zokudibanisa ikhowudi eziqokelelwe kangangeenyanga, kunye neemeko ezigcinwa ngokuqhubekayo kumyalelo omnye ngaphandle kokuhla komgangatho. Ngoku le modeli ixhasa iifestile zomxholo ezifikelela kwizigidi ezili-12 zeethokheni kwaye isebenza kakuhle kakhulu kwimilinganiselo yokuvavanya ukufumana ulwazi kumxholo omde neyokuqiqa ngokubanzi. Kwisigidi esi-1 seethokheni, i-SubQ ifuna izixhobo zokubala ezimbalwa kakhulu (kubangwa ukuba zimbalwa nge-64.5× kunezo zifunwa yingqalelo exineneyo) kwaye isebenza ngokukhawuleza kakhulu kuneendlela eziqhelekileyo zengqalelo.

Le nkampani igxile kumgaqo othi “ukusebenzisa izixhobo ngobuchule bubukrelekrele.” Ngokusombulula isithintelo sokukhula ngokwesikwere kwinqanaba loyilo, i-Subquadratic ijolise ekwenzeni isizukulwana esitsha see-arhente ze-AI ezinokusebenza ngokwenene ngomxholo omninzi endaweni yokuxhomekeka ekuwucandeni, ekuwushwankatheleni, okanye kumaqhinga okufumana ulwazi.

I-Subquadratic ifaneleke kakhulu kubaphuhlisi, ii-arhente zokubhala ikhowudi, amaqela amashishini, nabaphandi abafuna ukuqiqa okuthembekileyo ngomxholo omde — ingakumbi ukuhlalutya yonke ikhowudi yeprojekthi, ii-arhente ezisebenza ixesha elide, kunye nemisebenzi entsonkothileyo ebandakanya amaxwebhu amaninzi.

Amanqaku aphambili

• Uyilo lokuqala lwe-LLM olunokukhula okungaphantsi ngokupheleleyo kokwesikwere (ukukhula ngokomgca othe ngqo).
• Umxholo onezigidi ezininzi zeethokheni (ukuya kutsho kwi-12M yeethokheni).
• Ukufumana ulwazi kumxholo omde ngokugqwesileyo kusetyenziswa izixhobo ngobuchule obuphezulu.
• Yenzelwe oovimba abapheleleyo nemeko ye-arhente egcinwa ngokuqhubekayo.
• I-API ehambelana ne-OpenAI kunye nokudityaniswa nee-arhente zokubhala ikhowudi.

• Izinto ezinokufuna ukuqwalaselwa

• Njengemodeli yophando ephambili, inkqubo yayo ebanzi yezixhobo neenkonzo kunye neentlobo zayo ezinobukhulu obahlukeneyo zisakhula.
• Ifaneleke kakhulu ekusetyenzisweni okunomxholo omde kunakwincoko eqhelekileyo enemiyalelo emifutshane.

Isigqibo ngokubanzi: iinkwenkwezi ezi-4.7/5

I-Subquadratic ijongene nomnye weyona mida ibalulekileyo yee-LLM zangoku — iindleko zengqalelo ezikhula ngokwesikwere. Ngokubonelela ngokuqiqa okusebenzisekayo ngezigidi ezininzi zeethokheni kusetyenziswa izixhobo zokubala ezimbalwa kakhulu, i-SubQ imele inkqubela ebalulekileyo kuyilo lwee-arhente nezicelo ezisebenzisa umxholo omde.

Fumana umvuzo we-QUA ngokubhala uphononongo kwi-Quasa.io nawe!

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