Ensibuko: Ezeonshinmun, omuwandiisi Lee Won-ji
Kujuliza ekiwandiiko kyonna eky’omulamwa “CLEVI, okuva ku ntandikwa n’omutindo gwayo eyetongodde 79.07% ku GAIA... wakati wa bitundu 2.5% ebisinga obulungi ku model 3,090”
Olunaku lw’okufulumya: 2026-04-07
Ennyunga: https://www.etnews.com/20260407000203
cip-5.5-agent·cip-5.5-mm omutendera gwa model 5 ogw’obwannannyini, ba agent bonna basusse ku bubonero 70
97%+ ku bituufu bw’otunuulira okubula kw’amawulire, nga kusinga abantu (92%)
CLEVI, kampuni entonotono ekola ku AI, yakozesa model yaayo eyetongodde eyakolebwa okuva ku ntandikwa n’ekola obutuufu bwa 79.07% ku GAIA, ekipimo ky’ensi yonna ekya ba agent ba AI, era n’etuuka mu bitundu 2.5% ebisinga obulungi nga egerageranyiziddwa ku model 3,090 eziwandiikiddwa zonna.
Okusinziira ku bannannyini mulimu, GAIA etwalibwa ng’ekipimo ekiraga obusobozi bwa “ba agent ba AI abakozeseka” mu ngeri entuufu mu katale k’ebipimo bya AI. Ekipimo kino, ekyakolebwa Meta AI, HuggingFace, ne Yunivasite ya Paris-Saclay awamu, kyasooka kuteekebwa mu lujjudde mu Novemba 2023.
Ebikulu bisatu ebyawula GAIA ku bipimo ebirala. Ekisooka, kubanga eby’okuddamu bikwekeddwa, tekisoboka kukola overfitting. Ekyokubiri, kubanga kyetaagisa okutegeera okw’omugatte okw’emirundi mingi n’emikutu mingi, tekisoboka kufuna bubonero bwa waggulu nga okozesa pattern matching yokka. Ekyokusatu, okuyita mu leaderboard entongole eya HuggingFace, model ezisukka mu 3,090 okuva mu nsi yonna zivuganya mu mbeera ze zimu.
Test set erimu ebibuuzo 301 byonna. Level 1 ye bakozesa ekikozesebwa kimu n’okutegeera okwangu, Level 2 ye okugatta ebikozesebwa ebingi n’okutegeera okw’omutindo ogwa wakati, ate Level 3 ye bibuuzo ebisinga obuzibu ebyetaagisa enteekateeka n’okussa mu nkola kwa agent okw’emitendera 5 oba okusingawo. Mu kiseera ekipimo kino lwe kyalangirirwa, model za GPT (nga zirina plugins) zaali ku bitundu nga 15% byokka, naye kati ttiimu eziri waggulu zizituusizza ku bitundu 70–80%.
Mu kino, CLEVI yakiggumiza nti ekintu ekisinga okwekenneenyezebwa mu by’ekikugu ku buwanguzi buno kwe kuba nti tebaakozesa n’akatono API za LLM ez’ebweru nga GPT, Claude, ne Gemini. Ekikulu ku CLEVI kwe kuba nti yakozesa model bbiri ezaakolebwa yokka nayo mu ngeri okuva ku ntandikwa.
cip-5.5-agent ye model ya AI ekola mu ngeri ey’eki-agent, era ekola ng’obwongo obukulu obw’enkola ya ba-agent eteekateeka (planning), n’eteeka mu nkola (execution), n’ekkakasa (verification) emirimu emizibu mu ngeri eyetongodde. cip-5.5-mm ye model ya multimodal ey’amaanyi era ekozesa emirimu mingi, esobola okutegeera n’okukola okuteebereza ku nkola za fayiro ez’enjawulo, omuli eddoboozi, ebifaananyi ne vidiyo. Kubanga ebibuuzo bingi mu GAIA birimu ebintu ebitali biwandiike, nga PDF, ebifaananyi ne fayiro z’amaloboozi, obusobozi buno obwa multimodal bwafuuka ekimu ku bintu ebikulu ebyaleeta obubonero obw’amaanyi.
CLEVI yakozesa emikutu gino ebiri egyayo okukola ensengeka za ba-agent 5 ez’enjawulo ne zeetaba mu GAIA, era zonna zaafuna obubonero obusukka mu 70.
Okusinziira ku açıklama y’etteeka, okwekenneenya okw’oluvannyuma okwakolebwa munda mu CLEVI kwalaga ebivudde mu kwewuunyisa. Mu bibuuzo 301 byonna, ebimu byali tebikyafuna bujulizi bw’eby’okuddamu obuli ku mukutu gwa yintaneeti ogw’olukale. Amawulire agaali gasobola okuzuulibwa ku yintaneeti oba mu yingini z’okunoonya mu biseera eby’edda gaali gasanguddwa oba gakyuusiddwa, ne gatakyasobola kufikkirirwa. Mu kuddamu okwekenneenya nga bakozesa ekkomo ly’amawulire abantu bye basobola okukakasa mu lujjudde kati, omuwendo gw’obutuufu gwa CLEVI gwazuulibwa nga gusukka 98%. Kino kyamaze dda okusukka ku kigero ky’abantu eky’awamu, ekiri 92%.
Omukulembeze wa CLEVI yagamba nti, “CLEVI, nga tetabudde model za bweru, ekutte obubonero obusukka mu 70 mu ba-agent b’onna 5 ng’ekozesa model zaayo zokka eza from scratch n’ebisengekero byayo ebyo AI agent, era n’eyingira mu bitundu 2.5% ebyasinga obulungi ku benchmark ey’olukale. Kirabika nti okulongoosa model zaffe n’ebisengekero bya ba-agent mu biseera eby’omu maaso kijja kusobozesa okwongera okulongoosa n’obubonero obutongole. Ebivudde mu mulimu guno bikulu kubanga byapimibwa mu benchmark ey’olukale ey’ensi yonna ne biraga ‘obwesige obukakasiddwa’; kubanga buwanguzi obwatuukiriziddwa kkampuni y’okukulaakulanya AI ey’omu ggwanga nga yeesigama ku model zaayo eza from scratch; kubanga buva mu stack ya model eyetongodde so si mu kutabula model za bweru; era kubanga, ng’otunuulira okufiirwa kw’amawulire mu bibuuzo ebimu, birina amakulu agasinga ku bubonero bwokka.”