Fibea: Digital Times, Gu Bon-gyu kyɛwfo
“AI startup Klɛbi akɔ GAIA atifi 2.5% mu… ɛkyerɛ ahotoso a wɔasɔ ahwɛ” nsɛm no nyinaa a wɔafa aka
Da a wɔyii no adi: 2026-04-08
Link: https://n.news.naver.com/article/029/0003020511?sid=105
AI startup ‘Klɛbi (Clevi)’ a ɛwɔ ɔman no mu no aka sɛ, bere a wɔde from scratch ayɛ wɔn ankasa model ne wɔn ankasa agent solution no, wɔayeyɛ ɔman no mu deɛ edi kan a wɔakɔ GAIA benchmark no mu, na wɔakɔ model 3,090 a wɔakyerɛw wɔn din nyinaa mu atifi 2.5%.
Sɛnea nnwuma no mu animdefo kyerɛ no, GAIA a Meta AI na ɛyɛɛ ne nhyehyɛe na Hugging Face na ɛhwɛ so no, ɛnhwehwɛ sɛ AI nya ‘ahoɔden wɔ ade baako pɛ mu’; mmom ɛhwɛ sɛ ebia ɛwɔ ‘tumi sɛ ɛde ahoɔden ahorow bɛka abom de asiesie ɔhaw ankasa’ anaa. Ɛsɛ sɛ ɛhwehwɛ nsɛm wɔ wɛbsaet so, ɛkenkan pon a ɛwɔ PDF mu, ɛyɛ mfonini mu nhwehwɛmu, ɛde code yɛ akontaabu, na ɛka nea ɛde ba no bom de mmuae baako ma. Ɛyɛ nhyehyɛe a ɛwɔ nsɛmmisa 301 a wɔmmɔɔ no dawuru, denam den a ɛwɔ gyinapɛn abiɛsa so, na ɛfa nnyinasosɛm a ɛne sɛ wɔremmɔ mmuae no dawuru ho; enti ɛnyɛ yiye sɛ wɔde wɔn ho bɛhyɛ nsɛmmisa no mu dodo anaasɛ wɔbɛyɛ nnaadaa.
Nea ɛhia sɛ obi nya akontaabu a ɛkorɔn wɔ sɔhwɛ yi mu no yɛ mmienu. Edi kan ne kasa model no a ɛyɛ fapem no mu adwene a ɛtumi susuw nneɛma ho, na nea ɛto so mmienu ne agent solution a ɛde model no bata wiase ankasa mu nnwinnade ho ma ɛtumi yɛ adwuma no ankasa. Sɛ model no yɛ papa sɛnea ɛte biara, sɛ agent no yɛ mmerɛw a, ɛrentumi nni Level 3 so; na sɛ agent no yɛ pɛ sɛnea ɛte biara, sɛ model no mu adwene a ɛtumi susuw nneɛma ho no sua a, mmuae a ɛnteɛ bɛtrɛw afi gyinabea a ɛwɔ mfinimfini no.
Sɛ yɛhwɛ GAIA leaderboard no nhyehyɛe a ɛwɔ hɔ mprempren a, nhwɛso a ɛyɛ anigye bi pue. Agent dodow no ara a wɔwɔ atifi hɔ no agye ‘multi-model mix’ kwan no ato mu, a wɔde big tech models ahorow te sɛ GPT, Claude, Gemini ne afoforo ka bom. Ɛyɛ kwan a ɛfata, efisɛ ɛka model biara mu ahoɔden bom na ɛbɔ ne ho ban fi akontaabu a ɛbɛso atew esiane nsɛm a ɛyera ne nneɛma foforo nti.
Nanso Klɛbi anka wɔn ho. cip-5.5-agent (agentic AI) a wɔde from scratch ayɛ no yɛ adwuma a ɛyɛ den ho nhyehyɛe, sɛnea wɔde bedi dwuma na wɔahwɛ sɛ ɛyɛ nokware no ankasa, na cip-5.5-mm (high-performance general-purpose multimodal) no te nneɛma ahorow te sɛ nne, mfonini ne video fael ahorow ase na ɛsusuw ho. Wɔyɛɛ model mmienu yi nyinaa wɔ Klɛbi mu sɛ wɔn ankasa nneɛma, na wɔankɔ external LLM API so koraa.
Na wɔde stack a wɔn ankasa yɛe yi maa ɛnan agents gyinaa GAIA so, na wɔn nyinaa nyae bɔɔ 70+ marks. Efi nea ɛkorɔn paa a ɛyɛ 79.07% kosi 70.76% no, ɛdaa adi sɛ ɛnyɛ single result mu ahotɔ nko ara, na mmom stack no nyinaa yɛ den na ɛyɛ pɛpɛɛpɛ.
Sɛnea adwumakuw no kae no, akontaahyɛde foforo wɔ aban mu score a ɛyɛ 79.07% no akyi. Sɛnea Clevi mu nhwehwɛmu a wɔyɛe akyi kyerɛ no, wohuu sɛ wɔ nsɛmmisa 301 no mu, nsɛmmisa pii wɔ hɔ a mprempren wonnya mmuae no ho adanse wɔ wɛbsaet a obiara betumi ahu so. Bere a wɔsan susuw nsɛmmisa a mmuae no ho adanse da so wɔ wɛbsaet so no nkutoo ho no, Clevi nyaa mmuae a ɛyɛ nokware wɔ 98% anaa nea ɛboro saa mu. Eyi yɛ dodow a aboro nnipa dodow no average a ɛyɛ 92% dedaw.
Nanso, sɛ wɔde nnwuma foforo mu general-purpose models a ɛwɔ ahoɔden fraa mu a, anka wobetumi ama aban score no akɔ soro koraa. Nanso Clevi amfa saa nhyehyɛe no sɛ́ nea wɔbɛfa so. Nea enti ne sɛ, saa benchmark yi botae nyɛ sɛ wɔbɛpere sɛ hena na obenya score a ɛkorɔn paa, na mmom sɛ wɔbɛhwɛ sɛ wɔn ankasa model a wɔyɛe from scratch no duu gyinapɛn a ɛbɛma atumi ne afoforo asi akan wɔ wiase nyinaa anaa.
Nkyerɛw a ɛka sɛ GAIA leaderboard no so no wɔ wɔn din no kyerɛ sɛ wɔanya ahotoso a wɔde ɔmanfoɔ a wɔnyɛ wɔn adwuma no yɛe na wɔagye atom. Eyi yɛ adanse a ɛyɛ nokware a wobetumi de adi dwuma wɔ sikasɛm a wɔrehwehwɛ, amannɔne gua so kɔ, ne B2B aguadi mu nyinaa.
Clevi panyin bi kae sɛ, “Aban mu mmuae a ɛnteɛ 63 no mu pii fi output format a ɛnhyia, mfonini ne nsɛm a wɔde ani hu mu nkyerɛase mfomso, ne nsɛmmisa a ɛrentumi mma mmuae efisɛ nsɛm no ayera mu. Ɛnyɛ ɔhaw a ɛwɔ logical reasoning mu titiriw, na mmom ɛyɛ post-processing pɛpɛɛpɛyɛ ne sɛnea wobenya nsɛm a ɛwɔ abɔnten ho. Esiane sɛ eyi yɛ yɛn ankasa model nti, Clevi ankasa betumi ama ayɛ papa. Eyi ne mfaso a ɛwɔ yɛn ankasa model a yɛyɛe from scratch no nhyehyɛe mu. Clevi adeyɛ yi ma پو? AI adwuma wɔ Korea no bisa sɛ, ‘Ɛbɛyɛ da bɛn ansa na yɛagyae sɛ yɛde “adwene a yɛde fɛm” yɛ adwuma?’ Clevi apaw kwan a ɛyɛ den kɛse a wɔfrɛ no from scratch, na wɔpɛ sɛ wɔde wɔn mmuae no si so wɔ wiase no anim,” ɔde kaa ho.
