From-scratch Foundation Model CLEVI AI(ivy)

“K kampuni entono eyasobola etya okukola enkola y’olulimi ey’obunene obusinga obw’omutindo mu nsi yonna?”

K kampuni entono, yasobola etya okukola enkola y’olulimi ey’obunene obusukkiridde?

Kino kye kibuuza ekisinga okubuzibwa okuva CLEVI lwe yatangaza Foundation model ey’obunene bwa 1.4 trillion parameters.

Obutafaananako makampuni mangi agakola fine-tuning ennyangu ku nkola ez’ensibuko eziggule, CLEVI yakoze buli mutendera mu bwannannyini, okuva ku kukungaanya data, okukola dizayini ya model, okuyiga okw’ensasaanya ku kigero kinene, okutuuka ku kukakasa omutindo.

Wansi wano tunnyonnyola mu bujjuvu ekyama kino n’amaanyi agakiviiramu okuvuganya.

1. K kampuni entono esobola okukola enkola y’olulimi ey’obunene obusukkiridde?

Mu July 2025, CLEVI bwe yatangaza enkola y’olulimi ey’obunene obusukkiridde gye yakola yokka ku katale, bakasitoma (abakozesa) bangi baasamaalirira nga balina okusanyuka n’okubuusabuusa. “Ddala yakolebwa yokka?” “Si model ya open source eyakyusiddwaako katono?”

Mu butuufu, amakampuni mangi mu ggwanga n’ebweru gakoma ku kukola transfer learning (fine-tuning) ennyangu ku models za open source, ne galangirira nti zino models zaago zokka.

Naye CLEVI yayanjula **‘From-scratch Foundation Model’**.

Mu ngeri endala, yakola n’eteeka mu nkola buli mutendera mu bwannannyini, okuva ku kukungaanya data, okukola dizayini ya model, okuyiga okw’ensasaanya ku kigero kinene, okutuuka ku kukakasa omutindo.

2. Ensonga lwaki okukola enkola y’olulimi ey’obunene obusukkiridde kizibu

Okusobozesa From-Scratch Foundation Model okuyiga, ebintu bino byonna birina okuyigibwa.

Kyetaagisa dataset ennene ey’omutindo ogwa waggulu

• Obw’enjawulo: okukakasa obw’enjawulo mu nnimi, domains, n’ebika (ebiwandiiko, ebifaananyi n’ebirala)
• Okulongoosa: okuggyawo amaloboozi, okulabirira omutindo, n’okusekula data eziddiddwaamu/ez’obulabe
• Layisinsi: okutegeera obulungi eddembe ly’obuyiiya n’eddembe ly’okukozesa data

Ebikozesebwa mu kompyuta eby’obunene obusukkiridde

• Ekibinja kya GPU/TPU eky’omutindo ogwa waggulu: tekinologiya asobozesa okuyiga okw’ensasaanya okw’obungi ng’agatta ebyuma okuva ku nkumi okutuuka ku makumi g’enkumi aga nodes
• Framework y’okuyiga okw’ensasaanya ennungamu: DeepSpeed, Megatron-LM, Ray n’ebirala
• Storage/network: okuyingiza n’okufulumya data ennene, n’embeera ya network ey’amangu

Okukola dizayini ya model architecture

• Okulaga ensengeka ez’omulembe: Transformer, MoE(Mixture of Experts), multimodal n’ebirala
• Scalability: Okulowooza ku bugazi bw’omuwendo gwa parameters, layers, obuwanvu bw’ebiyingizibwa n’ebirala
• Obukugu: Okulongoosa obwangu bw’okutendeka/okuteebereza n’okukozesa memory

Enkola n’obukodyo bw’okutendeka

• Ekigendererwa ky’okutendeka nga bukyali: language model (okugeza, next token prediction), multimodal (okugeza, contrastive learning) n’ebirala
• Obukodyo bw’okulongoosa: mixed precision, gradient accumulation, learning rate schedule n’ebirala
• Normalization/okunyweza: dropout, layer norm, weight decay n’ebirala

Enkola y’okupima n’okukakasa

• Benchmark sets: Okukozesa datasets eza mutindo (Benchmarks)
• Okupima okw’omunda: Okupima nga kwesigamiziddwa ku mbeera z’okukozesa ez’amazima
• Okulondoola obutasalako: Okukebera omutindo, obukyamu n’obukuumi mu kiseera n’oluvannyuma lw’okutendeka

Obusobozi bw’abakozi n’ekitongole

• Abanoonyereza mu AI/ML: Okukola dizayini ya model n’okukola algorithms
• Data engineers: Okukung’aanya, okulongoosa n’okuddukanya data
• Infrastructure engineers: Enkola ezikola mu ngeri eyasaasaanizibwa, n’okuddukanya cloud/on-premises
• Project managers: Okuddukanya enteekateeka, embalirira n’omutindo

Okulowooza ku mpisa, amateeka n’obukuumi

• Empisa za AI: Obukyamu, obukuumi, obwerufu n’obuvunaanyizibwa
• Okugoberera amateeka: Ebikwata ku bikwata ku muntu, copyright n’obwannannyini bwa data
• Obukuumi: Okwewala okubikkula data/model n’okufuga okuyingira

Mu kiseera kino, omuwendo gw’abanoonyereza mu AI mu nsi yonna guli ku bantu nga 2,000, era n’abasinga obungi ku bo bakunganidde mu kkampuni ennene ez’obwannannyini bwa tekinologiya nga META, Google ne XAI. Mu Korea, ttiimu ezisobola okukola Foundation model nga zetegekera okuva ku kukung’aanya data okutuuka ku bikozesebwa eby’okutendekera eby’obunene obungi, ntono nnyo.

3. Okukola model y’olulimi ennene ennyo n’abakozi abatono.

Omukulu Lee Hwan-ho yatandikawo kkampuni (Clevi) emyaka esatu egiyise ng’alina obumanyirivu bw’okusoma deep learning ne machine learning okumala emyaka egisukka mu kkumi, era ng’alina ekigendererwa eky’okukola “AI esinga obulungi mu nsi yonna.”

Yategeka ye kennyini enkola zonna ezetaagisa okukola Foundation Model, omuli okuzimba data pipeline ya kkampuni, okukola dizayini y’ebikozesebwa bya deep learning ebisaasaaniziddwa mu kigero ekinene, okukola dizayini y’ensengeka ya model n’okukakasa omutindo.

Oluvannyuma lw’okutendeka models emirundi mingi, kati balina Clevi-5-x model esobola okuvuganya ne models ezisinga obulungi mu nsi yonna.

Okusobola okutendeka obulungi ebika by’enkola z’olulimi ennene ennyo, kyetaagisa okugatta GPU okuva ku bikumi okutuuka ku nkumi mu kiraasi, n’ebikozesebwa ebisobola okukola okubala okutuuka ku bukadde bwa bukadde buli sikonda mu kiseera kye kimu. Mu nkola eno, tekinologiya akola kinene mu kukendeeza obudde bw’okutambuza n’okukwataganya okubala mu mbeera y’okusaasaanya ennene. Okutuusa kati mu Korea, amakampuni agasinga obungi tegasobodde kutendeka bulungi bikola byabwe olw’obutaba na “algorithms ez’okufuga obutuufu”. CEO wa Clevi Lee Hwan-ho yasobola okutendeka model y’olulimi ennene ennyo eyasooka mu Korea, ng’erina parameters obukadde bwa bukadde 1.4 trillion (1.4 Trillion), ng’akola algorithm ye ey’obwannannyini ey’okufuga okubala okusaasaanidde.

4. Ensonga lwaki kyali kisoboka okukola ebika eby’enjawulo n’abakugu abatono mu kukola pulogulaamu

N’ebitongole ebisinga obungi eby’obwetebiiki obunene birina tekinologiya ow’okufuga ebikozesebwa ebinene n’okulongoosa data.

Okubifuga, kyetaagisa abakugu mu kunoonyereza okuva ku bikumi okutuuka ku nkumi.

Naye Clevi ekola era erina ebika eby’enjawulo ng’ekozesa abakugu abatono mu kunoonyereza.

Kino kyasoboka kitya?

Clevi ekola era n’erina ebika eby’enjawulo ng’eyita mu tekinologiya ya Teacher-Student Model(Knowledge Distilation), ng’ekozesa abakugu abatono.

Kati ka twekenneenye tekinologiya ya Teacher-Student.

Teacher-Student(okusengejja okumanya)

Tekinologiya ya Teacher-Student, mu ngeri ennyangu, ye tekinologiya ekkiriza okukola amangu ebika eby’enjawulo n’abantu abatono, ng’ekozesa enkola y’olulimi ennene ennyo (Teacher Model) okwekenneenya n’okuwabula model y’omwana (Student Model), bwe kityo n’esikira abakugu abawera amakhulu mu kunoonyereza.

• Mother Model(enkola ennene ennyo) yeekenneenya era n’ewa okuwabula ku bika eby’enjawulo eby’ebitongole, n’esikira abakugu abawera amakhulu mu kunoonyereza.
• Bw’otendeka ng’okozesa enkola eno okuyita mu Clevi-x-platform, osobola okutendeka n’okusaasaanya ebika eby’omutindo ogwa waggulu nga Reasoning, VLM, Physical AI, ne Coding Agent mu nnaku 10–15 zokka.

Clevi Coding Agent

Clevi Phsycal AI Learning Platform

VLM-Vision Language Model

Agent w’eby’obukuumi

5. Enjawulo mu tekinologiya n’omutindo gwa Clevi-x-platform

Obusobozi n’okugaziya model

• Obukulu bw’okutegeera (CoT/Reasoning): cip-5-x erina enkola y’okukola okuwerekeragana kw’ensonga n’okulaga obujulizi nga kimu ku nfiro zayo ez’okukola, era eraga obusobozi obwesigika obw’okubala n’okwekenneenya mu kugonjoola ebizibu bya sayansi, okubala n’obwannakyemalirizo. Nga kwesigamiziddwa ku bipimo by’omunda, ekolebwa era n’eddukanyizibwa ng’erina ekigendererwa ky’okutuuka ku busobozi obwenkana oba obusinga ku ddaala lya GPT-5, ng’okuddamu okukola n’okukakasa ebivaamu mu mbeera ya prompt y’emu bikuumibwa ng’ebipimo ebikulu eby’omutindo.

• Obusobozi obujjuvu obw’emifaliso mingi: kisoboka okukola n’okugatta mu pulatifomu emu VLM egendereddwa ku kigendererwa (cip-5-vision), omulembe gw’okukola ebirimu eby’emifaliso mingi mu bungi (ivy-4-mm), ne ivy-3-text etali nzito ekolera ku kyuma, okufuna okugatta okusookerwako okusinziira ku ngeri y’omulimu (okunoonya/okugabanya/okufunza vs. okutegeera/okunnyonnyola/okukola).

Obusobozi bw’okukozesebwa mu bitongole

• Obukuumi n’obwetaavu bw’okubeerawo ku byuma by’omukozesa: eddukanyizibwa mu mukutu ogw’enkugira mu mitendera gyonna (okuyingiza-okuyigiriza-obuweereza-okutereka obutereka), nteekateeka ya HA, okugaziya mu bitundu, ne SVCE (empeereza y’okukolera mu mbeera eyawula obukuumi) ebyawula era ne bikuuma data n’ebipimo bya model.
• Okutandika n’okugaziya amangu: Ivy Chat (emboozi/ba-agenti ezikola emirimu ezirina emifaliso mingi), API Platform (SaaS ey’omutindo gw’ensi yonna) biyamba okukozesa n’okuddukanya amangu.

Obw’enjawulo bwa Physical AI

• Nga tugatta simulation eyesigamiziddwa ku NVIDIA Isaac ne reinforcement learning eyesigamiziddwa ku nkulaakulana (Evolutionary RL), n’okuyigiriza okw’enkola emu ku data ezikoleddwa mu parallel ne Sim2Real transfer, twongedde obulungi bw’okuyigiriza n’obusobozi bw’okwetegeka embeera y’omu kifo. Obusobozi obukwata ku buli mutendera okuva ku digital ne robotic learning okutuuka ku deployment bulina abalala abatono nnyo abasobola okubuwagira.

Omutindo gw’obuweereza n’enkola y’okufuga

• Ennambika z’enkyusa za model/prompt/ebikozesebwa, evaluation suite (amanyi mu nnimi z’eggwanga n’Olungereza, emirimu egy’amakolero ag’enjawulo, ebipimo by’okuloota n’obukuumi), enkyukakyuka z’enkola (SLO/SLA), n’okulondoola enkola (obudde bw’okuddamu, omugerageranyo gw’obuwanguzi, TCO) bifugibwa mu nkola ya pulatifomu emu.

Kaakano mu Korea ey’Eddembe waliwo amakampuni agawera mu 300 agakola empeereza za AI, naye

CLEVI lyokka lisobola okuwa obuweereza bwa on-premises ne cloud mu kiseera kye kimu nga liyita mu foundation model yaayo ey’obusobozi obw’amaanyi era ey’enjawulo.

6. Engeri y’okukozesaamu language model mu makolero

Olw’okuba okuleeta AI kwetaaga ssente nnyingi n’okukakasa okw’amaanyi, kikulu nnyo okugeraageranya n’okugezesa ku bw’obwo oba solution eyo ddala eyamba kampuni.

• CLEVI tegikoma ku kukola by model zokka, wabula etuwa eby’okugonjoola ebya AI ebisobola okukozesebwa mu mirimu gy’amakolero egy’amazima.
• Okugeza, etuukiriza emikutu gya enterprise nga Ivy Chat Platform, Clevi API Platform, okukyusa okusinziira ku makolero, n’okukola ku byetaago by’obukuumi.
• Erina obusobozi mu by’ekikugu obutatera kusangibwa mu ggwanga oba ebweru, nga mulimu Physical AI, robotics, n’okukola ku data ya multimodal.

CLEVI etuwa eby’okugonjoola ebya AI ebikola ku mazima, ng’etwala AI ng’“ekintu ekikozesebwa” so si “ekigendererwa”, era nga biyamba okwongera obukugu mu mirimu n’okuleeta obuyiiya mu makolero. Weetegereze obwawufu bwa model ya Foundation eyakolebwa From-scratch ggwe kennyini.

Okubuuza n’okusaba demo: [email protected]

Copyright© 2025 Clevi Inc. Amal अधिकार gyonna gakuumiddwa.