Ụdị Ntọala e Si Malite na Ọ́dị̀mà CLEVI AI(ivy)

“Kedu ka obere ụlọ ọrụ si nwee ike ịmepụta nnukwu ụdị asụsụ AI nke kacha elu n’ụwa?”

Kedu ka mmalite azụmahịa pere mpe si mepụta nnukwu ụdị asụsụ?

Nke a bụ ajụjụ a jụrụ CLEVI ọtụtụ oge kemgbe ọ kpughere Ụdị Ntọala ya nke ruru páraamita tiriliọnụ 1.4.

N’adịghị ka ọtụtụ ụlọ ọrụ na-eme naanị nkwalite nta nke ụdị ndị mepere emepe, CLEVI rụrụ usoro niile n’onwe ya, site na ịnakọta data, imepụta nhazi ụdị, mmụta kesara n’ogo buru ibu, ruo na nkwenye ogo.

N’okpuru ebe a, anyị na-akọwa ihe nzuzo na ike asọmpi ya n’ụzọ zuru ezu.

1. Obere ụlọ ọrụ nwere ike ịmepụta nnukwu ụdị asụsụ?

Na Julaị 2025, mgbe CLEVI kpughere nnukwu ụdị asụsụ AI nke o mepụtara n’onwe ya n’ahịa, ọtụtụ ndị ahịa (ndị ọrụ) jupụtara n’ihe ijuanya na obi abụọ. “Ọ bụ n’ezie mmepe nke onwe ha?” “Ọ̀ bụ na ha agbanwebeghị naanị ntakịrị ụdị mepere emepe?”

N’ezie, ọtụtụ ụlọ ọrụ n’ime na n’èzí obodo na-eme naanị mmụta nnyefe (fine-tuning) n’ụdị ndị mepere emepe, ma na-akpọsa ha dị ka ụdị nke ha n’onwe ha.

Ma CLEVI kwadoro **‘Ụdị Ntọala e Si Malite na Ọ́dị̀mà’**.

Ya bụ, ọ haziri ma rụọ usoro niile n’onwe ya, site na ịnakọta data, imepụta nhazi ụdị, mmụta kesara n’ogo buru ibu, ruo na nkwenye ogo.

2. Ihe mere mmepe nnukwu ụdị asụsụ ji sie ike

Iji zụọ Ụdị Ntọala e Si Malite na Ọ́dị̀mà, a ga-amụta ihe niile dị n’okpuru.

A chọrọ nnukwu dataset nwere ogo dị elu

• Ụdị dị iche iche: Ịhụ na e nwere ụdị asụsụ, ngalaba, na usoro dị iche iche (ederede, foto, wdg.)
• Nhicha: Iwepụ mkpọtụ, ijikwa ogo, na nzacha data ugboro ugboro/data na-emerụ ahụ
• Ikikere: Ime ka ikike nwebisiinka na ikike iji data doo anya

Akụrụngwa kọmpụta buru ibu

• Ụyọkọ GPU/TPU nwere arụmọrụ dị elu: Teknụzụ na-ejikọta akụrụngwa ọtụtụ puku ruo iri puku ọnụ iji kwado mmụta kesara nke data buru ibu
• Usoro nhazi mmụta kesara nke ọma: DeepSpeed, Megatron-LM, Ray wdg.
• Nchekwa/Netwọk: Ntinye na mwepụta data buru ibu na gburugburu netwọk ngwa ngwa

Nhazi ihe owuwu ụdị

• Imelite ụkpụrụ ọhụrụ: Transformer, MoE(Mixture of Experts), ọtụtụ ụdị mgbasa ozi na ndị ọzọ
• Scalability: ịtụle ike ịgbasa ọnụ ọgụgụ parameters, layers, ogologo ntinye na ndị ọzọ
• Ịrụ ọrụ nke ọma: imezi ọsọ ọzụzụ/inference na ojiji ebe nchekwa

Atụmatụ na algọridim ọzụzụ

• Ebumnuche pretraining: language model (dịka next token prediction), multimodal (dịka contrastive learning) na ndị ọzọ
• Usoro optimization: mixed precision, gradient accumulation, learning rate schedule na ndị ọzọ
• Normalization/stabilization: dropout, layer norm, weight decay na ndị ọzọ

Usoro nyocha na nkwenye

• ชุด benchmarks: iji ọkọlọtọ datasets (Benchmarks)
• Nyocha nke ime: nyocha dabere na ọnọdụ ojiji n'ezie
• Nleba anya na-aga n'ihu: ịlele ịdị mma, ele mmadụ anya n’ihu na nchekwa n’oge na mgbe ọzụzụ gasịrị

Ikike ndị ọrụ na nzukọ

• Ndị nyocha AI/ML: imepụta model na mmepe algọridim
• Data engineers: ịnakọta, ihicha na ijikwa data
• Infrastructure engineers: sistemụ kesara na ijikwa igwe ojii/on-premises
• Project managers: ijikwa nhazi oge, mmefu ego na ịdị mma

Ntụle ụkpụrụ omume, iwu na nchekwa

• Ụkpụrụ omume AI: ele mmadụ anya n’ihu, nchekwa, nghọta na ibu ọrụ
• Ịgbaso iwu: ozi nkeonwe, ikike nwebisiinka na ikike data
• Nchekwa: igbochi mwepụ data/model na ịchịkwa ohere ịbanye

Ugbu a, ọnụ ọgụgụ ndị ọrụ nyocha AI n’ụwa niile dị ihe dị ka mmadụ 2,000, ọtụtụ n’ime ha nọkwa n’ụlọ ọrụ teknụzụ ukwu dịka META, Google na XAI. Otu ndị nwere ike ịhazi ma wuo Foundation Model n’ime obodo, site na nchịkọta data ruo na nnukwu akụrụngwa ọzụzụ, dị nnọọ ole na ole.

3. Iji obere ndị ọrụ wuo nnukwu asụsụ model.

Dabere na ahụmịhe karịrị afọ 10 n’ime nyocha deep learning na machine learning, CEO Lee Hwan-ho guzobere ụlọ ọrụ ahụ (Clevi) afọ 3 gara aga site n’ebumnuche nke “imepụta AI kacha mma n’ụwa.”

Site n’ichepụta n’onwe ya usoro data pipeline nke ụlọ ọrụ, imepụta nnukwu akụrụngwa deep learning kesara, imepụta model architecture na ịkwado ịdị mma, o wulitere ikike niile achọrọ maka mmepe Foundation Model.

Mgbe ọ zụrụ model ahụ ọtụtụ ugboro, ugbu a o nwere Clevi-5-x model nke nwere ike ịsọ mpi na model ndị kacha mma n’ụwa.

Iji zụọ nnukwu ụdị asụsụ nke ọma, a chọrọ akụrụngwa jikọtara ọtụtụ narị ruo ọtụtụ puku GPU n’otu klasta, nke nwere ike ịhazi ọtụtụ iri puku puku ijeri ọrụ mgbakọ n’otu sekọnd n’otu oge. N’usoro a, teknụzụ na-eme ka mmekọrịta mgbakọ na nkwukọrịta na gburugburu nkesa buru ibu dịrị ntakịrị bụ isi ihe dị mkpa. Ruo ugbu a, n’ime obodo, ọtụtụ ụlọ ọrụ enweghị ike ịzụ ụdị ha nke ọma n’ihi na ha enweghị “algọridim njikwa ziri ezi”. Site n’ịzụlite algọridim njikwa mgbakọ nkesa pụrụ iche a n’onwe ya, Onye isi CLEVI Lee Hwan-ho ghọrọ onye mbụ n’ime obodo ịga nke ọma n’ịzụ nnukwu ụdị asụsụ nwere nha paramita 1.4 trillion (1.4 Trillion).

4. Ihe mere o ji kwe omume iji ọnụ ọgụgụ ndị mmepe pere mpe zụlite ụdị dị iche iche

Ọtụtụ ụlọ ọrụ big tech nwekwara teknụzụ njikwa akụrụngwa buru ibu na nhicha data.

Iji jikwaa nke a, a chọrọ ọtụtụ narị ruo ọtụtụ puku ndị nyocha.

Ma CLEVI na-eji obere ndị nyocha zụlite ma nwee ụdị dị iche iche.

Olee otú nke a si kwe omume?

CLEVI na-eji teknụzụ Teacher-Student Model(Knowledge Distilation) zụlite ma nwee ụdị dị iche iche site n’iji obere ndị ọrụ.

Ka anyị leba anya na teknụzụ Teacher-Student.

Teacher-Student(ịgbasa ihe ọmụma)

Teknụzụ Teacher-Student, n’ikwu ya n’ụzọ dị mfe, bụ teknụzụ na-eji nnukwu ụdị asụsụ (Teacher Model) enyocha ma nye ụdị nwa (Student Model) nzaghachi, si otú a dochie ọtụtụ narị ndị nyocha ma jiri obere ndị ọrụ zụlite ụdị dị iche iche ngwa ngwa.

• Mother Model(nnukwu ụdị) na-enyocha ma na-enye ụdị ndị dị n’akụkụ dị iche iche nzaghachi, si otú a dochie ọtụtụ narị ndị nyocha.
• Ọ bụrụ na a na-eme ọzụzụ n’ụzọ a site na Clevi-x-platform, enwere ike ịzụ ma kesaa ụdị arụmọrụ dị elu dịka Reasoning, VLM, Physical AI, Coding Agent na naanị ụbọchị 10–15.

Clevi Coding Agent

Clevi Phsycal AI Learning Platform

VLM-Vision Language Model

Onye nnọchiteanya nchekwa

5. Ihe dị iche iche nke teknụzụ na ogo Clevi-x-platform

Arụmọrụ na mgbasawanye nke ụdị

• Ọkachamara na ntụgharị uche (CoT/Reasoning): cip-5-x gụnyere mmepe echiche ezi uche dị na usoro na igosipụta ihe akaebe n’ime nkà ihe ọmụma nhazi ya, na-egosikwa ikike ịgbakọ na ịkọwa nke ntụkwasị obi dị elu n’idozi nsogbu sayensị, mgbakọ na mwepụ, na injinia. Dabere na benchmark ime ụlọ, a na-ahazi ma na-arụ ya iji nweta ikike arụmọrụ ruru ma ọ bụ karịa ọkwa GPT-5, ebe a na-elegharị anya na nkwughachi na ikike nkwenye n’okpuru otu ọnọdụ prompt dị ka ihe nlele dị mkpa nke ogo.

• Ụdị multimodal zuru oke: site n’ịmepụta na ijikọta VLM dabere n’ebumnuche (cip-5-vision), ụdị nhazi multimodal nke nnukwu ikike (ivy-4-mm), na ụdị ederede dị fechaa maka ngwaọrụ (ivy-3-text) n’otu ikpo okwu, ọ ga-ekwe omume ịhọrọ ngwakọta kachasị mma dabere na njirimara ọrụ ahụ (ịchọ/ịkewa/nchịkọta vs. ntụgharị uche/nkọwa/ime ihe).

Ikike itinye n’ọrụ ụlọ ọrụ

• Nchekwa na nnweta n’ime ụlọ ọrụ: na-arụ ọrụ n’ime netwọk mechiri emechi n’akụkụ niile (ntinye-ọzụzụ-ọrụ-ndabere), yana nhazi HA, mgbasawanye nwere modul, na SVCE (gbara gburugburu ogbugbu nchekwa kewapụrụ iche) iji kewaa ma chebe data na parameters ụdị.
• Mbido na mgbasawanye ngwa ngwa: Ivy Chat (mkparịta ụka/agent multimodal maka ọrụ), na API Platform (SaaS ọkọlọtọ zuru ụwa ọnụ) na-akwado itinye n’ọrụ na ịrụ ọrụ ngwa ngwa.

Ọdịiche nke Physical AI

• Ejikọtala simulation dabere na NVIDIA Isaac na mmụta nkwado dabere na evolushọn (Evolutionary RL), ma kwalite arụmọrụ mmụta na ikike ime mgbanwe n’ebe ọrụ site na nnyefe Sim2Real na mmụta data sịntetik n’otu oge. Ebe ọ na-ekpuchi ihe niile site na mmụta dijitalụ na robotics ruo na nnyefe, ndị ọzọ dochie ya dị ole na ole.

Ogo ọrụ na ọchịchị

• A na-achịkwa njikwa ụdị nke model/prompt/ngwaọrụ, evaluation suite (ihe ọmụma n’asụsụ obodo na Bekee, ọrụ ụlọ ọrụ dị iche iche, ihe nlele hallucination na nchekwa), njikwa mgbanwe (SLO/SLA), na nleba anya ọrụ (latency/ọnụọgụ ihe ịga nke ọma/TCO) site n’usoro otu ikpo okwu.

Ugbu a, e nwere ihe karịrị ụlọ ọrụ ọrụ ụdị AI 300 n’obodo a, mana

CLEVI bụ naanị ebe nwere ike inye ọrụ on-premises na ọrụ igwe ojii n’otu oge site na foundation model nke ya nwere arụmọrụ dị elu.

6. Ụzọ e si eji ụdị asụsụ eme ihe n’ụlọ ọrụ

N’ihi na iwebata AI chọrọ itinye ego buru ibu na nkwenye siri ike, ọ dị mkpa iji aka gị tụnyere ma nwalee ma ihe ngwọta ahụ ọ̀ bụ nke ga-enyere ụlọ ọrụ aka n’ezie.

• CLEVI anaghị akwụsị naanị n’ịmepụta ụdị, kama ọ na-enye ngwọta AI ndị a pụrụ itinye n’ọrụ n’ezie n’ebe ọrụ ụlọ ọrụ.
• Dịka ọmụmaatụ, ọ nwere ọwa ụlọ ọrụ zuru oke, gụnyere Ivy Chat Platform, Clevi API Platform, nhazi ahaziri maka ụlọ ọrụ dị iche iche, na imezu ihe ndị nchekwa chọrọ.
• Ọ nwere ikike teknụzụ ndị siri ike, dị ka Physical AI, robotics, na nhazi data multimodal, ndị siri ike ịhụ n’ime na n’èzí obodo a.

CLEVI na-enye ngwọta AI bara uru nke na-ewere AI dị ka ‘ngwaọrụ’ kama ịbụ ‘ebumnuche’, na-enyere aka n’ịkwalite arụmọrụ ọrụ n’ezie na imepụta mgbanwe n’ụlọ ọrụ. Nweta ahụmịhe kpọmkwem banyere ọdịiche nke ezigbo From-scratch Foundation ụdị.

Ajụjụ na arịrịọ ngosi: [email protected]

Copyright© 2025 Clevi Inc. All rights reserved.