From-scratch Foundation Model CLEVI AI(ivy)

“Kodi kampani yaing’ono yakwanitsa bwanji kupanga chilankhulo cha AI chachikulu kwambiri padziko lonse?”

Kodi chinachitika bwanji kuti pakhale chilankhulo chachikulu kwambiri kuchokera ku kampani yaying’ono yoyambitsa bizinesi?

Ili ndi funso lomwe CLEVI yalandira kwambiri kuyambira pomwe idatulutsa Foundation model yokhala ndi ma parameter 1.4 thililiyoni.

Mosiyana ndi makampani ambiri omwe amangopanga fine-tuning ya ma model a open source, CLEVI idachita pawokha njira zonse kuyambira kusonkhanitsa deta, kupanga kapangidwe ka model, maphunziro akulu ogawidwa, mpaka kutsimikizira ubwino.

Pansipa tikufotokoza mwatsatanetsatane chinsinsi cha kupambana kwake ndi mpikisano wake.

1. Kodi kampani yaing’ono ingapange chilankhulo chachikulu kwambiri?

Mu Julayi 2025, CLEVI itatulutsa pamsika chilankhulo chachikulu kwambiri chomwe idapanga yokha, makasitomala (ogwiritsa ntchito) ambiri anadabwa komanso kukayikira. “Kodi chinapangidwadi ndi kampaniyo yokha?” “Kodi sanangosintha pang’ono model ya open source?”

M’chenicheni, makampani ambiri a m’dziko muno komanso akunja amangogwiritsa ntchito transfer learning (fine-tuning) pa ma model a open source, kenako n’kulengeza kuti ndi ma model awoawo.

Koma CLEVI inanena kuti ndi **‘From-scratch Foundation Model’**.

Izi zikutanthauza kuti idapanga ndikuchita yokha njira zonse kuyambira kusonkhanitsa deta, kupanga kapangidwe ka model, maphunziro akulu ogawidwa, mpaka kutsimikizira ubwino.

2. Chifukwa chake kupanga chilankhulo chachikulu kwambiri n’kovuta

Kuti muphunzitse From-Scratch Foundation Model, muyenera kuphunzira ndikukonzekera zinthu zonse zili pansipa.

Pakufunika dataset yayikulu yokhala ndi ubwino wapamwamba

• Zosiyanasiyana: Kuonetsetsa kuti pali zilankhulo, madera ogwiritsira ntchito, ndi mawonekedwe osiyanasiyana (zolemba, zithunzi, ndi zina)
• Kuyeretsa: Kuchotsa phokoso, kuyang’anira ubwino, komanso kusefa deta yobwerezedwa kapena yovulaza
• Malayisensi: Kufotokoza momveka bwino za ufulu wa kukopera ndi ufulu wogwiritsa ntchito deta

Zomangamanga zazikulu za kompyuta

• Gulu la ma GPU/TPU amphamvu kwambiri: Ukadaulo wophatikiza zida zambirimbiri, kuyambira masauzande mpaka makumi a masauzande a node, kuti uthandize maphunziro akulu ogawidwa
• Framework yothandiza pa maphunziro ogawidwa: DeepSpeed, Megatron-LM, Ray ndi zina
• Kusungira deta/Network: Kulowetsa ndi kutulutsa deta yambiri, komanso malo a network othamanga

Kapangidwe ka model

• Kuwonetsa kapangidwe katsopano: Transformer, MoE(Mixture of Experts), multimodal ndi zina
• Scalability: kuganizira za kukula kwa chiwerengero cha ma parameter, ma layer, kutalika kwa zolowetsa ndi zina
• Kuchita bwino: kukonza liwiro la kuphunzira/kulosera komanso kugwiritsa ntchito memory

Njira ndi ma algorithm ophunzirira

• Cholinga cha pretraining: language model (mwachitsanzo, next token prediction), multimodal (mwachitsanzo, contrastive learning) ndi zina
• Njira zokonzera: mixed precision, gradient accumulation, learning rate schedule ndi zina
• Normalization/kukhazikitsa: dropout, layer norm, weight decay ndi zina

Njira zowunikira ndi kutsimikizira

• Benchmark set: kugwiritsa ntchito ma dataset ovomerezeka (Benchmarks)
• Kuwunika kwamkati: kuwunika kutengera zochitika zenizeni zogwiritsira ntchito
• Kuwunika kosalekeza: kuyang'anira khalidwe, tsankho ndi chitetezo panthawi komanso pambuyo pa kuphunzira

Luso la ogwira ntchito ndi bungwe

• Ofufuza a AI/ML: kapangidwe ka model ndi chitukuko cha ma algorithm
• Data engineer: kusonkhanitsa, kukonza ndi kuyang'anira data
• Infrastructure engineer: distributed systems ndi kasamalidwe ka cloud/on-premises
• Project manager: kasamalidwe ka nthawi, bajeti ndi khalidwe

Zoganizira za ethics, malamulo ndi chitetezo

• AI ethics: tsankho, chitetezo, kuwonekera komanso kuyankha pa udindo
• Kutsatira malamulo: zinsinsi za anthu, copyright ndi ulamuliro wa data
• Chitetezo: kupewa kutayikira kwa data/model ndi kuwongolera mwayi wopeza

Pakali pano, chiwerengero cha ofufuza a AI padziko lonse chili pafupifupi 2,000, ndipo ambiri mwa iwo ali m'makampani akuluakulu a teknoloji monga META, Google ndi XAI. Ku Korea, magulu omwe angapange Foundation model mwa kupanga okha zonse kuyambira kusonkhanitsa data mpaka kumanga large-scale learning infrastructure ndi ochepa kwambiri.

3. Kupanga large language model pogwiritsa ntchito ogwira ntchito ochepa.

Potengera zaka zoposa 10 za kafukufuku wa deep learning ndi machine learning, zaka 3 zapitazo CEO Lee Hwan-ho anayambitsa kampaniyo (Clevi) ndi cholinga choti “apange AI yabwino kwambiri padziko lonse.”

Anapanga yekha njira zonse zofunika popanga Foundation Model, kuphatikizapo kumanga data pipeline ya kampaniyo, kupanga large-scale distributed deep learning infrastructure, kupanga model architecture ndi kutsimikizira khalidwe, motero kampaniyo inapeza luso lofunikira.

Pambuyo pophunzitsa mitundu kangapo, tsopano tili ndi mtundu wa Clevi-5-x womwe ungapikisane ndi mitundu yabwino kwambiri padziko lonse lapansi.

Kuti tiphunzitse bwino kwambiri mitundu ikuluikulu ya zilankhulo, pamafunika zomangamanga zolumikiza ma GPU kuyambira mazana angapo mpaka masauzande kukhala gulu limodzi, zomwe zingathe kukonza mawerengero okwana makumi a quadrillion pa sekondi imodzi nthawi imodzi. Pa ntchitoyi, ukadaulo wochepetsa kulumikizana kwa mawerengero ndi kuchedwa kwa mauthenga m’malo akulu ogawidwa ndi wofunika kwambiri. Mpaka pano, makampani ambiri ku Korea sanathe kuphunzitsa bwino mitundu chifukwa kunalibe “algorithmu yolamulira molondola”. Mtsogoleri wamkulu wa Clevi, Lee Hwan-ho, adapanga yekha algorithmu yapadera yolamulira mawerengero ogawidwa, ndipo anakhala woyamba ku Korea kuphunzitsa bwino mtundu waukulu wa chilankhulo wokhala ndi ma parameter okwana 1.4 thililiyoni (1.4 Trillion).

4. Chifukwa chomwe tinatha kupanga mitundu yosiyanasiyana ndi ofufuza ochepa

Ngakhale makampani ambiri akuluakulu aukadaulo ali ndi ukadaulo wolamulira zomangamanga zazikulu ndi kuyeretsa deta.

Kuti aziyendetsedwa, pamafunika ofufuza kuyambira mazana angapo mpaka masauzande.

Komabe, Clevi imapanga ndi kukhala ndi mitundu yosiyanasiyana pogwiritsa ntchito ofufuza ochepa.

Zinatheka bwanji?

Clevi imagwiritsa ntchito ukadaulo wa Teacher-Student Model(Knowledge Distilation) kupanga ndi kukhala ndi mitundu yosiyanasiyana pogwiritsa ntchito anthu ochepa.

Tsopano tiyeni tiwone za ukadaulo wa Teacher-Student.

Teacher-Student(kuwusitsa chidziwitso)

Ukadaulo wa Teacher-Student, mwachidule, ndi ukadaulo womwe umathandiza kupanga mwachangu mitundu yosiyanasiyana pogwiritsa ntchito anthu ochepa, popeza mtundu waukulu wa chilankhulo (Teacher Model) umawunika ndi kupereka ndemanga ku mtundu wa mwana (Student Model), motero umasinthira ntchito ya ofufuza mazana ambiri.

• Mother Model(mtundu waukulu kwambiri) imawunika ndi kupereka ndemanga pa mitundu ya magawo osiyanasiyana, motero imalowa m’malo mwa ofufuza mazana ambiri.
• Ngati maphunziro achitika kudzera mu Clevi-x-platform pogwiritsa ntchito njira imeneyi, mitundu yapamwamba monga Reasoning, VLM, Physical AI, ndi Coding Agent ingaphunzitsidwe ndi kutumizidwa mkati mwa masiku 10–15.

Clevi Coding Agent

Clevi Phsycal AI Learning Platform

VLM-Modeli ya Masomphenya ndi Chilankhulo

Agenti ya Chitetezo

5. Kusiyana kwa Clevi-x-platform pa luso ndi khalidwe

Kuchita bwino kwa modeli ndi kukulitsa

• Kupambana pa inference (CoT/Reasoning): cip-5-x inapangidwa ndi nzeru yakuti igwiritse ntchito kufotokoza kwa mfundo motsatizana ndi kupereka zifukwa. Izi zimathandiza kuti ikhale ndi luso lodalirika lowerengera ndi kutanthauzira pothetsa mavuto a sayansi, masamu ndi uinjiniya. Ikupangidwa ndi kuyendetsedwa ndi cholinga cha kuchita bwino kofanana kapena kuposa mulingo wa GPT-5 malinga ndi ma benchmark amkati, ndipo kubwerezabwereza ndi kutsimikizika pansi pa mikhalidwe yofanana ya prompt kumayendetsedwa ngati zizindikiro zazikulu za khalidwe.

• Mitundu yonse ya multimodal: VLM yochokera pa cholinga (cip-5-vision), modeli yayikulu yokonza multimodal (ivy-4-mm), ndi modeli yopepuka ya pa chipangizo (ivy-3-text) zimapangidwa ndi kuphatikizidwa pa nsanja imodzi, zomwe zimathandiza kusankha kuphatikiza koyenera malinga ndi mtundu wa ntchito (kusaka/kugawa mwamagulu/kufupikitsa motsutsana ndi kulingalira/kufotokoza/kuchita).

Kugwiritsa ntchito m'mabizinesi akulu

• Chitetezo ndi kupezeka kwa on-premises: Kuyendetsa ntchito mu netiweki yotsekedwa kuyambira koyika, kuphunzitsa, kupereka ntchito mpaka kusunga zosunga zobwezeretsera, kapangidwe ka HA, kukulitsa kwa ma module, ndi SVCE(mlingo wotetezeka wochitira ntchito wopatula) zimateteza deta ndi magawo a modeli mwapadera.
• Kuyambitsa ndi kukulitsa mwachangu: Ivy Chat(kukambirana kwa multimodal ndi agenti kwa ntchito), API Platform(SaaS ya muyezo wapadziko lonse) zimathandiza kugwiritsa ntchito ndi kuyendetsa mwachangu.

Kusiyana kwa Physical AI

• Kuphatikiza simulation yozikidwa pa NVIDIA Isaac ndi reinforcement learning yozikidwa pa chisinthiko (Evolutionary RL), komanso kuphunzira kofanana kwa kusamutsa Sim2Real ndi deta yopangidwa, kwawonjezera luso la kuphunzira ndi kusinthira ku malo enieni. Ochepa ndi njira zina zomwe zimaphimba kuyambira kuphunzira kwa digito ndi ma robotics mpaka kutumiza.

Khalidwe la ntchito ndi ulamuliro

• Kuwongolera mitundu ya modeli/prompt/zida, evaluation suite(chidziwitso cha Chikorea ndi Chingerezi, ntchito za mafakitale osiyanasiyana, zizindikiro za hallucination ndi chitetezo), change management(SLO/SLA), ndi kuyang'anira ntchito(latency, kuchuluka kwa kupambana, TCO) kumayendetsedwa kudzera mu njira imodzi ya nsanja.

Pakali pano, ku Korea kuli makampani opitilira 300 opereka ntchito za ma modeli a AI, koma

Clevi ndiye kampani yokhayo yomwe ingapereke nthawi imodzi ntchito za on-premises ndi za pa cloud kudzera mu modeli yake ya foundation yapamwamba komanso yodziyimira pawokha.

6. Njira zogwiritsira ntchito ma modeli a zilankhulo m'mafakitale

Popeza kuphatikiza AI kumafuna ndalama zambiri komanso kutsimikizira mosamalitsa, ndikofunikira kuyerekezera ndi kuyesa nokha ngati ndi yankho lomwe limathandizadi mabizinesi.

• CLEVI sikungopanga ma model okha, koma imapereka mayankho a AI omwe angagwiritsidwe ntchito m'malo enieni a mafakitale.
• Mwachitsanzo, ili ndi njira zonse za mabizinesi monga Ivy Chat Platform, Clevi API Platform, kusintha malinga ndi mafakitale, komanso kukwaniritsa zofunikira zachitetezo.
• Ili ndi luso laukadaulo losapezeka mosavuta mkati kapena kunja kwa dziko, kuphatikiza Physical AI, robotics, ndi kukonza deta ya multimodal.

CLEVI imapereka mayankho enieni a AI omwe amathandiza kukonza magwiridwe antchito ndi kusintha mafakitale, potenga AI ngati ‘chida’ osati ‘cholinga’. Dziwani nokha kusiyana kwa mtundu wa From-scratch Foundation.

Mafunso ndi pempho la demo: [email protected]

Copyright© 2025 Clevi Inc. Ufulu wonse ndi wosungidwa.