From-scratch Foundation Model CLEVI AI(ivy)

“Me pēhea e taea ai e tētahi kamupene iti te hanga i tētahi tauira reo AI tino nui o te taumata pai rawa atu o te ao?”

Me pēhea te hanga tauira reo tino nui mai i tētahi whakaoho iti?

Koinei te pātai e tino whiwhi nuitia ana mai i te wā i whakapuaki ai a CLEVI i tana tauira Foundation 1.4 trillion ngā tawhā te rahi.

Ahakoa he maha ngā kamupene ka whakatikatika noa i ngā tauira puna tuwhera, nā CLEVI i whakahaere takitahi ngā tukanga katoa, mai i te kohikohi raraunga, te hoahoa tauira, te whakangungu tohatoha nui, tae atu ki te manatoko kounga.

Ka whakaatuhia ngā mea ngaro me ngā painga whakataetae i raro nei.

1. Ka hangaia e tētahi kamupene iti he tauira reo tino nui?

I Hōngongoi 2025, i te wā i whakapuaki ai a CLEVI i tana tauira reo tino nui i whakawhanakehia e ia anō ki te mākete, ka ohooho, ka ruarua hoki ngā kaihoko (kaiwhakamahi) tokomaha. “He whanaketanga ā-roto tonu tēnei?” “Kāore rānei i whakarerekētia paku noa iho tētahi tauira puna tuwhera?”

Inā hoki, he maha ngā kamupene o roto, o tāwāhi hoki ka whakamahi noa i te ako whakawhiti (fine-tuning) ki ngā tauira puna tuwhera, kātahi ka whakatairanga i aua tauira hei tauira nō rātou ake.

Heoi, i whakatakoto a CLEVI i te **‘From-scratch Foundation Model’** hei aronga matua.

Arā, i hoahoa, i whakahaere tika hoki i ngā tukanga katoa, mai i te kohikohi raraunga, te hoahoa tauira, te whakangungu tohatoha nui, tae atu ki te manatoko kounga.

2. Ngā take he uaua ai te whakawhanake tauira reo tino nui

Hei whakangungu i tētahi From-Scratch Foundation Model, me ako ngā mea katoa kei raro iho nei.

Me nui te huinga raraunga kounga teitei

• Kanorautanga: Me whakarite te kanorautanga o ngā reo, ngā rohe whakamahi, me ngā whakatakotoranga (kuputuhi, atahanga, me ērā atu)
• Whakamā: Te tango i te haruru, te whakahaere kounga, me te tātari i ngā raraunga tārua, kino hoki
• Raihana: Me whakamārama ngā manatārua me ngā motika whakamahi raraunga

Hanganga rorohiko tino nui

• Huinga GPU/TPU mahi teitei: Te whakakotahi i ngā taputapu mai i ngā mano ki ngā tekau mano o ngā kōpuku hei tautoko i te hangarau whakangungu tohatoha raraunga nui
• Anga whakangungu tohatoha whaihua: DeepSpeed, Megatron-LM, Ray me ētahi atu
• Rokiroki/whatunga: Te tāuru me te whakaputa raraunga nui, me tētahi taiao whatunga tere

Te hoahoa hoahoanga tauira

• Te whakauru i ngā hanganga hou: Transformer, MoE(Mixture of Experts), multimodal me ērā atu
• Te āhei whakarahi: Te whai whakaaro ki te whakawhānui i te maha o ngā tawhā, ngā paparanga, te roa o ngā tāuru me ērā atu
• Te whaihua: Te arotau i te tere whakangungu/whakatau me te whakamahinga pūmahara

Ngā rautaki me ngā hātepe ako

• Ngā whāinga whakangungu tuatahi: Ngā tauira reo (hei tauira, next token prediction), ngā tauira multimodal (hei tauira, contrastive learning) me ērā atu
• Ngā tikanga arotau: mixed precision, gradient accumulation, learning rate schedule me ērā atu
• Te whakariterite/te whakapūmau: dropout, layer norm, weight decay me ērā atu

Te pūnaha aromatawai me te whakamana

• Ngā huinga benchmark: Te whakamahi i ngā huingararaunga paerewa (Benchmarks)
• Te aromatawai ā-roto: Te aromatawai i runga i ngā horopaki whakamahinga tūturu
• Te aroturuki haere tonu: Te tirotiro i te kounga, te rītaha me te haumaru i te wā me muri i te whakangungu

Ngā pūkenga o ngā kaimahi me te whakahaere

• Ngā kairangahau AI/ML: Te hoahoa tauira me te whakawhanake hātepe
• Ngā kaiwhakamahi hangarau raraunga: Te kohi, te pure me te whakahaere raraunga
• Ngā kaiwhakamahi hangarau hanganga: Ngā pūnaha tohatoha, te whakahaere kapua/kei runga rawa
• Ngā kaiwhakahaere kaupapa: Te whakahaere i te wātaka, te pūtea me te kounga

Ngā whakaaro matatika, ā-ture, ā-haumarutanga

• Ngā tikanga matatika AI: Te rītaha, te haumaru, te mārama me te kawenga
• Te ū ki te ture: Ngā raraunga whaiaro, te manatārua me te tino rangatiratanga raraunga
• Te haumaru: Te aukati i te turuturu o ngā raraunga/me ngā tauira, me te whakahaere urunga

Tata ki te 2,000 ngā kaimahi rangahau AI puta noa i te ao i tēnei wā, ā, ko te nuinga tonu o rātou e aro ana ki ngā kamupene hangarau nui pērā i META, Google me XAI. He tino onge ngā tīma i Aotearoa e āhei ana ki te hoahoa ā-ringa i ngā mea katoa mai i te kohinga raraunga ki te hanganga whakangungu nui, ā, ki te waihanga i tētahi Foundation model.

3. Te waihanga i tētahi tauira reo tino nui me te tokoiti o ngā kaimahi.

I runga i tana wheako neke atu i te 10 tau ki te rangahau i te deep learning me te machine learning, i whakatūria e te Tumuaki a Lee Hwan-ho te kamupene (클레비) i ngā tau e 3 ki muri, me te whāinga “kia waihanga i te AI pai rawa atu o te ao”.

Nāna anō i hoahoa ngā tukanga katoa e tika ana mō te whakawhanake i tētahi Foundation Model, tae atu ki te hanga i te raina tukatuka raraunga a te kamupene, te hoahoa i te hanganga deep learning tohatoha nui, te whakawhanake i te hanganga tauira me te whakamana kounga.

Whai muri i te whakangungu i te tauira mō ngā wā maha, kua whai tauira Clevi-5-x mātou ināianei, e taea ana te whakataurite ki ngā tauira taumata-runga o te ao.

Hei whakangungu tōtika i ngā tauira reo tino nui, me hono mai i ngā rau ki ngā mano tini o ngā GPU hei kāhui, me whai hanganga e taea ai te tukatuka ngātahi i ngā tātaitanga e tae ana ki te tekau mano hautoru ia hēkona. I roto i tēnei tukanga, he mea tino nui ngā hangarau e tukutahi ana i ngā tātaitanga, e whakaiti ana hoki i te tōmuri whakawhitiwhiti i ngā taiao tohatoha nui. Tae noa mai ki tēnei wā, nā te kore o tētahi “algorithm whakahaere tōtika”, kāore te nuinga o ngā kamupene o te motu i taea te whakangungu tika i ngā tauira. Nā te whakawhanake motuhake a te Tumuaki o Clevi, a Lee Hwan-ho, i tēnei algorithm whakahaere tātaitanga tohatoha, ko Clevi te kamupene tuatahi i te motu i angitu ki te whakangungu i tētahi tauira reo tino nui e 1.4 trillion (1.4 Trillion) ngā tawhā.

4. Te take i taea ai te whakawhanake i ngā tauira kanorau me te tokoiti o ngā kaimahi whakawhanake

Kei te nuinga o ngā kamupene hangarau-nui hoki ngā hangarau whakahaere hanganga nui me te purei raraunga.

Hei whakahaere i ēnei mea, me hia rau ki ngā mano tini o ngā kairangahau.

Heoi anō, e whakawhanake ana, e pupuri ana hoki a Clevi i ngā tauira kanorau me te tokoiti o ngā kairangahau.

I pēhea i taea ai?

Mā te hangarau Teacher-Student Model(Knowledge Distilation), e whakawhanake, e pupuri hoki a Clevi i ngā tauira kanorau me te tokoiti o ngā kaimahi.

Nō reira, me tirotiro tātou ki te hangarau Teacher-Student.

Teacher-Student(whakawhitinga mātauranga)

Ko te hangarau Teacher-Student, ki te whakamārama māmā, he hangarau e whakamahi ana i te tauira reo tino nui (Teacher Model) hei aromatawai, hei tuku urupare hoki ki te tauira tamariki (Student Model), kia tere ai te whakawhanake i ngā tauira kanorau, mā te whakakapi i ngā rau o ngā kairangahau ahakoa he tokoiti noa ngā kaimahi.

• Mā te Mother Model(tauira tino nui) e aromatawai, e tuku urupare hoki ki ngā tauira o ngā momo wāhanga, ā, ka whakakapia ngā rau o ngā kairangahau.
• Mēnā ka whakahaerehia te whakangungu mā Clevi-x-platform i tēnei ara, ka taea te whakangungu me te tuku ki te whakamahinga i ngā tauira mahi-nui pēnei i a Reasoning, VLM, Physical AI, Coding Agent i roto i te 10–15 rā.

Clevi Coding Agent

Clevi Phsycal AI Learning Platform

VLM-Vision Language Model

Kaitohu haumaru

5. Te rerekētanga hangarau me te kounga o Clevi-x-platform

Te kaha me te whakawhānuitanga o te tauira

• Te hiranga o te whakatauira (CoT/Reasoning): kua whakaurua e cip-5-x te whanaketanga arorau ā-taahiraa me te whakaatu taunakitanga ki roto i tōna rapunga whakaaro hoahoa, ā, e whakaatu ana i te kaha nui ki te tātai me te whakamāori pono i te whakaoti rapanga pūtaiao, pāngarau me te hangarau. E hoahoatia ana, e whakahaerehia ana hoki kia whai i te taumata mahi o GPT-5, neke atu rānei, e ai ki ngā paerewa aromātai ā-roto, ā, ko te tukurua me te taea te whakamana i raro i ngā tikanga akiaki ōrite ngā tohu kounga matua e whakahaerehia ana.

• Te whānuitanga katoa o te multimodal: mā te whakaputa me te whakakotahi i te VLM e ū ana ki te whāinga (cip-5-vision), te tauira tukatuka multimodal rahi (ivy-4-mm), me te tauira mama ki te taputapu (ivy-3-text) i runga i te tūāpapa kotahi, ka taea te whiriwhiri i te huinga tino pai e hāngai ana ki ngā āhuatanga o te mahi (rapu/whakarōpūtanga/whakarāpopototanga vs. whakatauira/whakamārama/whakatinana).

Te hāngaitanga ki ngā hinonga

• Haumarutanga me te wātea ki runga i te on-premises: ka whakahaerehia te whānuitanga katoa o te whatunga kati (tāuru-whakangungu-ratonga-pūrua), me te hoahoa HA, te whakawhānui ā-kōwae, me te SVCE (taiao whakahaere haumaru kua wehea), hei tiaki wehe i ngā raraunga me ngā tawhā tauira.
• Te whakatinanatanga me te whakawhānui tere: ka tautoko a Ivy Chat (kōrerorero multimodal/agent mō ngā mahi), me API Platform (SaaS paerewa ā-ao), i te whakamahinga me te whakahaere tere.

Te rerekētanga o te AI ā-tinana (Physical AI)

• Mā te whakakotahi i te whaihanga e takea mai ana i NVIDIA Isaac me te ako whakapakari e takea mai ana i te whanaketanga (Evolutionary RL), me te ako whakarara mā te whakawhiti Sim2Real me ngā raraunga waihanga, kua piki ake te pai o te ako me te kaha urutau ki te wāhi mahi. He iti noa ngā kōwhiringa whakakapi e kapi ana i te ako me te tuku mai i te ao matihiko ki te karetao.

Te kounga ratonga me te kāwanatanga

• Ka whakahaerehia te whakahaere putanga o ngā tauira/akiaki/taputapu, te huinga aromātai (mātauranga reo Koreana me te reo Ingarihi, ngā tūmahi ā-ahumahi, ngā tohu pohehe me te haumaru), te whakahaere panoni (SLO/SLA), me te mātakitanga whakahaere (te roa whakautu, te ōrau angitu, TCO) mā ngā tukanga o te tūāpapa kotahi.

Ahakoa neke atu i te 300 ngā kamupene ratonga tauira AI kei Korea i tēnei wā,

Ko Clevi anake te kamupene ka taea te tuku ratonga on-premises me ngā ratonga kapua i te wā kotahi mā roto i tētahi tauira tūāpapa motuhake, tino kaha te mahi.

6. Ngā huarahi ki te whakamahi i ngā tauira reo ki ngā ahumahi

Nā te mea he nui te haumi me tino āta whakamātau te whakatakinga o AI, he mea tino nui te whakataurite me te whakamātau tika i ngā rongoā kia mōhio ai mēnā ka tino whai hua ki te hinonga.

• Ehara i te mea ka hanga tauira noa iho a CLEVI; ka whakarato kē ia i ngā rongoā AI ka taea te whakamahi i ngā wāhi ahumahi tūturu.
• Hei tauira, kua oti i a ia ngā hongere hinonga, pēnei i te Ivy Chat Platform, Clevi API Platform, te whakarite motuhake mō ia ahumahi, me te urutau ki ngā whakaritenga haumarutanga.
• Kei a ia ngā āheinga hangarau kāore e kitea nuitia ana i roto, i waho hoki o te motu, pēnei i te AI ā-tinana, te karetao, me te tukatuka raraunga multimodal.

Ka whakarato a CLEVI i ngā rongoā AI tūturu, ehara i te mea ko AI te ‘whāinga’, engari ko te ‘taputapu’ hei whai wāhi atu ki te pai ake o ngā mahi me te auahatanga ahumahi. Whakamātauria te rerekētanga o te tauira From-scratch Foundation māu ake.

Ngā pātai me ngā tono whakaaturanga: [email protected]

Copyright© 2025 Clevi Inc. Kua rahuitia ngā motika katoa.