🤖 Kia ora, ko CLEVI tēnei.
I haere mātou ki te kaupapa 「2025 JBNU Research Fair Week」 i whakahaerehia e te Whare Wānanga o Jeonbuk mai i te 20 ki te 22 o Hānuere.
He hui whakahirahira tēnei i whakaatu i te aronga o ngā hangarau anamata e hono ana ki ngā rautaki ahumahi me te motu, e aro ana ki ngā hua rangahau hirahira a te Whare Wānanga o Jeonbuk, ā, i whakarite hoki i tētahi wāhi whakawhitiwhiti mō te ahumahi, te ao mātauranga, te rangahau me te kāwanatanga.
I pōhiritia hoki a Lee Hwan-ho, te Tumuaki o CLEVI, hei kaikōrero VIP mō te kauhau i te hui whakatuwhera i te rā 20! 👏
Kia rongo anō tātou i te wairua o taua rā!
Anei a Tumuaki Lee Hwan-ho e whakarite ana i tana kauhau i mua i te tīmatanga, e kitea ana tana whakaaroaro mō te huarahi kia whai kiko, kia ngāwari hoki te kauhau.
Ka tīmata te kaupapa, ka whakahaerehia te hui whakatuwhera.
I tīmata whakamīharo te kaupapa mā tētahi whakaaturanga toi ora i oti wawe i tētahi ākonga o te Whare Wānanga o Jeonbuk.
Whai muri i te kauhau matua tuatahi, i tuku kauhau a Lee Hwan-ho, te Tumuaki o CLEVI, mō te kaupapa “He aha i hiahiatia ai e AI tētahi reo tinana?” mō te kauhau matua tuarua. Me tirotiro poto tātou.
Me pēhea rawa e AI e mārama ai ki te ao,
e hanga ai i ngā rerenga kōrero, ā, e mahi ai hoki?
Ka tiro tahi tātou ki roto i te 'roro' o ngā tauira reo nui (LLM) pēnei i a ChatGPT me Claude e whakamahia ana e tātou ia rā, kia mārama ai tātou ki ngā mātāpono e mārama ai, e hanga ai hoki rātou i ngā rerenga kōrero.
- Te tuku tapanga ingoa (Token ID): Tuatahi, ka tuku a AI i tētahi tau motuhake pēnei i te '5' me te '8' ki ia kupu hei whakahaere i a ia.
- Te tohu wāhi ki te mahere (Vector): Engari, kāore e taea te mōhio ki te tikanga mā ngā tau noa iho, nō reira ka hoatu ki ia kupu tētahi 'wāhi taipitopito (taunga)' i hangaia mai i ngā rau tau.
- Te kimi hoa tata: Ko ngā kupu he ōrite te tikanga (hei tauira, ko 'the' me 'a') ka noho piri tata ki a rātou anō i runga i tēnei mahere, pēnei i ngā hoa tata.
Nō reira, me pēhea a AI e matapae ai i te kupu e whai ake nei?
Kua mārama tātou ki te panoni raraunga taketake, nō reira me tirotiro ināianei i te tukanga katoa e puta ai ēnei tau hei rerenga kōrero tūturu. Me whakaaro he uara tāuru tō tātou e kī ana "Paris is the city...".
Tokenizer (whakaritenga wāhanga): Ko ngā rerenga kōrero ka whakaurua e tātou ka hurihia mā te tokenizer hei tau (Token ID), pēnei i tā tātou i kite i mua. Ko "Paris" te nama 1652, ko "is" te nama 5... pēnā tonu.
LLM (te tātaitai nui): Ka uru ēnei huinga tau ki roto i te tātaitai nui e kīia ana ko LLM (te tauira Bloom kei te whakaahua). I konei, mā ngā mahi tātai 'vector' i ako ai tātou i mua, ka mārama ki te horopaki. Ka kī pea, 'Ā, e kōrero ana ināianei mō te tāone o Paris!'
Logits & Softmax (te tātai tūponotanga): Ka mutu ana te tātai, ka hoatu e AI he kaute ki ngā mano tini o ngā kupu ka whai ake. Kātahi ka hurihia mā tētahi tātari e kīia ana ko 'softmax' hei **'tūponotanga'** e māmā ana ki a tātou te kite.
Te kōwhiringa whakamutunga: Tirohia te kauwhata kei te taha matau.
- of → 33.7%
- that → 15.3%
- where → 9.4%
Nāwai rā, ka kōwhiri a AI i te **'of'** he nui rawa atu tōna tūponotanga, ā, ka oti i a ia te rerenga "Paris is the city of...". Mā te tukurua i tēnei tukanga, he mano tini ngā wā, ka hangaia te whakautu roa e kitea ana e tātou.
Hei whakatau, ko te maumahara a AI ki tētahi kupu ehara i te maumahara noa ki ngā pū reta, engari he maumahara ki tōna 'wāhitau'—ki te wāhi e noho ana taua kupu i runga i te mahere nui o ngā tikanga.
Ka taea te whakawhānui i te LLM ki te tirohanga, ki ngā karetao, me te ako whanonga.
He AI e mārama ana ki te tikanga, e kite ana, e mārama ana, ā, e mahi ana.
Ko te tuku ringa me ngā waewae ki AI he whakawhānui i te māramatanga mai i te reo ki te whakamātau, ā, ka anga whakamua ki te mahi.
Kua kore a AI e noho hei wāhanga o ngā whakaaro, o ngā wawata rānei; kua uru kē ki te wāhanga o te whakatinanatanga me te ao tūturu.
Te hīkoi whai ake a CLEVI
Hei te wā kei te heke mai, ka whakawhānui tonu a CLEVI i ngā whai wāhitanga pērā i ngā awheawhe ā-roto, ā-waho hoki i te whare wānanga, ngā kaupapa whakaharatau, me ngā wāhi whakamātau, kia tokomaha ake ai te hunga ka taea te wheako ā-ringa me pēhea te mahi a te hangarau AI i te ao tūturu.
Pērā i te noho o te kauhau ki te Whare Wānanga o Jeonbuk hei piriti whai tikanga e hono ana i ngā ākonga me mātou, ka haere tonu mātou ki te mahi hei piriti i waenga i ngā kaiwhakatinana hangarau me ngā kaiwhakamahi, i te ao mātauranga me te ao ahumahi.
