A takaice dai kalmar GPT ta zama, cikin 'yan shekaru kaɗan, alamar babban cikas a fasaha. A bayan waɗannan haruffa uku akwai wani canji na tsari: mun ƙaura daga kayan aikin kwamfuta da umarni ke sarrafawa zuwa ga masu haɗin gwiwa na dijital waɗanda muke tattaunawa da su. Don fahimtar girman wannan juyin juya hali, dole ne mutum ya duba ƙarƙashin rufin asiri ya kuma bincika dabarun wannan hankali.
I. Ma'anar Fasaha: Ginshiƙai Uku na Kalmomin Allon
GPT tana nufin Generative Pre-trained Transformer. Kowace kalma tana bayyana wani muhimmin mataki a cikin ƙirƙirar samfurin.
1. Mai samarwa
A duniyar fasahar kere-kere, ana bambanta tsakanin nau'ikan samfura guda biyu: masu nuna bambanci da kuma masu haifar da yanayi.
- Tsarin nuna wariya yana lura da bayanai kuma yana sanya masa lakabi (misali, "Wannan hoton kyanwa ne").
- Samfurin samarwa, kamar GPT, ya ci gaba da tafiya. Dangane da iliminsa, yana da ikon samar da sabbin bayanai gaba ɗaya waɗanda ba su wanzu a da ba. Idan ya yi nazarin miliyoyin waƙoƙi, ba zai haddace ɗaya ba; zai rubuta sabo, yana girmama ƙa'idodin waƙa da mita waɗanda ya "fahimta."
2. An riga an horar da shi
Nan ne inda ƙarfin lantarki yake. Kafin ya isa allonka, samfurin ya shiga wani babban matakin koyo. Ka yi tunanin ɗakin karatu wanda ke ɗauke da babban ɓangare na intanet na duniya: cikakken Wikipedia, ɗakunan karatu na dijital, dandalin tattaunawa (kamar Reddit), ma'ajiyar lambobin kwamfuta (GitHub), da kuma rumbun adana labarai.
Tsawon watanni, a kan dubban na'urori masu amfani da wutar lantarki, samfurin yana "karanta" wannan rubutun ba kawai don koyon takamaiman bayanai ba, har ma don koyon zurfin tsarin harshe da kansa. Wannan "horarwa kafin" yana bawa samfurin damar isa ga tushen al'adu da harshe na duniya gaba ɗaya kafin ma ya sami tambayarsa ta farko.
3. Transformer
Wannan shine ci gaban kimiyya na shekarar 2017, wanda ya samo asali daga wata takarda ta bincike ta Google mai taken "Hankali Duk Abinda Kake Bukata"Kafin Transformer, samfuran AI suna karanta jimloli kalma ɗaya bayan kalma, a cikin tsari (samfuran da ke maimaitawa). Duk da haka, Transformer yana kallon jumla gaba ɗaya kuma a lokaci guda.
Godiya ga wata dabara da ake kira Hankali, tana iya auna muhimmancin kowace kalma dangane da wasu. A cikin jumlar "Dabbar ba ta ketare titi ba saboda ta gaji sosai"," Transformer ya fahimci cewa "it"yana nufin dabbar. Idan jumlar ta ƙare da "saboda ya yi faɗi sosai"," nan take ya fahimci cewa "shi" yana nufin titi. Wannan ikon fahimtar mahallin da ke da sarkakiya shine abin da ke sa martanin GPT su ji kamar ɗan adam.
shafi: ChatGPT Down? Me yasa akwai matsala a AI da kuma yadda ake samun damar mafi kyawun GPT kyauta
II. Takaitaccen tarihin juyin halitta: Daga GPT-1 zuwa GPT-5
Labarin GPT yana ɗaya daga cikin ci gaban da kamfanin ke samu BABI.
Farkawa (GPT-1 da GPT-2)
A shekarar 2018, GPT-1 ta kafa harsashin ginin. Tana iya yin ayyuka masu sauƙi, amma ƙwaƙwalwarta ta yi ƙasa. A shekarar 2019, GPT-2 ta ƙaru sosai da sigogi biliyan 1.5 (haɗin cikin samfurin). Ta ƙware sosai wajen samar da rubutu mai ma'ana har OpenAI ta yi jinkirin fitar da shi, tana tsoron za a yi amfani da shi wajen ƙirƙirar bayanai marasa tushe a fannin masana'antu.
Fashewar (GPT-3 da 3.5)
A shekarar 2020, GPT-3 ta ketare wani shinge na tunani da ke da sigogi biliyan 175. Dabba ce ta lissafi. Ba wai kawai mai fassara ko marubuci ba ce; ta zama mai iya yin tunani mai ma'ana da kuma rubuta lambobi a cikin harsunan shirye-shirye da dama. Sigar GPT-3.5 ta yi aiki a matsayin injin ƙaddamar da ChatGPT ga jama'a a watan Nuwamba na 2022, wanda ya haifar da yanayi na duniya baki ɗaya.
Balaga (GPT-4 da Sama)
An fitar da GPT-4 a shekarar 2023, kuma yana da tsarin multimodal. Yana iya nazarin hotuna, magance matsalolin lissafi masu rikitarwa, da kuma nuna manyan matakan kerawa. An inganta daidaitonsa don rage "mafarki" (lokacin da AI ke ƙirƙiro gaskiya da tabbaci).
Related: ChatGPT-5 mara iyaka kuma kyauta akan AI kyauta akan layi
III. Ayyukan Cikin Gida: Injin Yiwuwa
Sabanin yadda aka yi imani da shi, GPT ba ta da "tunani" ko ra'ayi. Aikinta na ƙididdiga ne kawai, amma a kan babban sikelin da yake kwaikwayon hankali.
Lokacin da mai amfani ya rubuta tambaya, samfurin yana canza kowace kalma zuwa jerin lambobi da ake kira embeddings (ko vectorization). Waɗannan lambobi suna ba da damar sanya kalmomi a cikin sararin lissafi mai dubban girma. A cikin wannan sarari, kalmar "Sarki" tana kusa da "Sarauniya" ko "Kambi" a fannin lissafi.
Tsarin mayar da martani jerin hasashe ne:
- Samfurin yana nazarin da sauri.
- Yana ƙididdige wace kalma (ko kuma "Alama") yana da mafi girman damar fara amsawar.
- Da zarar an zaɓi wannan kalmar, za ta mayar da ita cikin nazarinta don yin hasashen kalma ta gaba.
- Yana maimaita wannan tsari har sai ya kai ga inda aka ƙare.
Abin da ya sa sakamakon ya burge ni shi ne cewa GPT ba koyaushe take zaɓar kalmar da ta fi dacewa ba (wanda zai sa ta zama mai maimaitawa da ban sha'awa), amma tana amfani da yanayin "zafin jiki" (wanda aka sarrafa bazuwar) don canza salonta da kerawa.
IV. Mataki mai mahimmanci: Daidaito tsakanin ɗan adam (RLHF)
Idan kawai ka ba wa na'ura dukkan intanet ɗin don karantawa, zai iya zama mai tayar da hankali ko kuma bayar da shawara mai haɗari, domin intanet ta ƙunshi mafi kyau da mafi muni na ɗan adam.
Don sanya GPT "mai kyau," OpenAI tana amfani da Reinforcement Learning from Human Feedback (RLHF). Dubban masu gwajin ɗan adam sun yi mu'amala da farkon sigar samfurin, suna kimanta amsoshin: "Wannan amsa ce mai kyau," "Wannan rashin kunya ce," "Wannan ƙarya ce."
Ta wannan tsari, AI ta koyi yadda ake ɗaukar matsayin mataimaki mai taimako, tsaka-tsaki, kuma mai ɗabi'a. Wannan matakin ƙarshe ya canza fasahar zamani zuwa samfurin kasuwa mai yawa.
V. Kalubale da ƙuntatawa
Duk da ƙarfinsa, GPT ta kasance kayan aiki mai wuraren makafi:
- Hallucination: Domin yana neman ya zama abin dogaro ga kididdiga, yana iya ƙirƙiro tarihin rayuwa ko tushen shari'a wanda babu shi, kawai saboda kalmomin "suna da sauti" tare.
- Rashin hankali: GPT ba ta jin komai. Ba ta da fahimtar duniyar zahiri. Idan ta yi bayani kan yadda ake dafa omelet, ba ta san menene zafi ko warin ƙwai ba; kawai ta san alaƙar harshe tsakanin waɗannan ra'ayoyin.
- Kudin muhalli: Horarwa da gudanar da waɗannan samfuran suna buƙatar wutar lantarki da ruwa mai mahimmanci (don sanyaya sabar).
Kammalawa
GPT ya fi software na sarrafa kalmomi kawai. Tsarin fahimta ne. Kamar yadda wutar lantarki ta ba da damar sarrafa aikin jiki, GPT ta fara sarrafa wasu nau'ikan aikin tunani ta atomatik.
Ta hanyar fahimtar cewa GPT tana nufin Generative (ƙirƙira), Pre-trained (babban ilimi), da Transformer (fahimtar mahallin), za mu iya fahimtar dalilin da ya sa wannan fasaha ke nuna wani sauyi. Muna farkon binciken abin da waɗannan samfuran masu sauƙi da hankali za su kawo wa kimiyya, ilimi, da ƙirƙirar fasaha.
