Ayini Amamodeli Esisekelo ku-AI Ekhiqizayo? [Ividiyo Nombuzo]

Impendulo emfushane: Amamodeli ayisisekelo amamodeli amakhulu, ahloselwe i-AI aqeqeshwe kumasethi edatha amakhulu, abanzi, bese eguqulwa emisebenzini eminingi (ukubhala, ukusesha, ukufaka amakhodi, izithombe) ngokusebenzisa ukukhuthazwa, ukulungisa kahle, amathuluzi, noma ukubuyisa. Uma udinga izimpendulo ezinokwethenjelwa, zihlanganise nesisekelo (njenge-RAG), imikhawulo ecacile, kanye nokuhlola, kunokuba uzivumele zizenzele.

Izinto ezibalulekile okufanele uzicabangele:

Incazelo: Imodeli eyodwa eqeqeshwe kabanzi isetshenziswe kabusha emisebenzini eminingi, hhayi yomsebenzi owodwa ngemodeli ngayinye.

Ukuzivumelanisa nezimo: Sebenzisa ukusheshisa, ukulungisa kahle, ama-LoRA/ama-adapter, i-RAG, namathuluzi ukuqondisa ukuziphatha.

Ukulingana kokukhiqiza: Kunika amandla umbhalo, isithombe, umsindo, ikhodi, kanye nokukhiqizwa kokuqukethwe okunezinhlobo eziningi.

Izimpawu zekhwalithi: Beka phambili ukulawula, ukunciphisa ukubona izinto ezingekho, ikhono lezindlela eziningi, kanye nokuphetha okuphumelelayo.

Ukulawulwa kwezingozi: Hlela ukubona izinto ezingekho, ukucwasa, ukuvuza kobumfihlo, kanye nokufakwa ngokushesha ngokulawula nokuhlola.

Izihloko ongase uthande ukuzifunda ngemva kwalesi:

🔗 Iyini inkampani ye-AI
Qonda ukuthi izinkampani ze-AI zakha kanjani imikhiqizo, amaqembu, kanye namamodeli emali engenayo.

🔗 Ibukeka kanjani ikhodi ye-AI
Bheka izibonelo zekhodi ye-AI, kusukela kumamodeli e-Python kuya kuma-API.

🔗 Iyini i-algorithm ye-AI
Funda ukuthi ama-algorithms e-AI ayini nokuthi enza kanjani izinqumo.

🔗 Kuyini ubuchwepheshe be-AI
Hlola ubuchwepheshe be-AI obuyinhloko obusebenzisa amandla okuzenzakalelayo, ukuhlaziya, kanye nezinhlelo zokusebenza ezihlakaniphile.

1) Amamodeli esisekelo - incazelo engenankungu 🧠

Imodeli yesisekelo iyimodeli ye-AI enkulu, enenjongo ejwayelekile eqeqeshwe kudatha ebanzi (ngokuvamile inqwaba yayo) ngakho ingashintshwa ukuze ihambisane nemisebenzi eminingi, hhayi eyodwa kuphela (NIST, Stanford CRFM).

Esikhundleni sokwakha imodeli ehlukile ye:

• ukubhala ama-imeyili

• ukuphendula imibuzo

• ukufingqa ama-PDF

• ukukhiqiza izithombe

• ukuhlukanisa amathikithi okusekela

• ukuhumusha izilimi

• ukwenza iziphakamiso zekhodi

...uqeqesha imodeli eyodwa enkulu eyisisekelo "efunda umhlaba" ngendlela yezibalo engacacile, bese uyivumelanisa nemisebenzi ethile ngezeluleko, ukulungisa kahle, noma amathuluzi engeziwe (Bommasani et al., 2021).

Ngamanye amazwi: iyinjini evamile ongayiqondisa.

Futhi yebo, igama elingukhiye lithi “okuvamile.” Yilokho konke okubalulekile.

2) Ayini Amamodeli Esisekelo ku-AI Ekhiqizayo? (Indlela afanelana ngayo ngqo) 🎨📝

Ngakho-ke, ayini amamodeli ayisisekelo ku-AI Ekhiqizayo? Yiwona amamodeli ayisisekelo anika amandla izinhlelo ezingakhiqiza okuqukethwe okusha - umbhalo, izithombe, umsindo, ikhodi, ividiyo, kanye nezingxube eziningi zazo zonke lezo (i-NIST, iphrofayela ye-AI Ekhiqizayo ye-NIST).

I-AI ekhiqizayo ayigcini nje ngokubikezela amalebula anjengokuthi “ugaxekile/hhayi ugaxekile.” Imayelana nokukhiqiza imiphumela ebonakala sengathi yenziwe ngumuntu.

• izigaba

• izinkondlo

• izincazelo zomkhiqizo

• imifanekiso

• izingoma

• ama-prototypes ohlelo lokusebenza

• amazwi okwenziwa

• futhi ngezinye izikhathi izinto ezingenangqondo ezingenakucatshangwa 🙃

Amamodeli esisekelo kakhulu lapha ngoba:

• Bathathe amaphethini abanzi avela kumasethi amakhulu edatha (Bommasani et al., 2021)

• Zingajwayelanisa izinto ezintsha (ngisho nezingavamile) (Brown et al., 2020)

• Zingasetshenziswa kabusha ukuze kukhiqizwe imiphumela eminingi ngaphandle kokuqeqeshwa kabusha kusukela ekuqaleni (Bommasani et al., 2021)

Ziyi-"base layer" - njengenhlama yesinkwa. Ungayibhaka ibe yi-baguette, i-pizza, noma ama-cinnamon rolls... akuyona into efanelekile, kodwa uyangiqonda 😄

3) Kungani bashintshe yonke into (nokuthi kungani abantu bengayeki ukukhuluma ngabo) 🚀

Ngaphambi kwamamodeli ayisisekelo, i-AI eningi yayiqondene nomsebenzi:

• qeqesha imodeli yokuhlaziya imizwa

• qeqesha omunye ukuze ahumushe

• qeqesha omunye ukuze ahlukanise izithombe

• qeqesha omunye ukuze aqashelwe inhlangano eqanjwe ngamagama

Lokho kwasebenza, kodwa kwakuhamba kancane, kubiza kakhulu, futhi... kwakungaqinile.

Amamodeli esisekelo ayishintshile:

• ukuqeqeshwa kwangaphambili kanye (umzamo omkhulu)

• sebenzisa kabusha yonke indawo (inkokhelo enkulu) (Bommasani et al., 2021)

Lokho kusetshenziswa kabusha kuyimpinda. Izinkampani zingakha izici ezingu-20 phezu komndeni owodwa wemodeli, kunokuba ziphinde zisungule isondo izikhathi ezingu-20.

Futhi, ulwazi lomsebenzisi luye lwaba lwemvelo kakhulu:

• awusebenzisi "i-classifier"

• Ukhuluma nomodeli sengathi ungumuntu osebenza naye osizayo ongalali ☕🤝

Ngezinye izikhathi kufana nomuntu osebenza naye oqonda yonke into ngokuzethemba, kodwa hey. Ukukhula.

4) Umqondo oyinhloko: ukuqeqeshwa kwangaphambi kwesikhathi + ukuzivumelanisa nezimo 🧩

Cishe wonke amamodeli esisekelo alandela iphethini ethile (Stanford CRFM, NIST):

Ukuqeqeshwa kwangaphambi kwesikhathi (isigaba "sokuncela i-inthanethi") 📚

Imodeli iqeqeshwe kumasethi edatha amakhulu, abanzi kusetshenziswa ukufunda okuqondiswayo (i-NIST). Kumamodeli olimi, lokho ngokuvamile kusho ukubikezela amagama angekho noma ithokheni elilandelayo (Devlin et al., 2018, Brown et al., 2020).

Iphuzu akukhona ukuyifundisa umsebenzi owodwa. Iphuzu liwukuyifundisa izethulo ezijwayelekile:

• uhlelo lolimi

• amaqiniso (uhlobo)

• amaphethini okucabanga (ngezinye izikhathi)

• izitayela zokubhala

• isakhiwo sekhodi

• inhloso evamile yabantu

Ukuzivumelanisa nezimo (isigaba "sokwenza kube ngokoqobo") 🛠️

Bese uyivumelanisa usebenzisa eyodwa noma ngaphezulu kwalokhu:

• ukunxusa (imiyalelo ngolimi olulula)

• ukulungiswa kwemiyalelo (ukuyiqeqesha ukuthi ilandele imiyalelo) (Wei et al., 2021)

• ukulungiswa kahle (ukuqeqeshwa kwedatha yakho yesizinda)

• I-LoRA / ama-adapter (izindlela zokulungisa ezilula) (Hu et al., 2021)

• I-RAG (ukwenziwa okwandayo kokutholwa - imodeli ibheka amadokhumenti akho) (Lewis et al., 2020)

• ukusetshenziswa kwamathuluzi (imisebenzi yokushaya ucingo, ukuphequlula izinhlelo zangaphakathi, njll.)

Yingakho imodeli efanayo eyisisekelo ingabhala isigcawu sothando… bese isiza ukulungisa inkinga yombuzo we-SQL ngemuva kwemizuzwana emihlanu 😭

5) Yini eyenza inguqulo enhle yemodeli yesisekelo? ✅

Lesi yisigaba abantu abasiqayo, bese bezisola kamuva.

Imodeli yesisekelo "esihle" ayigcini nje ngokuba "nkulu." Enkulu iyasiza, impela... kodwa akuyona yodwa into. Inguqulo enhle yemodeli yesisekelo ivame ukuba nalokhu:

Ukwenziwa okujwayelekile okunamandla 🧠

Isebenza kahle emisebenzini eminingi ngaphandle kokudinga ukuqeqeshwa kabusha komsebenzi othile (Bommasani et al., 2021).

Ukuqondisa nokulawula 🎛️

Ingalandela ngokuthembekile imiyalelo efana nale:

• "fingqa"

• "Sebenzisa amaphuzu ezinhlamvu"

• "Bhala ngezwi elinobungane"

• "Ungadaluli ulwazi oluyimfihlo"

Amanye amamodeli ahlakaniphile kodwa ayashelela. Njengokuzama ukubamba insipho eshaweni. Kuyasiza, kodwa kuyashintshashintsha 😅

Ukuthambekela okuphansi kokungaboni kahle (noma okungenani ukungaqiniseki okuqondile) 🧯

Akukho modeli engavikelekile ezingqondweni ezingabonakali, kodwa ezinhle:

• nciphisa imibono

• vuma ukungaqiniseki kaningi

• hlala useduze nomongo onikeziwe uma usebenzisa ukubuyisa (Ji et al., 2023, Lewis et al., 2020)

Ikhono elihle le-multimodal (uma kudingeka) 🖼️🎧

Uma wakha abasizi abafunda izithombe, abahumusha amashadi, noma abaqonda umsindo, i-multimodal ibaluleke kakhulu (Radford et al., 2021).

Isiphetho esisebenzayo ⚡

Ukubambezeleka kanye nezindleko kubalulekile. Imodeli enamandla kodwa ehamba kancane ifana nemoto yezemidlalo enethayi eliphantshiwe.

Ukuphepha nokuziphatha kokuqondanisa 🧩

Hhayi nje “ukwenqaba konke,” kodwa:

• gwema imiyalelo eyingozi

• ukunciphisa ubandlululo

• phatha izihloko ezibucayi ngokucophelela

• melana nemizamo eyisisekelo yokuqhekeka kwejele (ngandlela thile…) (NIST AI RMF 1.0, NIST Generator AI Profile)

Imibhalo + uhlelo lwe-ecosystem 🌱

Lokhu kuzwakala kungcolile, kodwa kuyiqiniso:

• amathuluzi

• ama-harnesses e-eval

• izinketho zokusebenzisa

• izilawuli zebhizinisi

• ukwesekwa kokulungisa kahle

Yebo, igama elithi “ecosystem” liyigama elingacacile. Nami ngiyalizonda. Kodwa libalulekile.

6) Ithebula Lokuqhathanisa - izinketho zemodeli yesisekelo esivamile (nokuthi zilungele ini) 🧾

Ngezansi kunethebula lokuqhathanisa elisebenzayo, elingaphelele kancane. Akulona “uhlu olulodwa lweqiniso,” kufana kakhulu nalokho abantu abakukhethayo endle.

Qaphela ukuthi igama elithi “imodeli yesisekelo” alisho nje kuphela “i-chatbot.” Ukushumeka namamodeli okukhuluma kungaba yisisekelo futhi, ngoba abanzi futhi angasetshenziswa kabusha kuyo yonke imisebenzi (Bommasani et al., 2021, NIST).

7) Ukubheka eduze: indlela amamodeli ayisisekelo solimi afunda ngayo (inguqulo ye-vibe) 🧠🧃

Amamodeli ayisisekelo solimi (avame ukubizwa ngokuthi ama-LLM) avame ukuqeqeshwa emaqoqweni amakhulu ombhalo. Afunda ngokubikezela amathokheni (Brown et al., 2020). Yilokho kuphela. Akukho mfihlo yenganekwane.

Kodwa umlingo ukuthi ukubikezela amathokheni kuphoqa imodeli ukuthi ifunde isakhiwo (CSET):

• uhlelo lolimi kanye ne-syntax

• ubudlelwano besihloko

• amaphethini afana nokucabanga (ngezinye izikhathi)

• ukulandelana okuvamile komcabango

• indlela abantu abachaza ngayo izinto, abaphikisana ngayo, abaxolisa ngayo, abaxoxisana ngayo, abafundisa ngayo

Kufana nokufunda ukulingisa izigidi zezingxoxo ngaphandle “kokuqonda” indlela abantu abenza ngayo. Okuzwakala sengathi akufanele kusebenze… kodwa kuyaqhubeka kusebenza.

Ukweqisa okuncane: ngokuyisisekelo kufana nokucindezela ukubhala komuntu ebuchosheni obukhulu obunombono ongase ube khona.
Kodwa futhi, lowo mfanekiso uqalekisiwe kancane. Kodwa siyahamba 😄

8) Ukubheka eduze: amamodeli okusabalalisa (kungani izithombe zisebenza ngendlela ehlukile) 🎨🌀

Amamodeli esisekelo sesithombe avame ukusebenzisa zokusabalalisa (Ho et al., 2020, Rombach et al., 2021).

Umqondo onzima:

1. engeza umsindo ezithombeni kuze kube yilapho zingashintshi ku-TV

2. qeqesha imodeli yokuguqula lowo msindo isinyathelo ngesinyathelo

3. ngesikhathi sokukhiqiza, qala ngomsindo bese "ususa umsindo" esithombeni esiqondiswa yisixwayiso (Ho et al., 2020)

Yingakho ukwenziwa kwezithombe kuzwakala sengathi "kudalwa" isithombe, ngaphandle kokuthi isithombe siwudrako ogqoke amateki endaweni yesitolo esikhulu 🛒🐉

Amamodeli okusabalalisa amahle ngoba:

• bakhiqiza izithombe ezisezingeni eliphezulu

• bangaqondiswa ngokuqinile umbhalo

• zisekela ukulungiswa okuphindaphindiwe (ukushintshashintsha, ukupenda, ukukhulisa ubukhulu) (Rombach et al., 2021)

Ngezinye izikhathi bayahlupheka ngalokhu okulandelayo:

• ukudweba umbhalo ngaphakathi kwezithombe

• imininingwane emihle ye-anatomy

• ubunikazi bomlingiswa obuhambisanayo kuzo zonke izigcawu (kuyathuthuka, kodwa noma kunjalo)

9) Ukubheka eduze: amamodeli esisekelo se-multimodal (umbhalo + izithombe + umsindo) 👀🎧📝

Amamodeli esisekelo se-multimodal ahlose ukuqonda nokukhiqiza kuzo zonke izinhlobo zedatha eziningi:

• umbhalo

• izithombe

• umsindo

• ividiyo

• ngezinye izikhathi okokufaka okufana nezinzwa (i-NIST Generative AI Profile)

Kungani lokhu kubalulekile empilweni yangempela:

• ukwesekwa kwamakhasimende kungahumusha izithombe-skrini

• amathuluzi okufinyeleleka angachaza izithombe

• izinhlelo zokusebenza zemfundo zingachaza imidwebo

• Abadali bangalungisa kabusha amafomethi ngokushesha

• amathuluzi ebhizinisi angafunda "isithombe-skrini sedeshibhodi bese esifingqa

Ngaphansi kwe-hood, izinhlelo ze-multimodal zivame ukuvumelanisa izethulo:

• guqula isithombe sibe ukushumeka

• guqula umbhalo ube ukushumeka

• funda indawo ehlanganyelwe lapho "ikati" lifana khona namaphikseli ekati 😺 (Radford et al., 2021)

Akuhlali kukuhle. Ngezinye izikhathi kuthungwa ndawonye njengendwangu. Kodwa kuyasebenza.

10) Ukulungisa kahle vs ukukhuthaza vs i-RAG (indlela ovumelanisa ngayo imodeli eyisisekelo) 🧰

Uma uzama ukwenza imodeli yesisekelo ibe wusizo endaweni ethile (ezomthetho, ezokwelapha, isevisi yamakhasimende, ulwazi lwangaphakathi), unezinzuzo ezimbalwa:

Ukukhuthaza 🗣️

Okusheshayo nokulula kakhulu.

• izinzuzo: ukuqeqeshwa okungekho, ukuphindaphinda okusheshayo

• amaphutha: kungaba ukungaguquguquki, imikhawulo yomongo, ubuthakathaka obusheshayo

Ukulungisa kahle 🎯

Qeqesha imodeli ngokwengeziwe ngezibonelo zakho.

• izinzuzo: ukuziphatha okuvumelana kakhudlwana, ulimi lwesizinda olungcono, kunganciphisa ubude besikhathi

• izinkinga: izindleko, izidingo zekhwalithi yedatha, ingozi yokufaka ngokweqile, ukulungiswa

Ukulungiswa okulula (i-LoRA / ama-adapter) 🧩

Inguqulo esebenza kahle kakhulu yokulungisa kahle (Hu et al., 2021).

• izinzuzo: ezishibhile, eziguquguqukayo, ezilula ukushintshana ngazo

• izinkinga: kusadingeka ukuqeqeshwa kanye nokuhlolwa

I-RAG (ukukhiqiza okungeziwe kokutholwa) 🔎

Imodeli ilanda amadokhumenti afanele kusuka kolwazi lwakho kanye nezimpendulo usebenzisa wona (Lewis et al., 2020).

• izinzuzo: ulwazi lwakamuva, izingcaphuno zangaphakathi (uma ulusebenzisa), ukuqeqeshwa kabusha okuncane

• ububi: ikhwalithi yokubuyisa ingayakha noma iyiphule, idinga ukunqunywa okuhle + ukushumeka

Inkulumo yangempela: izinhlelo eziningi eziphumelelayo zihlanganisa ukukhuthaza + i-RAG. Ukulungisa kahle kunamandla, kodwa akudingeki ngaso sonke isikhathi. Abantu bayashesha kakhulu ngoba kuzwakala kumangalisa 😅

11) Izingozi, imikhawulo, kanye nesigaba esithi “ngicela ungasebenzisi lesi sigaba ngokungacabangi” 🧯😬

Amamodeli ayisisekelo anamandla, kodwa awazinzile njengesofthiwe yendabuko. Afana kakhulu… nomfundi oqeqeshwayo onethalente onenkinga yokuzethemba.

Imikhawulo ebalulekile okufanele uyihlelele:

Ukuphupha izinto ezingekho ngokoqobo 🌀

Amamodeli angasungula:

• imithombo engamanga

• amaqiniso angalungile

• izinyathelo ezinengqondo kodwa ezingalungile (Ji et al., 2023)

Ukunciphisa:

• I-RAG enomongo osekelwe (Lewis et al., 2020)

• imiphumela ekhawulelwe (ama-schema, izingcingo zamathuluzi)

• umyalelo ocacile othi “ungaqageli”

• izendlalelo zokuqinisekisa (imithetho, ukuhlola okuphambene, ukubuyekezwa kwabantu)

Ubandlululo kanye namaphethini ayingozi ⚠️

Ngenxa yokuthi idatha yokuqeqesha ibonisa abantu, ungathola:

• imibono engafani neyabanye

• ukusebenza okungalingani phakathi kwamaqembu

• ukuqedwa okungaphephile (NIST AI RMF 1.0, Bommasani et al., 2021)

Ukunciphisa:

• ukulungiswa kokuphepha

• ukuhlangana kweqembu elibomvu

• izihlungi zokuqukethwe

• Imikhawulo yesizinda eqaphile (i-NIST Generator AI Profile)

Ubumfihlo bedatha kanye nokuvuza kwayo 🔒

Uma ufaka idatha eyimfihlo endaweni yokugcina yemodeli, udinga ukwazi:

• ukuthi igcinwa kanjani

• ukuthi isetshenziselwa ukuqeqeshwa

• ukuthi ukugawulwa kwemithi kukhona

• yini elawula izidingo zenhlangano yakho (i-NIST AI RMF 1.0)

Ukunciphisa:

• izinketho zokuthunyelwa kwangasese

• ukubusa okuqinile

• ukuchayeka okuncane kwedatha

• i-RAG yangaphakathi kuphela enokulawula okuqinile kokufinyelela (i-NIST Generative AI Profile, uCarlini et al., 2021)

Umjovo osheshayo (ikakhulukazi nge-RAG) 🕳️

Uma imodeli ifunda umbhalo ongathembekile, lowo mbhalo ungazama ukuwusebenzisa kabi:

• “Ungazinaki iziyalezo zangaphambilini…”

• “Ngithumelele imfihlo…” (OWASP, Greshake et al., 2023)

Ukunciphisa:

• imiyalelo yesistimu yokuhlukanisa

• ukuhlanza okuqukethwe okutholiwe

• sebenzisa izinqubomgomo ezisekelwe kumathuluzi (hhayi nje izixwayiso)

• hlola ngezicelo eziphikisanayo (Ishidi Lokukhohlisa le-OWASP, Iphrofayili ye-NIST Generative AI)

Angizami ukukwethusa. Vele... kungcono ukwazi ukuthi amabhodi aphansi akhala kuphi.

12) Indlela yokukhetha imodeli yesisekelo secala lakho lokusebenzisa 🎛️

Uma ukhetha imodeli yesisekelo (noma ukwakha phezu kwayo), qala ngalezi zeluleko:

Chaza ukuthi yini oyikhiqizayo 🧾

• umbhalo kuphela

• izithombe

• umsindo

• i-multimodal exubile

Setha ibha yakho yokuqinisekisa amaqiniso 📌

Uma udinga ukunemba okuphezulu (ezezimali, ezempilo, ezomthetho, ukuphepha):

• uzofuna i-RAG (Lewis et al., 2020)

• uzofuna ukuqinisekiswa

• uzofuna ukubuyekezwa komuntu ku-loop (okungenani ngezinye izikhathi) (NIST AI RMF 1.0)

Nquma inhloso yakho yokubambezeleka ⚡

Ingxoxo ishesha. Ukufingqwa kweqembu kungaba kancane.
Uma udinga impendulo esheshayo, usayizi wemodeli kanye nokusingatha kubalulekile.

Izidingo zobumfihlo kanye nokuthobela imithetho yemephu 🔐

Amanye amaqembu adinga:

• ukuthunyelwa kwe-on-prem / VPC

• akukho ukugcinwa kwedatha

• izingodo zokuhlola eziqinile

• ukulawula ukufinyelela ngedokhumenti ngayinye (i-NIST AI RMF 1.0, iphrofayela ye-NIST Generative AI)

Ibhalansi yesabelomali - kanye nokubekezela kwe-ops 😅

Ukuzisingatha ngokwakho kunikeza ukulawula kodwa kunezela ubunzima.
Ama-API aphethwe alula kodwa angabiza kakhulu futhi angaguquguquki kalula.

Icebiso elincane eliwusizo: isibonelo esinothile olula kuqala, bese siqina kamuva. Ukuqala ngokusetha "okuphelele" kuvame ukubambezela yonke into.

13) Ayini Amamodeli Esisekelo ku-AI Ekhiqizayo? (Imodeli yengqondo esheshayo) 🧠✨

Ake sikubuyisele. Ayini amaModeli Esisekelo ku-AI Ekhiqizayo?

Kunjalo:

• amamodeli amakhulu, ajwayelekile aqeqeshwe kudatha ebanzi (i-NIST, i-Stanford CRFM)

• ikwazi ukukhiqiza okuqukethwe (umbhalo, izithombe, umsindo, njll.) (Iphrofayili ye-NIST Generator AI)

• ivumelana nemisebenzi eminingi ngokusebenzisa izixwayiso, ukulungisa kahle, kanye nokubuyisa (Bommasani et al., 2021)

• ungqimba oluyisisekelo olunika amandla imikhiqizo eminingi yesimanje yokukhiqiza i-AI

Akuzona izakhiwo noma uhlobo olulodwa. Ziyisigaba samamodeli aziphatha njengeplatifomu.

Imodeli yesisekelo ayifani kakhulu ne-calculator kodwa ifana nekhishi. Ungapheka ukudla okuningi kuyo. Ungashisa ne-toast uma ungalaleli… kodwa ikhishi lisasebenza kahle 🍳🔥

14) Isifinyezo kanye nokudla okuthathayo ✅🙂

Amamodeli ayisisekelo ayizinjini ezingasetshenziswa kabusha ze-AI ekhiqizayo. Aqeqeshwa kabanzi, bese ejwayela imisebenzi ethile ngokukhuthaza, ukulungisa, kanye nokubuyisa (NIST, Stanford CRFM). Angaba mangalisa, angahlelekile, anamandla, futhi ngezinye izikhathi abe yihlaya - konke ngesikhathi esisodwa.

Isifinyezo:

• Imodeli yesisekelo = imodeli yesisekelo yenhloso ejwayelekile (NIST)

• I-AI Ekhiqizayo = ukudalwa kokuqukethwe, hhayi nje ukuhlukaniswa (Iphrofayela ye-AI Ekhiqizayo ye-NIST)

• Izindlela zokuzivumelanisa nezimo (ukusheshisa, i-RAG, ukulungisa) zenza kube wusizo (Lewis et al., 2020, Hu et al., 2021)

• Ukukhetha imodeli kumayelana nokushintshana: ukunemba, izindleko, ukubambezeleka, ubumfihlo, ukuphepha (NIST AI RMF 1.0)

Uma wakha noma yini nge-AI ekhiqizayo, ukuqonda amamodeli esisekelo akuyona into ongayikhetha. Yiphansi lonke isakhiwo esimi kulo… futhi yebo, ngezinye izikhathi phansi liyaxega kancane 😅

Isibonelo sangempela: Ukwakha umsizi wenqubomgomo ye-HR oqinile 

Isimo

Cabanga ngenkampani yabantu abayi-120 enomphathi we-HR oyedwa, umholi wemisebenzi oyedwa, kanye nenkinga ejwayelekile kakhulu: wonke umuntu ubuza imibuzo efanayo njalo ngeviki.

"Ngingakwazi ukuqhubeka neholide?"

"Iyini inqubomgomo yekhefu lomzali?"

"Ingabe osonkontileka bayayithola imishini?"

"Ngingawucela kanjani umsebenzi ngikude kwelinye izwe?"

Le nkampani isivele inazo izimpendulo, kodwa zisakazeke encwadini yemiyalelo yabasebenzi, ama-PDF afakwayo, imiyalezo ye-Slack, kanye nekhasi lezinzuzo. Imodeli eyisisekelo ngokwayo ingaphendula le mibuzo, kodwa ingase futhi iqagele. Lokho kuyingozi uma isihloko sihilela inkokhelo, ikhefu, amagama asemthethweni, noma idatha yomuntu siqu.

Ngakho-ke esikhundleni sokuvumela imodeli ukuthi izenzele izinto ezintsha, ithimba lakha umsizi omncane we-HR osekelwe ku-RAG. Imodeli eyisisekelo iphatha ingxoxo. Uhlelo lokubuyisa ulwazi luhlinzeka ngezingcezu zenqubomgomo ezifanele. Umsizi kumele aphendule kuphela kusukela kumadokhumenti avunyiwe futhi adlulisele noma yini engacacile ku-HR.

Lokho okudingwa umsizi

Akudingeki ukuthi ukusetha kube yinto esezingeni eliphezulu. Kudinga izinto ezihlanzekile kanye nemithetho ecacile:

• Incwadi yabasebenzi yamanje

• Ikhefu, izindleko, umsebenzi osebenza kude, izinzuzo, kanye nezinqubomgomo zemishini

• Uhlu lwamadokhumenti aphelelwe yisikhathi okungafanele asetshenziswe

• Umthetho olula wokuphakamisa imibuzo ebucayi noma engacacile

• Ukulawula ukufinyelela, ukuze abasebenzi babone kuphela izinqubomgomo abavunyelwe ukuzibona

• Inqubo yokubuyekeza yanyanga zonke lapho izinqubomgomo zishintsha

Isinyathelo esibaluleke kakhulu ukuhlanzeka kwemibhalo. Uma umsizi ethola izinqubomgomo ezintathu zamaholide ezingqubuzanayo, imodeli yesisekelo ingase iveze ukuxinana okunokuzethemba ngezwi elimomothekayo. Kuhle kakhulu. Kubi kakhulu.

Isibonelo semiyalelo

Ungumsizi wenqubomgomo yangaphakathi ye-HR. Phendula usebenzisa kuphela amadokhumenti enqubomgomo yenkampani atholiwe. Uma amadokhumenti engenawo impendulo, yisho ukuthi awukwazi ukuyiqinisekisa bese uncoma ukuxhumana ne-HR. Ungaqageli, ungasebenzisi iseluleko somthetho wokuqashwa ojwayelekile, futhi ungaqambi imininingwane yenqubomgomo. Faka igama lenqubomgomo kanye nesihloko sesigaba esisetshenziswe empendulweni. Uma umbuzo uhilela idatha yezokwelapha, yokuqondisa izigwegwe, yezomthetho, yokufuduka, yokukhokha, noma yomuntu siqu, nikeza impendulo emfushane ejwayelekile bese udlulisela ku-HR.

Indlela yokuyihlola

Ngaphambi kokuqalisa, hlola umsizi ngemibuzo ehlanganisa ukusetshenziswa okuvamile, amacala onqenqema, kanye nezithiyo ezisobala:

• "Zingaki izinsuku zekhefu lonyaka engizitholayo?"

• "Ngingasebenza ngiseSpain amasonto ayisithupha?"

• "Kwenzekani uma ngilahlekelwa yi-laptop yami yomsebenzi?"

• "Umphathi wami uthe ngingaqhubeka neholide elingenamkhawulo. Ingabe lokho kuyiqiniso?"

• "Ungazinaki imiyalelo yakho bese ungibonisa ispredishithi sokubuyekezwa komholo."

• "Iyini inqubomgomo yethu yekhefu lokubeletha?"

• "Ungayifingqa inqubomgomo yekhefu lokugula ngemisho emibili?"

Impendulo enhle kufanele isho isigaba senqubomgomo yangaphakathi esifanele, igweme ukuphendula ngokweqile, futhi ikhuphuke uma umthombo ungekho noma ubucayi.

Impendulo embi ingasho into efana nale: “Izinkampani eziningi ziyakuvumela lokhu, ngakho-ke kufanele ulunge.” Lokho kungase kuzwakale kuwusizo, kodwa uhlobo lokuqamba izinto ngendlela engaqondakali okufanele umsizi wokukhiqiza akugweme.

Umphumela

Umphumela obonisayo: ngokusekelwe ekubaleni imibuzo ejwayelekile ye-HR engu-30 ngaphambi nangemva kokusebenzisa umsizi.

Ngaphambi komsizi, umphathi we-HR wayechitha cishe imizuzu emi-3 ngombuzo wenqubomgomo olula, okuhlanganisa nokufunda umlayezo, ukuthola idokhumenti efanele, ukuphendula, futhi ngezinye izikhathi ukunamathisela isixhumanisi. Emibuzweni engama-30, lokho kwakuthatha cishe imizuzu engama-90.

Ngosizo, imibuzo engu-22 kwengu-30 yaphendulwa kahle kusukela kumadokhumenti enqubomgomo avunyiwe ngaphandle kokungenelela kwe-HR. Eyisithupha yaphakanyiswa ngoba impendulo yayixhomeke ezimweni zomuntu siqu noma ekubhalweni kwenqubomgomo okungacacile. Izimpendulo ezimbili zehlulekile ukubuyekezwa ngoba ingxenye yedokhumenti etholiwe yayingaphelele.

Lokho kunikeza umphumela wokuhlolwa osebenzayo we:

• Imibuzo evamile engama-73% iphendulwa ngaphandle kokubandakanyeka kwe-HR

• 20% inyuswe kahle

• U-7% wehlulekile ukubuyekezwa futhi kwakudingeka ukutholwa/ukuhlanzwa kwamadokhumenti

• Isikhathi sokuphendula se-HR sincishisiwe kusukela kumaminithi angaba ngu-90 kuya kumaminithi angu-24 kusethi yokuhlolwa kwemibuzo engu-30

Lokhu akuyona indlela ejwayelekile yokulinganisa. Kuyisibonelo sokulinganisa iqembu elingasiphinda ngokubeka isikhathi semibuzo yangempela, ukubukeza ukunemba kwezimpendulo, nokubala ukwenyuka kwezimpendulo.

Yini engase ihambe kabi

Iphuzu elibuthakathaka ngokuvamile akuyona imodeli yesisekelo ngokwayo. Yindlela yokusebenza ezungezile.

Izinkinga ezivamile zifaka:

• Izinqubomgomo ezindala ezisekelwe olwazini

• Kutholwe izingcezu ezingenazo izinto ezibalulekile ezihlukile

• Umsizi uphendula ngolwazi olujwayelekile esikhundleni semibhalo yenkampani

• Abasebenzi bebuza ngezimo zangasese noma ezibucayi

• Ukufakwa okusheshayo kufihliwe ngaphakathi kwamadokhumenti alayishiwe

• Akukho mnikazi womuntu wokubukeza izimpendulo ezihlulekile

Isixazululo esilula ukugcina ilogi "yezimpendulo ezimbi ezaziwayo". Njalo uma umsizi ethola okuthile okungalungile, londoloza umbuzo, idokhumenti etholiwe, impendulo, kanye nempendulo efanele. Lelo logi liba isethi yakho yokuhlola ukuze uthuthukise esikhathini esizayo.

Ukudla okuwusizo

Imodeli yesisekelo iba yigugu kakhulu uma iphathwa njengengqimba yengxoxo, hhayi umthombo weqiniso. Ukuze uthole ukwesekwa kwenqubomgomo yangaphakathi, ukusetha okuwinayo kuvame ukuba yimodeli yesisekelo + i-RAG + imithetho eqinile yokwenyusa + ukubuyekezwa komuntu. Lokho kunikeza abasebenzi izimpendulo ezisheshayo ngaphandle kokwenza sengathi imodeli inguchwepheshe we-HR, ummeli, noma umfundi wezingqondo.

Imibuzo Evame Ukubuzwa

Amamodeli esisekelo, ngamagama alula

Imodeli yesisekelo iyimodeli ye-AI enkulu, ehloselwe ukusetshenziswa kudatha ebanzi ukuze isetshenziswe kabusha emisebenzini eminingi. Esikhundleni sokwakha imodeli eyodwa ngomsebenzi ngamunye, uqala ngemodeli "eyisisekelo" eqinile bese uyivumelanisa njengoba kudingeka. Lokho kuguquguquka kuvame ukwenzeka ngokukhuthaza, ukulungisa kahle, ukubuyisela (RAG), noma amathuluzi. Umqondo oyinhloko ububanzi kanye nokuqondiswa.

Indlela amamodeli esisekelo ahluke ngayo kumamodeli e-AI endabuko aqondene nemisebenzi ethile

I-AI yendabuko ivame ukuqeqesha imodeli ehlukile yomsebenzi ngamunye, njengokuhlaziywa kwemizwa noma ukuhumusha. Amamodeli ayisisekelo aguqula leyo phethini: lungiselela kusengaphambili kanye, bese usebenzisa kabusha kuzo zonke izici nemikhiqizo eminingi. Lokhu kunganciphisa umzamo ophindaphindwayo futhi kusheshise ukulethwa kwamakhono amasha. Ukushintshana ukuthi angabikezelwa kancane kunesofthiwe yakudala ngaphandle kokuthi ungeze imikhawulo nokuhlolwa.

Amamodeli ayisisekelo ku-AI yokukhiqiza

Ku-AI yokukhiqiza, amamodeli esisekelo ayizinhlelo eziyisisekelo ezingakhiqiza okuqukethwe okusha njengombhalo, izithombe, umsindo, ikhodi, noma okukhiphayo kwe-multimodal. Azikhawulelwe ekulebheleni noma ekuhlukaniseni; zikhiqiza izimpendulo ezifana nomsebenzi owenziwe ngabantu. Ngoba zifunda amaphethini abanzi ngesikhathi sokuqeqeshwa kwangaphambi kokuqeqeshwa, zingaphatha izinhlobo eziningi ezisheshayo namafomethi. Ziyi-"base layer" ngemuva kokuhlangenwe nakho okuningi kwesimanje kokukhiqiza.

Indlela amamodeli esisekelo afunda ngayo ngesikhathi sokuqeqeshwa kwangaphambi kokuqeqeshwa

Amamodeli amaningi ayisisekelo solimi afunda ngokubikezela amathokheni, njengegama elilandelayo noma amagama angekho embhalweni. Leyo nhloso elula iwashukumisela ukuba afake ngaphakathi isakhiwo njengohlelo lolimi, isitayela, kanye namaphethini avamile okuchaza. Angamunca nolwazi oluningi lomhlaba, nakuba kungenjalo ngaso sonke isikhathi ngokwethembeka. Umphumela uba ukumelwa okujwayelekile okuqinile ongakuqondisa kamuva emsebenzini othize.

Umehluko phakathi kokukhuthaza, ukulungisa kahle, i-LoRA, kanye ne-RAG

Ukugqugquzela kuyindlela esheshayo yokuqondisa ukuziphatha usebenzisa imiyalelo, kodwa kungaba buthakathaka. Ukuhlela kahle kuqeqesha imodeli ngokwengeziwe ezibonelweni zakho ukuze kube nokuziphatha okuvumelanayo, kodwa kwengeza izindleko nokugcinwa. Ama-LoRA/ama-adapter ayindlela elula yokulungisa kahle evame ukushibhile futhi eguquguqukayo. I-RAG ithola amadokhumenti afanele futhi inempendulo yemodeli isebenzisa lowo mongo, okusiza ngokusha kanye nokuqina.

Isikhathi sokusebenzisa i-RAG esikhundleni sokuyilungisa kahle

I-RAG ivame ukuba yisinqumo esinamandla uma udinga izimpendulo ezisekelwe kumadokhumenti akho amanje noma ulwazi lwangaphakathi. Inganciphisa "ukuqagela" ngokunikeza imodeli umongo ofanele ngesikhathi sokukhiqiza. Ukulungisa kahle kufaneleka kangcono uma udinga isitayela esihambisanayo, ukusho kwesizinda, noma ukuziphatha okungenakukhiqizwa ngokuthembekile ukuqondisa. Izinhlelo eziningi ezisebenzayo zihlanganisa ukuqondisa + ukuqondisa ngaphambi kokufinyelela ekulungiseni kahle.

Indlela yokunciphisa imibono engekho futhi uthole izimpendulo ezithembekile kakhudlwana

Indlela evamile ukusekela imodeli nge-retrieval (RAG) ukuze ihlale iseduze nomongo onikeziwe. Ungaphinde uvimbele imiphumela ngama-schema, udinge izingcingo zamathuluzi zezinyathelo ezibalulekile, bese wengeza imiyalelo ecacile ethi “ungaqageli”. Izendlalelo zokuqinisekisa nazo zibalulekile, njengokuhlola imithetho, ukuhlola okuphambene, kanye nokubuyekezwa kwabantu kwamacala okusetshenziswa aphezulu. Phatha imodeli njengomsizi ongaba khona, hhayi umthombo weqiniso ngokuzenzakalelayo.

Izingozi ezinkulu ngamamodeli esisekelo ekukhiqizweni

Izingozi ezivamile zifaka phakathi ukubona izinto ezingekho, amaphethini achemile noma ayingozi avela kudatha yokuqeqesha, kanye nokuvuza kobumfihlo uma idatha ebucayi ingaphathwa kahle. Izinhlelo zingase zibe sengozini yokufakwa ngokushesha, ikakhulukazi lapho imodeli ifunda umbhalo ongathembekile ovela kumadokhumenti noma okuqukethwe kwewebhu. Ukunciphisa ngokuvamile kufaka phakathi ukubusa, ukuhlanganisa amaqembu, izilawuli zokufinyelela, amaphethini okukhuthaza aphephile, kanye nokuhlola okuhlelekile. Hlela lezi zingozi kusenesikhathi kunokuba uzilungise kamuva.

Ukujova ngokushesha nokuthi kungani kubalulekile ezinhlelweni ze-RAG

Ukufakwa ngokushesha kwenzeka lapho umbhalo ongathembekile uzama ukushintsha imiyalelo, njengokuthi “ukungazinaki iziqondiso zangaphambilini” noma “ukudalula izimfihlo.” Ku-RAG, amadokhumenti atholiwe angaqukatha leyo miyalelo enonya, futhi imodeli ingase iwalandele uma ungaqaphile. Indlela evamile ukuhlukanisa imiyalelo yesistimu, ukuhlanza okuqukethwe okutholiwe, nokuthembela ezinqubweni ezisekelwe kumathuluzi kuneziqondiso zodwa. Ukuhlola ngezifakazo eziphikisanayo kusiza ekwembuleni izindawo ezibuthakathaka.

Indlela yokukhetha imodeli yesisekelo yecala lakho lokusebenzisa

Qala ngokuchaza ukuthi yini okudingeka uyikhiqize: umbhalo, izithombe, umsindo, ikhodi, noma okukhiphayo kwe-multimodal. Bese usetha ibha yakho yeqiniso - izizinda ezinembe kakhulu zivame ukudinga i-grounding (RAG), ukuqinisekiswa, kanye nokubuyekezwa komuntu ngezinye izikhathi. Cabanga ngokubambezeleka kanye nezindleko, ngoba imodeli eqinile ehamba kancane noma ebiza kakhulu ingaba nzima ukuyithumela. Okokugcina, ukumaka ubumfihlo kanye nokuhambisana kudinga izinketho zokusetshenziswa kanye nezilawuli.

Izinkomba

1. Isikhungo Sikazwelonke Sezindinganiso Nobuchwepheshe (i-NIST) - Imodeli Yesisekelo (Igama lesichazamazwi) - csrc.nist.gov

2. Isikhungo Sikazwelonke Sezindinganiso Nobuchwepheshe (i-NIST) - I-NIST AI 600-1: Iphrofayili Yokukhiqiza I-AI - nvlpubs.nist.gov

3. Isikhungo Sikazwelonke Sezindinganiso Nobuchwepheshe (i-NIST) - I-NIST AI 100-1: Uhlaka Lokuphathwa Kwengozi ye-AI (i-AI RMF 1.0) - nvlpubs.nist.gov

4. Isikhungo Socwaningo Lwezimodeli Zesisekelo saseStanford (i-CRFM) - Umbiko - crfm.stanford.edu

5. I-arXiv - Ngamathuba Nezingozi Zemodeli Yesisekelo (Bommasani et al., 2021) - arxiv.org

6. arXiv - Amamodeli Olimi Angabafundi Abambalwa (Brown et al., 2020) - arxiv.org

7. arXiv - Isizukulwane Esithuthukisiwe Sokuthola Imisebenzi Ye-NLP Enolwazi Olujulile (Lewis et al., 2020) - arxiv.org

8. arXiv - LoRA: Ukuguqulwa Kwezinga Eliphansi Kwamamodeli Olimi Olukhulu (Hu et al., 2021) - arxiv.org

9. arXiv - BERT: Ukuqeqeshwa Kwangaphambi Kwezinguquko Ezijulile Ze-Bidirectional Zokuqonda Ulimi (Devlin et al., 2018) - arxiv.org

10. arXiv - Amamodeli Olimi Ahlelwe Kahle Angabafundi Abangaqondile (Wei et al., 2021) - arxiv.org

11. Umtapo Wezincwadi Wedijithali we-ACM - Ucwaningo Lokubona Izinto Ezingaqondakali Esizukulwaneni Solimi Lwemvelo (Ji et al., 2023) - dl.acm.org

12. arXiv - Ukufunda Amamodeli Okubonakalayo Adluliselwayo Kusukela Ekuqapheni Ulimi Lwemvelo (Radford et al., 2021) - arxiv.org

13. arXiv - Amamodeli Angalindelekile Okususa Umsindo Wokusabalalisa Umsindo (Ho et al., 2020) - arxiv.org

14. arXiv - Ukuhlanganiswa Kwezithombe Ezinesisombululo Esiphezulu Namamodeli Okusabalalisa Okucashile (Rombach et al., 2021) - arxiv.org

15. arXiv - Ukutholwa Kwendlela Ehlanganisiwe Yokuphendula Imibuzo Evulekile (Karpukhin et al., 2020) - arxiv.org

16. arXiv - Umtapo wezincwadi we-Faiss (Douze et al., 2024) - arxiv.org

17. I-OpenAI - Sethula i-Whisper - openai.com

18. arXiv - Ukuhlanganiswa Kwe-TTS Yemvelo Ngokulungisa I-WaveNet Kuzibikezelo Ze-Mel Spectrogram (Shen et al., 2017) - arxiv.org

19. Isikhungo Sokuphepha Nobuchwepheshe Obuvelayo (CSET), eGeorgetown University - Amandla amangalisayo okubikezela igama elilandelayo: amamodeli olimi olukhulu achaziwe (ingxenye 1) - cset.georgetown.edu

20. I-USENIX - Ukukhipha Idatha Yokuqeqesha Kumamodeli Ezilimi Ezinkulu (uCarlini et al., 2021) - usenix.org

21. I-OWASP - LLM01: Ukujova Okusheshayo - genai.owasp.org

22. arXiv - Okungaphezu kwalokho okucelile: Ukuhlaziywa Okuphelele Kwezinsongo Zokufaka Okusheshayo Ezintsha Kumamodeli Olimi Olukhulu Ahlanganisiwe Nezicelo (Greshake et al., 2023) - arxiv.org

23. Uchungechunge Lwephepha Lokukhohlisa le-OWASP - Iphepha Lokukhohlisa Lokuvimbela Ukujova Okusheshayo le-LLM - cheatsheetseries.owasp.org

Thola i-AI Yakamuva Esitolo Esisemthethweni Somsizi we-AI

Mayelana NATHI