- Incazelo yamamodeli ayisisekelo njengezinhlelo eziqeqeshwe ngezinga ezisebenza njengesisekelo semisebenzi eminingi ethile.
- Ukuhlaziywa kokwakheka kwe-Transformer kanye nenqubo yokuziqeqesha ngaphambi kokuqeqeshwa kanye nokuzilungisa kahle.
- Ukuhlolwa komthelela womphakathi, izingozi zokuziphatha, kanye nezinselele zokubala zokusetshenziswa kwayo okukhulu.
Uma ubulokhu ulandela impikiswano yakamuva mayelana nobuhlakani bokwenziwa, kungenzeka ukuthi uke wezwa ngamathuluzi abonakala enza konke kusukela ekubhaleni izinkondlo kuya ekuhleleni ikhodi eyinkimbinkimbi. Ngemuva kwalo lonke lomlingo kukhona umqondo oyinhloko: amamodeli ayisisekelo . Empeleni, afana ne-chassis yemoto; isakhiwo esiqinile, esijwayelekile esingabe sesishintshwa sibe yimoto yezemidlalo, iloli, noma imoto yosizo, kuye ngokuthi yini esiyidingayo nganoma yisiphi isikhathi.
Okwenza lezi zinhlelo ziphazamiseke kangaka ukuthi asisaklami i-AI yomsebenzi owodwa, kodwa kunalokho sakha i- giant yedijithali eqeqeshwe ngenani elikhulu ledatha esingayibumba "bese siyibumba." Lokhu kushintsha kwendlela kuvumele i-AI ukuthi iguquke isuke kokuthile okuqinile kakhulu iye kuthuluzi eliguquguqukayo elifunda amaphethini ajwayelekile emhlabeni bese liwasebenzisa ezinkingeni ezithile, lisheshise ukusungula izinto ezintsha ngesivinini, empeleni, esishiya wonke umuntu engenalutho.
Iyini ngempela imodeli eyisisekelo?
Kalula nje, imodeli ye-AI eyisisekelo yinoma yiluphi uhlelo lwe-AI oluqeqeshwe ngenani elikhulu ledatha (ngokuvamile ngokuziqapha ngokwakho) olungavumelanisa, ngenqubo ebizwa ngokuthi i-fine-tuning, nezinhlobo eziningi zezicelo. Azintsha ngokuphelele, njengoba zisebenzisa ukufunda okujulile nokudluliswa kolwazi, kodwa isilinganiso esingakaze sibonwe sedatha namandla ekhompyutha siholele ekuveleni kwamakhono okungekho muntu owayewalindele.
Kubalulekile ukungawadidanisi nama- Large Language Models (LLMs) . Nakuba evame ukusetshenziswa ngokushintshana, ubudlelwano bungobohlobo nohlobo: wonke ama-LLM ayizibonelo eziyisisekelo, kodwa akuwona wonke ama-LLM ayisisekelo. Ngenkathi ama-LLM egxila embhalweni nasekhodini, ama-basement model angaba yi -multimodal , acubungule izithombe, umsindo, ividiyo, noma ngisho nezimpawu zokuzwa ezivela kumarobhothi.
Izinsika zobuchwepheshe: I-Transformer nokuqeqeshwa
Ukwakheka okwenze konke lokhu kwaba nokwenzeka yi- Transformer . Lolu hlelo lusebenzisa indlela ebizwa ngokuthi "ukuzinaka," evumela imodeli ukuqonda ukubaluleka kwezingxenye ezahlukene zochungechunge, kungakhathaliseki ukuthi ziqhelelene kangakanani. Ngokuvamile luqukethe i-encoder kanye ne-decoder ekhiqiza izethulo ze-vector (ukushumeka) ezibamba incazelo eyisisekelo yolimi noma izithombe.
Inqubo yokudala ivame ukuhlukaniswa ngezigaba eziningana:
- Ngaphambi kokujima: Lesi yisigaba esibiza kakhulu. Imodeli "idla" i-inthanethi, izincwadi, kanye nezizindalwazi ukuze ifunde isakhiwo esijwayelekile solwazi. Yilapho-ke inani lamapharamitha, inani lamathokheni kanye nefasitela lomongo.
- Ukuhleleka Kwayo: Uma imodeli isijwayelekile, iqeqeshwa ngemininingwane ethile futhi iqondiswa abantu ukuze ibe uchwepheshe endaweni ethile, njengezokwelapha noma umthetho.
- Ukuzivumelanisa nomsebenzi: Lesi yisinyathelo sokugcina, lapho imodeli ilungiselelwe khona umsebenzi othize kakhulu, njengokudala injini yokusesha yomthetho wamacala noma umkhiqizi wemibiko yobuchwepheshe.
Amamodeli avelele ashintshe umlando
Kusukela ngo-2018, sibone uchungechunge lwamamodeli ashintshe izinto. U-BERT wayengomunye wabasunguli, owaziwa ngekhono lakhe lokuqonda umongo ngezindlela ezimbili. Kwabe sekufika umndeni we-OpenAI we-GPT , oguquka kusuka ku-GPT-1 kuya ku-GPT-4, okubonisa ukuthi lapho ubukhulu bukhula, kulapho amakhono avela khona, njengokufunda okungadingi zithombe.
Kodwa ababodwa. SinoClaude ovela ku-Anthropic , onezinguqulo ezifana ne-Sonnet, i-Opus, ne-Haiku ezilwela ukulinganisela phakathi kobuhlakani nesivinini. Ngakolunye uhlangothi, i-Amazon Nova inikeza ububanzi obuqala ekuqondeni okuningi kuya ekudalweni kwamavidiyo okudala. Asikwazi ukukhohlwa i-Stable Diffusion , eyaletha amandla amamodeli ayisisekelo ezweni lezithombe, noma i-BLOOM , umzamo wezilimi eziningi nowokubambisana obonisa ukuthi umthombo ovulekile nawo unendawo yawo.
Amakhono kanye nezinhlelo zokusebenza zangempela
Lawa mamodeli awabhali ama-imeyili kuphela. Umthelela wawo ufinyelela emikhakheni ebalulekile:
- Impilo: Bayasiza ku- ukutholakala kwezidakamizwa kanye nokuhlaziywa kwezithombe zezokwelapha, yize zidinga incazelo egcwele ukuze kugwenywe amaphutha abulalayo.
- Kwesokudla: Zenza kube lula ukubuyekezwa kwezinkontileka kanye nokuhlaziywa kwezindaba zomthetho ezinde, kunciphisa izindleko zokufinyelela ubulungiswa.
- Imfundo: Bavumela i- ngezifiso yokufunda futhi ezivumelana nezimo, yize ziletha izinselelo maqondana nokukopisha kanye negebe lobuchwepheshe.
- Amarobhothi: Inhloso ukudala amarobhothi ajwayelekile angadingeki ukuthi ahlelwe kusukela ekuqaleni komsebenzi ngamunye, kodwa esikhundleni salokho asebenzise imodeli eyisisekelo ukuze sebenzisana nemvelo ngokomzimba
Uhlangothi Olumnyama: Izingozi, Ukuziphatha, kanye Nemvelo
Akusikho konke ukukhanya kwelanga nama-rose. Ukuhlangana kwe-homogenization kudala ingozi enkulu: uma wonke umuntu esebenzisa imodeli efanayo yesisekelo, noma yikuphi ukubandlulula noma iphutha elingokwemvelo kuzosakazeka kuzo zonke izinhlelo zokusebenza ezisuselwe kuzo, kudale "i-algorithmic monoculture." Ngaphezu kwalokho, kukhona ukubona izinto ezingekho - lezo zikhathi lapho i-AI iqamba idatha ngokuqiniseka okumangalisayo, okudinga ukuqinisekiswa okuqhubekayo komuntu.
Ngokombono wokuziphatha nowomthetho, kukhona impi evulekile mayelana nempahla yobuhlakani , njengoba amamodeli amaningi eqeqeshwa ngedatha ngaphandle kwemvume ecacile yabadali bawo. Okungeziwe kulokhu kulimala kwemvelo; ukuqeqesha la ma-giants kudla amandla amaningi namanzi, okuphoqa ukusesha izakhiwo ezisebenza kahle kakhulu kanye nezikhungo zedatha ezisimeme.
Ikusasa kanye nokuphathwa kwe-AI
Indlela eya phambili ilele ekufinyeleleni ngentando yeningi . Njengamanje, ukuqeqesha amamodeli anamandla kakhulu kugxile ezinkampanini ezimbalwa ngenxa yezindleko zehadiwe (ama-GPU). Kubalulekile ukuthi amanyuvesi nohulumeni batshale imali engqalasizinda yomphakathi ukuvimbela amandla e-AI ukuthi angaweli ezandleni ze- oligopoly yezobuchwepheshe.
Isihluthulelo kuzoba ukucaca kanye nokuhlolwa okuzimele. Akwanele ngenkampani ukuthi nje ithi imodeli yayo iphephile; sidinga izinhlaka zomthetho (njengoMthetho we-AI we-EU) kanye nezindinganiso zobungcweti eziqinisekisa ukuthi la mathuluzi asetshenziselwa ukuthuthukisa umphakathi, hhayi ukwenza kube lula ukuqapha okukhulu noma ukwaziswa okungamanga.
Uhlelo lwe-artificial intelligence lushintshele esakhiweni lapho amamodeli ambalwa ayisisekelo esekela izinkulungwane zezinhlelo zokusebenza ezikhethekile, ehlanganisa amandla okucubungula okukhulu nokunemba kokulungiswa kwezinga lomuntu. Lokhu kuthuthuka, ngenkathi kuphakamisa izinkinga ezinkulu zokuziphatha kanye nokusetshenziswa kwamandla, kuchaza kabusha ubudlelwano phakathi kwabantu nemishini ngokuguqula i-AI ibe yingqalasizinda yenhloso ejwayelekile ekwazi ukucabanga, ukudala, nokufunda kuzo zonke izizinda eziningi ngasikhathi sinye.
