Isikhulu se-Lightning AI sigxeka iphepha le-OpenAI GPT-4 “njengelizenza ucwaningo”

Ngokushesha ngemva kwe-OpenAI’s ukukhululwa okumangazayo yalo osekunesikhathi ilindelwe I-GPT-4 imodeli izolo, kube nenqwaba yokugxekwa okuku-inthanethi mayelana nokuthi yini ehambisana nesimemezelo: Amakhasi angama-98 umbiko wezobuchwepheshe mayelana “nokuthuthukiswa kwe-GPT-4.”

Abaningi bathi lo mbiko uphawuleka kakhulu ngalokho okwenzile hhayi hlanganisa. Esigabeni esibizwa ngokuthi I-Scope and Limitations of this Technical Report, ithi: “Uma kubhekwa kokubili isimo sokuncintisana kanye nemithelela yokuphepha yamamodeli amakhulu njenge-GPT-4, lo mbiko awuqukethe imininingwane eyengeziwe mayelana nezakhiwo (okuhlanganisa nosayizi wamamodeli), i-hardware, ikhompuyutha yokuqeqesha, ukwakhiwa kwedathasethi, indlela yokuqeqesha, noma okufanayo.”

“Ngicabanga ukuthi singayibiza ngokuthi ivaliwe ku-‘Open’ AI: iphepha elinamakhasi angu-98 elethula i-GPT-4 ngokuziqhenya limemezela ukuthi abadaluli *lutho* mayelana nokuqukethwe kwesethi yabo yokuqeqeshwa,” ku-tweet U-Ben Schmidt, i-VP yokwakhiwa kolwazi kwa-Nomic AI.

Futhi uDavid Picard, umcwaningi we-AI kwa-Ecole des Ponts ParisTech, ku-tweet: “Ngiyacela @OpenAI shintsha igama lakho NGOKUSHESHA. Kuwukudelela ubuhlakani bethu ukuzibiza ngokuthi ‘uvuliwe’ futhi ukhiphe lolo hlobo ‘lombiko wezobuchwepheshe’ ongaqukethe nhlobo ulwazi lobuchwepheshe.”

Omunye umgxeki ophawulekayo walo mbiko uWilliam Falcon, oyi-CEO ye-Lightning AI nomsunguli we-PyTorch Lightning, umtapo wezincwadi we-Python ovulekile ohlinzeka ngesixhumi esibonakalayo esisezingeni eliphezulu sohlaka lokufunda okujulile oludumile lwe-PyTorch. Ngemuva kokuthi ethumele i-meme elandelayo, ngafinyelela ku-Falcon ukuze aphawule. Le nhlolokhono ihleliwe futhi yafinyezwa ukuze icace.

I-VentureBeat: Kunokuningi ukugxekwa njengamanje mayelana nephepha locwaningo le-GPT-4 elisanda kukhishwa. Yiziphi izinkinga ezinkulu?

UWilliam Falcon: Ngicabanga ukuthi okukhathaza wonke umuntu ukuthi i-OpenAI yenza iphepha lonke elifana namakhasi angamashumi ayisishiyagalolunye ubude. Lokho kwenza kuzwakale sengathi iwumthombo ovulekile futhi iyimfundo, kodwa akunjalo. Azichazi lutho ngempela lapho. Uma iphepha lezemfundo lisho ama-benchmarks, lithi ‘Hey, senze kangcono kunalokhu futhi nansi indlela yokuthi uqinisekise lokho.’ Ayikho indlela yokuqinisekisa lokho lapha.

Lokho akuyona inkinga uma uyinkampani futhi uthi, ‘Into yami ishesha izikhathi ezingu-10 kunalokhu.’ Sizothatha lokho ngohlamvu lukasawoti. Kodwa uma uzama ukuzenza ucwaningo, yinkinga leyo.

Uma ngishicilela, noma noma ubani emphakathini eshicilela iphepha, ngiyaliqhathanisa nezinto abantu asebenazo kakade, futhi zisesidlangalaleni futhi ngibeka ikhodi lapho futhi ngibatshele kahle ukuthi iyini idatha. Ngokuvamile, kunekhodi ku-GitHub ongayigijima ukuze ukhiqize lokhu.

VB: Ingabe lokhu kwehlukile kunalokho okwakuyikho ngenkathi i-ChatGPT iphuma? Noma i-DALL-E? Ingabe labo ababezenza ucwaningo babefana ngendlela efanayo?

I-Falcon: Cha, babengekho. Khumbula, i-GPT-4 isuselwe ekwakhiweni kwe-Transformer eyavulwa iminyaka eminingi yi-Google. Ngakho sonke siyazi ukuthi yilokho kanye abakusebenzisayo. Ngokuvamile babenekhodi okufanele bayiqinisekise. Bekungaphindaphindeki ngokuphelele, kodwa ubungakwenza kwenzeke uma bewazi okwenzayo. Nge-GPT-4, awukwazi ukukwenza.

Inkampani yami ayiqhudelani ne-OpenAI. Ngakho asinandaba ngempela. Iningi labanye abantu ababhala ku-tweet bayizimbangi. Ngakho-ke inyama yabo yenkomo iningi ukuthi ngeke bakwazi ukuphinda imiphumela. Okufanelekile ngokuphelele— I-OpenAI ayifuni ukuthi uqhubeke ukopisha amamodeli abo, lokho kunengqondo. Ninalo lonke ilungelo lokwenza lokho njengenkampani. Kodwa uzenza ucwaningo. Inkinga ke leyo.

Ukusuka ku-GPT kuya ku-ChatGPT, into eyenze yasebenza kahle kakhulu i-RLHF, noma ukuqinisa ukufunda kusuka kumpendulo yomuntu. I-OpenAI ibonise ukuthi lokho kuyasebenza.Abazange badinge ukubhala iphepha mayelana nokuthi isebenza kanjani ngoba leyo yindlela yocwaningo eyaziwayo. Uma sipheka, kufana nokuthi sonke siyazi ukuthi sisosa kanjani, ngakho-ke ake sizame lokhu. Ngenxa yalokho, kunezinkampani eziningi ezifana ne-Anthropic empeleni eziphindaphinde imiphumela eminingi ye-OpenAI, ngoba zazazi ukuthi yini iresiphi. Ngakho-ke ngicabanga ukuthi i-OpenAI ezama ukuyenza manje, ukuvikela i-GPT-4 ukuthi ingakopishwa futhi, ngokungakutsheli ukuthi yenziwa kanjani.

Kodwa kukhona okunye abakwenzayo, uhlobo oluthile lwe-RLHF olungavuliwe, ngakho akekho owaziyo ukuthi kuyini lokho. Kungenzeka ukuthi indlela ethile ehluke kancane eyenza isebenze. Eqinisweni, angazi noma isebenza kangcono yini. Kuzwakala sengathi kwenza. Ngizwa imiphumela exubile mayelana ne-GPT-4. Kodwa iphuzu liwukuthi, kukhona isithako esiyimfihlo lapho abangatsheli muntu ukuthi siyini. Lokho kudida wonke umuntu.

I-VB: Ngakho-ke esikhathini esidlule, nakuba kwakungenakuphindaphindeka, okungenani wawazi ukuthi yiziphi izithako eziyisisekelo zeresiphi. Kodwa manje nasi isithako esisha okungekho muntu ongasikhomba, njengeresiphi eyimfihlo ye-KFC?

UFalcon: Yebo, yilokho kanye okuyikho. Kungaba ngisho idatha yabo. Mhlawumbe alukho ushintsho. Kodwa cabanga nje uma ngikunikeza iresiphi yenkukhu ethosiwe – sonke siyayazi indlela yokwenza inkukhu ethosiwe. Kodwa ngokuzumayo ngenza okuthile okuhlukile futhi ufana nokulinda, kungani lokhu kuhluke? Futhi awukwazi ngisho nokukhomba isithako. Noma mhlawumbe akuthosiwe. Kwazi bani?

Kufana nokuthi kusukela ngo-2015-2019 besizama ukuthola njengenkundla yocwaningo ukuthi yikuphi ukudla abantu abafuna ukukudla. Sithole ukuthi ama-burgers ayeyi-hit. Kusukela ngo-2020-2022 safunda ukuwapheka kahle. Futhi ngo-2023, ngokusobala manje sengeza amasoso ayimfihlo kuma-burgers.

I-VB: Ingabe ukwesaba ukuthi yilapho siya khona – ukuthi izithako eziyimfihlo ngeke zize zabiwe, ingasaphathwa imodeli ngokwayo?

UFalcon: Yebo, kuzobeka isibonelo esibi. Ngiphatheke kabi ngalokhu. Sonke sasivela ezifundweni. Ngingumcwaningi we-AI. Ngakho izindinganiso zethu zisekelwe emthonjeni ovulekile kanye nasezifundweni. Ngivela elebhu ka-Yann LeCun ku-Facebook, lapho yonke into abayenzayo ingumthombo ovulekile futhi uqhubeka nokwenza lokho futhi ubekwenza lokho kakhulu kwa-FAIR. Ngicabanga ukuthi i-LLaMa, kuneyakamuva eyethulwe eyisibonelo esihle salokho kucabanga. Iningi lomhlaba we-AI likwenzile lokho. Inkampani yami ingumthombo ovulekile, konke esikwenzile kuwumthombo ovulekile, ezinye izinkampani ziwumthombo ovulekile, sinika amandla amaningi alawo mathuluzi e-AI. Ngakho-ke sonke sinikeze lokho okuningi emphakathini ukuze i-AI ibe lapho ekhona namuhla.

Futhi i-OpenAI ibilokhu isekela lokho ngokuvamile. Badlale kahle. Manje, ngenxa yokuthi banale ngcindezi yokwenza imali, ngicabanga ukuthi namuhla usuku lapho baba umthombo ovaliwe ngempela. Bavele bazihlukanise nomphakathi. Bafana, asinandaba nezifundiswa, sidayisa eSilicon Valley. Sonke sinoxhaso lwe-VC, kodwa sonke sisalondoloza ubuqotho ezifundweni.

I-VB: Ngakho-ke ungasho ukuthi lesi sinyathelo siya kude kunanoma yini evela ku-Google, noma i-Microsoft, noma i-Meta?

I-Falcon: Yebo, i-Meta iyona evuleke kakhulu – angichemi, ngivela lapho, kodwa zisavuleke kakhulu. I-Google isenawo amamodeli ayimfihlo kodwa ahlala ebhala amaphepha ongawaphinda. Manje kungase kube nzima ngempela, njengompheki noma indawo yokudlela ehlanyayo ebhala iresiphi lapho abantu abane emhlabeni bengaphinda leyo recipe, kodwa ilapho uma ufuna ukuzama. I-Google ihlezi ikwenza lokho. Zonke lezi zinkampani zine. Ngicabanga ukuthi okokuqala lapho ngibona lokhu akunakwenzeka, ngokusekelwe kuleli phepha.

VB: Yiziphi izingozi zalokhu mayelana nezimiso zokuziphatha noma i-AI enomthwalo wemfanelo?

I-Falcon: Okokuqala, kunezinkampani eziningi eziqala ukuphuma ezingaphumi emphakathini wezifundiswa. Ziyizinhlobo zokuqala ze-Silicon Valley eziqala izinkampani, futhi azilethi ngempela lezi zindinganiso zocwaningo lwe-AI zokuziphatha. Ngicabanga ukuthi i-OpenAI ibabekela isibonelo esibi. Empeleni bathi, kuhle, yenza nje into yakho, asinandaba. Ngakho-ke uzoba nazo zonke lezi zinkampani ezingeke zisakhuthazwa ukuthi zenze izinto zibe umthombo ovulekile, ukutshela abantu ukuthi zenzani.

Okwesibili, uma le modeli ingahambi kahle, futhi izoba, usuvele uyibone ngemibono futhi ikunikeza imininingwane engamanga, umphakathi kufanele usabele kanjani? Ngabe abacwaningi bokuziphatha kufanele bahambe kanjani futhi empeleni baphakamise izixazululo futhi bathi, le ndlela ayisebenzi, mhlawumbe ukuyilungisa ukwenza le enye into? Umphakathi ulahlekelwa yikho konke lokhu, ngakho lawa mamodeli angaba yingozi kakhulu ngokushesha, ngaphandle kokuthi abantu abawaqaphe. Futhi kunzima ngempela ukuhlola. Kufana nebhange okungelona ele-FINRA, njengokuthi kufanele ulilawule kanjani?

VB: Ucabanga ukuthi kungani i-OpenAI yenza lokhu? Ingabe ikhona enye indlela ababengavikela ngayo i-GPT-4 ekuphindaphindeni futhi bayivule?

UFalcon: Kungase kube nezinye izizathu, ngiyamazi uSam, kodwa angikwazi ukufunda ingqondo yakhe. Ngicabanga ukuthi bakhathazeke kakhulu ngokwenza umkhiqizo usebenze. Impela banokukhathazeka ngezimiso zokuziphatha futhi baqinisekise ukuthi izinto azilimazi abantu. Ngicabanga ukuthi bebecabangile ngalokho.

Kulokhu, ngicabanga ukuthi empeleni kumayelana nabantu abangaziphindaphindi ngoba, uma uqaphela, njalo uma bethula okuthile okusha. [it gets replicated]. Ake siqale nge-Stable Diffusion. I-Stable Diffusion yaphuma eminyakeni eminingi edlule yi-OpenAI. Kuthathe iminyaka embalwa ukuphindaphinda, kodwa kwenziwa emthonjeni ovulekile yi-Stability AI. Yabe isiphuma i-ChatGPT futhi inezinyanga ezimbalwa nje ubudala futhi sesivele sinenguqulo enhle kakhulu engumthombo ovulekile. Ngakho isikhathi siya siba sifushane.

Ekupheleni kosuku, kuzokwehla ukuthi iyiphi idatha onayo, hhayi imodeli ethile noma amasu owasebenzisayo. Ngakho into abangayenza ukuvikela idatha, okuyinto asebevele beyenza. Abakutsheli ngempela ukuthi baziqeqesha ngani. Ngakho lolo uhlobo lwento esemqoka abantu abangayenza. Ngicabanga nje ukuthi izinkampani ngokuvamile zidinga ukuyeka ukukhathazeka kakhulu ngamamodeli ngokwawo angumthombo ovaliwe futhi zikhathazeke kakhulu ngedatha kanye nekhwalithi ukuthi iyona nto oyivikelayo.