NgoLwesibili olwedlule, i-OpenAI ikhiphe umbiko omusha wobuchwepheshe onamakhasi angu-98 njengengxenye yokwethulwa komodeli wayo wakamuva wolimi lwe-GPT-4 . Unjiniyela uBen Schmidt wenkampani yokumapha ebizwa ngokuthi i-Nomic uthe ngeke baveze lutho mayelana nendlela abayiqeqesha ngayo i-GPT entsha, ukuze bavimbele amanye amapulatifomu ukuthi ayikopishe.
Ngakho-ke, i-GPT-4 ifihliwe , esizokufunda njengoba ithuthukiswa. I-OpenAI ibilokhu iqaphela kakhulu ukuvimbela izimbangi zayo eziyinhloko ukuthi zisebenzise umsebenzi wayo kanye nomzamo wayo . Ngaphezu kwalokho, bathi bebelokhu becacile ekulungiseleleni umbiko.
I-GPT-4 ye-OpenIA ingumthombo ovaliwe
Lo mbiko unombhali oyedwa, i-OpenAI, okusho ukuthi inkampani ayikadalulanga ukuthi isebenzise ini ukuqeqesha lo modeli omusha, nokuthi ikwenze kanjani lokho . Akucaci futhi ukuthi zaziyini izindleko zehadiwe kanye namandla ekudalweni kwawo.
Isizathu salokhu yisimo samanje sokuncintisana kanye nezizathu zokuphepha, njengoba bekholelwa ukuthi ukwembula lolu lwazi kungaphindaphindwa ngabancintisana nabo. Iqiniso liwukuthi, lo mbiko uqukethe imininingwane embalwa kakhulu mayelana nenqubo kanye nokuqeqeshwa okuhilelekile.
Kwaphawulwa ukuthi ikhasi lesibili lombiko likhuluma kafushane ngobubanzi kanye nemikhawulo yalo modeli. Ngaphezu kwalokho, babonise ukuthi, naphezu kokunikeza idatha eyengeziwe ngamakhono ayo, ngeke badalule imininingwane yangaphakathi yentuthuko yalo modeli ukuze bagweme ukwenza kube lula ngabancintisana nabo ukuwukopisha.
Konke lokhu kwenze ukwethulwa kwe-GPT-4 kwaba yinto eyimfihlo kakhulu inkampani eyake yayibona. Kodwa-ke, abanye ochwepheshe bakholelwa ukuthi lesi simo esisha esivela ku -OpenAI singabeka engcupheni ukuphepha kwala mamodeli ngoba, ngaphandle kokubonakala ngokuphelele , alahlekelwa ikhono lokuhlolwa ezinye izinhlangano noma ochwepheshe abazimele.
Ngokusho kuka -Ilya Sutskever, omunye wabasunguli be-OpenAI, okuphambene nalokho kuyiqiniso. Ukholelwa ukuthi lezi zinhlobo zamamodeli zingabangela umonakalo omkhulu njengoba amakhono azo ethuthuka , ngakho-ke kunengqondo ukungadaluli ulwazi mayelana nentuthuko ukuze kuvinjelwe abenzi bobubi ukuthi basizakale ngakho.
"Ngenxa yokuncintisana kanye nemiphumela yokuphepha kumamodeli amakhulu njenge-GPT-4, lo mbiko awunayo imininingwane eyengeziwe mayelana nokwakhiwa, ihadiwe, ukubalwa koqeqesho, ukwakhiwa kwesethi yedatha, indlela yokuqeqesha, nezinye izici ezifanayo."
U-Ilya Sutskever, omunye wabasunguli be-OpenAI
Ngaphezu kwalokho, u-Ilya Sutskever uthe izinkampani eziningi ezincintisanayo zifuna ukwenza into efanayo, ngakho ukholelwa ukuthi lesi sinyathelo esisha se- OpenAI sivuthiwe kakhulu esifisweni sayo sokuvikela omunye wemikhiqizo yayo eyinhloko . Ngakho-ke, inkampani isishintshe ekwabelaneni ngolwazi , njengaseminyakeni edlule, yaya ekusebenziseni indlela yokuvikela kakhulu ekuthuthukisweni kwayo.
USutskever uphinde wakhuluma ngokuthi i-OpenAI ibilokhu inephutha, yingakho kwadingeka benze ushintsho oluphelele. Esikhathini esidlule, babelane ngolwazi oluningi mayelana namamodeli abo. Yingakho umsunguli wenkampani ekholelwa ukuthi uma ubuhlakani bokwenziwa bunamandla , khona-ke akunangqondo ukuvula ikhodi yomthombo.
"Kungenzeka kakhulu ukuthi eminyakeni embalwa izinkampani eziningi zizoqonda ukuthi ubuhlakani bokwenziwa obuvulekile akuwona umqondo omuhle kunoma ubani; kufanele buthuthukiswe ngasese futhi busetshenziswe ezinhlelweni ezivaliwe."
U-Ilya Sutskever, omunye wabasunguli be-OpenAI
Iqiniso liwukuthi ukwethulwa kwe-GPT-4 kungenye yezikhali zakamuva ze-OpenAI emncintiswaneni wobuhlakani bokwenziwa nezinye izinkampani ezifana ne -Google, i-Microsoft , noma i-Meta, ezisanda kwethula i-LLaMA, umncintiswano womthombo ovulekile otholakala ukuyifaka kumakhompyutha.
Ngakolunye uhlangothi, i-Google iphinde yamemezela ngoLwesithathu olwedlule ukuthi imodeli yayo yolimi lwe-PaLM izokhipha i-API ezosetshenziswa amabhizinisi nabathuthukisi. I-Microsoft ayishiywanga ngemuva, njengoba ichithe inani elikhulu labantu eqenjini layo le-AI futhi iqhubeka nokuzibophezela kwayo emikhiqizweni ye-OpenAI.
Ngicabanga ukuthi singakubiza ngokuthi kuvaliwe ku-'Vula' AI: iphepha elinamakhasi angu-98 elethula i-GPT-4 limemezela ngokuziqhenya ukuthi alivezi *lutho* mayelana nokuqukethwe yisethi yabo yokuqeqeshwa. pic.twitter.com/dyI4Vf0uL3
— Ben Schmidt / @ [email protected] (@benmschmidt) Mashi 14, 2023
