Isiqulatho[Fihla][Bonisa]
Imifuziselo yolwimi olukhulu lolona phuhliso luchukumisayo lwamva nje kwicandelo lenkqubo yolwimi lwendalo kunye nothungelwano lwe-neural.
I-OpenAI's GPT-3 igqame njengenye yeemodeli eziqhuba kakuhle phaya. Imveliso yemodeli ihlala ingabonakali kwisicatshulwa esivela ebantwini.
Nangona kunjalo, i-GPT-3 iseyimodeli yomthombo ovaliweyo. Ngelixa inamandla amakhulu, kukho imida ethile enokuthi iyenze ingafaneleki kwiimeko ezithile zokusetyenziswa.
Kweli nqaku, siza kuhamba phezu ezimbalwa ezinkulu iimodeli zolwimi enokukhuphisana nokusebenza kwe-GPT-3 ekrwada.
Kutheni ujonge enye indlela ye-OpenAI GPT-3?
Imodeli ye-OpenAI ye-GPT-3 isebenzisa i-Advanced ukufunda okunzulu iimodeli zokuvelisa umbhalo ofana nomntu. Yimodeli yoqikelelo yolwimi lwesizukulwana sesithathu ukusuka kwilabhoratri yophando ye-OpenAI.
Imodeli yakhululwa ekuqaleni njenge-beta evaliweyo ngaphambi kokuba i-OpenAI ivule i-API eluntwini ngasekupheleni kuka-2021.
Okwangoku, i-GPT-3 ineemodeli ezine ezisisiseko onokuthi ukhethe kuzo. I-Ada, imodeli yexabiso eliphantsi kunye nekhawulezayo ixabisa i-$ 0.0004 kuphela ngamathokheni angama-1000. Imodeli enamandla kakhulu ye-OpenAI, i-Davinci, ixabisa i-$ 0.02 ngeetokheni ze-1000, okanye malunga namaxesha angama-50 abiza kakhulu.
I-OpenAI ikwafuna ukuba umphuhlisi alandele eyakhe izikhokelo zokusetyenziswa. Umphuhlisi uya kubonelela ngesabelo esilinganiselweyo sosetyenziso esinokunyuswa nje ukuba isicelo somphuhlisi samkelwe ngenkqubo yokuphonononga ngesandla.
Ngelixa imveliso ye-GPT-3 isaziwa kakhulu ngemveliso yayo ekumgangatho ophezulu, asiyiyo yodwa imodeli yoqikelelo lolwimi olukhoyo ukuze uyisebenzise.
Makhe sijonge ezinye iimodeli ezikhuphisanayo onokuzisebenzisa njengenye indlela ye-GPT-3.
1. GPT-J
I-GPT-J yimodeli yolwimi oluvulelekileyo liqela le-Eleuther AI.
Ukusebenza kwe-zero-shot kuhambelana ngokuhambelana ne-GPT-3 kwaye isondele kakhulu ekusebenzeni kunezinye ezininzi zokuphunyezwa kwe-GPT.
I-6-billion parameter autoregressive text generation model iqeqeshwe kwisethi yedatha eyaziwa ngokuba yi "The Pile".
Imfumba ngenene yindibaniselwano ye 22 yedatha encinci edityanisiweyo kunye. Inobungakanani befayile edibeneyo ye-825 GiB kwaye ibonwe ukuba igxininise kakhulu kwimithombo yezemfundo kunye nezobugcisa.
Ungayivavanya imodeli ngokwakho ngale nto isicelo sewebhu sasimahla.
Ndikwazile ukuvavanya imodeli nge-prompt elula. I-GPT-J iphumelele ekubaleni “iindlela ezizezona zokufunda ulwimi olutsha namhlanje”.
Nangona kunjalo, ukusebenza kunendawo ethile xa ndizama ukuyibuza ukuba ichaze ukuba yintoni imodeli yokuvelisa umbhalo ozimeleyo.
Ngelixa isiphumo sasivakala, asikhange siphendule ngokukhawuleza ngendlela enentsingiselo.
namaxabiso
Kuba i-GPT-J iyimodeli yomthombo ovulekileyo, ungaqhuba owakho umzekelo ngokwakho. Ngokutsho kwe indawo yokugcina esemthethweni, imodeli yenzelwe ukuba isebenze kwiyunithi yokucubungula i-tensor (TPU). Ngelixa ilungile, le isenokungabi lolona khetho lunexabiso eliphantsi kuba iGoogle inexabiso eliphantsi TPUs ilifu iindleko malunga ne-$ 4.50 / iyure.
Kusenokungabizi kakhulu ekuhambeni kwexesha ukusebenzisa eyakho i-GPU okanye urente iseva ye-GPU ezinikeleyo ngeenkonzo ezifana Vast.ai or FluidStack.
2. Jurassic-1
I-Jurassic-1 imodeli yolwimi ekhutshwe yi-AI21 Labs, inkampani ye-AI ye-Israel egxile kwi-NLP. Njenge-OpenAI, banikezela nge-API ekuvumela ukuba ufikelele kwimodeli yolwimi lwabo.
Unokwenza i-akhawunti kubo website ukufikelela kwi-app yendawo yokudlala ukuze uvavanye imodeli yakho.
I-AI21 Studio ikwabandakanya inqaku apho unokuqeqesha kwaye ubuze iinguqulelo zakho zesiko kwiimodeli zabo zeJurassic-1. Ngokutsho kwe post blog blog, imifuziselo yesiko enemizekelo embalwa nje engamashumi amahlanu inokugqwesa ubunjineli obukhawulezileyo kusetyenziswa imodeli yoqobo.
namaxabiso
Banikezela ngamaxabiso aguquguqukayo asekelwe kusetyenziso kwimodeli nganye yesiseko sabo ezintathu. Ngokomzekelo, bahlawula i-$ 0.25 kuzo zonke iithokheni ze-1000 ezenziwe yimodeli. Ngokomyinge, umqondiso ngamnye umalunga negama eli-1 okanye amagama amathandathu.
Oku kuthetha ukuba ungasebenzisa eyona modeli ilungileyo ye-AI21 ukwenza uxwebhu lwamagama angama-4000 nge-$1 kuphela. Inye into ekufuneka uyikhumbule kukuba kusafuneka uhlawule ubuncinci be-29 yeedola rhoqo ngenyanga ukuze usebenzise imodeli.
3. UmbhaloSynth
I-TextSynth yenye inkonzo yewebhu ye-NLP onokuyisebenzisa ukwenza umbhalo. Ngokungafaniyo nemizekelo emibini yangaphambili, i-TextSynth ayisiyiyo imodeli ezimeleyo. Le nkonzo isebenza ngokunika umsebenzisi ufikelelo kwezinye iimodeli zolwimi ezinkulu ezinomthombo ovulekileyo njengeGPT-NeoX, M2M100 kunye neGPT-J.
Abaphuhlisi banokusebenzisa zabo I-API yokuphinda ukudibanisa iimodeli zolwimi kwizicelo zabo. Ungazama ukukhangela ngaphandle kwabo simahla iphepha lokudlala ukubona ukuba imodeli nganye ekhoyo isebenza njani.
namaxabiso
Isicwangciso sabo sasimahla sikunika ukufikelela kuyo yonke imifuziselo yolwimi kunye nemida ethile yomlinganiselo. Inkonzo ithintela isicelo ngasinye ubude beempawu ezingama-200.
Isicwangciso esiqhelekileyo sisusa umda kwinani lamathokheni avelisiweyo. Imodeli yamaxabiso isekelwe kwikhredithi ukuphepha iindleko ezingalindelekanga. Ubuncinci benani leekhredithi ozithengayo yi-20 yeedola. Iikhredithi ezingasetyenziswanga azivumelekanga emva konyaka.
Ixabiso lesicelo ngasinye lisekelwe kwinani legalelo kunye namathokheni avelisiweyo. Ngokusekwe kwitheyibhile kwiwebhusayithi yabo esemthethweni, unokulindela ukuhlawula malunga ne-0.75 yeedola ukuya kwi-1.25 yeedola ngokusebenzisa iimodeli zabo ezingabizi kakhulu.
isiphelo
Ngethemba, eli nqaku linokukunceda ufumane imodeli yolwimi efikelelekayo nesebenzayo onokuyisebenzisa njenge Enye indlela ye-OpenAI GPT-3.
Iimodeli zeelwimi ezinkulu zinamandla kakhulu kwaye zinokusetyenziswa kwimisebenzi eyahlukeneyo. Zingasetyenziselwa ukuvelisa umbhalo, ukuguqulela phakathi kweelwimi, nokuqonda nokuphendula kulwimi lwendalo.
Ngokusekwe kuphando lwam esikhaleni kunye novavanyo endilwenzileyo, i-GPT-3 isagqwesa zonke ezinye imodeli yolwimi olukhulu Ndizamile. Nangona kunjalo, oku kunokutshintsha kwixesha elizayo njengoko abaphandi bephuhlisa kwaye bakhuphe iimodeli ezintsha.
Abaphandi kuGoogle, kuFacebook, nakwezinye iilabhoratri ze-AI basazoqhubeka nokusebenza ukuqhubela phambili ii-LMM zabo. Ngokuqinisekileyo kunokwenzeka ukuba elinye lala maqela e-AI liza kuphuma nemodeli ephezulu kune-GPT-3.
Shiya iMpendulo