Izihloko

I-GPT4 vs ChatGPT: Sihlaziya izindlela zokuqeqesha, ukusebenza, amakhono kanye nemikhawulo

Kulindeleke ukuthi imodeli entsha yolimi iguqule zonke izimboni, okuhlanganisa abezindaba, ezemfundo, umthetho kanye nobuchwepheshe.

Ezinyangeni ezisanda kwedlula, isivinini okukhishwe ngaso amamodeli ezilimi ezinkulu siyamangaza. Kulesi sihloko, sizofaka ukufana okuyinhloko nomehluko phakathi kwe-GPT4 vs ChatGPT, okuhlanganisa izindlela zokuqeqesha, ukusebenza, amakhono kanye nemikhawulo.

Inkomba yokuqukethwe

I-GPT4 vs I-ChatGPT: Ukufana nokwehluka kwezindlela zokuqeqesha

I-GPT4 ne-ChatGPT zakhela ezinguqulweni ezindala zamamodeli e-GPT ngokuthuthukiswa kwezakhiwo eziyimodeli, zisebenzisa izindlela zokuqeqesha eziyinkimbinkimbi, kanye nenani eliphakeme lemingcele yokuqeqeshwa.

Yomibili imiklamo isekelwe ekwakhiweni kwe-transformer, esebenzisa isishumeki sekhodi ukucubungula ukulandelana kokokufaka kanye nesikhiphi khodi ukuze sikhiqize ukulandelana kokuphumayo. I-encoder ne-decoder zixhunywe ngomshini, ovumela isikhiphi khodi ukuthi sibhekisise kakhulu ukulandelana kokokufaka okubaluleke kakhulu.

Umbiko wezobuchwepheshe we-GPT4 ye-OpenAI inikeza ukuqonda okuncane ngesakhiwo semodeli kanye nenqubo yokwakheka kwe-GPT4, icaphuna “competitive landscape and the safety implications of large-scale models“. Esikwaziyo ukuthi i-GPT4 ne-ChatGPT cishe ziqeqeshwe ngendlela efanayo, okuhluke kakhulu ezindleleni zokuqeqesha ezisetshenziselwa i-GPT-2 ne-GPT-3. Sazi okuningi mayelana nezindlela zokuqeqesha ze-ChatGPT kune-GPT4, ngakho-ke sizoqala lapho.

I-ChatGPT

I-ChatGPT iqeqeshwe ngamasethi edatha yengxoxo, okuhlanganisa idatha yedemo, lapho izichasiselo ezingabantu zibonisa okukhiphayo okulindelekile komsizi wengxoxo ekuphenduleni izicelo ezithile. Le datha isetshenziselwa ukushuna i-GPT3.5 ngokufunda okugadiwe, okukhiqiza imodeli yenqubomgomo, esetshenziselwa ukukhiqiza izimpendulo eziningi lapho izicelo zinikezwa. Izichasiselo ezingabantu zibe sezihlukanisa ukuthi yiziphi izimpendulo zokwaziswa okunikeziwe eziveze imiphumela engcono kakhulu, esetshenziselwa ukuqeqesha imodeli yomklomelo. Imodeli yokuklomelisa ibe isisetshenziswa ukulungisa ngokuphindaphindiwe imodeli yenqubomgomo kusetshenziswa ukufunda okuqiniswayo.

I-ChatGPT iqeqeshelwe ukusebenzisa Ukufunda Okuqiniswayo Kumpendulo Yomuntu (RLHF), indlela yokuhlanganisa impendulo yomuntu ukuze kuthuthukiswe imodeli yolimi ngesikhathi sokuqeqeshwa. Lokhu kuvumela okukhiphayo kwemodeli ukuthi kuhambisane nomsebenzi ocelwe umsebenzisi, kunokumane ubikezele igama elilandelayo emshweni ngokusekelwe kudatha yokuqeqeshwa ejwayelekile, njenge-GPT-3.

I-GPT4

I-OpenAI isazodalula imininingwane yokuthi iyiqeqeshe kanjani i-GPT4. Umbiko wabo wezobuchwepheshe awufaki “details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar“. Esikwaziyo ukuthi i-GPT4 iwuhlobo oluqeqeshiwe lwe-transformer-style generative multimode. Kokubili kudatha etholakala esidlangalaleni nakudatha yenkampani yangaphandle enelayisensi futhi kamuva kwalungiswa kahle kusetshenziswa i-RLHF. Kuyathakazelisa ukuthi i-OpenAI yabelane ngemininingwane emayelana namasu abo athuthukisiwe e-RLHF ukwenza izimpendulo zemodeli zinembe kakhudlwana futhi mancane amathuba okuthi zikhukhulekele ngaphandle kwezivikelo zokuphepha.

Ngemva kokuqeqesha imodeli yenqubomgomo (njengaku-ChatGPT), i-RLHF isetshenziswa ekuqeqesheni okuphambene, inqubo eqeqesha imodeli ngezibonelo ezinonya ezihloselwe ukukhohlisa imodeli ukuthi iyivikele ezibonelweni ezinjalo esikhathini esizayo. Endabeni ye-GPT4, ochwepheshe bahlola izimpendulo zemodeli yezepolitiki ezidingweni eziphikisanayo. Lezi zimpendulo zibe sezisetshenziselwa ukuqeqesha amamodeli emivuzo engeziwe ahlunga ngokuphindaphindiwe imodeli yenqubomgomo, okuholela kumodeli okungenzeka ukuthi inikeze izimpendulo eziyingozi, ezibalekelayo, noma ezingalungile.

I-GPT4 iqhathaniswa ne-ChatGPT ukufana nokwehluka ngokuya ngokusebenza namandla

Amandla

Mayelana nokusebenza, i-ChatGPT ne-GPT4 zifana kakhulu kunokuhlukile. Njengomanduleli wayo, i-GPT-4 iphinde ihlanganyele ngesitayela sengxoxo esihlose ukuvumelanisa nomsebenzisi. Njengoba ungabona ngezansi, izimpendulo phakathi kwamamodeli amabili ombuzo obanzi zifana kakhulu.

I-OpenAI iyavuma ukuthi umehluko phakathi kwamamodeli ungaba ocashile futhi uthi "umehluko uphuma lapho ubunzima bomsebenzi bufinyelela embundwini owanele". Uma kubhekwa izinyanga eziyisithupha zokuqeqeshwa kwezitha imodeli eyisisekelo ye-GPT4 etholwe esigabeni sayo sangemva kokuqeqeshwa, lokhu cishe kuwuphawu olunembile.

Ngokungafani ne-ChatGPT, eyamukela umbhalo kuphela, i-GPT4 yamukela kokubili izithombe nokwaziswa kombhalo, ibuyisela izimpendulo zombhalo. Kusukela kulokhu kubhalwa, ngeshwa, ikhono lokusebenzisa okokufaka kwesithombe alikakatholakali esidlangalaleni.

Ukusebenza

Njengoba kushiwo ngenhla, i-OpenAI ibika ukuthuthuka okuphawulekayo ekusebenzeni kokuvikeleka kwe-GPT4, uma kuqhathaniswa ne-GPT-3.5 (lapho i-ChatGPT ishunwe khona). Nokho, okwamanje akukacaci ukuthi:

ukuncishiswa kwezimpendulo ezicelweni zokuqukethwe okungavunyelwe,
ukuncishiswa kokukhiqizwa kwezinto ezinobuthi e
ukuthuthukisa izimpendulo ezihlokweni ezibucayi

kungenxa yemodeli ye-GPT4 ngokwayo noma ezinye izivivinyo eziphikisanayo.

Ukwengeza, i-GPT4 idlula i-CPT-3.5 ezivivinyweni eziningi zezemfundo nezomsebenzi ezithathwa abantu. Ngokuphawulekayo, i-GPT4 ithola amaphuzu ku-90th percentile ekuhlolweni kwe-Uniform Bar uma kuqhathaniswa ne-GPT-3.5, ethola amaphuzu ku-10th percentile. I-GPT4 futhi idlula ngokuphawulekayo eyandulelayo kumabhentshimakhi emodeli yolimi lwendabuko namanye amamodeli e-SOTA (yize ngezinye izikhathi kancane kancane).

I-GPT4 vs ChatGPT: umehluko kanye nemikhawuloi

Kokubili i-ChatGPT ne-GPT4 inemikhawulo ephawulekayo kanye nezingozi. Ishidi lesistimu ye-GPT-4 lihlanganisa imininingwane evela ekuhloleni okuningiliziwe kwalezo zingozi ezenziwa yi-OpenAI.

Lezi ezinye zezingozi ezihlotshaniswa nawo womabili amamodeli:

Ama-hallucinations (ukuthambekela kokukhiqiza okuqukethwe okungenangqondo noma okungelona iqiniso)
Khiqiza okuqukethwe okulimazayo okwephula izinqubomgomo ze-OpenAI (isb. inkulumo enenzondo, ukugqugquzela udlame)
Khulisa futhi uqhubekisele phambili imibono yabantu abancishwe inyumbazane
Khiqiza ukwaziswa okungelona iqiniso okuhloselwe ukukhohlisa

Nakuba i-ChatGPT ne-GPT-4 ilwa nemikhawulo nobungozi obufanayo, i-OpenAI yenze imizamo ekhethekile, okuhlanganisa nokuhlola okuningi okungqubuzanayo, ukuzehlisa ku-GPT-4. Nakuba lokhu kukhuthaza, ishidi lesistimu ye-GPT-4 ekugcineni libonisa ukuthi i-ChatGPT yayisengozini kangakanani (futhi mhlawumbe namanje isesengozini). Ukuze uthole incazelo eningiliziwe yemiphumela eyingozi engahlosiwe, ngincoma ukufunda ishidi lesistimu ye-GPT-4, eliqala ekhasini 38 Umbiko wezobuchwepheshe we-GPT-4 .

isiphetho

Nakuba sazi okuncane mayelana nemodeli yezakhiwo, nezindlela zokuqeqesha ngemuva kwe-GPT4, kubonakala kunenguqulo ecwengisisiwe ye-ChatGPT. Eqinisweni, okwamanje i-GPT4 iyakwazi ukwamukela izithombe nokufaka umbhalo, futhi imiphumela iphephile, inembe kakhudlwana futhi inobuciko. Ngeshwa, kuzodingeka sithathe igama le-OpenAI ngakho, njengoba i-GPT4 itholakala kuphela njengengxenye yokubhaliselwe kwe-ChatGPT Plus.

Ukuhlala unolwazi mayelana nenqubekelaphambili, ubungozi kanye nemikhawulo yalezi zibonelo kubalulekile njengoba sizulazula kulo mkhakha ojabulisayo kodwa oshintshayo wamamodeli wezilimi ezinkulu.

BlogInnovazione.it