The 2m2k3r
Chero munhu anobvunzwa kuti sei Chat GPT, iyo inonyanya kufarirwa yeAI mumiririri nhasi inoshanda, vazhinji vachava nemhinduro pamuromo werurimi rwavo: njere dzekugadzira. Asi mhinduro iyoyo haina kujeka. Kunyangwe iri imwe yeminda yekufunda ine tsvakiridzo huru uye budiriro muComputing nhasi, hungwaru hwekugadzira hunosanganisira akati wandei madingindira esainzi. 4u3f73
Imwe yacho ndiyo kiyi ki kwemaitiro eChatGPT uye mazhinji maAI vamiririri anowanikwa pawebhu basa: LLMs. Muchikamu chino tichaongorora zvakadzama kuti pfungwa iyi yakashandura sei hungwaru hwekugadzira uye nyika yedu.
Chii chinonzi Mitauro Yakakura yemitauro (LLMs)? 2o5u16
Mienzaniso Yakakura Mutauro (LLMs, Makuru Mutauro Models, muchiPutukezi) ari algorithms e Deep Learning (Deep Learning, in Portuguese) inokwanisa kuita nhevedzano yemabasa Natural Language Processing (Natural Language Processing, muchiPutukezi). Phew, maacronyms akawanda, handiti?
MaLLM anoshandisa mamodhiyo eshanduko uye anodzidziswa achishandisa ma datasets makuru. Mimwe mienzaniso yedatasets dzakakurumbira ndeidzi: LAION-2B-en, CCAW e WikiText-103. A transformer modhi inogona kuita serobhoti inoshandura kuita mota, asi mumunda weAI ndiyo inonyanya kuvakwa yeLLM.
Transformer ine a encoder (encoder, muchiPutukezi) uye a Decoder (decoder, muchiPutukezi). Chaizvoizvo, iyo encoder ine basa rekuparadzanisa mazwi emutsara kana zvinyorwa muzvikamu zvidiki zvinonzi tokens, uye decoder inoita masvomhu kuti ione hukama pakati pematokeni aya.
Musiyano mukuru pakati pevashanduri uye chivakwa chakashandiswa makore apfuura, LSTM (Yakareba Yenguva Ipfupi Memory, kana Yenguva Yakareba Yenguva Yekurangarira), ndeyekuti vashanduri vanoshanda nemaitiro ekuzvitarisisa, ndiko kuti, vanokwanisa kudzidza nekukurumidza kana vachifunga zvikamu zvemutsara kana kunyange mamiriro ayo, kugadzira kufanotaura.
MaLLM ane hunyanzvi hweAI masisitimu ayo, pamusoro pekugona kugadzirisa mutauro wevanhu, anogona zvakare kuita mamwe mabasa akadai sekuongorora mapuroteni zvimiro uye kugadzira kodhi yehurongwa. Kuti ishande nemazvo, maLLM anoda pre-kudzidziswa uye kunyatsorongedza kubata mabasa akadai sekuronga zvinyorwa, muchidimbu, uye kupindura mibvunzo, zvichiita kuti zvive zvakakosha kumaindasitiri akadai sehutano, mari, uye varaidzo.
Zvikamu zvakakosha y434r
MaLLM anoumbwa neakawanda akaturikidzana neural network. Mune neural network (Neural Network, muchiRungu), chinosiyana chinoshandiswa sekuisa, chinogadziriswa nehuremu hwakasiyana uye masvomhu equation nechikamu chimwe kana kupfuura, uye kukosha kwekubuda kunogadzirwa.
Mhando yekutanga yeneural network iripo muLLMs ndiyo embedding layer (embedding layer, muchirungu). Iyo ine basa rekuisa mukati, kutora semantics uye syntactic zvinoreva zvekuisa, kuti modhi inzwisise mamiriro.
Ipapo isu tine feedforward layer (FFN, Feedforward Network, muChirungu) iyo inoumbwa neakawanda akabatana akaturikidzana anoshandura ekumisikidza ekuisa. Mukuita uku, ma layers aya anobvumira modhi kuti iunganidze zvepamusoro-soro, kureva kunzwisisa chinangwa chemushandisi nemanyorerwo emavara.
Tevere, tine dura rinodzokororwa rinoturikira mazwi ari muzvinyorwa zvekupinza achitevedzana. Inoita basa rekutora hukama huripo pakati pemazwi mumutsara.
Chekupedzisira asi chisiri chidiki, isu tine nzira yekutarisisa inobvumira iyo LLM kuti itarise pazvikamu zvimwechete zvezvinyorwa zvekupinza zvinoenderana nebasa rakapihwa. Iyi nhanho inobvumira modhi kuti ibudise zvakafanira uye zvakaringana zvinobuda.
Mashandiro avanoita 4d1g4v
Iye zvino zvatava kuziva kuti maLLM chii uye kuti ndezvipi zvakakosha zvikamu, tinogona kunzwisisa zvakajeka mashandiro avanoita. Chaizvoizvo, transformer-based LLMs inotora yekuisa, incode iyo, uye wobva waigadzirisa kuti ibudise zvakafanotaurwa. Nekudaro, iyo LLM isati yatora mameseji ekuisa uye kugadzira yakafanotaurwa inobuda, inoda kudzidziswa kuita mabasa akajairwa uye kugadzirisa kwakanaka kuti ikwanise kuita mamwe mabasa.
Pre-training (Pre-training, in English) chirongwa chekare mumunda we Machine Learning (Machine Learning, in English) mukati meArtificial Intelligence. Kuita uku, sekureva kunoita zita, kunosanganisira pre-training LLMs vachishandisa mameseti makuru ezvinyorwa zvematrillion emazwi kubva kumawebhusaiti senge. Wikipedia, GitHub, pakati pevamwe. Mushure mezvose, iyo LLM inoda kudzidza kubva kumwe kunhu, semwana mudiki, handiti?
Munguva iyi, iyo LLM inoita zvinodaidzwa kuti kudzidza kusingatarisirwe (Kudzidzira Kusina Kutarisirwa, muChirungu) - maitiro ekuti data seti inongoverengwa pasina mirairo yekunyengedza. Mune mamwe mazwi, pasina "murayiridzi", iyo LLM's yega AI algorithm ine basa rekudzidza zvinoreva izwi rega rega uye hukama pakati pavo. Pamusoro pezvo, LLM inodzidzawo kusiyanisa mazwi zvichienderana nemamiriro ezvinhu. Somuenzaniso, anodzidza kunzwisisa kana “kurudyi” kuchireva “kururamisa” kana kuti “kungopesana nekuruboshwe.”
Iye zvino maitiro ekugadzirisa zvakanaka (Kugadzirisa zvakanaka, in English) inoshanda “kunatsa” iyo LLM kuti iite nemazvo mabasa chaiwo, akadai sekushandura zvinyorwa, kukwenenzvera kuita kwayo. Kugadzirisa zvinokurudzira (mibvunzo nemirayiridzo yakapiwa kuLLM) inoshanda semhando yekugadzirisa zvakanaka, sezvo ichikwanisa kudzidzisa modhi kuita rimwe basa.
Kuti mhando yemutauro muhombe iite basa chairo, rakadai sekushandura, inofanirwa kugadziridzwa kuitira basa chairo. Fine-tuning inogonesa kuita kwemamwe mabasa.
Kurumidza tuning inoshanda basa rakafanana pakugadzirisa zvakanaka, kudzidzisa modhi kuita rimwe basa kuburikidza nediki-yeyedzo yekurudziro, kana zero-yeyedzo yekurudziro. Pazasi pane muenzaniso we "sensiment analysis" uchishandisa mashoma-pfuti kukurumidza:
Texto de entrada: Essa casa é linda!
Sentimento da frase: Positivo
Texto de entrada: Essa casa é horrível!
Sentimento da frase: NegativoZvichienderana nemhedzisiro yakawanikwa mumuenzaniso uyu, LLM yaizonzwisisa, kuburikidza nechirevo chechirevo che "zvinotyisa" uye nekuti mumwe muenzaniso wakapesana wakapihwa, kuti manzwiro emushandisi mumuenzaniso wechipiri "asina kunaka".
Maitiro ekushandisa 1g5050
Sezvatakambotaura, maLLM anogona kushandiswa kune akati wandei zvinangwa:
- Kudzosa ruzivo: Mune ino kesi tinogona kufungidzira kushandiswa kwayo mumawebhu ekutsvaga injini, seGoogle kana Bing. Kana mushandisi akashandisa iyo yekutsvaga ficha yemasevhisi aya, ivo vari kushandisa maLLM kuburitsa ruzivo muchimiro chemhinduro kuchikumbiro chavo. MaLLM anokwanisa kutora ruzivo, kupfupisa, uye kutaurirana mhinduro nenzira yehurukuro nemushandisi.
- Zvinyorwa uye purogiramu yekodhi yekugadzira: LLMs ndiyo huru "injini" ki kweGenerative AI seChatGPT, uye inogona kugadzira zvinyorwa uye kodhi yepurogiramu zvichienderana nezvakaiswa uye zvinokurudzira. Semuenzaniso, chatGPT inokwanisa kunzwisisa mapatani uye inogona kunyatsopindura kune zvikumbiro zvevashandisi senge "nyora nhetembo nezvemaruva muchimiro chaManuel Bandeira" kana "nyora kodhi yePython inokwanisa kuronga runyoro rwemafirimu mune alfabheti".
- Chatbots uye Kukurukurirana maAI: MaLLM atove kukwanisa kupa sevhisi yevatengi kuburikidza nechatbot vamiririri vanotaura nevatengi, kududzira zvinoreva mibvunzo yavo uye zvinovanetsa, uye kupa mhinduro dzakakodzera kana gwara.
Pamusoro pezviitiko zvekushandisa izvi, maLLM ari kuratidza kuve chishandiso cheAI chinovimbisa muminda yetekinoroji, hutano nesainzi, kushambadzira, mutemo uye zvakare kushandiswa mumabhanga masisitimu. Kuti ndikupe iwe zano, maLLM parizvino anokwanisa kufanotaura nehupamhi hwechokwadi kuitika kenza yemazamu kungoongorora seti dzemaserosa ane mwero wepamusoro wekururama kupfuura varapi vakawanda vane ruzivo.
LLMs uye Generative Pre-Trained Transformer (GPT) i62j
O Generative Pre-Trained Transformer (GPT) imhando chaiyo yeLLM inoshandisa dhizaini yekuvaka uye yakagadziriswa nekambani OpenAI. Yakagadzirirwa kunzwisisa, kugadzira uye kushandisa mutauro wechisikigo (wakaita sePutukezi kana Chirungu) nenzira ine hunyanzvi uye yechokwadi.
Kuputsa zita, tinogona kunzwisisa zviri nani kuti GPT chii:
- Kugadzira (Kugadzira, muchiPutukezi): inoratidza kuti modhi inogadzira zvinyorwa, ndiko kuti, inokwanisa kuburitsa mitsara mitsva, mhinduro, pfupiso, makodhi, nezvimwe.
- Pre-Trained (Akadzidziswa, muchiPutukezi): Izvi zvinoreva kuti inofanodzidziswa pahuwandu hwakawanda hwemavara kubva painternet, senge mabhuku, zvinyorwa, mawebhusaiti nezvimwe. Zvino inogona kugadziriswa kune mamwe mabasa.
- Transformer: Sezvatakambotaura, iyi ndiyo neural network architecture inopa hwaro hweiyo modhi. Iyo inofananidzwa zvakanyanya (inogona kuita akawanda mabasa panguva imwe chete) uye inoshanda pakubata kutevedzana kwakareba kwemavara.
Musiyano mukuru pakati peGPT nemamwe maLLM idanho rayo rekudzidzisa, iro rine maitiro matatu akasiyana:
- Pre-training: Huwandu hwedata hunotorwa kubva paInternet, mabhuku, uye kunyange mavhidhiyo nemimhanzi, uye wozogadziriswa kuita tokeni.
- Mirayiridzo yekugadzirisa zvakanaka: Pano muenzaniso "unodzidziswa" kuti unofanira kupindura sei kune mirairo chaiyo, kuenzanisa mhinduro dzayo kuitira kuti dzive dzakarurama.
- Kusimbisa Kudzidza Nemhinduro Yevanhu: zvakafanana nekugadzirisa zvakanaka, pano "kudzidzisa" kunoitwa kuburikidza nemhinduro dzevanhu dzinoita kuti nzira ye "kusimbisa kudzidza", apo AI inodzidza "chakarurama" uye chii "chisina kururama" kuburikidza nekudzokorora uye ruzivo rwakapiwa nemumiririri wekunze, munyaya iyi, mushandisi anoshandisa AI.
Nhoroondo: kubva kumabhiriyoni emashoko kusvika kumagwaro akaoma 2b6n5
Kunyangwe kuwedzera kwemhando dzemitauro kwakaitika chete muna 2017, kubvira 1990 mamodheru eiyo IBM aive mapiyona muhuwandu hwemitauro yemhando. Muna 2001, modhi yakadzidziswa pamashoko emamiriyoni matatu yakawana "State-of-the-art" maererano nekururama mukupirikira zvinyorwa nekuvaka mitsara yakabatana.
Kubva 2012 zvichienda mberi Neural Networks yakawana mukurumbira munyika yeAI uye nenguva isipi yakatanga kushandiswa pamabasa emitauro. Muna 2016, Google yakagamuchira iyo Neural Machine Shanduro (Neural Machine Translation, muchiPutukezi) uchishandisa mamodheru anobva pane iyi pfungwa. Muna 2018, kambani OpenAI yakapinda mukati mekuvandudzwa kweAI agents kwakavakirwa paLLMs uye yakatanga GPT-1 yekuyedza, uye raingove gore rakatevera apo GPT-2 yakatanga kukwezva veruzhinji nekuda kwekushandisa kwayo zvisina kunaka.
Muna 2020 the GPT-3 yakasvika ine zvidziviriro yekuwana chete kuburikidza neAPI, asi maingove muna 2022 apo ChatGPT (iyo AI mumiririri "inopihwa simba" neGPT-3) yakabata kutarisisa kweveruzhinji pasirese.
GPT-4 yakagadzirirwa kutanga muna 2023 iine multimodal kugona, kunyangwe ruzivo rwehunyanzvi haruna kuburitswa. Muna 2024, OpenAI yakatanga iyo muenzaniso o1, yakanangidzirwa pakugadzira cheni dzakareba dzokufunga. Zvishandiso izvi zvakafambisa kutorwa kwakapararira kweLLMs munzvimbo dzakasiyana siyana dzekutsvagisa.
Kubva muna 2024, ese makuru uye anoshanda zvakanyanya maLLM akavakirwa pane inoshandura dhizaini, paine vamwe vaongorori vanoedza uye kuyedza nemamwe mavakirwo, senge. Recurrent Neural Networks (Recurrent Neural Networks, muchiPutukezi).
Mabhenefiti uye Maganhuriro eLLMs 3t1154
Nemashandisirwo akasiyana-siyana, maLLM anobatsira zvakanyanya pakugadzirisa matambudziko sezvo achipa ruzivo mune yakajeka uye yakapusa maitiro ari nyore kuti vashandisi vanzwisise. Pamusoro pezvo, anogona kushandiswa kushandura mutauro, kupedzisa mutsara, kuongorora manzwiro, kupindura mibvunzo, masvomhu equation, nezvimwe.
Kuita kweLLMs kuri kuramba kuchivandudza sezvo ichikura sezvo mamwe data uye ma parameter achiwedzerwa. Nemamwe mashoko, paunowedzera kudzidza, unowana zviri nani. Uyezve, mienzaniso mikuru yemitauro inogona kuratidza izvo zvinonzi "kudzidza mumamiriro ezvinhu." Kana iyo LLM yakambodzidziswa, mashoma-kupfurwa anobvumira modhi kudzidza kubva mukuchimbidza pasina mamwe maparamita. Nenzira iyi, ari kuramba achidzidza.
Nekuratidza kudzidza mumamiriro ezvinhu, maLLM anodzidza nekukurumidza nekuti haadi humwe huremu, zviwanikwa, uye paramita pakudzidziswa. Vanokurumidza mupfungwa yokuti havadi mienzaniso yakawanda kuti vave "vakangwara".
Chinhu chakakosha cheLLMs kugona kwavo kupindura mibvunzo isingafungidzike. Purogiramu yekombuta yechinyakare, semuenzaniso, inogamuchira mirairo mune yayo inogamuchirwa syntax kana kubva kune yakapihwa seti yezvinopinza mushandisi. Kune rimwe divi, iyo LLM inogona kupindura kumutauro wevanhu uye kushandisa ongororo yedata kupindura mubvunzo usina kurongeka kana chikumbiro nenzira ine musoro. Kunyange chirongwa chekombuta chakajairika chaisazoziva kukurumidza senge "Ndeapi mapoka mashanu makuru edombo munhoroondo?", LLM inogona kupindura nerunyoro rwemabhendi mashanu akadaro uye nyaya inogutsa yekuti sei ivo vakanyanya kunaka.
Nekudaro, maererano neruzivo rwavanopa, maLLM anogona kungovimbika senge data ravanogamuchira. Kana vakagamuchira ruzivo rwenhema muchikamu chekutanga chekudzidzisa, vanozopa ruzivo rwenhema mukupindura mibvunzo yemushandisi. Dzimwe nguva maLLM anogona zvakare "kuona" nekugadzira mhinduro uye kunyangwe zvinyorwa zvenhema kana vasingakwanise kuburitsa mhinduro chaiyo.
Semuenzaniso, muna 2022, iyo nhau agency Fast Company yakabvunza ChatGPT nezve chikamu chemari chekambani chekare Tesla. Nepo ChatGPT yakapa chinyorwa chenhau chakabatana mukupindura, ruzivo rwakawanda rwuri mukati mayo rwakagadzirwa. Sezvo iri AI-based system, inozivikanwa kuti inogara ichivandudza, asi ichiri kukanganisa kuvimba ne100% yemhinduro dzakagadzirwa neLLMs.
Panyaya yechengetedzo, mashandisirwo akatarisana nevashandisi anoenderana neLLMs anowanzo kune bugs sechero imwe application. MaLLM anogona zvakare kushandiswa kuburikidza neruvengo rwekupinza kupa mamwe marudzi emhinduro pamusoro pevamwe, kusanganisira mhinduro dzine njodzi kana dzisina tsika.
Chekupedzisira, imwe yenyaya dzekuchengetedza neLLMs ndeyekuti vashandisi vanogona kurodha yakachengeteka uye yakavanzika data kuti vawedzere chibereko chavo. Asi maLLM anoshandisa mapimendi avanogashira kuti aenderere mberi nekudzidzisa mamodheru avo, uye haana kugadzirwa kuti ave mavhairi akachengeteka, sezvo achigona kufumura data rakadzama mukupindura mibvunzo kubva kune vamwe vashandisi.
LLMs uye huchenjeri huri shure kwemashoko 5el4f
Semwana anosunungurwa muraibhurari hombe, maLLM ane hungwaru maAI masisitimu anodzidza kunzwisisa uye kubereka mutauro wechisikigo wevanhu zvichibva pahuwandu hwedata. Ndichiri kupa akawanda mabhenefiti kune vakajairwa vashandisi uye nekuva chishandiso chine simba chekubatsira munzvimbo yehunyanzvi, kugona uye nenjodzi dzeLLMs zvichiri kuda kunyatso dzidza.
Uye iwe, wakafungei nezve tsananguro mune ino chinyorwa nezveLLMs? Siya maonero ako mumashoko.
ona zvimwe
Source: ElasticSearch, CloudFare, IBM
Yakaongororwa na Tiago Rodrigues musi wa 16/04/2025
