Ingabe Intuthuko ye-AI Iyahamba Kancane? I-Scaling Debate OpenAI Ayifuni Ukuba Nayo

Ngicabanga ukuthi i-OpenAI ayithembekile mayelana nembuyiselo enciphayo yokukala i-AI ngedatha futhi ibale iyodwa. Ngicabanga ukuthi nabo babeka umnotho omningi, umhlaba kanye nayo yonke le mboni engcupheni ngokungakhulumi ngokukhululeka ngesihloko.

Ekuqaleni, ngangikukholwa ababesitshela kona, ukuthi okudingeka ukwenze nje ukwengeza amandla ekhompiyutha engeziwe nedatha eyengeziwe, futhi ama-LLM kanye namanye amamodeli azoba ngcono. Ukuthi lobu budlelwano phakathi kwamamodeli, ikhompuyutha yawo nedatha kungakhula ngokulandelana kuze kube sekupheleni kwesikhathi. Ukweqa okuvela ku-GPT-3 naku-GPT-3.5 bekukukhulu. Ukweqa kusuka ku-GPT-3.5 kuya ku-GPT-4 kubonakala njengobufakazi obucacile bokuthi lokhu kucabangela kwakuyiqiniso. Kodwa-ke izinto zaba yinqaba.

Esikhundleni sokukhulula imodeli ebizwa nge-GPT-5 noma i-GPT-4.5, bakhiphe i-GPT-4-turbo. I-GPT-4-turbo ayihlakaniphile njenge-GPT-4 kodwa ishesha kakhulu futhi ishibhile. Konke lokho kunengqondo. Kodwa-ke, lo mkhuba waqhubeka uqhubeka.

Ngemuva kwe-GPT-4-turbo, ukukhishwa okulandelayo kwe-OpenAI kwaba yi-GPT-4o (sitrobheli). I-GPt-4o ihlakaniphe kakhulu njenge-GPT-4-turbo, kodwa ishesha kakhulu futhi ishibhile. Umsebenzi owasithengisa ngempela, nokho, bekuyikhono layo lokukhuluma nokuqonda izinto ngomsindo nesivinini sawo. Nokho, qaphela, kulesi sikhathi endabeni yethu, i-GPT-4-turbo ayihlakaniphe kakhulu kune-GPT-4 futhi i-GPT-4o ayihlakaniphe kakhulu kune-GPT-4-turbo. Futhi akekho kubo ohlakaniphe kakhulu kune-GPT-4.

Ukukhishwa kwabo okulandelayo nokwakamuva kwaba yi-GPT-o1. I-GPT-o1 ingenza kangcono kune-GPT-4 kweminye imisebenzi. Kodwa lokho kungenxa yokuthi i-o1 akuyona imodeli eyodwa ngempela. I-GPT-o1 empeleni iyibhokisi elimnyama lamamodeli amaningi we-LLM angasindi asebenza ndawonye. Mhlawumbe i-o1 ichazwa kangcono njenge-software noma i-middleware kunemodeli yangempela. Uyinika umbuzo, iqhamuke nempendulo, bese iphinda isebenzisa amanye amamodeli anikezwe umsebenzi wokuhlola impendulo ukuze iqinisekise ukuthi ilungile, futhi ifihla yonke le misebenzi. Ikwenza konke lokhu ngokushesha okukhulu.

Kungani ungavele wenze i-LLM enamandla kune-GPT-4? Kungani uphendukela kumasu anjalo wengubo-nensangu ukuze uzuze ukukhishwa okusha? I-GPT-4 yaphuma eminyakeni engu-2 edlule, kufanele sibe ngaphezu kwamandla ayo okwamanje. Nokho, uNoam Brown, umcwaningi kwa-OpenAI ubenokuthile angakusho ukuthi kungani behambe lo mzila no-o1 e-TED AI. Uthe: “Kuvele ukuthi ukuba ne-bot think imizuzwana engama-20 esandleni se-poker kusebenze ngendlela efanayo njengokukhuphula imodeli ngo-100,000x nokuyiqeqeshela izikhathi eziyi-100,000 ubude,”

Manje yima futhi ucabange ngempela ngalokho okushiwo lapho. Ukucabanga kwe-bot imizuzwana engu-20 kuhle njenge-bot eqeqeshwe izikhathi ezingu-100,000 isikhathi eside enamandla aphindwe ka-100,000 ekhompiyutha amaningi. Uma imithetho yokukala ingapheli, leyo zibalo ayinakwenzeka. Kukhona okungalungile lapha noma kukhona oqamba amanga.

Kungani konke lokhu kunendaba? I-OpenAI ibiza amadola ayizigidi eziyizinkulungwane ezingu-150 futhi ingxenye enkulu yaleyo makethe isekelwe ekuqageleni okuncike ekuthuthukisweni kwamamodeli ngokuhamba kwesikhathi. Uma i-AI iyinhle njengoba injalo namuhla, lelo kuseyikusasa elithakazelisayo, kodwa akusikho lokho okuthengiswa kubatshalizimali yizinkampani ze-AI ezine-IP yazo yonke imodeli yazo. Lokho futhi kushintsha umgwaqo womkhiqizo wezinye izinkampani eziningi ezincike ekuthuthukisweni okuqhubekayo kwama-LLM azo ukuze zakhe imikhiqizo yazo. Umgomo we-OpenAI kanye nezifiso ze-AGI zibambezeleka kakhulu uma konke lokhu kuyiqiniso.

I-hypothesis

Isizathu sokuthi ama-LLM amangalisa kangaka kungenxa yezinga eliphezulu lefilosofi esingakaze siyicabange, lolo limi ngokwemvelo lunomongo omkhulu kakhulu kanye nedatha emayelana nomhlaba ngaphakathi kwezingxenye ezincane zombhalo. Ngokungafani namaphikseli esithombeni noma kuvidiyo, amagama emshweni achazana ngokungagunci. Umusho ohlangene ngokuphelele ngokwencazelo, “okunengqondo”. Ukuthi kuyiqiniso noma cha kuyindaba ehluke kakhulu nenkinga edlula ulimi lodwa. Noma ngabe udla umbhalo ongakanani, “iqiniso” kanye “namanga” akuwona nje imiqondo yolimi. Ungasho ukuthi okuthile kunengqondo ngokuphelele kodwa akulona neze “iqiniso”. Kungalesi sikhathi ama-LLM azohlala eshaya udonga lwezitini. Kulezi zinyanga eziyi-12 ezidlule, ngingathanda ukuqagela ngokusemthethweni ukuthi ngemuva kweminyango evaliwe akubanga khona ukugxuma okukhulu kuma-LLM e-OpenAI, GrokAI, noma kwa-Google. Ukucacisa angicabangi ukuthi ukhona umuntu, noma yikuphi owenze noma iyiphi i-LLM eyi-1.5X engcono kune-GPT-4.

Kwa-OpenAI kubonakala sengathi abasebenzi bezinga eliphezulu bayayeka. Njengamanje bathi kungenxa yokuphepha kodwa ngizofaka isigqoko sami se-tinfoil manje bese ngiphonsa umbono lapho. Bayalwazi lolu daba futhi bagxumela umkhumbi kungakephuzi kakhulu.

Ukuqinisekisa

Ngaqala ukuxoxa ngalokhu kukhathazeka nabangane 3 izinyanga ezedlule. Ngabizwa ngamagama amaningi haha.

Kodwa emavikini angu-3 adlule, abezindaba abaningi sebeqalile ukuhogela into eshisayo:

• I-OpenAI ayisakhiphi i-Orion (GPT-5) ngenxa yokuthi ayizange ihlangabezane namabhentshimakhi okusebenza alindelwe futhi ibona ukuncipha kwembuyiselo. ( https://www.theinformation.com/articles/openai-shifts-strategy-as-rate-of-gpt-ai-improvements-slows )

  
  

• I-Bloomberg ibika ukuthi i-OpenAI, i-Google kanye ne-Anthropic bonke banenkinga yokwenza i-AI ethuthuke kakhulu. ( https://www.bloomberg.com/news/articles/2024-11-13/openai-google-and-anthropic-are-struggling-to-build-more-advanced-ai )
  

Yini esingayenza ngakho?

Kunzima ukuncoma isisombululo esisodwa. Ubuchwepheshe obungemuva kwe-o1 buwubufakazi bokuthi ngisho namamodeli asebenza kancane angaphinde ahloselwe ukwenza imisebenzi eyinkimbinkimbi. Kepha leso akusona isixazululo senkinga yokukala kwe-AI. Ngicabanga ukuthi kufanele kube nokutshalwa kwezimali okukhulu kanye nokuhlolwa okusheshayo kwezakhiwo ezintsha zamamodeli. Sesiphelelwe yidatha futhi sidinga izindlela ezintsha zokwengeza idatha esebenzisekayo ukuze ama-LLM aqeqeshwe ngazo. Mhlawumbe ukusebenzisa ilebula ye-multidimensional esiza ukuqondisa izithenjwa zayo ukuze uthole ulwazi oluyiqiniso ngokuqondile. Omunye umqondo omuhle kungaba ukumane uqhubeke nokulungisa kahle ama-LLM ezimweni ezithile zokusetshenziswa njengezibalo, isayensi, nokunakekelwa kwezempilo okusebenzayo nokusebenzisa ukuhamba komsebenzi kwe-ejenti ye-AI, efana ne-o1. Kungase kunikeze izinkampani eziningi ithuba lokunyakazisa kuze kube yilapho kuvela isakhiwo esisha. Le nkinga yimbi impela kodwa ngicabanga ukuthi ubuhlakani bokufunda komshini kanye nokuthuthukiswa kwesoftware okuzoyigqugquzela kuzoba kukhulu. Uma sesidlulile lesi singqinamba, sizobe sesisebenza kahle ohlelweni lwe-AGI futhi mhlawumbe ne-ASI.