I-ZeroSearch: Inguquko ka-Alibaba yokuqeqesha i-AI kahle nangokuzimele

Ikhava » Intelligence yokufakelwa » I-ZeroSearch: Inguquko ka-Alibaba yokuqeqesha i-AI kahle nangokuzimele

Ukuqamba okusha emkhakheni wobuhlakani bokwenziwa kuye kwaqhuma eminyakeni yamuva, ikakhulukazi maqondana nezinhlobo zezilimi ezinkulu (LLMs). Enye yempumelelo ebaluleke kakhulu yango-2025 kube i-ZeroSearch, ubuchwepheshe obuthuthukiswe i-Alibaba obunyakazisa izisekelo zokuthi lawa mamodeli aqeqeshwa kanjani. Imayelana nani ngempela i-ZeroSearch, futhi kungani idala umsindo ongaka embonini? Kulesi sihloko, sibheka kabanzi le ndlela entsha, okuhlanganisa ukuthi isebenza kanjani, yiziphi izinzuzo ezinikezayo kunezindlela zendabuko, nokuthi ingashintsha kanjani ukuthuthukiswa kwe-AI kuwo wonke amazinga.

Emibuthanweni yobuchwepheshe, inkulumo imayelana nakho: I-ZeroSearch ithembisa ukunciphisa izindleko zokuqeqesha zamamodeli obuhlakani bokwenziwa ngokungekho ngaphansi kuka-88%.. Lokhu kweqa ekusebenzeni kahle, kunokuba kube iqhinga lokumaketha nje, kunemithelela ejulile kumabhizinisi amakhulu namancane, abathuthukisi, futhi, vele, ekuthuthukisweni kobuhlakani obujwayelekile bokwenziwa.

Iyini i-ZeroSearch futhi ivela kuphi?

I-ZeroSearch iyindlela entsha yokuqinisa esekelwe ekufundeni eklanyelwe ukuqeqesha amamodeli olimi ngaphandle kokuthembela ezinjinini zokusesha zangaphandle zangempela phakathi nenqubo yokuqeqeshwa. Lokhu okusha kuvela elabhorethri yase-Alibaba's Tongyi, ngenhloso yokuxazulula izinkinga ezimbili ezivamile ekuqeqesheni amamodeli e-AI asebenzisa ukusesha kuwebhu: i izindleko zezomnotho eziphezulu ukusetshenziswa kwama-API kanye ukungaqiniseki ngekhwalithi yamadokhumenti atholiwe.

Kuze kube manje, ukwakhiwa kwabasizi abathuthukile, ama-chatbots, noma izinjini zokuncoma kudinga ukuthumela amashumi ezinkulungwane zemibuzo ezinjinini zokusesha ezifana ne-Google ngamasevisi akhokhelwayo, kukhushulwe izindleko nokukhawulela ukukaleka, ikakhulukazi ezinkampanini ezinesabelomali esiqinile.

I-ZeroSearch ishintsha imithetho yomdlalo ngokubheja ohlelweni lapho I-LLM ngokwayo ifunda ukulingisa ukusebenza kwenjini yokusesha, ekhiqiza amadokhumenti abalulekile noma anomsindo (angabalulekile) ephendula imibuzo futhi ngaleyo ndlela avumele ukuqeqeshwa ngaphandle kokuxhumana kwangaphandle.

Isebenza kanjani i-ZeroSearch? Incazelo yezobuchwepheshe enemininingwane

Enhliziyweni ye-ZeroSearch wuhlaka lokufunda oluqinisayo (RL) olususa isidingo sosesho lwangempela lwewebhu phakathi nokuqeqeshwa. Ake sibheke le nqubo isinyathelo ngesinyathelo, ngokusekelwe endleleni ka-Alibaba kanye nokuhlaziywa okushicilelwe okushicilelwe kwesu.

1. Ukushuna okugadiwe okungasindi ukuze kulingise ukusesha

Konke kuqala kokukodwa ukulungiswa kahle okugadiwe (SFT) lapho i-LLM iqeqeshelwa khona ukuziphatha njengemojula yokubuyisa ulwazi. Ngalokhu kulungiswa, ifunda ukwenza imibhalo yezimpendulo zemibuzo, ilingise isitayela sombhalo kanye nohlobo lokuqukethwe okunganikezwa injini yokusesha yangempela. Phakathi nalesi sigaba sokuqala, izindlela zokusebenzisana phakathi kwemodeli nenjini yokusesha ziyaqoqwa, kutholwe amarekhodi emibuzo kanye namadokhumenti atholakele.

Izindlela eziphumelelayo, okungukuthi, lezo eziholela empendulweni efanele, zilebulwe njengeziqondile (imibhalo ewusizo), kuyilapho lezo ezibangela amaphutha noma izimpendulo ezingalungile zimakwa njengezinegethivu (imibhalo enomsindo). Lo mehluko kamuva uzosiza imodeli ukuthi iqonde futhi ikhiqize kabusha ukuguquguquka kokusesha okungokoqobo, okuhlanganisa imiphumela efanelekile kanye neminye engasebenziseki kangako.

2. Iqhaza lokuqinisa ukufunda ngokulingisa kwekharikhulamu

Ngemva kokushuna okugadiwe, imodeli idlulela esigabeni sokuqeqesha sokuqinisa, lapho izenzo ezinhle ziqiniswa khona futhi amaphutha ajeziswa. Lapha, i-LLM elungisiwe ngokwayo isebenza njengenjini yokusesha, iphendula imibuzo ekhiqizwe imodeli yenqubomgomo kanye nokubuyisela amadokhumenti angase abe usizo noma abe nomsindo.

Ubunzima bemodeli buyanda ngokuqhubekayo, kulandela isu lekharikhulamu elehlisa kancane kancane ikhwalithi yamadokhumenti akhiqizwayo, ukuze Uhlelo kuqala lufunda ezindaweni ezilawulwayo futhi, njengoba luqhubeka, lubhekana nezibonelo ezinomsindo ezikhulayo noma eziyinkimbinkimbi.. Le ndlela isiza imodeli ukuthi ithuthukise amakhono okusesha aqinile kanye nokucabanga ngaphansi kwezimo ezingokoqobo.

3. Idizayini yemiklomelo namamethrikhi okuhlaziya

Ukuqondisa ukufunda, i-ZeroSearch isebenzisa umsebenzi wokuthi umvuzo osuselwe kumphumela we-F1, okulinganisa ukunemba nokukhumbula ngokucabangela ukufanisa igama phakathi kwesibikezelo nempendulo efanele. Umgomo uwukukhulisa ukunemba kwezimpendulo zokugcina imodeli ekwaziyo ukuzikhiqiza, ngaphandle kokukhathazeka ngokweqile ngokufometha, njengoba ama-LLM ngokuvamile akhiqiza imibhalo efomethiwe kahle ngokwemvelo.

4. Ukusebenzisana kwama-Multi-turn kanye nezifanekiso zokucabanga

Ngesikhathi sokuqeqeshwa, izifanekiso zokusebenzisana zisetshenziswa ezihlukanisa inqubo ibe izigaba ezintathu: ukucabanga kwangaphakathi (ihlukaniswe phakathi kwamathegi afana <think>...</think>), eqhuba ukubonisana (<search>...</search>) futhi isizukulwane sokuphendula (<answer>...</answer>). Lokhu kuvumela imodeli ukuthi ithuthukise amandla ayo okwenza imibuzo efanele futhi inikeze izimpendulo ezinesisekelo esihle.

5. Ukuhambisana kanye scalability

I-ZeroSearch isekela izinhlobo zezilimi ezinkulu, ezifana nomndeni wakwaQwen-2.5, Qwen-2.5, LLaMA-3.2 kanye nokuhlukahluka okuyisisekelo noma okucushwe imiyalelo. Ngaphezu kwalokho, ingasetshenziswa ngama-algorithms ahlukene okuqinisa (i-PPO, i-GRPO, phakathi kokunye), okusiza ukwamukelwa kwayo ezindaweni ezihlukahlukene zokuthuthukiswa.

Idatha yomhlaba wangempela: I-ZeroSearch ilondoloza malini futhi isebenza kanjani?

Ukuhlola okwenziwa i-Alibaba futhi okubikwe ekushicilelweni okukhethekile namakhosombe kubonisa ukuthi i-ZeroSearch ifinyelela ukusebenza okuqhathaniswa, futhi nangaphezu kwalokho, okutholwe ngezinjini zokusesha zangempela zezentengiso.. Ukonga izindleko kuphawuleka kakhulu:

• Ukwenza imibuzo engu-64.000 usebenzisa i-Google Search API kungabiza cishe Amadola ka-586,70 (cishe. €540).
• Ivolumu yombuzo ofanayo, ekhiqizwe futhi yaphathwa ngepharamitha eyizigidi eziyizinkulungwane eziyi-14.000 LLM kusetshenziswa i-ZeroSearch, yehlisa izindleko zibe nje. Amadola ka-70,80 (cishe u- € 65).
• Lo mehluko usho ukuthi a 88% ukonga ezindlekweni zokuqeqesha, ukususa ukuncika kuma-API angaphandle nokuvumela ukukala okukhulu.

Ngakolunye uhlangothi, imiphumela yekhwalithi iyamangalisa: ukuhlolwa kubonisa ukuthi imojuli yokubuyiswa kwepharamitha engu-7B ifana nokusebenza kwezinhlelo ezisuselwe Kusesho lwe-Google, kuyilapho ngamapharamitha angu-14B imodeli idlula ngisho emisebenzini yemibuzo nezimpendulo, isebenzisa kokubili i-single-hop kanye nedathasethi eyinkimbinkimbi ye-inference.

Izinzuzo ezibalulekile kanye nomthelela embonini yezobunhloli bokwenziwa

Ukufika kwe-ZeroSearch kumelela ushintsho olukhulu endleleni izinkampani nabathuthukisi abangabhekana ngayo nokuqeqeshwa kwamamodeli athuthukile.:

• Ukwehliswa okukhulu kwesithiyo somnotho: Isiza ukufinyelela kumasu athuthukile e-AI yama-SME, iziqalo, nonjiniyela abazimele ababekade begodliwe izindleko zama-API okuhweba.
• Ukulawula okukhulu ekuqeqesheniNgokukhiqiza amadokhumenti alingisa, amaqembu angachaza ngqo ukuthi yiluphi ulwazi olutholwa yimodeli, alungise ubunzima nekhwalithi ukuze ivumelane nezidingo zabo.
• Ukuthuthukisa ukuzimela kwezobuchwepheshe: Yehlisa ukuncika kumapulatifomu amakhulu obuchwepheshe bangaphandle, iphromotha ukuthuthukiswa kwendawo kwezixazululo ze-AI ezenziwe ngokwezifiso.
• Ukuzivumelanisa nezimo kanye ne-modularityI-ZeroSearch ingasatshalaliswa kumamodeli ahlukahlukene futhi ihambisane nokugeleza komsebenzi okuhlukene kanye nezidingo zebhizinisi.

Umehluko kumasu wangaphambilini: i-RAG, ukusesha kwangempela kanye nokulingiswa

Ngaphambi kwe-ZeroSearch, isixazululo esivame kakhulu sokuhlinzeka ngolwazi lwakamuva nolunembile kuma-LLM kwakuwukusetshenziswa kwe-RAG (Retrieval-Augmented Generation), lapho imodeli ibuza imithombo yangaphandle isebenzisa ukusesha komhlaba wangempela. Nokho, lokhu kuveza ezinye izinkinga ezisobala:

• Izindleko eziphezulu: Ukusetshenziswa okuqhubekayo kwama-API kungakhuphula ibhajethi.
• Ikhwalithi eguquguqukayo: Amadokhumenti atholiwe angase ahluke kakhulu kuye ngosesho kanye ne-API ngokwayo.
• Imikhawulo yezomthetho nemfihlo: Ukuthembela ezinsizeni zezinkampani zangaphandle kuhilela ubungozi bezomthetho nezepolitiki, ikakhulukazi uma uqeqesha ngolwazi olubucayi.

I-ZeroSearch iqeda isidingo sokuqhubeka ubhekisela emithonjeni yangaphandle, ivumela imodeli ukuthi ifunde ukusesha "ngaphakathi kwayo" njengoba ilingisa ulwazi lokusebenzelana nenjini yokusesha.

Umthelela kanye nezicelo zangempela zempilo: kusuka ku-Quark kuya ku-democratization ye-AI

I-Alibaba isivele ihlanganise i-ZeroSearch emikhiqizweni yezentengiso. Uhlelo lwabo lwe-Quark, oluxhaswe amamodeli akwa-Qwen, lubone ukuthuthuka okubalulekile ekucabangeni nasekuphenduleni okunembile emibuzweni eyinkimbinkimbi ngenxa yale nqubo. Kodwa mhlawumbe into ebaluleke kakhulu yilokho I-ZeroSearch ivula umnyango wezinkampani ezincane ukuthi zenze amamodeli azo athuthukile ngaphandle kwesidingo sengqalasizinda yangaphandle ebizayo..

I-athikili ehlobene:

Konke odinga ukukwazi ngoManus, umenzeli we-AI ofuna ukwenza umsebenzi wakho

Umphakathi wocwaningo uyakwazi ukufinyelela inqolobane yekhodi, amasethi edatha, namamodeli aqeqeshwe kusengaphambili kukho kokubili i-GitHub ne-Hugging Face, okukhuthaza ukutholwa nokuhlolwa komhlaba wonke.

Lizobukeka kanjani ikusasa lokuqeqeshwa kwe-AI ngenxa yeZeroSearch?

Njengoba la masu ekhula, sizobona ukwanda kwabasizi abahlakaniphile abanamakhono okusesha athuthukisiwe ngaphandle kokuthembela ku-Google, i-Bing, noma okufanayo. Lokhu kuvula amathuba amasha kwezemfundo, ibhizinisi, kanye nocwaningo, kuyilapho kucekela phansi ukubusa kwezinjini ezinkulu zokusesha emkhakheni wezobunhloli bokwenziwa.

E-Spain naseYurophu, lokhu kumelela amathuba okukhula okuzimele, ukunciphisa ukuncika kwezobuchwepheshe kanye nezindleko, kanye nokulawula kwamasu okukhulu ezinhlelweni zolwazi olubalulekile.

Ukunyuka kwe-ZeroSearch kuphawula ukuqala kwenkathi entsha lapho amamodeli e-AI okuqeqesha azoyeka ukuba ubumnandi obutholakala kwabambalwa abakhethiwe futhi abe ithuluzi elifinyeleleka kalula, elinwebekayo, nelithuthukayo. Ngokufundisa i-AI ukusesha ngaphandle kokushiya indawo yayo, i-Alibaba ithathe isinyathelo esikhulu ekuthuthukiseni izinhlelo ezizimele, ezisebenza kahle ezivumelana nanoma yisiphi isidingo.. Akuseyona nje eyokwehlisa izindleko, kodwa mayelana nokusungula kabusha imithetho yegeyimu yayo yonke imboni yezobunhloli bokwenziwa.

I-athikili ehlobene:

Kuyini i-e-commerce: Izihluthulelo eziyi-10 Zokuqonda Ukuhweba Kwe-elekthronikhi