Iyini i-Z.AI, isebenza kanjani, futhi kungani idlondlobala?

I-Z.AI ingene endaweni yesigameko njengendawo yokuxoxa ukuhlakanipha okufakelwayo Ifinyeleleka kusiphequluli, ibeka phambili ukuzizwisa okuqondile, ngaphandle kokubhalisa noma ubunzima. Inzuzo yayo eyinhloko ukuthi isebenza nayo amamodeli siqu ochungechunge lwe-GLM (GLM-4.5 ne-GLM-4.6)Idizayinelwe ukucabanga okuthuthukile, ukukhiqizwa kwekhodi, nemisebenzi ye-ejenti. Uma ufuna enye indlela engancikile kubathengisi bendabuko, uzozizwa usekhaya lapha.

Isiphakamiso sigqama ngenxa yaso ukufinyelela mahhala ngemikhawulo evulekileInikeza izimpendulo ezisheshayo namathuluzi awusizo njengokuhlaziya ifayela nosesho lwewebhu. Ngaphezu kwalokho, ihlinzeka nge-API ekhokhelwayo eqondiswe kubathuthukisi nokuhlanganiswa, kodwa igcina ifilosofi evulekile kumamodeli ayo, evumelana nalabo abafuna ukugwema izithiyo ze-ecosystem kanye ukusebenza ngobuchwepheshe bomthombo ovulekile ngaphansi kwelayisensi ye-MIT.

Yini i-Z.AI futhi ubani ongemuva kwayo?

Empeleni, i-Z.AI iyingxoxo exhaswe amamodeli olimi esizukulwane esilandelayo, egxile kakhulu ukucabanga okuhlelekile futhi uhleloNgokungafani nezinye izinkundla, akudingi ukubhalisa ukuze uqalise: ufinyelela kuyo ngesiphequluli sakho, ubuze umbuzo wakho, bese kuba lokho—akukho zinyathelo zokulinda noma eziphakathi. Lokhu kushesha kuheha abasebenzisi abazisa ubulula futhi Ubumfihlo bususelwa ekungabhalisini.

Umkhiqizo ophambili uxhunywe ku-Zhipu AI (igama eligcwele: I-Beijing Zhipu Huazhang Technology), inkampani eyasungulwa eNyuvesi yaseTsinghua futhi kamuva yaba isiqalo esizimele. E-China, ithathwa njengenye ye "AI Tigers," futhi ngokusho kwe-International Data Group, isivele ibambe iqhaza. isikhundla sesithathu emakethe ye-LLM ezweni, ingqophamlando ebalulekile emkhakheni oqhudelana kakhulu.

Umkhondo wakhe wezezimali nawo uphawulekile: ngo-2023 wakhulisa ngaphezu kwalokho 2,5 billion yuan ngoxhaso lwemidondoshiya efana ne-Alibaba ne-Tencent, kwathi ngo-2024 i-Prosperity7 Ventures yajoyina nokutshalwa kwezimali kwe- Ama-dollar ayi-400ibeka inani elilinganiselwa ku-$3.000 billion. Ngokuhambisanayo, i-Zhipu AI igcine isu elivulekile: amamodeli ayo e-GLM angasetshenziswa ukuthengisa ngaphansi kwelayisensi ye-MIT, alandwe kumakhosombe afana Ubuso Obugonayo noma I-ModelScope ukuze ukubulawa kwendawo.

Omunye umniningwane obalulekile ukuzimela kwezobuchwepheshe. I-Z.AI ayincikile kungqalasizinda ye I-OpenAI noma ku-Cloudflare futhi umemezele ukuhambisana nayo Huawei Ascend kanye nama-Cambricon chipsukuqondanisa nezinye izingxenye zekhompuyutha ku-ecosystem ephethwe ama-Western GPUs. Kubasebenzisi namabhizinisi, lokhu kusho enye indlela yokusebenzisa IA ngaphandle kokuboshelwa kubahlinzeki abangabodwa.

Amamodeli amakhulu namakhono

Inkundla incike kakhulu kuyo I-GLM-4.5 kanye ne-GLM-4.6Lokhu kwakamuva kumenyezelwe ngokwesekwa kwama-chips asekhaya aseShayina kanye nokuthuthukiswa kokucabanga kanye nekhodi. Womabili amamodeli amukela ukwakheka nesibambo se-Mixture of Experts (MoE). amawindi anomongo obanziAmathokheni okufaka afinyelela kwangu-128.000 kanye namathokheni okukhiphayo angaba ngu-96.000, okukuvumela ukuthi uphathe imibhalo emide, izingxoxo ezibanzi, nokuhlaziya okujulile ngaphandle kokulahlekelwa ukuhambisana.

Ezinhlobonhlobo ezithile zokucabanga ezijulile, i-GLM-4.6 iyakwazi ukuhlanganisa okuqukethwe okubanzi kakhulu (ikhono lokusebenza nakho okuqukethwe okungafika kumathokheni angama-200K ekufinyezeni imisebenzi). Ngaphezu kwalokho, inezici eziningi: iqonda umbhalo, ilawula ikhodi, futhi ingasebenza nayo izithombe ekugelezeni kokuhlaziyaephindaphinda izinhlelo zokusebenza.

Esinye isici esihlukanisayo isici se i-agentIsistimu ingaphula imisebenzi eyinkimbinkimbi, ihlele izinyathelo, ishayele ama-API, futhi isebenzise ukulandelana kwezenzo ngokuzenzakalelayo. Lokhu kwenza kube lula ukuhlela imisebenzi ephindaphindwayo noma ukugeleza komsebenzi obekudinga izikripthi ze-ad hoc ngaphambilini, ukuhlinzeka "nge-copilot" esebenza ngalé kwengxoxo elula.

Ngaphakathi kwe-Z.ai ecosystem, amamodeli epharamitha angu-32B nawo avela ukuze ahlolwe futhi aqhathaniswe nezimo, njenge I-GLM-4-32B (isisekelo)Imodeli ye-Z1-32B yokucabanga nokuhluka okucatshangelwayo kwe-Z1-Rumination-32B kufakiwe. Lokhu kunikezwa kuhlose ukuhlinzeka ngomuzwa othi "zero barrier to entry" ukuze noma ubani akwazi ukuhlola amakhono okhiye ngaphandle kwanoma yiziphi izinyathelo zokusetha.

Inkundla ye-Z.ai: ukufinyelela kuwebhu kanye nesithiyo esiziro kulwazi lokungena

Isizinda esiyinhloko sithi chat.z.ailapho ungaxoxa khona ngokukhululeka esipheqululini sakho. Azikho izidingo zokubhalisa ukuze uqalise, futhi isixhumi esibonakalayo sewebhu siyinto encane kakhulu: uthayipha umbuzo wakho ebhokisini lombhalo, uwuthumele, futhi uthole impendulo. Ngaphezu kwengxoxo, izinketho eziwusizo zifakiwe, njenge Ukulayishwa kwefayela nokuhlaziya (TXT, PDF, iZwi) kanye nosesho lwe-inthanethi lokunothisa izimpendulo ngolwazi lwakamuva.

Inkundla akukona nje okokuhlola: onjiniyela "bangagxumela" endaweni ye-Zhipu's MaaS ngokuchofoza okukodwa. bigmodel.cn ukuze uthole okhiye be-API futhi uhlanganise amamodeli kwewakho Izinhlelo zokusebenzaLeyo "switch esheshayo" yakhelwe ukufinyeza indlela ye- test ekuthuthukisweni kanye nokunika amandla amaqembu ochwepheshe ukuthi ashintshe ukusuka kumodi yokuhlola aye ekukhiqizeni okunokungqubuzana okuncane.

Isevisi yesipiliyoni eku-inthanethi imahhala ngaphakathi kwemikhawulo ethile yansuku zonke, evulekile kubasebenzisi abaningi. Uma ufinyelela umkhawulo, kuzodingeka ulinde kuze kube usuku olulandelayo ukuze uqhubeke usebenzisa ingxoxo, ngakho-ke kufanelekile ukuthatha ikhefu. thuthukisa ukusetshenziswa kwamathokheni ukukhetha imodeli efanele yomsebenzi ngamunye.

Ungayiqala kanjani ingxoxo yakho yokuqala

Ukuqalisa kulula njengokuya kusayithi, ukukhetha imodeli, bese ufaka umbuzo wakho. Ayikho i-akhawunti edingekayo, ngakho-ke oxhumana naye wokuqala uyashesha futhi, kwabaningi, kulula kakhulu ngokugwema ukubhaliswa. Ngezinyathelo ezine, uzobe uvuka futhi uyasebenza: Ngena ngemvume, khetha imodeli, buza umbuzo wakho, bese uqhubeka nengxoxo..

1. Finyelela ku-Z.AI kusuka kusiphequluli sakhoAkudingi ukubhaliswa, ngakho-ke ungazama "ngokundiza." Lokhu kuntuleka kwamaphepha okuqala kuvame ukuhlotshaniswa nolwazi oluyimfihlo futhi oluqondile.
2. Khetha imodeli efaneleUmndeni we-GLM unikeza okuhlukile kokucabanga, ikhodi, ukusesha kwewebhu, nezifinyezo. Ukukhetha okuhlukile okulungile kuthuthukisa ikhwalithi yokukhipha futhi kunciphisa ukusetshenziswa kwamathokheni.
3. Yisho ngokucacile isicelo.Nikeza umongo, chaza izinjongo, cela isitayela, futhi, uma kufanele, faka izibonelo ezimfushane. Isaziso esakhiwe kahle senza wonke umehluko.
4. Iphinda emculweni ofanayoIngxoxo igcina umongo, ngakho-ke ungacela ukulungiswa futhi ujule ngaphandle kokulahlekelwa uchungechunge noma ukuhambisana.

Iyiphi imodeli elungele umsebenzi ngamunye

Ukukhetha okuhlukile kusiza ukukhulisa ukusebenza kahle futhi ugweme ukweqa isibonelelo sakho sansuku zonke. Nazi iziqondiso ezisebenzayo zemisebenzi evamile, kucatshangelwa GLM-4.5 kanye GLM-4.6 okukhethekile.

• Izifinyezo ezindeI-GLM-4.6 enokucabanga okujulile iyikhandidethi elingcono kakhulu lamadokhumenti amade (okuqukethwe kucashunwe kufika kumathokheni angu-200K). Ngokuvamile ihlanganisa ngokunembile futhi igcina isakhiwo.
• Ukuhlela nokulungisa iphuthaUkwehluka okwenziwe ngekhodi kwe-GLM-4.6 kufakazele ukuthi yenza kahle kakhulu izimbangi ezimeni ezingama-74 ezisebenzayo, cishe ukusebenza kahle okungama-30%..
• Imibuzo ejwayelekile kanye nokubhala okuyisisekeloI-GLM-4.5 ibhalansisa ikhwalithi yethokheni nezindleko, ilungele imibuzo yansuku zonke nemibhalo evamile.
• Ukuhlaziywa kwefayela nosesho lwewebhuI-GLM-4.6 enomsebenzi Wosesho Lwewebhu ihlanganisa ukuhlaziya idokhumenti nemibuzo eku-inthanethi futhi inikeza indlela ye-multimodal.

Khumbula ukuthi ukusetshenziswa kwamathokheni Iyashintsha ngokuya ngemodeli nomsebenziNgakho-ke, kuhle ukubukeza uhlobo lombuzo, ithoni, nobude obufunayo ngaphambi kokuthumela, ukuze "ungashi" i-quota ekusebenzisaneni okungadingekile.

Izinhlelo, imikhawulo, nama-API

Ingxoxo eyisisekelo imahhala futhi inemikhawulo emikhulu. Uma izidingo zakho zidinga ukuhlanganiswa kuzinqubo noma imikhiqizo, i-API itholakala ngohlelo lwe ukubhalisa (hhayi ngamathokheni)ngohlelo lwezinga lokungena olungabizi kakhulu oluqala cishe ku-$3 ngenyanga. Ukusuka lapho, ulungisa uhlelo ngokusekelwe kumodeli nomthwalo.

Iphaneli ye-API ikuvumela ukuthi udale okhiye, usethe imikhawulo yokusetshenziswa, futhi ulawule izinkokhelo. Ukuhamba komsebenzi okuklanyelwe ukusuka ukuhlolwa komshayeli Ngaphandle kwezithiyo, kokubili kubasebenzisi ngabanye abanezifiso zokuzenzakalela kanye namathimba afuna ukulawula izindleko nokubonakala.

Umuzwa wedeskithophu

Uma uthanda ukuba namathuluzi akho kumawindi azinikele, ungakwazi "ukusonga" i-Z.AI njengohlelo lokusebenza lwedeskithophu ku Windows noma i-macOS nge-wrapper engekho emthethweni. Ayingezi izici zomdabu, kodwa inikeza ukusebenziseka kwayo ama-akhawunti amaningi kanye nokugxila ngaphandle kweziphazamiso zesiphequluli.

Imikhuba emihle ukuze uthole okuningi kuyo

Imikhuba embalwa yenza umehluko omkhulu ngekhwalithi kanye nezindleko. Cabanga ngale mihlahlandlela njengekhithi yakho "yokuphepha" ongasebenza nayo Amamodeli e-GLM ngokuphumelelayo kusukela osukwini lokuqala.

• Ukwaziswa ngomongoFaka umgomo, iphimbo, ubude, izethameli okuqondiwe kuzo kanye nesibonelo. Ngokucacile kafushane, kuba lula ukubethela umphumela.
• Hlola-isilinganiso-lungisaImijikelezo emifushane enama-KPI alula (usizo olubonwayo, isikhathi esilondoloziwe, izindleko ngomsebenzi ngamunye). Phinda ufunde ngokushesha ukuze uthuthuke.
• Olunye uhleloGcina izinketho ezi-2-3 (imodeli noma umphakeli) ngokuhambisana ngaphambi kokuzibophezela okungu-100%.
• Isibuyekezo sokuhlelaEzingxenyeni zomphakathi, yengeza ukufundwa komuntu kwesibili nokulawula umthombo ukuze kuqiniswe ukulandeleka.

Ukusuka ekuzameni ukuya ku-MVP emizuzwini eyi-15

Uma ufuna ukusuka kumbono uye ku-prototype ngokuphazima kweso, zama lo mephu womgwaqo. Idizayinelwe ukuqinisekisa ngokushesha ukuthi i-Z.AI iyangena yini ekuhambeni komsebenzi wakho futhi uma ubudlelwano bayo izinga/isivinini/izindleko yeqa isisekelo sakho.

1. Izame kuwebhu: Yakha imibuzo yangempela emi-3–5 (ukumaketha, umkhiqizo noma ikhodi) bese uhlola ikhwalithi nokubambezeleka.
2. Isifinyezo esincane: cela uhlaka nokusalungiswa (ikhasi lokubikezela, i-imeyili noma amazwibela) futhi wenze iziphindaphindo ezimbili zokulinganisa ukucaca namaphutha.
3. I-mini-pilot ene-APIDala ukhiye futhi wenze ngokuzenzakalelayo inqubo ephindaphindayo (isb., isifinyezo sethikithi). Gada amathokheni nezikhathi.
4. Ikala ukusuka kumzuzu wokuqalaUkusebenza okubonwayo, iphesenti lezimpendulo ezivumelekile, isikhathi esilondoloziwe, nezindleko zomsebenzi ngamunye. Uma ingahlangabezani nebhentshimakhi, misa isilingo bese uzama enye indlela.

Ukuqhathanisa okusheshayo namanye ama-AI

Ukuqhathanisa I-ChatGPT, Claude noma GeminiI-Z.AI iqhudelana ngokugcizelela okucacile kokuthi ukucabanga kanye nezindlekoInzuzo yangempela incike olimini, amathuluzi atholakalayo (ukusesha, amafayela, abenzeli), nokubambezeleka endaweni yangakini. Lapho kukhanya khona ngempela kusemisebenzini ehlelekile nesekelwe kukhodi, sibonga i-GLM-4.6.

Ngokufanayo, uma kuqhathaniswa nezinketho ezifana I-DeepSeekI-Z.AI ingazimisa njengenye indlela ukonga kakhulu Ezimweni ezithile, kodwa isinqumo sokugcina sincike kumthamo wakho, izivumelwano, kanye nezakhiwo. Ngokwezilinganiso ezishiwo emithonjeni, uchungechunge lwe-GLM ludlula amamodeli esizukulwane sakamuva se-GPT ekucabangeni nasekubhaleni ikhodi; thatha lokhu njengenkomba ewusizo futhi Iqinisekise esimweni sakho sokusebenzisa ngaphambi kokushaya ihlombe noma ukuxoshwa.

Izici ezisebenzayo: amafayela, iwebhu kanye nendlela ye-multimodal

Ukulayisha amafayela (TXT, PDF, Word) kuyisici esisebenziseka kalula: ulayisha idokhumenti nesifanekiso hlaziya okuqukethwe kwayo ukuphendula ngokusekelwe kukho. Iwusizo kumibiko, amamanyuwali, izisekelo zolwazi lwangaphakathi, noma imibhalo yezomthetho edinga umongo ongashintshi.

Ukukhetha kwe sesha iwebhu Yengeza isendlalelo solwazi olusha, ikakhulukazi uma udinga amaqiniso akamuva noma imithombo ephikisanayo. Khumbula ukuthi umbuzo wosesho ngamunye ungadla izinsiza eziningi, ngakho-ke uwusebenzise kuphela uma ungeza inani ngempela.

Ukusekelwa kwe-Multimodal kuvula umnyango wezinkomba ezibukwayo nemisebenzi lapho izithombe zihambisana nombhalo. Akuwona umhleli wezithombe, kodwa kunalokho i-AI eqondayo izinkomba ezibonakalayo ngaphakathi kohlaka lwenhloso yakho yokuxhumana noma yobuchwepheshe.

Ikhathalogi yamamodeli kumodi "yomhloli".

Ngokungeziwe ku-GLM-4.5/4.6, inkundla inikeza "umhloli" onamodeli we-32B wokuqhathanisa ukuziphatha kwesikhathi sangempela. Lezi zihlanganisa I-GLM-4-32B (isisekelo)I-Z1-32B iqondiswe ekucabangeni, futhi i-Z1-Rumination-32B inokugxila kokucabangela/okujulile. Umgomo uwukuba uzame ngokuhlukile okuhlukile bese ukhetha. akukho mali yokungena leyo elingana kahle kakhulu.

Uma ngemva kokuhlola ufuna ukuthuthela ekukhiqizweni, ungaxhuma endaweni ye-Zhipu's MaaS ngokuchofoza okukodwa kokuthi bigmodel.cn ukuphatha ama-API, imikhawulo, nezinkokhelo. Lokhu kuqhubeka kunciphisa ukungqubuzana phakathi kokuthi "Ngithanda engikubonayo" nokuthi "asikuhlanganise."

Ibhizinisi, uxhaso kanye namasu omthombo ovulekile

I-Zhipu AI ithole uxhaso olusezingeni eliphezulu, ngokutshalwa kwezimali okuvela ku-Alibaba, Tencent, kanye nesikhwama se-Prosperity7. Ngokungeziwe ekulinganiseni kwayo, [ulwazi olulahlekile - okungenzeka inkundla ethile noma isu] iyagqama. ukubheja kumthombo ovulekileAmamodeli e-GLM angadawunilodeka futhi asetshenziswe ngaphansi kwe-MIT, okuvumela ukuthunyelwa kwendawo ngaphandle kokuboshelwa kumasevisi avaliwe.

Ukuhambisana nenye ihadiwe (Huawei I-Ascend, Cambricon) kanye neqiniso lokuthi abancikile ku-OpenAI noma ku-Cloudflare kugcizelela ukulandisa kwabo kokuzimela. Ezinkampanini eziningi, lokhu kuyisihluthulelo: ahlukanise abahlinzeki futhi anciphise ubungozi ubungcweti bobuchwepheshe nohwebo sebuyisidingo.

Izibonelo kanye namacala okusetshenziswa

Ekusetshenzisweni kwansuku zonke, i-Z.AI isebenza kahle njengomsizi we phendula achaze okuqukethwe imibhalo yezobuchwepheshe noma ejwayelekile, yokubhala nokuhlela (izihloko, amanothi angaphakathi, Ama-imeyili), futhi njengokwesekwa kwentuthuko (ukukhiqiza, ukulungisa iphutha nokuchaza ikhodi, ukuphakamisa ukuhlolwa, ukuphakamisa iziqinisi).

Futhi ingena ekusekelweni kwangaphakathi: ukuguqula amadokhumenti abe yizimpendulo zohlobo lwesisekelo solwazi, ukudala izifinyezo eziphezulu zemibiko, noma ukulungiselela. kafushane kanye FAQs emaqenjini angewona awobuchwepheshe. Imodi yomenzeli wayo isiza ukuhlanganisa izinyathelo ndawonye, ​​ikuvumela ukuthi wenze ngokuzenzakalela izinqubo ezilula ngaphandle kokuhlela okudlulele.

Njengesibonelo esisebenzayo, ukwaziswa kwesitayela "Dala i-a ideshibhodi esebenzisanayo okukhombisa ukunyuka komoya ogwini lwase-Andalusia,” futhi kwabonwa ngeso lengqondo imiphumela ewusizo ye-prototyping.” Lolu hlobo lokuhlola okusheshayo lusebenza ukulinganisa uku izinga elibonakalayo nesivinini ngaphambi kokuhlanganisa ithuluzi ekukhiqizeni.

I-Z.AI ibekwe njengomsizi oguquguqukayo ohlanganisa ukufinyelela kwamahhala ngokushesha, amamodeli aphezulu we-GLMInikeza ukusebenza komenzeli kanye nebhuloho elicacile lama-API wamaqembu. Uma uzazisa izindleko, ukusebenza, kanye nenkululeko yezobuchwepheshe, kufanele ukucatshangelwa ngokujulile; linganisa kusukela ekuqaleni bese unamathela kulokho okufanelana nezidingo zakho zangempela.

I-athikili ehlobene:

I-Intel Gaudi 3 manje isiyatholakala ku-IBM Cloud: enye indlela enamandla nengabizi yebhizinisi i-AI.

Isaka

Umbhali oshisekayo ngomhlaba wamabhayithi nobuchwepheshe ngokujwayelekile. Ngiyathanda ukwabelana ngolwazi lwami ngokubhala, futhi yilokho engizokwenza kule bhulogi, ngikubonise zonke izinto ezithakazelisayo kakhulu ngamagajethi, isofthiwe, ihadiwe, izitayela zobuchwepheshe, nokuningi. Inhloso yami ukukusiza ukuthi uzulazule emhlabeni wedijithali ngendlela elula nejabulisayo.