Okwe-BMS, BUS, Industrial, Instrumentation Cable.
U-Elon Musk kanye nethimba le-xAI bethule ngokusemthethweni inguqulo yakamuva ye-Grok, Grok3, ngesikhathi sokusakaza bukhoma. Ngaphambi kwalo mcimbi, inani elibalulekile lolwazi oluhlobene, oluhlanganiswe ne-Musk's 24/7 hype yokukhangisa, kukhuphule okulindelwe emhlabeni jikelele kwe-Grok3 kumazinga angakaze abonwe. Esontweni eledlule, uMusk washo ngokuzethemba ngesikhathi esakaza bukhoma ngenkathi ephawula ku-DeepSeek R1, "i-xAI isizokwethula imodeli ye-AI engcono." Kusukela kudatha ethulwe bukhoma, i-Grok3 kubikwa ukuthi idlule kuwo wonke amamodeli ajwayelekile amanje ezilinganiso zezibalo, isayensi, nezinhlelo, kanti u-Musk waze wathi i-Grok3 izosetshenziselwa imisebenzi yekhompyutha ehlobene nomsebenzi we-SpaceX's Mars, ebikezela "impumelelo ezingeni Lomklomelo KaNobel phakathi neminyaka emithathu." Kodwa-ke, lokhu okwamanje kungukugomela kukaMusk. Ngemva kokwethulwa, ngihlole inguqulo yakamuva ye-beta ye-Grok3 futhi ngabuza umbuzo wobuqili wamamodeli amakhulu: "Iyiphi enkulu, 9.11 noma 9.9?" Ngokudabukisayo, ngaphandle kwanoma yiziphi iziqu noma amamaki, okuthiwa i-Grok3 ehlakaniphe kakhulu namanje ayikwazanga ukuphendula lo mbuzo ngendlela efanele. I-Grok3 yehlulekile ukukhomba ngokunembile incazelo yombuzo.
Lokhu kuhlola kusheshe kwadonsa ukunaka okukhulu kubangani abaningi, futhi kwaqondana, ukuhlola okufanayo okuhlukahlukene phesheya kwezilwandle kubonise i-Grok3 idonsa kanzima nemibuzo eyisisekelo ye-physics/mathematics njengokuthi "Iliphi ibhola eliwa kuqala ukusuka ku-Leaning Tower of Pisa?" Ngakho-ke, kuye kwabhalwa ngokuhlekisayo ngokuthi "ungqondongqondo ongazimisele ukuphendula imibuzo elula."
I-Grok3 yinhle, kodwa ayingcono kune-R1 noma i-o1-Pro.
I-Grok3 ihlangabezane "nokwehluleka" ezivivinyweni eziningi zolwazi olujwayelekile ekusebenzeni. Phakathi nomcimbi wokwethulwa kwe-xAI, u-Musk wabonisa esebenzisa i-Grok3 ukuze ahlaziye amakilasi abalingiswa kanye nemiphumela evela kumdlalo we-Path of Exile 2, athi uwudlala njalo, kodwa eziningi zezimpendulo ezinikezwe i-Grok3 zazingalungile. U-Musk ngesikhathi sokusakaza bukhoma akazange ayibone le nkinga esobala.
Leli phutha aligcinanga nje ngokunikeza ubufakazi obengeziwe kubasebenzisi baphesheya kwezilwandle ukuthi bagcone u-Musk "ngokuthola umuntu obambele" emdlalweni kodwa futhi laphakamisa ukukhathazeka okukhulu mayelana nokwethembeka kwe-Grok3 ekusebenziseni okungokoqobo. "Kungqondongqondo" enjalo, kungakhathaliseki ukuthi ingakanani ikhono layo langempela, ukwethembeka kwayo ezimweni eziyinkimbinkimbi kakhulu zohlelo lokusebenza, njengemisebenzi yokuhlola i-Mars, kuhlala kukungabaza.
Njengamanje, abahloli abaningi abathole ukufinyelela ku-Grok3 amasonto adlule, nalabo abasanda kuhlola amakhono emodeli amahora ambalwa izolo, bonke bakhomba esiphethweni esivamile: "I-Grok3 inhle, kodwa ayingcono kune-R1 noma i-o1-Pro."
Umbono Obalulekile "wokuphazamisa i-Nvidia"
Ku-PPT eyethulwe ngokusemthethweni phakathi nokukhishwa, i-Grok3 iboniswe “ikude lee” ku-Chatbot Arena, kodwa lawa masu asetshenziswa ngobuhlakani: i-eksisi emile ebhodini labaphambili ibhale kuphela imiphumela ebangeni lamaphuzu angu-1400-1300, okwenza umehluko wangempela ongu-1% emiphumeleni yokuhlolwa ubonakale ubaluleke kakhulu kule phrezentheshini.
Emiphumeleni yamaphuzu yemodeli yangempela, i-Grok3 ingaphambili ngo-1-2% kune-DeepSeek R1 ne-GPT-4.0, ehambisana nolwazi lwabasebenzisi abaningi ekuhloleni okungokoqobo okuthole "awukho umehluko obonakalayo." I-Grok3 idlula kuphela abalandeli bayo ngo-1% -2%.
Nakuba i-Grok3 ithole amaphuzu aphezulu kunawo wonke amamodeli ahlolwe esidlangalaleni okwamanje, abaningi abakuthathi ngokungathi sína lokhu: ngemva kwakho konke, i-xAI ike yagxekwa "ngokukhohlisa amaphuzu" esikhathini se-Grok2. Njengoba ibhodi labaphambili lijezisa isitayela sobude bempendulo, amaphuzu ehla kakhulu, okuholele abangaphakathi embonini ukuthi bavamise ukugxeka isenzeko "sokuthola amaphuzu aphezulu kodwa ikhono eliphansi."
Kungakhathaliseki ukuthi "ngokukhohlisa" kwebhodi labaphambili noma amaqhinga okuklama emifanekisweni, aveza ukuthatheka kuka-xAI no-Musk ngomqondo "wokuhola iphekhi" ngamakhono emodeli. U-Musk ukhokhe inani eliphakeme kakhulu lawa mamajini: ngesikhathi sokwethulwa, waqhosha ngokusebenzisa ama-GPU angu-200,000 H100 (efuna "ngaphezu kwe-100,000" ngesikhathi sokusakaza bukhoma) futhi wazuza isikhathi esiphelele sokuqeqeshwa amahora ayizigidi ezingu-200. Lokhu kuholele ekutheni abanye bakholelwe ukuthi imele enye inzuzo ebalulekile embonini ye-GPU futhi babheke umthelela we-DeepSeek emkhakheni "njengobuwula." Ngokuphawulekayo, abanye bakholelwa ukuthi amandla okuhlanganisa azoba ikusasa lokuqeqeshwa okuyimodeli.
Nokho, abanye abasebenzisi benethiwekhi baqhathanise ukusetshenziswa kwama-GPU angu-2000 H800 phakathi nezinyanga ezimbili ukuze kukhiqizwe i-DeepSeek V3, bebala ukuthi ukusetshenziswa kwangempela kwamandla okuqeqeshwa kwe-Grok3 kuphindwe izikhathi ezingu-263 kune-V3. Igebe phakathi kwe-DeepSeek V3, ethole amaphuzu angu-1402, kanti i-Grok3 ingaphansi kwamaphuzu ayi-100. Ngemva kokukhishwa kwale datha, abaningi baqaphela ngokushesha ukuthi ngemva kwesihloko se-Grok3 njengokuthi "iqine kakhulu emhlabeni" kukhona umphumela ocacile wensiza—umqondo wamamodeli amakhulu akhiqiza ukusebenza okuqinile usuqalile ukukhombisa imbuyiselo enciphayo.
Noma "inamaphuzu aphezulu kodwa ikhono eliphansi," i-Grok2 yayinenani elikhulu ledatha yenkampani yokuqala yekhwalithi ephezulu evela ku-X (Twitter) yesikhulumi ukusekela ukusetshenziswa. Kodwa-ke, ekuqeqeshweni kwe-Grok3, i-xAI ihlangabezane ngokwemvelo "nophahla" i-OpenAI ebhekene nayo njengamanje—ukuntuleka kwedatha yokuqeqeshwa kwe-premium kuveza ngokushesha ukusetshenziswa okuseceleni kwamakhono emodeli.
Abathuthukisi be-Grok3 ne-Musk kungenzeka bangabokuqala ukuqonda nokuhlonza la maqiniso ngokujulile, yingakho u-Musk elokhu esho njalo ezinkundleni zokuxhumana ukuthi abasebenzisi benguqulo ababhekene nayo manje "kuseyi-beta nje" nokuthi "inguqulo ephelele izokhishwa ezinyangeni ezizayo." U-Musk uthathe indima yomphathi womkhiqizo we-Grok3, ephakamisa ukuthi abasebenzisi banikeze impendulo ngezinkinga ezihlukahlukene okuhlangatshezwane nazo esigabeni sokuphawula. Angase abe umphathi womkhiqizo olandelwa kakhulu Emhlabeni.
Nokho, phakathi nosuku, ukusebenza kwe-Grok3 ngokungangabazeki kwaphakamisa ama-alamu kulabo abanethemba lokuthembela “esicubungwini esikhulu sokuhlanganisa” ukuze baqeqeshe amamodeli amakhulu anamandla: ngokusekelwe olwazini olutholakala esidlangalaleni lwe-Microsoft, i-OpenAI’s GPT-4 inosayizi wepharamitha wamapharamitha ayizigidi eziyizinkulungwane eziyi-1.8, ngaphezu kokuphindwe kashumi kune-GPT-3. Amahemuhemu aphakamisa ukuthi usayizi wepharamitha we-GPT-4.5 ungaba mkhulu nakakhulu.
Njengoba ubukhulu bepharamitha bemodeli bukhuphuka, izindleko zokuqeqesha nazo ziyakhuphuka. Ngobukhona be-Grok3, abaqhudelanayo abafana ne-GPT-4.5 nabanye abafisa ukuqhubeka “beshisa imali” ukuze bathole ukusebenza okungcono kwemodeli ngosayizi wepharamitha kufanele bacabangele uphahla oselubonakala ngokucacile futhi bacabange ukuthi bangalunqoba kanjani. Ngalesi sikhathi, u-Ilya Sutskever, owayengusosayensi oyinhloko e-OpenAI, wayeke wathi ngoDisemba odlule, "Ukuqeqeshwa kwangaphambili esikujwayele kuzofika ekupheleni," okuye kwavela ezingxoxweni, okwenza imizamo yokuthola indlela yangempela yokuqeqesha amamodeli amakhulu.
Umbono ka-Ilya uzwakalise umsindo embonini. Ubone kusengaphambili ngokunembile ukukhathala okuseduze kwedatha entsha efinyelelekayo, okuholela esimweni lapho ukusebenza kungeke kuqhubeke kuthuthukiswa ngokutholwa kwedatha, okufanisa nokuphelelwa amandla kwezimbiwa phansi. Ubonise ukuthi "njengowoyela, okuqukethwe okukhiqizwa umuntu ku-inthanethi kuwumthombo olinganiselwe." Ezibikezelweni zikaSutskever, isizukulwane esilandelayo samamodeli, ngemva kokuqeqeshwa kwangaphambili, sizoba "nokuzimela kweqiniso" namandla okucabanga "afana nengqondo yomuntu."
Ngokungafani namamodeli anamuhla aqeqeshwe kusengaphambili ancike ngokuyinhloko ekufanisweni kokuqukethwe (okusekelwe kokuqukethwe kwemodeli okufundiwe ngaphambilini), amasistimu e-AI esikhathi esizayo azokwazi ukufunda nokusungula izindlela zokuxazulula izinkinga ngendlela efana "nokucabanga" kobuchopho bomuntu. Umuntu angakwazi ukuzuza ubungcweti obuyisisekelo esifundweni ngemibhalo yobungcweti eyisisekelo nje, kuyilapho imodeli enkulu ye-AI idinga izigidi zamaphoyinti edatha ukuze izuze ukusebenza ngempumelelo kwezinga lokungena. Ngisho nalapho amagama ashintshwa kancane, le mibuzo eyisisekelo ingase ingaqondwa kahle, okubonisa ukuthi imodeli ayizange ithuthuke ngempela ekuhlakanipheni: imibuzo eyisisekelo kodwa engaxazululeki okukhulunywe ngayo ekuqaleni kwesihloko imelela isibonelo esicacile salesi senzakalo.
Isiphetho
Kodwa-ke, ngaphezu kwamandla anonya, uma i-Grok3 iphumelela ngempela ukudalula embonini ukuthi "amamodeli aqeqeshwe ngaphambilini asondela ekupheleni kwawo," kungaba nomthelela omkhulu kulo mkhakha.
Mhlawumbe ngemva kokudamba kancane kancane kwesiyaluyalu esizungeze i-Grok3, sizobona izimo eziningi njengesibonelo sika-Fei-Fei Li "sokulungisa amamodeli asebenza kahle kakhulu kudathasethi ethile ngama-$50 kuphela," ekugcineni sithole indlela yeqiniso eya ku-AGI.
Control Cables
Uhlelo Lwekhebula Oluhlelekile
Inethiwekhi&Idatha, Ikhebula Le-Fiber-Optic, Intambo Yesiphesheli, Amamojula, I-Faceplate
Apr.16th-18th, 2024 Middle-East-Energy e-Dubai
Ephreli 16-18, 2024 Securika e-Moscow
May.9th, 2024 UMCIMBI EMISHA NOBUCHWEPHESHE WETHULA UMCIMBI e-Shanghai
Oct.22nd-25th, 2024 SECURITY CHINA e-Beijing
Nov.19-20, 2024 CONNECTED WORLD KSA
Isikhathi sokuthumela: Feb-19-2025