Ye-BMS, BUS, Industrial, Instrumentation Cable.
U-Elon Musk kunye neqela le-xAI baphehlele ngokusemthethweni inguqulelo yamva nje yeGrok, Grok3, ngexesha lokudlala. Ngaphambi kwesi siganeko, isixa esibalulekileyo solwazi olunxulumeneyo, kunye ne-Musk's 24/7 yokuthengisa i-hype, yaphakamisa ulindelo lwehlabathi jikelele kwi-Grok3 ukuya kumazinga angazange abonwe ngaphambili. Kwiveki nje edlulileyo, uMusk wathetha ngokuzithemba ngelixa wayethetha nge-DeepSeek R1, "i-xAI iza kumisela imodeli engcono ye-AI." Ukususela kwidatha evezwe ngokuphilayo, i-Grok3 ixelwe ukuba igqithise yonke imodeli ekhoyo ngoku kwiibenchmarks zemathematika, isayensi, kunye neprogram, kunye no-Musk nokubanga ukuba i-Grok3 iya kusetyenziselwa imisebenzi yekhompyutha enxulumene nemishini ye-SpaceX ye-Mars, iqikelela "impumelelo kwinqanaba leNobel Prize kwisithuba seminyaka emithathu." Nangona kunjalo, ezi ziingcinga nje zikaMusk. Emva kokusungulwa, ndiye ndavavanya inguqulelo ye-beta yamva nje yeGrok3 kwaye ndabuza umbuzo wobuqhinga kwiimodeli ezinkulu: "Yeyiphi enkulu, 9.11 okanye 9.9?" Ngelishwa, ngaphandle kweemfaneleko okanye amanqaku, iGrok3 ebizwa ngokuba ngoyena ukrelekrele akakwazanga ukuphendula lo mbuzo ngokuchanekileyo. UGrok3 usilele ukuchonga ngokuchanekileyo intsingiselo yombuzo.
Olu vavanyo lwakhawuleza lwatsalela ingqalelo enkulu kubahlobo abaninzi, kwaye ngempazamo, iimvavanyo ezahlukeneyo ezifanayo phesheya kolwandle zibonise iGrok3 esokola nemibuzo esisiseko yefiziksi/yemathematika efana nethi "Yeyiphi ibhola ewela kuqala kwiNqaba yePisa eLeaning?" Yiyo ke loo nto ibhalwe ngokuhlekisayo ngokuthi "yingqondi engafuni ukuphendula imibuzo elula."
I-Grok3 ilungile, kodwa ayikho ngcono kune-R1 okanye i-o1-Pro.
I-Grok3 ifumene "iintsilelo" kwiimvavanyo ezininzi zolwazi oluqhelekileyo ekusebenzeni. Ngethuba lomcimbi wokuqaliswa kwe-xAI, u-Musk wabonisa ukusebenzisa i-Grok3 ukuhlalutya iiklasi zomlinganiswa kunye nemiphumo evela kumdlalo weNdlela ye-Exile 2, owathi wayidlala rhoqo, kodwa uninzi lweempendulo ezibonelelwe nguGrok3 zazingalunganga. UMusk ngexesha lokuhamba ngokuphila akazange aqaphele lo mbandela ocacileyo.
Le mpazamo ayizange ibonelele kuphela ubungqina obungakumbi kubasebenzisi be-intanethi baphesheya ukuba bahlekise uMusk "ngokufumana indawo ebambeleyo" kumdlalo kodwa iphinde yaphakamisa inkxalabo ebalulekileyo malunga nokuthembeka kweGrok3 kwizicelo ezisebenzayo. "Kwingqondo" enjalo, nokuba bubuphi na ubunakho bayo, ukuthembeka kwayo kwiimeko ezintsonkothileyo zosetyenziso, njengemisebenzi yokuhlola iMars, kuhlala kungamathandabuzo.
Okwangoku, abaninzi abavavanyi abafumene ukufikelela kwiiveki ze-Grok3 ezidlulileyo, kunye nabo bavavanya nje izakhono zemodeli kwiiyure ezimbalwa izolo, bonke babhekisela kwisigqibo esifanayo: "I-Grok3 ilungile, kodwa ayingcono kune-R1 okanye i-o1-Pro."
Umbono obalulekileyo malunga "nokuphazamisa iNvidia"
Kwi-PPT enikezelwe ngokusemthethweni ngexesha lokukhululwa, i-Grok3 ibonakaliswe "ikude phambili" kwi-Chatbot Arena, kodwa le ndlela yobugcisa esetyenziswe ngobuchule: i-axis ethe nkqo kwibhodi yabaphambili idweliswe kuphela iziphumo kwi-1400-1300 yamanqaku, okwenza i-1% yangaphambili umehluko kwiziphumo zovavanyo zibonakala zibaluleke kakhulu kule nkcazo.
Kwiziphumo zamanqaku omzekelo wangempela, i-Grok3 i-1-2% nje ngaphambi kwe-DeepSeek R1 kunye ne-GPT-4.0, ehambelana namava amaninzi abasebenzisi kwiimvavanyo eziphathekayo ezifumene "akukho mahluko obonakalayo." I-Grok3 idlula kuphela abalandela bayo nge-1% -2%.
Nangona iGrok3 ifumene amanqaku aphezulu kunazo zonke iimodeli ezivavanyiweyo ngoku esidlangalaleni, uninzi aluyithathi ngqalelo le nto: emva koko, i-xAI ibikhe yagxekwa ngaphambili "ngokukhohlisa amanqaku" kwixesha leGrok2. Njengoko ibhodi yabaphambili isohlwaya isimbo sobude bempendulo, amanqaku ehle kakhulu, ekhokelela abantu bangaphakathi kwishishini ukuba basoloko begxeka into "yokufumana amanqaku aphezulu kodwa amandla aphantsi."
Nokuba ngebhodi yabaphambili "yokukhohlisa" okanye amaqhinga oyilo kwimifanekiso, ityhila i-xAI kunye ne-Musk's obsession ngembono "yokukhokela ipakethi" kubuchule bemodeli. U-Musk uhlawule ixabiso eliphakamileyo kule mida: ngexesha lokuqaliswa, waqhayisa ngokusebenzisa i-200,000 H100 GPUs (ibanga "ngaphezu kwe-100,000" ngexesha lokuhamba ngokuphila) kunye nokufezekisa ixesha loqeqesho elipheleleyo leeyure ze-200 yezigidi. Oku kuye kwakhokelela ekubeni abanye bakholelwe ukuba imele enye inzuzo ebalulekileyo kwishishini le-GPU kwaye bathathele ingqalelo impembelelo ye-DeepSeek kwicandelo "njengobudenge." Ngokucacileyo, abanye bakholelwa ukuba amandla apheleleyo okubala aya kuba likamva loqeqesho lwemodeli.
Nangona kunjalo, abanye abasebenzisi be-intanethi bathelekisa ukusetyenziswa kwe-2000 H800 GPUs kwisithuba seenyanga ezimbini ukuvelisa i-DeepSeek V3, bebala ukuba ukusetyenziswa kwamandla okwenene kweGrok3 kukuphindaphinda ngama-263 e-V3. Isithuba phakathi kwe-DeepSeek V3, efumene amanqaku e-1402, kunye ne-Grok3 ingaphantsi kwamanqaku angama-100. Ukulandela ukukhutshwa kwale datha, abaninzi baqonda ngokukhawuleza ukuba emva kwesihloko seGrok3 njengeyona "eyona yomeleleyo yehlabathi" ilele isiphumo esicacileyo esiluncedo-ingqiqo yeemodeli ezinkulu ezivelisa ukusebenza okunamandla sele iqalile ukubonisa imbuyekezo ehlayo.
Nokuba "namanqaku aphezulu kodwa amandla aphantsi," i-Grok2 yayinexabiso elikhulu ledatha yeqela lokuqala elikumgangatho ophezulu ukusuka kwiqonga le-X (Twitter) ukuxhasa ukusetyenziswa. Nangona kunjalo, kuqeqesho lwe-Grok3, i-xAI ngokwemvelo idibene "nesilingi" i-OpenAI ejongene nayo ngoku-ukunqongophala kwedatha yoqeqesho lweprimiyamu ngokukhawuleza kubhentsisa ukusetyenziswa komda kwesakhono semodeli.
Abaphuhlisi beGrok3 kunye neMusk kusenokwenzeka ukuba ngabokuqala ukuziqonda kwaye bachonge ezi nyaniso ngokunzulu, yiyo loo nto uMusk ethe gqolo ukukhankanya kumajelo eendaba ezentlalo abajongana nazo abasebenzisi benguqulelo ngoku “iseyi-beta” kwaye “inguqulelo epheleleyo iya kukhutshwa kwiinyanga ezizayo.” U-Musk uthathe indima yomphathi wemveliso ye-Grok3, ecebisa abasebenzisi ukuba banike impendulo kwimibandela eyahlukeneyo edibene nayo kwicandelo lamagqabaza. Usenokuba ngoyena mphathi wemveliso ulandelwa kakhulu emhlabeni.
Nangona kunjalo, ngaphakathi kosuku, ukusebenza kweGrok3 ngokungathandabuzekiyo kwaphakamisa i-alamu kwabo banethemba lokuthembela "kwi-massive computational muscle" ukuqeqesha iimodeli ezinkulu ezomeleleyo: ngokusekelwe kulwazi olufumaneka esidlangalaleni lweMicrosoft, i-OpenAI's GPT-4 inobungakanani beparameter ye-1.8 yeetriliyoni zeeparamitha, ngaphezu kwezihlandlo ezilishumi ze-GPT-3. Amahemuhemu acebisa ukuba ubungakanani beparameter ye-GPT-4.5 inokuba nkulu ngakumbi.
Njengoko ubungakanani beparameter yemodeli bunyuka, iindleko zoqeqesho nazo ziyenyuka. Ngobukho beGrok3, abantu abakhuphisanayo abanje nge-GPT-4.5 kunye nabanye abanqwenela ukuqhubeka "betshisa imali" ukuze bafezekise ukusebenza kwemodeli engcono ngobungakanani beparamitha kufuneka bathathele ingqalelo isilingi esele ibonakala ngokucacileyo kwaye bacinge ngendlela yokuyoyisa. Ngeli xesha, u-Ilya Sutskever, owayengusosayensi oyintloko kwi-OpenAI, wayetshilo ngaphambili ngoDisemba ophelileyo, "Uqeqesho lwangaphambili esiluqhelileyo luya kufikelela esiphelweni," oluye lwaphinda lwavela kwiingxoxo, lukhuthaza iinzame zokufumana indlela yokwenyani yokuqeqesha iimodeli ezinkulu.
Umbono ka-Ilya uvakalise i-alamu kwishishini. Wabona ngokuchanekileyo ukutshatyalaliswa okukhawulezayo kwedatha entsha efikelelekayo, ekhokelela kwimeko apho ukusebenza kungenakuqhubeka kuphuculwe ngokufunyanwa kwedatha, kufanise nokuphela kwamafutha e-fossil. Ubonise ukuba "njengeoli, umxholo owenziwe ngumntu kwi-intanethi ngumthombo olinganiselweyo." Kuqikelelo lukaSutskever, isizukulwana esilandelayo seemodeli, emva koqeqesho lwangaphambili, siya kuba “nenkululeko yokwenyani” kunye namandla okuqiqa “afana nengqondo yomntu.”
Ngokungafaniyo neemodeli zanamhlanje eziqeqeshwe kwangaphambili ezixhomekeke ikakhulu kumxholo wokuthelekisa (ngokusekwe kumxholo ofundiweyo wemodeli), iinkqubo ze-AI zexesha elizayo ziya kukwazi ukufunda nokuseka iindlela zokusombulula iingxaki ngendlela efana "nokucinga" kwengqondo yomntu. Umntu unokufikelela kwisakhono esisisiseko kwisifundo esinoncwadi olusisiseko nje, ngelixa imodeli enkulu ye-AI ifuna izigidi zamanqaku edatha ukufezekisa eyona ndlela isisiseko yokungena. Kwanaxa igama litshintshwa kancinane, le mibuzo isisiseko isenokungaqondwa ngokuchanekileyo, nto leyo ebonisa ukuba imodeli ayiphucukanga ngokunyanisekileyo kubukrelekrele: imibuzo esisiseko kodwa engaphendulekiyo ekhankanywe ekuqaleni kwenqaku imela umzekelo ocacileyo wale nto.
Ukuqukumbela
Nangona kunjalo, ngaphaya kwamandla akhohlakeleyo, ukuba i-Grok3 ngokwenene iyaphumelela ekuboniseni ishishini ukuba "iimodeli eziqeqeshwe kwangaphambili zisondela esiphelweni sazo," iya kuba nempembelelo ebalulekileyo kwintsimi.
Mhlawumbi emva kokuba isiphithiphithi esijikeleze iGrok3 sithothe kancinci, siya kungqina iimeko ezininzi ezinje ngomzekelo kaFei-Fei Li "wokulungisa iimodeli zokusebenza okuphezulu kwidathasethi ethile nge-50 yeedola," ekugqibeleni sifumene indlela yokwenyani eya kwi-AGI.
Iintambo zokulawula
Inkqubo yeCabling eyakhiweyo
Inethiwekhi kunye neDatha, iFiber-Optic Cable, iPatch Cord, iiModyuli, iFaceplate
Epreli.16th-18th, 2024 Middle-East-Energy eDubai
Epreli.16-18th, 2024 Securika eMoscow
May.9th, 2024 IIMVELISO EZINTSHA KUNYE NETEKNOLOJI ZIQALISA UMNYATHO eShanghai
Oct.22nd-25th, 2024 UKHUSELEKO CHINA e Beijing
Nov.19-20, 2024 IHLABATHI ELIQHELEKILEYO KSA
Ixesha lokuposa: Feb-19-2025