Gabatarwa

Kuna tsammanin Grok3 zai zama "makon ƙarshe" na samfuran da aka riga aka horar?

Elon Musk da ƙungiyar xAI a hukumance sun ƙaddamar da sabon sigar Grok, Grok3, yayin wani raye-raye. Kafin wannan taron, mahimman bayanai masu alaƙa, haɗe tare da tallan talla na 24/7 na Musk, ya ɗaga tsammanin duniya don Grok3 zuwa matakan da ba a taɓa gani ba. Mako guda da suka gabata, Musk ya bayyana da gaba gaɗi yayin raye-raye yayin da yake yin tsokaci kan DeepSeek R1, "xAI na gab da ƙaddamar da mafi kyawun samfurin AI." Daga bayanan da aka gabatar kai tsaye, Grok3 ya bayar da rahoton cewa ya zarce duk samfuran yau da kullun na yau da kullun a cikin ma'auni na lissafi, kimiyya, da shirye-shirye, tare da Musk har ma da iƙirarin cewa za a yi amfani da Grok3 don ayyukan ƙididdigewa da suka shafi ayyukan SpaceX na Mars, yana hasashen "nasarar a matakin lambar yabo ta Nobel a cikin shekaru uku." Koyaya, waɗannan a halin yanzu ikirari ne kawai Musk. Bayan ƙaddamarwa, na gwada sabon nau'in beta na Grok3 kuma na gabatar da tambayar dabarar gargajiya don manyan samfura: "Wane ne ya fi girma, 9.11 ko 9.9?" Abin takaici, ba tare da wasu masu cancanta ko alamomi ba, abin da ake kira smartest Grok3 har yanzu ya kasa amsa wannan tambayar daidai. Grok3 ya kasa gano daidai ma'anar tambayar.

Wannan gwajin ya jawo hankali da sauri daga abokai da yawa, kuma kwatsam, gwaje-gwaje iri-iri iri-iri a ƙasashen waje sun nuna Grok3 yana fama da ainihin tambayoyin kimiyyar lissafi/mathematics kamar "Wace ball ce ta fara fadowa daga Hasumiyar Leaning na Pisa?" Don haka, an yi masa lakabi da raha a matsayin "mai hazaka wanda ba ya son amsa tambayoyi masu sauki."

Grok3 yana da kyau, amma bai fi R1 ko o1-Pro ba.

Grok3 ya sami "rashin nasara" akan gwaje-gwajen ilimin gama gari da yawa a aikace. A yayin taron ƙaddamar da xAI, Musk ya nuna ta yin amfani da Grok3 don nazarin azuzuwan halayen da tasiri daga hanyar wasan Exile 2, wanda ya yi iƙirarin yin wasa sau da yawa, amma yawancin amsoshin da Grok3 ya bayar ba daidai ba ne. Musk a lokacin raye-raye bai lura da wannan fitowar fili ba.

Wannan kuskuren ba wai kawai ya ba da ƙarin shaida ba ga masu amfani da yanar gizo na ketare don yin ba'a ga Musk don "neman madadin" a cikin wasa amma kuma ya haifar da damuwa mai mahimmanci game da amincin Grok3 a aikace-aikace masu amfani. Don irin wannan “hazaka,” ba tare da la’akari da ainihin iyawarsa ba, amincinsa a cikin yanayin aikace-aikace masu sarƙaƙƙiya, kamar ayyukan binciken Mars, ya kasance cikin shakka.

A halin yanzu, yawancin masu gwadawa waɗanda suka sami damar zuwa Grok3 makonni da suka wuce, da waɗanda kawai suka gwada ƙarfin samfurin don 'yan sa'o'i a jiya, duk suna nuna ƙarshen gamawa: "Grok3 yana da kyau, amma bai fi R1 ko o1-Pro ba."

Mahimman ra'ayi akan "Rikicin Nvidia"

A cikin PPT da aka gabatar a hukumance yayin sakin, an nuna Grok3 ya kasance "na gaba" a cikin Chatbot Arena, amma wannan dabarar da aka yi amfani da ita ta hanyar fasaha: axis a tsaye akan allon jagora kawai an jera sakamakon a cikin kewayon maki 1400-1300, wanda ya sa ainihin 1% bambanci a sakamakon gwajin ya bayyana na musamman a cikin wannan gabatarwar.

A cikin ainihin sakamakon ƙima na ƙirar ƙira, Grok3 shine kawai 1-2% gaba da DeepSeek R1 da GPT-4.0, wanda yayi daidai da ƙwarewar masu amfani da yawa a cikin gwaje-gwaje masu amfani waɗanda suka sami "babu wani bambanci mai ban mamaki." Grok3 kawai ya wuce magadansa da 1% -2%.

Kodayake Grok3 ya zira kwallaye sama da duk samfuran da aka gwada a yanzu a bainar jama'a, da yawa ba sa ɗaukar wannan da mahimmanci: bayan haka, xAI a baya an soki shi don "maganin maki" a zamanin Grok2. Kamar yadda jagororin ke azabtar da salon tsayin amsa, makin ya ragu sosai, yana jagorantar masana'antun masana'antu sau da yawa suna sukar lamarin "babban maki amma rashin iyawa."

Ko ta hanyar "manipulation" na jagora ko ƙirƙira dabaru a cikin misalai, suna bayyana xAI da Musk sha'awar ra'ayin "jagoranci fakitin" a cikin iyawar ƙira. Musk ya biya farashi mai tsada don waɗannan iyakokin: yayin ƙaddamarwa, ya yi alfahari da yin amfani da 200,000 H100 GPUs (yana da'awar "sama da 100,000" a lokacin raye-raye) da samun jimlar lokacin horo na sa'o'i miliyan 200. Wannan ya sa wasu suyi imani cewa yana wakiltar wani muhimmin al'amari ga masana'antar GPU kuma suyi la'akari da tasirin DeepSeek akan sashin a matsayin "wauta." Musamman ma, wasu sun yi imanin cewa ƙarfin lissafin ƙididdiga zai zama makomar horon samfurin.

Koyaya, wasu netizens idan aka kwatanta amfani da 2000 H800 GPUs sama da watanni biyu don samar da DeepSeek V3, ƙididdige cewa ainihin ƙarfin horo na Grok3 shine sau 263 na V3. Rata tsakanin DeepSeek V3, wanda ya sami maki 1402, kuma Grok3 yana ƙarƙashin maki 100. Bayan fitowar wannan bayanan, mutane da yawa da sauri sun gane cewa a bayan taken Grok3 a matsayin "mafi ƙarfi a duniya" yana da fa'ida mai fa'ida mai amfani - dabarar manyan samfuran da ke haifar da aiki mai ƙarfi ya fara nuna raguwar dawowa.

Ko da tare da "babban maki amma ƙananan iyawa," Grok2 yana da adadi mai yawa na ingantattun bayanai na ɓangare na farko daga dandalin X (Twitter) don tallafawa amfani. Koyaya, a cikin horarwar Grok3, xAI a zahiri ta ci karo da "rufin" da OpenAI ke fuskanta a halin yanzu-rashin bayanan horarwa cikin sauri yana fallasa ƙarancin amfanin samfurin.

Masu haɓaka Grok3 da Musk wataƙila su ne farkon waɗanda suka fara fahimta da gano waɗannan gaskiyar sosai, wanda shine dalilin da ya sa Musk ya ci gaba da ambata a kan kafofin watsa labarun cewa masu amfani da sigar yanzu shine "har yanzu kawai beta" kuma "za a fitar da cikakken sigar a cikin watanni masu zuwa." Musk ya ɗauki nauyin mai sarrafa samfur na Grok3, yana ba da shawarar masu amfani su ba da ra'ayi game da batutuwa daban-daban da aka fuskanta a cikin sashin sharhi. Maiyuwa ne shi ne mai sarrafa samfur da aka fi bi a Duniya.

Amma duk da haka, a cikin rana guda, aikin Grok3 babu shakka ya ɗaga ƙararrawa ga waɗanda ke fatan dogaro da "babban tsokar lissafi" don horar da manyan ƙira: dangane da bayanan Microsoft da ake samu a bainar jama'a, GPT-4 na OpenAI yana da girman sigina na sigogin tiriliyan 1.8, sama da sau goma na GPT-3. Jita-jita sun nuna cewa girman siga na GPT-4.5 na iya zama ma fi girma.

Yayin da girman siginar ƙirar ke ƙaruwa, farashin horon kuma yana ƙaruwa. Tare da kasancewar Grok3, masu fafatawa kamar GPT-4.5 da sauran waɗanda ke son ci gaba da “ƙona kuɗi” don cimma kyakkyawan aikin ƙirar ƙira ta hanyar girman sigina dole ne suyi la’akari da rufin da yake gani a sarari kuma suyi la’akari da yadda za a shawo kan shi. A wannan lokacin, Ilya Sutskever, tsohon babban masanin kimiyya a OpenAI, ya bayyana a baya a watan Disambar da ya gabata, "Tsarin horon da muka saba da shi zai zo ƙarshe," wanda ya sake farfadowa a cikin tattaunawa, wanda ya haifar da ƙoƙari na gano hanyar gaskiya don horar da manyan samfurori.

Ra'ayin Ilya ya haifar da ƙararrawa a cikin masana'antar. Ya tsinkayi daidai ga gajiyawar sabbin bayanai da za a iya samu, wanda ya kai ga yanayin da ba za a iya ci gaba da inganta aikin ta hanyar sayan bayanai ba, yana kamanta shi da gajiyar mai. Ya yi nuni da cewa "kamar man fetur, abubuwan da mutane ke samarwa a yanar gizo suna da iyakacin albarkatu." A cikin tsinkayar Sutskever, ƙarni na gaba na ƙira, bayan horo, za su mallaki "yancin kai na gaskiya" da ikon tunani "mai kama da kwakwalwar ɗan adam."

Ba kamar samfuran da aka riga aka horar da su na yau waɗanda ke dogara da madaidaicin abun ciki (dangane da abin da aka koya a baya), tsarin AI na gaba zai iya koyo da kafa hanyoyin magance matsaloli ta hanyar da ta dace da “tunanin” kwakwalwar ɗan adam. Mutum na iya samun ƙwarewa ta asali a cikin wani batu tare da adabi na ƙwararru kawai, yayin da babban samfurin AI yana buƙatar miliyoyin maki don cimma ainihin ingantaccen matakin shigarwa. Ko da lokacin da aka ɗan canza kalmomin, waɗannan tambayoyi masu mahimmanci ba za a iya fahimtar su daidai ba, suna nuna cewa samfurin bai inganta da gaske ba a hankali: ainihin tambayoyin da ba a warware su ba da aka ambata a farkon labarin suna wakiltar misali bayyananne na wannan sabon abu.

Kammalawa

Koyaya, fiye da ƙarfin ƙarfi, idan da gaske Grok3 yayi nasarar bayyanawa masana'antar cewa "samfuran da aka riga aka horar suna gabatowa ƙarshensu," zai haifar da tasiri mai mahimmanci ga filin.

Wataƙila bayan tashin hankalin da ke kewaye da Grok3 ya ragu a hankali, za mu shaidi ƙarin shari'o'i kamar misalin Fei-Fei Li na "daidaita ƙirar ƙira akan takamaiman bayanan dala kawai $ 50," a ƙarshe gano hanyar gaskiya zuwa AGI.