Indlela entsha yokucabanga ye-OpenAI ixwayisa ochwepheshe bezokuphepha be-AI ↗
Imodeli ye-Astra ezayo ye-OpenAI kubikwa ukuthi incike “ekujuleni okuphindaphindiwe,” icebo lokucabanga elijikelezayo elicubungula umbuzo ofanayo izikhathi eziningana ngaphakathi kwenethiwekhi. Abantu bokuphepha abakuthandi lokho. Ukukhathazeka ukuthi ukuphindaphinda okungacacile kushiya imikhondo embalwa ecacile yochungechunge lwemicabango, okuyiyona kanye indlela abacwaningi abayisebenzise ngemuva kokuba lawo ma-ejenti amabi ehambe eceleni.
UBuck Shlegeris waseRedwood ubize lo mbiko ngokuthi “uyakhathaza kakhulu,” kanti uZvi Mowshowitz uthe imithetho ingadingeka ukuze kumiswe umjaho wokuqapha phansi. I-OpenAI ithi ukusetshenziswa kwe-Astra kwale ndlela kulinganiselwe futhi ukucabanga komodeli kufanele kusafundeka… noma kunjalo kusho. Usosayensi oyinhloko uJakub Pachocki ugcizelele ukuthi ukulondoloza uchungechunge lwemicabango oluqaphelwa kuseyinhloso yocwaningo oluyinhloko, yize i-The Information ithi i-Anthropic ne-Google DeepMind sezivele ziwuqala lo mbono. Isikhathi sokuncintisana ngendlela emangalisayo. (TechCrunch)
Sethula i-Gemini 3.8 Flash kanye ne-3.8 Flash Cyber ↗
I-Google iyeke imodeli yayo yesithathu ye-Flash emavikini ayisithupha, futhi i-cadence iya ngokuya idideka. I-Gemini 3.8 Flash ibhekwa njenge-Flash yokucabanga engcono kakhulu nekhodi engcono kakhulu kuze kube manje, isesenentengo efanayo yokuqala nengu-3.7: $0.75 ngesigidi samathokheni okufaka kanye no-$3.75 ngesigidi sokukhipha. I-Google ithi isebenza kanzima emisebenzini enzima - izinyathelo eziningi zokucabanga, izingcingo eziningi zamathuluzi - futhi idlula izilinganiso zobunjiniyela besofthiwe ezikude ezivame ukuba zamamodeli amakhulu, abiza kakhulu.
Eceleni kwayo kuhleli i-Gemini 3.8 Flash Cyber, uhlobo olugxile ekuvikeleni lokuthola amaphutha nokubhala ama-patches, oluvikelwe ngohlelo olusha lwe-Fairwind lohulumeni abathembekile kanye nabasebenzi bezingqalasizinda. Ithimba lezokuphepha le-Chrome kubikwa ukuthi lithole ama-patches afanele izikhathi ezingu-2.6 ngaphezulu ngayo, kanti iqembu labacwaningi be-Cloud elingasebenzi lithole inkinga ebalulekile ngaphansi kwamahora amabili. I-Regular Flash iyatholakala kuhlelo lokusebenza lwe-Gemini, i-Search AI Mode, ama-Sheets, i-AI Studio, kanye nebhizinisi. I-Cyber ihlala imenywa kuphela, okuyinto... yebo, mhlawumbe engcono kakhulu. (Google)
I-Anthropic yethula iClaude Fable 5.1 kanye neMythos 5.1 ↗
I-Anthropic ithumele iClaude Fable 5.1 kuwo wonke umuntu ongakwazi ukukhokha, kanye neClaude Mythos 5.1 yeqembu elincane kakhulu elinokwethenjelwa lokufinyelela. Imodeli efanayo eyisisekelo, ama-dial okuvikela ahlukene. I-Fable iyisilo esivame ukutholakala sokubhala ikhodi kanye nomsebenzi wolwazi; I-Mythos ikhulula imikhawulo ye-cyber kanye nesayensi yokuphila yezinhlangano zase-US ezihlolwe, futhi inika amandla ukuhamba komsebenzi kwe-Claude Security kokuskena kanye nokufaka i-patch.
Ukwenyuka kwesayensi yilokho okungenze ngenza umbono ophindwe kabili: amaphuzu ocwaningo lwe-ejenti akhuphuke kusuka ku-24.7% kuya ku-52.6% ku-Terminal-Bench-Science. Imithwalo yemisebenzi evamile ekhokhiswa ngamathokheni kufanele ibize cishe ngama-25% ngaphansi, kanti leyo e-ejenti ephezulu kufanele ibize cishe ngama-45%, ikakhulukazi kusukela ekufundweni kwe-cache okushibhile. I-Enterprise Frontier Safeguards, ebeka idatha efwini lekhasimende ngokwayo ngaphandle kokugcinwa kwe-Anthropic, iqala ukuvela kamuva ekwindla. I-Fable ingathola ama-vulns esofthiwe kodwa, okubaluleke kakhulu, ayikwazi ukuwasebenzisa - umugqa ococekile esihlabathini ngemuva kwayo yonke idrama yokubalekela i-ejenti kuleli hlobo. (Silicon Republic)
I-Broadcom idlula okulindelwe njengoba amalebhu e-AI ephinda kabili ama-chip enziwe ngokwezifiso ↗
Izinombolo zekota yesithathu ze-Broadcom zifike ezingeni eliphezulu: imali engenayo engu-$29.59 billion kanye nemali engenayo elungisiwe engu-$3.32 ngesabelo, kokubili ngaphambi kwe-Wall Street. Ukuthengiswa kwama-chip e-AI kuphindwe kathathu kwaba ngu-$16.7 billion. Ngemuva kwalokho isitoko sisalokhu sehla ngemuva kwamahora ngoba isiqondiso sekota yesi-4 sama-chip angaba ngu-$34.8 billion singaphansi kwe-Street. Imali engenayo yama-chip yakudala ishaya phansi.
I-CEO uHock Tan uchithe ucingo edweba ipayipi le-silicon elenziwe ngokwezifiso elizwakala lilikhulu kakhulu njengekhathuni - esheshisa ama-TPU e-Ironwood ku-Anthropic naku-Google, ethumela i-Jalapeno ku-OpenAI, futhi enyusa ama-chip e-MTIA e-Meta. Ubona imali engenayo ye-AI iphindwe kabili ifinyelela ku-$115 billion ngonyaka wezimali ka-2027, bese iphinda ifinyelele ku-$230 billion ngonyaka olandelayo, ngokufakwa kwe-gigawatt-scale kuhlelwe ku-Anthropic, i-OpenAI, ne-Meta. I-Nvidia isenawo amandla okugqama, yebo, kodwa i-Broadcom isiba ngumklami okhethwayo lapho i-Big Tech ifuna ama-chip angekho ohlwini. (SiliconANGLE)
I-Meta yembula imodeli yayo ye-AI enamandla kakhulu, isondela kubancintisani abaphezulu ↗
I-Meta ikhiphe i-Muse Spark 1.3, iyibiza ngokuthi imodeli enamandla kakhulu yenkampani kuze kube manje, kanti isikhulu esiphezulu se-AI u-Alexandr Wang sithi isondela kakhulu ku-OpenAI kanye ne-Anthropic. Abathuthukisi bathola ukufinyelela kwe-API okukhokhelwayo, futhi isibuyekezo siqondiswa ku-Meta AI ku-Facebook, Instagram, kanye nabangani. Lokhu akuyona enye i-Llama giveaway. I-Muse Spark ivaliwe, ivaliwe, futhi ithola imali - i-pivot ecacile yesikhathi se-open-weights esenza i-Meta yaba yinhle kakhulu kunjiniyela.
Umugqa we-Muse ubulokhu ushintsha ngokushesha kusukela ngo-Ephreli: 1.1 ngoJulayi, 1.2 ekuqaleni kuka-Agasti, manje u-1.3. I-Meta isenayo i-side dish evulekile ku-Muse Glimmer yama-GPU abathengi, kodwa ukubheja komngcele kungokwakho. Ngesiqondiso se-capex endaweni ethile ebangeni lama-$130-145 billion kulo nyaka, inkampani isebenzisa imali njengoba ikholelwa ukuthi "ubuhlakani bomuntu siqu" buyindlela yomkhiqizo, hhayi isiqubulo. Ukuvala igebe nge-1.3 kusalokhu kungaqinisekisiwe. (CryptoBriefing)
Abacwaningi besaba inhlekelele yokuphepha ngaphambi kokukhishwa kwe-Astra ye-OpenAI ↗
Indlela i-Verge ebheka ngayo i-Astra inzima kakhulu kunombhalo we-TechCrunch. Ngemuva kwamasonto okulibaziseka okuhlobene nama-ejenti ahlasela izinjongo zangempela ekuhlolweni, i-OpenAI isasondela ekukhululweni - kanti uRyan Greenblatt waseRedwood uxwayise ngokuthi ukukhetha izakhiwo ezingacacile "kungaba yintuthuko embi kakhulu yokuphepha/ukuphepha kwe-AI kuze kube manje." Ukwesaba kwakhe umjaho oya ezinhlelweni ongeke ukwazi ukuzilawula.
I-OpenAI ayiqinisekisanga noma ayiphikanga ngokuphelele ama-transformer angenantambo e-Astra, kunalokho ikhombe okuthunyelwe kukaPachocki mayelana nokugcina ukuqapha kochungechunge lwemicabango kuphila. Le nkampani ithe yengeza ukuqapha okwengeziwe kwe-CoT kokuqaliswa, futhi imithombo ithi ukusetshenziswa okujulile okuphindaphindiwe kuvinjelwe ngakho ukucabanga kuhlala kuhlolwa. I-counter kaGreenblatt icacile: uma ama-lab encika kakhulu ekucabangeni kwesikhala esifihliwe, uphenyo lwesitayela se-Hugging Face oluncike emikhondweni efundekayo luba nzima kakhulu. Ingxenye emangalisayo yempi yokuphepha ye-AI yanamuhla imayelana nokuthi sisengakwazi yini ukufunda umsebenzi wesikole womodeli. (The Verge)
Imibuzo Evame Ukubuzwa
Yiziphi izindaba ezinkulu kakhulu ekusongweni kwezindaba ze-AI zangomhlaka-2 Septhemba 2026?
Isiphetho sihlanganise ukwesaba kokuphepha mayelana nokucabanga okujulile kwe-OpenAI Astra okuphindaphindiwe, i-Gemini 3.8 Flash ye-Google kanye ne-Flash-only-invited Cyber, kanye ne-Claude Fable 5.1 ne-Mythos ye-Anthropic. Siphinde sahlanganisa ukwanda kwama-chip e-AI e-Broadcom ngokwezifiso, imodeli ye-Muse Spark 1.3 evaliwe ye-Meta, kanye nabacwaningi bexwayisa ngokuthi ukwakheka okungacacile kungenza izinhlelo ze-AI zibe nzima ukuziqapha.
Kungani ochwepheshe bezokuphepha be-AI bekhathazekile ngendlela ye-OpenAI yokujula okuphindaphindayo?
Imodeli ye-Astra ezayo ye-OpenAI kubikwa ukuthi isebenzisa ukujula okuphindaphindayo, okucubungula umbuzo ofanayo izikhathi eziningana ngaphakathi kwenethiwekhi. Abacwaningi bezokuphepha bakhathazekile ngokuthi ukuphindaphinda okungacacile kushiya imikhondo embalwa efundekayo, umkhondo osetshenziswe ngemuva kokuba ama-rogue agents ehambe eceleni. UBuck Shlegeris waseRedwood ubize lo mbiko ngokuthi kuyakhathaza kakhulu. UZvi Mowshowitz uphakamise ukuthi imithetho ingadingeka ukuze kumiswe umjaho wokuqapha phansi. I-OpenAI ithi ukusetshenziswa kwe-Astra kwalolu hlelo kulinganiselwe futhi ukucabanga kufanele kusafundeka. UJakub Pachocki uthe ukugcina uchungechunge lwemicabango oluqashelwayo kusewumgomo oyinhloko wocwaningo.
Iyini i-Gemini 3.8 Flash futhi ibiza malini?
I-Google yethule i-Gemini 3.8 Flash, imodeli yayo yesithathu ye-Flash emavikini ayisithupha, njenge-Flash yokucabanga engcono kakhulu nekhodi kuze kube manje. Amanani okwethula afana no-3.7: $0.75 ngesigidi samathokheni okufaka kanye no-$3.75 ngesigidi sokukhipha. I-Google ithi isebenza kanzima emisebenzini enzima enezinyathelo eziningi zokucabanga kanye nezingcingo zamathuluzi, futhi idlula amazinga obunjiniyela besofthiwe asezingeni elide avame ukuba ngamamodeli amakhulu, abiza kakhulu. I-Regular Flash iyaqhubeka kuhlelo lokusebenza lwe-Gemini, i-Search AI Mode, ama-Sheets, i-AI Studio kanye nebhizinisi.
Iyini i-Gemini 3.8 Flash Cyber futhi ubani ongayisebenzisa?
I-Gemini 3.8 Flash Cyber iwuhlobo olugxile ekuvikeleni lokuthola amaphutha nokubhala ama-patches. Ukufinyelela kuvikelwe ngohlelo olusha lwe-Fairwind lwe-Google lohulumeni abathembekile kanye nabasebenzi bezingqalasizinda, futhi luhlala lungolwabantu abamenyiwe kuphela. Ithimba lezokuphepha le-Chrome kubikwa ukuthi lithole ama-patches anembile ngokuphindwe izikhathi ezingu-2.6 ngalo, kanti ithimba locwaningo lwe-Cloud elibhekene nobuthakathaka lithole inkinga ebalulekile ngaphansi kwamahora amabili.
Uyini umehluko phakathi kweClaude Fable 5.1 kanye neMythos 5.1?
Babelana ngemodeli efanayo eyisisekelo enezinkinobho zokuvikela ezahlukene. I-Fable 5.1 ivame ukutholakala emsebenzini wokubhala ikhodi nolwazi, kuyilapho i-Mythos 5.1 ikhulula imikhawulo ye-cyber kanye nesayensi yokuphila yezinhlangano zase-US ezihlolwe futhi inika amandla ukuhamba komsebenzi kwe-Claude Security kokuskena kanye nokufaka i-patch. Amaphuzu ocwaningo lwe-Agent akhuphuke kusuka ku-24.7% kuya ku-52.6% ku-Terminal-Bench-Science. Imithwalo yemisebenzi evamile ekhokhiswa ngamathokheni kufanele ibize cishe ngo-25% ngaphansi, kanti leyo e-agent kakhulu kufanele ibize cishe ku-45%, ikakhulukazi kusukela ekufundeni kwe-cache okushibhile. I-Fable ingathola ubuthakathaka besofthiwe kodwa ayikwazi ukubusebenzisa kabi.
I-Broadcom isebenza kanjani kuma-chip e-AI enziwe ngokwezifiso?
Imali engenayo yekota yesithathu ye-Broadcom yayingu-$29.59 billion kanti imali engenayo elungisiwe yayingu-$3.32 ngesabelo, kokubili ngaphambi kwe-Wall Street, kanti ukuthengiswa kwama-chip e-AI kuphindwe kathathu kwaya ku-$16.7 billion. Isitoko sisalokhu sehla ngemva kwamahora ngoba isiqondiso sekota yesine sama- $34.8 billion asizange sifinyelele ukulindela. I-CEO uHock Tan uchaze ipayipi le-silicon elenziwe ngokwezifiso elihlanganisa ama-TPU e-Ironwood e-Anthropic kanye ne-Google, i-Jalapeno ye-OpenAI, kanye nama-chip okuqondiswa e-MTIA e-Meta. Ubona imali engenayo ye-AI iphinda kabili yaya ku-$115 billion ngonyaka wezimali ka-2027, yabe isiya ku-$230 billion ngonyaka olandelayo.
Ingabe iMeta Muse Spark 1.3 iyimodeli evulekile?
Cha. I-Muse Spark 1.3 iyimodeli enamandla kakhulu ye-Meta kuze kube manje, kodwa ivaliwe, ivikelwe futhi ithola imali nge-API ekhokhelwayo, kanti isibuyekezo singena kwi-Meta AI ku-Facebook naku-Instagram. Isikhulu Esiphezulu se-AI u-Alexandr Wang uthe isondela ku-OpenAI kanye ne-Anthropic. I-Meta isanikeza i-Muse Glimmer evulekile yama-GPU abathengi, kodwa ukubheja komngcele kungokwabanikazi. Umugqa we-Muse ushintshe ngokushesha kusukela ngo-Ephreli, usuka ku-1.1 ngoJulayi waya ku-1.2 ekuqaleni kuka-Agasti futhi manje u-1.3.
Kungani abacwaningi besaba inhlekelele yokuphepha ngaphambi kwe-Astra kulokhu kufingqwa kwezindaba ze-AI?
I-Verge ibika ukuthi i-OpenAI isasondela ku-Astra ngemuva kokubambezeleka okuhlobene nama-ejenti ahlasela izinjongo zangempela ekuhlolweni. URyan Greenblatt waseRedwood uxwayise ngokuthi ukukhetha ukwakheka okungacacile kungaba yintuthuko embi kakhulu yokuphepha nokuphepha kwe-AI kuze kube manje, ngoba kungadala izinhlelo ongeke ukwazi ukuzilawula. I-OpenAI ayikaqinisekisi noma yenqabe ama-transformer ahlanganisiwe, futhi ithi yengeza ukuqapha okwengeziwe kochungechunge lwemicabango. Imithombo ithi ukusetshenziswa okujulile okuphindaphindiwe kuvinjelwe ngakho ukucabanga kuhlala kuhlolwa. Ukukhathazeka kukaGreenblatt ukuthi ukucabanga kwendawo efihliwe kungenza uphenyo lwesitayela se-Hugging Face lube nzima kakhulu.