I-Grok ye-xAI 3: Zonke i-GPU, None of the Breakthroughs

Kwangathi ngoFebruwari, uElon waya imveliso yayo esitsha. Kwakhona, i-"i-best in the world."

I-AI ye-Intelligent yeYurophu yintoni?

Ngexesha elizayo, uMask waya i-hype train. Kodwa akukho idatha ezininzi ezinxulumene kwi-launch. i-xAI short i-blog post wabhala ukuba i-beta yaye iimodeli zangaphakathi kwi-training.

Bhalisa ezinye iibhanki ezibonisa i-Grok 3 phambi. Nangona kunjalo, abafumanisa ukufikelela kwi-API. Oku kubalulekile ngenxa yokuba ama-benchmarks ezihlabathi zisetyenziswe.

Ngoko ke, i-Elon ibonisa ukuba i-Grok 3 yintoni yintoni "ngokuqhelekanga" kwaye ibonelela yonke into elinye. Kodwa iindlela ezininzi yokuvavanya kubandakanya ngokufanelekileyo okanye ukhangela ama-benchmarks zabo.

Yaye izibane ezininzi? Hlola:

Sibona le indawo elula kwakhona? Yinto i-Grok yokukhuthaza ngokufanisa amandla ye-computing (i-test-time compute) ukufumana iimpendulo ezininzi. I-not exactly a fair fight.

Ngaba uyazi ukuba iimodeli ze-AI ziyafumana imibuzo emibini emibini emibini - emibini engcono, emibini emibini emibini. Izixhobo ezininzi zixazulule le ubungakanani, ukucacisa kuphela imibuzo yokuqala (pass@1). Yinto elula kwaye ilinganiselwe indlela yethu ngokwenene ukusetyenziswa kwe-AI - sincoma imibuzo efanelekileyo ngexesha lokuqala.

But imiphumela ye-Grok yaziwa zonke ngokusebenzisa i-cons@64. Yeyiphi na ingxaki ze-64 yaye i-xAI ibonelela imibuzo embi. Emva koko, i-xAI ibonelela ukuba i-score yaye i-pass@1 yeengcali.

Ngoko ke, ngoko ke, zibonisa ukuba iyimodeli ye-next-gen. Ngoko ke, zibonisa izixhobo ezininzi ezininzi ezininzi.

Ukuba uqhagamshelane, kwindawo efanelekileyo, zonke iilebhu zibonise iinkqubo. Zibonisa izibambiso ze-Cherry-pick okanye zihlanganisa iimodeli ezininzi ezininzi zihlanganisa - kodwa ngokufanelekileyo.

OK, ama-benchmarks ngoko. Yintoni abasebenzisi abavela emva kokusetyenziswa ngokwenene? I-consensus epheleleyo:

Iimodeli enkulu kodwa ayikwazanga iziphumo. Kwakhona i-hallucinates kunye nokuphendula kwiingxaki ezininzi ezininzi.

Kwenziwe ngempumelelo, i-Grok 3 ibekwe phantsi kweemveliso ze-OpenAI ezihlangeneyo, ngoko ke ngexabiso engaphezulu kwe-DeepSeek kunye ne-Google ngexesha lokuvelisa.

Ngozi, iinyanga ezimbini emva, i-Gemini 2.5, i-Claude 3.7, kunye ne-GPT-4o entsha ziye ziye zithunyelwa. Kwakhona sikugqibela sifumane ukufikelela kwe-API ye-Grok 3 kunye ne-mini-version yayo. Ngokuqhelekileyo, kuphela i-mini-version lithunyelwe i-think mode kwi-API.

Ngoko ke namhlanje sinokufuneka ukuba i-expensive kwaye ngoko ke engabonakali kakhulu.

Ndiya, kodwa kunokuba kukho ngaphezulu kwi-story.

Imodeli yinto enzima kwaye kunzima ukubonisa. Kwaye kufuneka uqhagamshelane kubo, Elon kunye ne-xAI zangena kwimarike ngokukhawuleza, kwangena umdlali wayo ngexesha elidlulileyo.

1 – I-Hardware

Icebiso elikhulu apha?

Nge-2024, i-xAI ibekwe i-computer cluster eningi. Thina ubhale i-100,000 i-Nvidia H100 i-GPU ezisebenza kwiminyaka engama-4 kuphela. Emva koko zithuba ukuba i-200,000 i-cards kwiminyaka engama-3.

I-CEO yeNvidia, Jensen Huang, ngathi oku kuthatha malunga ne-4 iminyaka.

Iyi yaba inguqulelo elikhulu yobuchwepheshe. Kwaye ngexesha, akukho ibhizinisi emangalisayo – le data center elikhulu ehlabathini. Ngaba nayiphi na umntu awukwazi ukuqhagamshela i-GPU ezininzi kwindawo efanayo.

Ngokuqhelekileyo, i-clusters ezininzi ziquka i-data centers ezininzi ezivela kwi-Infiniband kabini ezininzi ezininzi. Kwi-training, i-centre ezininzi zihlanganisa iitoni zebhanki ngokugqithisileyo. Ukuba ukuxhumanisa kunzima, i-GPU ezininzi ezininzi zihlala, nto leyo iindaba ezininzi.

Icebiso yedatha elifanelekileyo ingaba i-10,000-20,000 i-GPU, evimbela i-20-30 megawatt ye-power. Umzekelo,, i-Microsoft (i-OpenAI) isebenza kwi-100k i-GPU network e-Arizona, kwaye i-Meta isebenza kwi-128k.

Ndinga izakhiwo ezimbini kwi-H-shaped? Enye izakhiwo ezimbini ze-Meta data centers ezihambelana.

Ixabiso lwekhompyutha yeengcali ezihlangene kwi-10x ukususela ngo-2022. Ngoku siza kuxhomekeka malunga ne-150 MW ngexabiso ngexabiso. Kuyinto efana nokukhuthaza iiyunithi omnqweno. Oku kwenza isibambiso esikhulu kwinethiwekhi zangaphakathi. Kwimeko ezininzi, kuxhomekeke ngexabiso yokukhuthaza kunokuba ukunikela ngenxa yokungabikho kwizilwanyana zangaphakathi.

Ngoko ke, Elon ivela kwimarike yaye ... yenza "I-Elon thing." Ukukhusela i-tweets yayo yonke into, umntu uyazi ukuba ukwakha iimveliso njengoko akukho enye.

Ukuvela i-Electrolux yobugcisa e-Memphis yaye wahlala ukwakha i-datacenter elikhulu kunokuba yi-network efana nabanye.

Wagqibeleleyo, amandla kwangaphambili.

I-factory yaba kuphela i-7MW ukusuka kwi-grid ye-local - ezininzi i-4,000 i-GPU. I-utility yendawo, i-Tennessee Valley Authority, ibonelela i-50MW ezininzi, kodwa ngaphandle kwe-Agasti. Kwaye isakhiwo se-substation ye-150MW ye-xAI yaba kuxhomekeke, engaphantsi lokugqibela kwiminyaka.

But waiting is not Musk's style.

Dylan Patel ( ukusuka Semianalysis) ngathi via satellite images that Elon just brought in 14 massive mobile diesel generators from VoltaGrid. Hooked them up to 4 mobile substations and powered the data center. Literally trucked in electricity.

I-Patel ibonisa ukuba bakwazi ukuthenga i-30% yeemarike yeMzantsi yaseMelika ngenxa yale generators (ngaphandle kokuba awukwazi ukufumana nayiphi na nto).

Ukuvavanyo, i-data center isetyenziselwa ukuchithwa kwe-liquid. I-Google kuphela iye yenza oku kwakhona ngexesha elandelayo. Kuyinto ingxaki elikhulu ngenxa yokuba i-Nvidia ye-chips ye-next generation, i-Blackwell B200s, kufuneka isetyenziswa kwe-liquid. Wonke omnye kufuneka ufakele i-datacenters zayo ezidlulileyo.

Ukuba ufumana iiyure ezidlulileyo le ngevidiyo ukubona into efanayo ngaphakathi. Ndifumene umnqweno malunga ne-hyped yaye malunga neengxaki ze-grey kunye ne-cable:

I-engineering enhle kakhulu - bheka ukulawulwa kwe-cable.

Ngaba nayiphi na umdla owenziwe ngexesha elininzi.

2 – Uninzi Hardware!

I-Elon uthi ngexesha le-2025, iya kuba i-GPU ye-300k kunye ne-Blackwell B200 chips. Ngokusho i-habit ye-exaggeration ye-Musk, sinokufundisa ukuba kuxhomekeke phakathi kwe-200-400k i-chips ezintsha ngexesha le-2025.

I-Musk inikezela ukwakha isakhiwo se-2.2 GW. Yintoni i-power engaphezulu kwe-city ye-middle-size.

Ndiya akuyona nje – zonke abadlali ezinkulu zenza into efanayo:

• I-Meta ibekwe kwiiyunithi ezimbini ze-gas eLouisiana.
• OpenAI/Microsoft ibeka into efanayo eTexas.
• I-Amazon kunye ne-Google ibekwe kwiziko ze-gigawatt.

Ukuba akuyona yombane? Kukho amandla, kodwa ukwakhiwa kwizakhiwo ze-nuclear kunzima kakhulu. Unako nje ukhangela elinye kwi-data center yakho kwiminyaka eyodwa. I-wind and solar farms kunye ne-batteries ziyafumaneka, kodwa ziquka ixesha elide ukuba zithunyelwe kwimveliso efunyenweyo.

Ngokusho, i-Microsoft kunye ne-Meta ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe.

3 – Grok 3 is Huge

Ngoko ke, Elon yenza le ibhokisi elikhulu, ezininzi. Ngoko ke?

Ukuhlaziywa kubonisa ukuba i-Grok 2 ifunyenwe kwi-20k H100s, kwaye i-Grok 3 isetyenziswe kwi-100k. Ngokutsho kwe-context, i-GPT-4 ifunyenwe ngexesha le-90-100 kwi-25k i-A100 chips ezidlulileyo, kunye ne-H100 i-2.25x engaphezulu.

Ukwenza imathematika, i-Grok 2 yaba malunga nexabiso lokucoca kulinganiswa ne-GPT-4. Kwaye i-Grok 3 yaba ngexesha elincinane kunokuba i-Grok 2. I-Google's Gemini 2.0 ingasetyenziswa ngamanani elifanelekileyo ye-hardware (100k ye-TPUv6 chips yayo), kodwa i-model ngokufanelekileyo ingabizi.

Ngokuqhelekileyo, ixabiso rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo rhoqo

Ngoko ke baye zithunyelwe izisombululo ezininzi kwi-mega-cluster yokwakha, kwaye iimodeli efanelekileyo ... ngokufanelekileyo kunye nabasebenzi. Ngokuqhelekileyo, akukho league engcono.

Ngokuqhelekileyo iinkcukacha ze-xAI kwi-training ziquka kwi-OpenAI, i-Google, okanye i-Anthropic. Zininzi ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye ziye zibe ziye ziye zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe zibe

But kukho umgangatho. Epoch AI isibuyekeza ukuba ngexesha elidlulileyo, iimveliso ze-algorithmic zithintela malunga neentlawulo yeemveliso. Iintlawulo ezininzi ezimbini ziye zithunyelwa nje ukusetyenziswa kwe-hardware kunye needatha kwiimodeli ezininzi.

I-Grute Force yenzelwe kwi-Grok 3 ngexesha elide, kodwa iindleko ziya kukunyuka ngokubanzi nangokufumana ukunyaniseka okungenani. Kwaye i-xAI kufuneka ukufikelela kwi-algorithm side. Iindaba elungileyo yinto ukuba ngoku ziyaziwa ukuba zikhuthaza i-frontier, ngoko kufuneka kulula kakhulu ukufumana i-talent ephezulu.

4 – Yintoni iingubo ye Grok?

1. Wagqibeleleyo ngokupheleleyo (ngokuthi ngexesha lokugqibela).

Ngoya ngaphandle kweengxaki ze-Anthropic, iingxaki ze-DeepSeek, okanye iingxaki ze-OpenAI.

Ukuba zonke iimodeli ezintsha ziye zithunyelwe kwiinyanga ezidlulileyo, i-Grok ibekwe kwi-top of the Chatbot Arena leaderboard.

Ndiya kwakhona i-benchmarking engaphandle ye- I-EpochAI:

And by LiveBench:

2. I-Reasoning & Deep Research Mode

   
   

Kwangathi ngoFebruwari, i-Free Deep Research ifunyenwe ngokubanzi ngaphandle kwe-Perplexity. Ngoku, i-Google kunye ne-OpenAI zinikeza ezinye kwi-tier base-ngokuthi i-Grok yandisa?

I-mode yenza ngokuzenzakalelayo i-links ye-30-100 (i-Google ingayenza ngaphezulu) kwimizuzu kwaye ibonisa inkxaso oluthe ngexabiso (yaye ebomvu) leyo nje kufuneka ukunxibelelana kunye nokukhawuleza iimeko. I-mode elula kunokuba i-research ye-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni-ni.

3. Ukuqhagamshelwano kunye neX

   
   

Yinto ingaba yintoni yayo: ukufuna semantic ayikho kuphela iingoma yokubhalisa, kodwa ngenxa yintoni oya kufuneka. Uyakwazi kwakhona ukhangela iimpawu kwi-theme ukucacisa iintlobo. Okanye ukufumana iimpawu ezidlulileyo evela kumasebenzisi eyodwa.

I-Twitter yintloko ekugqibeleni kwi-real-time information platform, ngoko kuhle. Kodwa ngexesha elide i-Grok isibindi, ukuthatha idatha evela kwiintsuku ezidlulileyo ezidlulileyo.

4. I-Unfiltered Stuff

   
   

Nge-grand finale, i- 18+ mode. I-Grok i-jailbreak yehlabathi ngokufanelekileyo ngaphandle kweemfuno emininzi. Uyakwazi ukufumana oku ... kakuhle, nangokufuneka, ukusuka kwi-voices ze-flirty ukuya kwi-recipes ezininzi ezininzi. Iimeko ze-voice ziyafumaneka kakhulu.

Sikufunda ukuya ekupheleni, kuhle!

I-Ironic, i-Grok ngokufanayo ayibonakalisa ukuba i-Musk (okanye i-Trump) ayinxalenye kakhulu. Xa oku kuthetha, i-xAI iye yandisa isisombululo - ngokucacileyo i-hardcoding isisombululo apho i-Grok ayikwazi ukucacisa i-Elon. Xa yaye i-exploded, baye zithintela umenzi we-OpenAI wokuqala ukuba "ukungabikho kwi-culture yebhizinisi."

Icebiso efanelekileyo yi-Grok iingcebiso zibonakalisa kuphela iinkcukacha zayo zokusebenza (i.e., i-internet), kwaye akukho iingcebiso eziluncedo. Ukusabela ukulungiselela iingcebiso ezininzi ngaphandle kokubonakaliswa kwimodeli epheleleyo.

5 - Uya kufuneka uqhagamshelane?

Ukuza oku, kodwa njenge-pilot yakho yesibini.

Ukulungiselela:

• Izixhobo ezininzi ukuqeqesha kunokuba iimodeli zeengcali.

• Ngaphandle kokuba, umphumela iye phantse ngexabiso.

Ndiya i-super fast and free (ngoku).

• I-Deep Research mode yinto efanelekileyo-ukutshintsha ukuba awunayo.

I-accessive data ye-Twitter.











































i-xAI ibonelela ukuba inokufunda i-infrastructure yehlabathi ehlabathini ngexesha elidlulileyo. Kodwa kwizinto ze-AI zayo zibonakalayo, zibonakalisa ukufikelela kwindawo ephakamileyo nge-computer power.




Ini inikeza umdlali omnye omnqweno omnqweno we-OpenAI, i-Google, kunye ne-Anthropic, ukutshintshisa i-AI kwimveliso kwi-commoditization. I-competition ikhulisa kwaye i-exclusivity yeemodeli ezihlangene.






Ngithanda oku? Nceda uqhagamshelane na i-newsletter yam. Ndingathanda!