At the end of February, Elon rolled out his latest model. Of course, it was "the best in the world."
Yintoni i-intelligent AI ehlabathini?
Njengomdla, uMusk waya i-hype train. Kodwa akukho idatha emangalisayo kakhulu ekuqaleni.Blog UkubalaUkubonisa ukuba i-beta yaye iimodeli zangaphakathi kwi-training.
Zifumaneka i-benchmarks ezininzi ezibonisa i-Grok 3 elandelayo. Nangona kunjalo, abafumanisa ukufinyelela kwi-API. Oku kubalulekile ngenxa yokusebenzisa i-benchmarks eyahlukileyo yokuhlola.
Ngoko ke, i-Elon ibonisa ukuba i-Grok 3 yintoni yintoni yintoni "ngokuqhelekanga" kwaye ibonelela yonke into. Kodwa iindlela ezininzi zokucacisa ziye zixhomekeke ngokufanelekileyo okanye ukhangela i-benchmarks zayo.
Yintoni iinkcukacha? Hlola:
Ukubona le indawo elula kwi-right? Yinto i-boost ye-Grok efunyenwe ngexesha elikhulu lokucoca (i-test-time compute) ukuze ufumane imibuzo efanelekileyo. Ayikho ngexesha elifanelekileyo.
Ingaba uyazi iimodeli ze-AI ziquka imibuzo emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala emibala.
Kodwa i-Grok iimiphumo ziye ziboniswe zonke ngokusebenzisa i-cons@64. Yintoni, i-inthanethi ye-64 yeenkcukacha kunye nokukhetha umxholo we-commonest. Emva koko, i-xAI iye yahlanganisa i-score kwi-pass@1 yeengcali.
Kwaye ngoko ke, bakwazi ukuba iimodeli ye-next-gen. Ngoko ke, zisebenzisa iinqwelo ezininzi eziluncedo.
Ukubala ngokufanelekileyo, kwindawo efanelekileyo, zonke iilebhu zihlanganisa iimveliso. Zibonisa iimveliso ezininzi okanye zihlanganisa iimveliso ezininzi ezininzi zihlanganisa - kodwa ngokufanelekileyo.
Okay, ama-benchmarks ngaphandle. Yintoni abasebenzisi abavela emva kokusetyenziswa ngokwenene? I-consensus epheleleyo:
Imodeli enkulu kodwa ayikwazanga iziphumo. Kwakhona i-hallucinates kunye nokuphendula kwiingxaki ezininzi ezininzi.
Ukusebenzisana kwimveliso, i-Grok 3 ibekwe phantsi kweemveliso ze-OpenAI ezihlangeneyo, ngoko ke ngexabiso engaphezulu kwe-DeepSeek kunye ne-Google ngexesha lokuvelisa.
Nangona iinyanga ezimbini emva, i-Gemini 2.5, i-Claude 3.7, kunye ne-GPT-4o entsha zilungele. Kwaye ngexesha lokugqibela sinike ukufikelela kwe-API ye-Grok 3 kunye ne-mini-version yayo. Ngokuqhelekileyo, kuphela i-mini-version ilungele i-think mode kwi-API.
Ngoko ke namhlanje sinokufuneka ukuba i-cost and definitely not the absolute best.
Nceda uqhagamshelane, kukho ngaphezulu kwi-story.
Kwaye kufuneka uqhagamshelane kubo, Elon kunye ne-xAI zangena kwimarike ngokukhawuleza, kwangena umdlali wokufanisa ixesha elidlulileyo.
1 - I-Hardware
Yintoni ingxoxo enkulu?
Ngo-2024, i-xAI ibekwe i-computer cluster eningi. Sibonelela i-100,000 i-Nvidia H100 i-GPU ezisetyenziswa kwiminyaka engama-4 kuphela. Emva koko zibandakanya ukuba kwi-200,000 i-cards kwiminyaka engama-3.
CEO kaNvidia, Jensen Huang,UkucaciswaOku kuthatha ngokuvamile malunga neenyanga ezimbini.
Kwaye ngexesha elide, akukho ibhizinisi yokuzonwabisa – le datacenter elikhulu ehlabathini. Akukho omnye awukwazi ukuqhagamshela i-GPU ezininzi kwindawo eyodwa.
Ngokuvamile, i-clusters ezininzi ziquka i-datacenter ezininzi ezivela kwi-Infiniband kabini ezininzi ezininzi. Kwi-training, i-centre ezininzi zihlanganisa iintlobo zebhanki ngokugqithisileyo. Ukuba i-connectivity i-slow, i-GPU ezininzi ezininzi zihlala, nto leyo iindaba ezininzi.
I-datacenter ye-typical ingaba i-10,000-20,000 i-GPU, i-sugging i-20-30 megawatt ye-power.Umzekelo, I-Microsoft (i-OpenAI) isebenza kwi-100k i-GPU ye-network e-Arizona, kwaye i-Meta isebenza kwi-128k.
Ukubona izakhiwo ezimbini H-shapha? Ezi zihlanganisa izakhiwo ezimbini ze-Meta data centers.
Ukusetyenziswa kwamanzi kwi-cluster ye-top-tier ziphuma ukuya kwi-10x ukususela kwi-2022. Ngoku siza kuxhomekeke kwi-150 MW ngalinye ngalinye. Yinto efana ne-powering ye-city encinane. Oku kwenza isibambiswano esikhulu kwinethiwekhi ze-power. Kwimeko ezininzi, kuxhomekeke kakhulu ukuvelisa amandla kunokuba ukunikezela ngenxa yokungabikho kwizilwanyana ze-power.
Ngoko ke, u-Elon ivela kwimarike. Yaye... uyenza i-"I-Elon thing." Ukukhangisa i-tweets yayo yonke into, umntu uyazi ukuvelisa iimveliso njengoko akukho enye.
I-Electrolux iye yenza iimveliso ezidlulileyo e-Memphis yaye iye yenzelwe ukwakha i-datacenter elikhulu kunokuba yi-network eyenziwe ngexesha elinye.
Ngokutsho, amandla kwangaphambili.
I-factory yaba kuphela i-7 MW ukusuka kwi-grid yendawo - ezininzi i-4,000 i-GPU. I-utility yendawo, i-Tennessee Valley Authority, ibonelela i-50 MW eyongezelelweyo, kodwa ngaphandle kwe-Agasti. Kwaye i-xAI ye-substation ye-150 MW yaye kwakhona isakhiwo, engaphantsi lokugqibela kwiminyaka.
Kodwa ukhangela ayikho style ye-Musk.
I-Dylan Patel ( ukusuka kwi-Semianalysis)Ukucingangokusebenzisa iifoto ze-satellite i-Elon iye yenza i-14 i-massive mobile diesel generators ukusuka kwi-VoltaGrid. I-hooked ukuya kwi-4 i-substation ye-mobile kunye ne-powered ye-datacenter. I-trucks i-electricity.
U-Patel wabhala ukuba bafumene i-30% yeemarike ye-US ngokupheleleyo ngenxa yeengenerators (ngaphandle kokuba awukwazi ukufumana nayiphi na into).
Kubaluleke, i-datacenter isetyenziselwa ukutshisa kwe-liquid. I-Google kuphela iye yenzelwe ngexesha elandelayo. Oku kubaluleke kakhulu ngenxa yokuba i-Nvidia ye-chips yeentsuku elandelayo, i-Blackwell B200s, kufuneka i-liquid cooling. Wonke umntu uya kufuneka ufakele iinkqubo ze-data centers zangaphakathi.
Ingaba ufumana iiyure ezimbini zokuqala le ngevidiyo ukuze ubone yintoni. Ndifumene i-chockled how hyped the guy is about grey boxes and cables:
I-engineering enhle kakhulu - bheka kuphela ukulawulwa kwe-cable.
Ngaba nayiphi na umntu owenziwe ngempumelelo ngexesha elininzi.
2 - Ngaphezu kwe-Hardware!
Elon says by summer 2025, they'll have a 300k GPU cluster with Blackwell B200 chips. Given Musk’s habit of exaggeration, let's say it's realistically somewhere between 200-400k new chips by the end of 2025. B200 is roughly 2.2 times better than H100 for model training (based on Nov 2024 estimates).
Musk ukhangela ukwakha isakhiwo se-2.2 GW, nto leyo i-power more than a medium-sized city consumes.
Kwaye ayikho nje - zonke abadlali ezinkulu zenza into efanayo:
- Ukucinga
- I-Meta ibekwe kwiiyunithi ezimbini ze-gas e-Louisiana. Ukucinga
- I-OpenAI / i-Microsoft ihamba into efanayo eTexas. Ukucinga
- I-Amazon kunye ne-Google zibonisa kwiziko ze-gigawatt. Ukucinga
Yintoni i-nuclear? Kufuneka amandla, kodwa ukwakhiwa kwizakhiwo ze-nuclear iya kuthatha ixesha elide. Ungayifaka nje phantsi kwicandelo yakho ye-datacenter ngonyaka. I-wind and solar farms kunye ne-batteries ziyafumaneka, kodwa ziya kuthatha ixesha elide ukuyisebenzisa kwi-scale efunekayo.
Ngenxa yoko, i-Microsoft kunye ne-Meta baye baye baye baye baye baye baye baye baye baye baye baye baye baye baye baye baye baye.Zifunyenza imvakalelo yayo ukwandisa uMoloch ukuya eHaveni!
I-Grok 3 I-Gigant
Ngoko ke, i-Elon yenza le ibhokisi elikhulu, elikhulu. Ngoko ke?
Iingcebiso zibonisa 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 ngosuku kwi-25k i-A100 chips ezidlulileyo, kunye ne-H100 i-approximately i-2.25x faster.
Ukwenza i-mathematics, i-Grok 2 ibonelela malunga nexabiso ye-computing eyenziwa kulinganiswa ne-GPT-4. Kwaye i-Grok 3 ibonelela kwiiyure ezincinane kunokuba i-Grok 2. I-Google's Gemini 2.0 ingasetyenziselwa inani elifanayo le-hardware (100k ye-TPUv6 chips yayo), kodwa i-model ngokwemvelo ibonakalayo.
Ngokutsho, ukuphephaIindleko ze-computeri-Grok 3 yintoni ubukhulu (10 amaxesha!) engaphezulu kunokuba i-competitor yayo engaphezulu. Ngokuqhelekileyo, sinokufuneka iinkcukacha ezisemthethweni kwi-GPT-4.5 okanye i-Gemini 2.5.
Ngoko ke zithunyelwe iimveliso ezininzi ezininzi ekubunjweni le mega-cluster, kwaye iimodeli efanelekileyo ... ngokufanelekileyo kunye neengcali. Kwangathi akukho league engcono.
Ukubonisa ukuba iingcali ye-xAI kwi-training inokufutshane kwi-OpenAI, i-Google, okanye i-Anthropic. Zininzi i-brut-forced wayeko kwi-top-tier. Akukho iingcingo ze-magic ezibonakalayo, nje: "Ukuba i-brut-force ayixazisa inkinga yakho, ungenza kakhulu."
Kodwa kunokuba i-catch nge-approach. Epoch AIUkucaciswaKwiminyaka elidlulileyo, iimveliso ze-algorithmic zihlanganisa malunga ne-third of the progress in model capabilities. Iintlobo ezine-third zithunyelwa nje ukusetyenziswa kwe-hardware kunye needatha kwiimodeli ezininzi.
I-Brute Force yasebenza kwi-Grok 3 ngexesha elide, kodwa izindleko ziya kwandisa kwi-exponentially kwaye zithunyelwe ngokunciphisa kakhulu. Kwaye i-xAI kufuneka ukufikelela kwi-algorithm side. Iindaba elungileyo yinto ukuba ngoku ziyaziwa ukuba zithunyelwe kwi-frontier, ngoko kufuneka kube lula kakhulu ukufumana i-talent ephezulu.
4 - Yintoni i-good about Grok?
- Ukucinga
- Yinto epheleleyo (ngokuthi ngexesha lokugqibela lokugqibela). Ukucinga
Kwaye ngaphandle kweengxaki ze-Anthropic, iingxaki ze-DeepSeek, okanye iingxaki ze-OpenAI.
Nangona zonke iimodeli ezintsha ezidlulileyo eminyakeni ezidlulileyo, i-Grok ibekwe kwi-top of the top.I-Chatbot Arenaumqhubi.
Kwakhona nathi i-benchmarking eyahlukileyoUkucingaiimveliso
KwakhonaI-LiveBenchiimveliso
- Ukucinga
- I-Reasoning & i-Deep Research Mode Ukucinga
Kwangathi ngoFebruwari, i-function ye-Deep Research yamahhala ngokubanzi i-Perplexity. Ngoku, i-Google kunye ne-OpenAI ibonelela ezinye kwi-tier base-mhlawumbi i-Grok iye yenza?
Le mode automatically analyzes 30-100 links (i-Google inokufumana ngaphezulu) kwimizuzu kwaye itshintshe i-summary esifunyenweyo (kuquka ebomvu) ukuba kufuneka nje ukubuyekeza kwaye ukubuyekeza iimeko. It is way easier than researching anything from scratch. I has found Grok's version works faster than the others, so I've started using it when I need to research something. Like, xa ukuthenga i-headset entsha.
- Ukuhlanganiswa nge-X Ukucinga
Yintoni ingaba yindlela yayo yokuzalwa: ukufuna semantic ayikho kuphela iisombululo, kodwa ngenxa yintoni oya kufuneka. Uyakwazi ukhangela iimpawu kwi-theme ukucacisa iintlobo. Okanye ukufumana iimpawu ezidlulileyo evela kumasebenzisi eyodwa.
I-Twitter yintoni efikelelekayo kwi-real-time information platform, ngoko iyona elungileyo. Kodwa ngexesha elandelayo i-Grok isibindi, ukuthatha idatha evela kwiintsuku ezidlulileyo ezidlulileyo.
- Ukucinga
- Izixhobo ze-unfiltered Ukucinga
Kwaye ngenxa ye-grand finale, i-mode ye- 18+. I-Grok i-jailbreak ye-notoriously lula ngaphandle kwexesha elide. Uyakwazi ukwenza... kakuhle, nangona kunjalo, ukusuka kwi-voices ze-flirty ukuya kwiingcebiso ze-recipes. Iingcebiso ze-voice ziquka kakhulu.
Ngaba ushiye ngexesha lokugqibela, i-hillary!
I-Ironically, i-Grok ngokufanayo ayibonakalisa ukuba i-Musk (okanye i-Trump) ayinxalenye kakhulu. Xa oku kuthetha, i-xAI iye yenza isisombululo - ngokucacileyo i-hardcoding isisombululo apho i-Grok ayikwazi ukuchithisa i-Elon. Xa yaye i-exploded, zithintela abasebenzi we-OpenAI yaye "ukungasebenzi kwi-culture yekhompyutha."
Iingxaki yokuba iingxaki zeGrok zibonakalisa kuphela iinkcukacha zayo zokusebenza (i.e., i-internet), kwaye akukho iingxaki zokusetyenziswa. Ukusabela ukulungiselela iingxaki zayo ngaphandle kokutshintshisa iimodeli epheleleyo.
I-5 - Uya kufuneka utshintshe?
Nceda uxolo, kodwa njenge-pilot yakho yokuqala.
Ukucinga:
- Ukucinga
- Iimveliso ezininzi iimveliso ze-train kunokuba iimveliso ze-competitors. Ukucinga
- Nangona kunjalo, imveliso kumnqweno malunga neengcono. Ukucinga
- Kodwa i-super fast and free (ngoku). Ukucinga
- I-Deep Research mode yinto efanelekileyo – utshintshe ukuba awunayo. Ukucinga
- I-Hallucinations kunye ne-Hallucinations kunye ne-Hallucinations.
- Ukusabela ngokuvamile ngokufanelekileyo kodwa ngokufanelekileyo. Ukucinga
- Ukufinyelela kuphela kwi-Twitter data. Ukucinga
xAI kuboniswa ekugqibeleni kwe-infrastructure yamazwe ngamazwe ngexesha elidlulileyo. Kodwa kwizinto ze-AI zayo zibonakalayo, zibonakalisa ukufikelela kwi-top kunye ne-computing power.
Yinto inikeza umdlali omnye oqinileyo enikezela i-OpenAI, i-Google kunye ne-Anthropic, ukutshintshisa i-IA kwimveliso kwi-commoditization. I-competition ikhulisa kwaye i-exclusivity yeemodeli ezihlangene.
Ngaba uyakufuneka? Faka i-upvote okanye ubhaliseleI-Newsletter yethuNdingathanda ukuba!
