graph LR
classDef perf fill:#ffd4d4, font-weight:bold, font-size:14px;
classDef scale fill:#d4ffd4, font-weight:bold, font-size:14px;
classDef tool fill:#d4d4ff, font-weight:bold, font-size:14px;
classDef multi fill:#ffffd4, font-weight:bold, font-size:14px;
classDef demo fill:#ffd4ff, font-weight:bold, font-size:14px;
classDef voice fill:#d4ffff, font-weight:bold, font-size:14px;
classDef api fill:#ffd4d4, font-weight:bold, font-size:14px;
classDef use fill:#d4ffd4, font-weight:bold, font-size:14px;
classDef future fill:#d4d4ff, font-weight:bold, font-size:14px;
Main --> P1[Grad exams crushed
near perfect scores. 1]:::perf
Main --> S2[Compute 100Ă— leap
via 100k H100. 2]:::scale
Main --> R3[RL enables
self correction reasoning. 3]:::scale
Main --> H4[Humanity Exam jumps
to 50 % multi-agent. 4]:::perf
Main --> T5[Native tools live
code search markets. 5]:::tool
Main --> M6[Multi-agent Heavy
spawns debating researchers. 6]:::multi
Main --> B7[Browser shows
black hole collisions. 7]:::demo
Main --> P8[Polymarket predicts
MLB odds live. 8]:::demo
Main --> X9[X finds employee
weirdest profile photo. 9]:::demo
Main --> V10[Voice latency halved
cinema prosody voices. 10]:::voice
Main --> C11[Live Diet Coke
opera creative fluency. 11]:::voice
Main --> A12[€300 API tier
256k context realtime. 12]:::api
Main --> V13[VendingBench sim
virtual business stable. 13]:::use
Main --> A14[ARC automates
CRISPR logs x-rays. 14]:::use
Main --> F15[Fin firms gain
live market insight. 15]:::use
Main --> F16[Four hour FPS
built from scratch. 16]:::use
Main --> V17[Next model adds
sharp video gen. 17]:::future
Main --> T18[100k GB200
train next video weeks. 18]:::future
Main --> K19[Elon envisions
Kardashev-1 AI loops. 19]:::future
Main --> S20[Sim fidelity goal
indistinguishable from reality. 20]:::future
Main --> D21[Data-center race
UAE vs China. 21]:::future
Main --> S22[Spain courts 80 M€
yet lacks strategy. 22]:::future
Main --> S23[Earth carpeted by
solar feeding data-life. 23]:::future
Main --> G24[GPU temples may
guide human destiny. 24]:::future
Main --> H25[Host invites hire
Grok agents lab. 25]:::future
Main --> M26[Madrid summer interview
teased after 300k. 26]:::future
Main --> A27[Alicante Chamber
AI business reinvention. 27]:::future
Main --> F28[Free content pledge
donations optional. 28]:::future
Main --> C29[Discord open for
early agent collab. 29]:::future
Main --> R30[Reflection: hybridize
or worship AI gods. 30]
Resume:
The broadcast opens with host Plácido Domènech framing the episode as a calm, Spanish-language deep-dive into Grok 4, insisting that X-Habai’s slower, analytical lens is more valuable than headline-chasing. He positions the release amid an “AI war” between U.S. and Chinese giants, while Europe remains a distracted bystander, and promises reflections on Elon Musk’s political moves plus a surprise cameo from Ilya Sutskever.
Domènech then walks through the polished launch video: Grok 4 is branded as the world’s smartest AI, able to ace any unseen graduate exam, from the SAT to the brutal “Humanity’s Last Exam.” He pauses to highlight the metaphor of AI consciousness evolving faster than human childhood, noting that the model’s breadth—PhD-plus skill across every discipline—renders specialized human experts redundant.
Technical gains are underlined: a 100× jump in pre-training compute since Grok 2, reinforcement-learning at unprecedented scale on the 100k-GPU Colossus cluster, and plans to reach one million GPUs. The host stresses that this is not mere scaling; new verifiable-reward techniques let the model self-correct and “reason from first principles,” pushing accuracy on HLE from single digits to 40 % text-only and over 50 % with multi-agent Heavy.
Tool use is showcased next. Unlike Grok 3’s brittle prompt-based tool calls, Grok 4 natively integrates calculators, code execution, search and future physics-grade simulators. Live demos include a browser-based black-hole collision, a Polymarket World-Series forecast, and an X integration that hunts for the employee with the “weirdest profile photo,” all executed in real time.
Voice upgrades close the product segment: latency halved, new naturalistic voices like “Sal” and “Eve” demonstrated through opera about Diet Coke, plus a playful trolling of OpenAI’s voice mode. API pricing is pitched at €300 for premium tiers; Domènech argues the cost is trivial for any serious business use.
In the final stretch the discussion widens to geopolitics. Europe’s “AI factories” and Spain’s 80-million-euro data-center land-grab are dismissed as folkloric distractions beside the trillion-dollar Stargate alliance of the U.S. and UAE. Invoking Sutskever’s vision of earth blanketed by solar panels and data centers that become “non-human life,” Domènech warns that only nations with GPU megaclusters will shape the coming Kardashev-scale economy. He signs off inviting viewers to hire Grok 4 agents for his new lab channel and to ponder whether humanity will merge with or merely worship the emergent superintelligence.
30 Key Ideas:
1.- Grok 4 surpasses every unseen graduate exam with near-perfect scores across all disciplines.
2.- Training compute jumped 100Ă— from Grok 2 through 100k H100 GPUs on Colossus cluster.
3.- Reinforcement learning at scale enables self-correction and first-principle reasoning breakthroughs.
4.- Humanity’s Last Exam accuracy rises from single digits to 40 % text-only, 50 % with multi-agent.
5.- Native tool integration allows live code, search, simulation, and market forecasting.
6.- Multi-agent Grok 4 Heavy spawns parallel researchers that debate and refine joint answers.
7.- Browser demo visualizes colliding black holes with post-Newtonian physics fidelity.
8.- Polymarket demo predicts MLB World Series odds by browsing, calculating, and comparing data.
9.- X integration locates employee with weirdest profile photo purely from public web context.
10.- Voice latency halved; new Sal and Eve voices deliver cinema-grade prosody and emotion.
11.- Opera about Diet Coke performed live to showcase creative fluency and humor.
12.- API priced at €300 tier for developers, offering 256 k context and real-time tool access.
13.- VendingBench simulation shows Grok running virtual vending business without performance decay.
14.- ARC Institute adopts Grok to automate CRISPR experiment log analysis and chest-x-ray review.
15.- Financial firms leverage Grok with live data feeds for rapid market insight generation.
16.- Four-hour first-person shooter built via Grok sourcing assets, code, and 3D models.
17.- Future version 7 foundation model will add sharp video understanding and generation.
18.- Training planned on 100k GB200 GPUs for next-gen video model within weeks.
19.- Elon Musk envisions Kardashev-1 economy powered by AI-driven innovation loops.
20.- Simulation fidelity goal: tests indistinguishable from physical reality for science.
21.- Data-center arms race pits U.S.–UAE Stargate against China; Europe lags with token factories.
22.- Spain courts 80 M€ data-center sites yet lacks exascale strategy and sovereign compute.
23.- Sutskever foresees earth carpeted by solar panels feeding self-aware data-center life.
24.- Superintelligence may reside in GPU temples guiding or replacing human destiny.
25.- Host invites audience to hire Grok agents for new Spanish AI lab channel experiments.
26.- Summer Madrid interview teased after prior 300k-view session on InversiĂłn Racional.
27.- Alicante Chamber event will explore business reinvention with AI in three-hour talk.
28.- Free content pledge reaffirmed; donations optional via WinMeACoffee, PayPal, or SuperChat.
29.- Community Discord open for researchers wanting early Grok 4 agent collaboration.
30.- Reflection ends on whether humanity hybridizes with or merely worships emerging AI gods.
Interviews by Plácido Doménech Espà & Guests - Knowledge Vault built byDavid Vivancos 2025