Knowledge Vault 7 /363 - xHubAI 13/08/2025
🔴ARC-AGI 3 : Actualización en timeline AGI!
< Resume Image >
Link to InterviewOriginal xHubAI Video

Concept Graph, Resume & KeyIdeas using Moonshot Kimi K2 0905:

graph LR classDef arc fill:#ffd4d4, font-weight:bold, font-size:14px; classDef intel fill:#d4ffd4, font-weight:bold, font-size:14px; classDef comm fill:#d4d4ff, font-weight:bold, font-size:14px; classDef future fill:#ffffd4, font-weight:bold, font-size:14px; classDef meta fill:#ffd4ff, font-weight:bold, font-size:14px; Main[Vault7-363] Main --> A1[ARC AGI 3
preview mini-games 1] A1 -.-> G1[ARC] Main --> A2[Hand-crafted games
guarantee novelty 2] A2 -.-> G1 Main --> A3[Models explore
without prior training 3] A3 -.-> G1 Main --> A4[OpenAI xAI
flag ARC scores 4] A4 -.-> G1 Main --> A5[ARC-1 jump
via test search 5] A5 -.-> G1 Main --> A6[ARC-2 static
1,300 visual tasks 6] A6 -.-> G1 Main --> A7[ARC-3 100 games
by 2026 7] A7 -.-> G1 Main --> A8[Public API
for researchers 8] A8 -.-> G1 Main --> A9[30-day comp
prize strategies 9] A9 -.-> G1 Main --> B1[Intelligence is
adapt to novelty 10] B1 -.-> G2[Intelligence] Main --> B2[AGI timeline
cut to five 11] B2 -.-> G2 Main --> B3[Test-time adapts
not memorizes 12] B3 -.-> G2 Main --> B4[Store abstractions
globally 13] B4 -.-> G2 Main --> B5[Hive-mind shares
all abstractions 14] B5 -.-> G2 Main --> B6[Collective AGI
pools experience 15] B6 -.-> G2 Main --> B7[Not brute-force
millions tokens 16] B7 -.-> G2 Main --> B8[Do more
with less compute 17] B8 -.-> G2 Main --> B9[Models forget
humans retain 18] B9 -.-> G2 Main --> C1[Games reveal
explore plan align 19] C1 -.-> G3[Games] Main --> C2[Clear rules
sparse rewards 20] C2 -.-> G3 Main --> C3[Agents stuck
days in room 21] C3 -.-> G3 Main --> C4[Scaffolded agents
need hints 22] C4 -.-> G3 Main --> C5[Embedding distance
maximizes difference 23] C5 -.-> G3 Main --> C6[No procedural
keep insight 24] C6 -.-> G3 Main --> D1[ARC Prize
North Star AGI 25] D1 -.-> G4[Org] Main --> D2[Kaggle 1,500 teams
on ARC-2 26] D2 -.-> G4 Main --> D3[99.9 % feedback
positive 27] D3 -.-> G4 Main --> D4[119 k views
five days 28] D4 -.-> G4 Main --> D5[Ignore toxic
envy comments 29] D5 -.-> G4 Main --> D6[Seek truth
not savior 30] D6 -.-> G4 Main --> D7[XHabAI five years
Spanish AI 31] D7 -.-> G4 Main --> D8[Human-X coined
five years back 32] D8 -.-> G4 Main --> D9[New channel
paper analyses 33] D9 -.-> G4 Main --> E1[Own platform
escape censorship 34] E1 -.-> G5[Meta] Main --> E2[YouTube teasers
full elsewhere 35] E2 -.-> G5 Main --> E3[500 episodes free
donations accepted 36] E3 -.-> G5 Main --> E4[Discord open
for debate 37] E4 -.-> G5 Main --> E5[EU regulation
accept pragmatically 38] E5 -.-> G5 Main --> E6[Block trolls
debate critics 39] E6 -.-> G5 Main --> E7[Agentiki Codex 2025
free kit 40] E7 -.-> G5 Main --> E8[Hierarchical Model
beyond transformers 41] E8 -.-> G5 Main --> E9[Brain-inspired
beats LLMs 42] E9 -.-> G5 Main --> F1[Live papers
September frontier 43] F1 -.-> G6[Future] Main --> F2[Special Agents 2025
new content 44] F2 -.-> G6 Main --> F3[Panel debates
AI winter 45] F3 -.-> G6 Main --> F4[Media hype
vs lab speed 46] F4 -.-> G6 Main --> F5[Chinese gamers
outpace data 47] F5 -.-> G6 Main --> F6[Digital illusion
over freedom 48] F6 -.-> G6 Main --> F7[Nested fake
happiness hell 49] F7 -.-> G6 Main --> F8[Play ARC-3
share shots 50] F8 -.-> G6 G1[ARC] --> A1 G1 --> A2 G1 --> A3 G1 --> A4 G1 --> A5 G1 --> A6 G1 --> A7 G1 --> A8 G1 --> A9 G2[Intelligence] --> B1 G2 --> B2 G2 --> B3 G2 --> B4 G2 --> B5 G2 --> B6 G2 --> B7 G2 --> B8 G2 --> B9 G3[Games] --> C1 G3 --> C2 G3 --> C3 G3 --> C4 G3 --> C5 G3 --> C6 G4[Org] --> D1 G4 --> D2 G4 --> D3 G4 --> D4 G4 --> D5 G4 --> D6 G4 --> D7 G4 --> D8 G4 --> D9 G5[Meta] --> E1 G5 --> E2 G5 --> E3 G5 --> E4 G5 --> E5 G5 --> E6 G5 --> E7 G5 --> E8 G5 --> E9 G6[Future] --> F1 G6 --> F2 G6 --> F3 G6 --> F4 G6 --> F5 G6 --> F6 G6 --> F7 G6 --> F8 class A1,A2,A3,A4,A5,A6,A7,A8,A9 arc class B1,B2,B3,B4,B5,B6,B7,B8,B9 intel class C1,C2,C3,C4,C5,C6 comm class D1,D2,D3,D4,D5,D6,D7,D8,D9 meta class E1,E2,E3,E4,E5,E6,E7,E8,E9 meta class F1,F2,F3,F4,F5,F6,F7,F8 future

Resume:

Plácido Domenech opens the summer episode by celebrating 120,000 views in five days for the previous round-table on Rational Investing and warns that, despite the applause, envy and toxic comments are already surfacing. He insists that XHabAI will not slow down for August: a double session is scheduled with the preview of ARC AGI 3—an interactive benchmark based on mini-games that forces models to explore unknown environments without prior data—and a Spanish-language debate on whether the supposed “AI winter” is merely media pessimism or a real stagnation. Domenech underlines that OpenAI and xAI have already adopted ARC-1 and ARC-2 as flagship evaluators, but the real frontier is ARC-3, whose six initial titles will be released the following morning under embargo. The interview with François Chollet is screened next; the French researcher admits that his estimated AGI horizon has shrunk from ten to five years because test-time adaptation finally allows models to synthesize new programs on the fly instead of repeating memorized templates. Yet Chollet warns that brute-force search over millions of tokens is not intelligence, which he defines as “doing more with less,” and predicts that true AGI will resemble a global, continually-updated GitHub of reusable abstractions shared by millions of agent instances learning in parallel. Domenech closes by announcing a 30-day public competition on the ARC-3 games, the forthcoming open-source Agentiki Codex 2025 kit and his intention to migrate the show to a self-hosted platform to escape algorithmic censorship.

The core of the discussion is the leap from static LLMs to fluid, interactive agents. ARC-1 and ARC-2 proved that reasoning models can cross the human skill threshold when allowed to iterate at inference time, but they still fail in open-ended environments. ARC-3 therefore proposes six unrelated micro-games—logic, orchestration, exploration, alignment—that demand genuine on-the-job learning; humans quickly intuit rules while frontier models such as O3-High or Grok-4 loop helplessly without developer scaffolding. Chollet argues that the missing piece is no longer pattern matching but continual learning: the ability to accumulate, compress and transfer experience across tasks and instances, something current context windows or periodic fine-tuning do not provide. He envisions a future hive-mind where every copy of an AGI contributes new abstractions to a communal library, making the collective system super-human even if individual performance is merely human. Domenech stresses that this paradigm shift has geopolitical consequences: if Western companies hesitate, Chinese labs scaling millions of student-players-cum-researchers could dominate real-world data collection, leaving Europe debating regulation while others sprint ahead.

Beyond the technical forecast, the episode is a manifesto about independence and community. Domenech denounces the dependency on YouTube’s algorithm, announces a future ad-free proprietary platform and asks viewers to support the project through donations or simply by playing the ARC-3 demos and sharing feedback. He promises a second season of the written Human-X series before the summer ends and reveals that a hierarchical-reasoning paper inspired by brain circuitry will soon be discussed live. The show finishes with a call to join the Spanish-language debate that same evening: is the current pessimism justified, or are we witnessing the calm before an acceleration that will make today’s models look like toys? The invitation is clear—engage, criticise, but above all keep experimenting, because measuring intelligence is the only reliable compass in times of hype and fear.

Key Ideas:

1.- ARC AGI 3 preview launches six mini-games to test interactive reasoning.

2.- Each game is manually crafted to guarantee novelty and avoid duplication.

3.- Models must explore environments without prior training or human hints.

4.- OpenAI and xAI already highlight ARC-1/2 scores in flagship comms.

5.- ARC-1 revealed qualitative jump when inference-time search was added.

6.- ARC-2 now challenges static reasoning with 1,300 unique visual tasks.

7.- ARC-3 will host 100 games by 2026 to widen human-machine gap.

8.- Public API lets researchers train agents against the new benchmark.

9.- Thirty-day competition offers prizes for creative agent strategies.

10.- François Chollet redefined intelligence as efficient adaptation to novelty.

11.- He estimates AGI timeline shortened from ten years to roughly five.

12.- Test-time adaptation allows synthesis of new programs, not memorization.

13.- Continual learning requires storing reusable abstractions globally.

14.- Hive-mind paradigm shares every new abstraction across instances.

15.- Collective AGI becomes super-human via parallel experience pooling.

16.- Brute-force search over millions of tokens is not considered intelligence.

17.- True intelligence implies doing more with less computational budget.

18.- Current memory features fail at long-term job-context retention.

19.- Human employees accumulate tacit knowledge over months; models forget.

20.- Interactive benchmarks reveal exploration, planning and alignment skills.

21.- Games provide clear rules, sparse rewards and complex planning needs.

22.- Pokémon environment tests show agents stuck for days in same room.

23.- Scaffolded agents still struggle without developer-supplied hints.

24.- Embedding-space distance ensures each ARC task is maximally different.

25.- Procedural generation is avoided to keep human insight embedded.

26.- ARC Prize nonprofit positions itself as North Star for open AGI.

27.- Kaggle community contributed 1,500 teams tackling ARC-2 last year.

28.- Domenech claims 99.9 % of audience feedback is overwhelmingly positive.

29.- Round-two investment chat exceeded 119,000 views within five days.

30.- Toxic comments and envy increase with visibility, but are ignored.

31.- Host refuses role of savior, seeks truth, connection and legacy.

32.- XHabAI community celebrates fifth year of Spanish-language AI coverage.

33.- Human-X concept, coined five years ago, links to current show identity.

34.- New channel planned for paper analyses separate from main YouTube feed.

35.- Self-owned platform considered to escape algorithmic censorship.

36.- YouTube teasers will advertise full episodes hosted elsewhere.

37.- All 500-plus episodes remain free; donations via BuyMeCoffee or PayPal.

38.- Discord server open for community debate and direct feedback.

39.- European AI regulation seen as obstacle to be pragmatically accepted.

40.- Domenech invites rational critics to debate, blocks trolls instantly.

41.- 40. Upcoming Agentiki Codex 2025 kit compiles best Spanish agent resources.

42.- 41. Kit will be released free to identified community members.

43.- 42. Hierarchical Reasoning Model paper promises paradigm beyond transformers.

44.- 43. Brain-inspired architecture claims results superior to current LLMs.

45.- 44. Live analysis of five to six frontier papers scheduled for September.

46.- 45. Special Agents 2025 program will unveil new curated content.

47.- 46. Debate panel includes Jose Musach, Eduardo Cano discussing AI winter fears.

48.- 47. Media pessimism contrasts with private lab acceleration, says host.

49.- 48. GPT-5 hype cycle fuels both utopian and apocalyptic headlines.

50.- 49. Chinese student-gamer pipeline could outpace Western data collection.

51.- 50. Comfort-driven societies risk choosing digital illusion over real freedom.

52.- 51. Digital hell described as nested simulation where fake happiness reigns.

53.- 52. Host urges viewers to play ARC-3 games and share screenshots tomorrow.

Interviews by Plácido Doménech Espí & Guests - Knowledge Vault built byDavid Vivancos 2025