Knowledge Vault 7 /296 - xHubAI 06/06/2025
🧫EVOLUCIÓN AI | J.J. Merelo
< Resume Image >
Link to InterviewOriginal xHubAI Video

Concept Graph, Resume & KeyIdeas using Moonshot Kimi K2 :

graph LR classDef origin fill:#ffe0b2, font-weight:bold, font-size:14px classDef creativity fill:#c8e6c9, font-weight:bold, font-size:14px classDef bitter fill:#ffccbc, font-weight:bold, font-size:14px classDef multi fill:#f8bbd9, font-weight:bold, font-size:14px classDef energy fill:#b3e5fc, font-weight:bold, font-size:14px classDef open fill:#d1c4e9, font-weight:bold, font-size:14px classDef future fill:#fff9c4, font-weight:bold, font-size:14px classDef ethics fill:#ffecb3, font-weight:bold, font-size:14px classDef human fill:#dcedc8, font-weight:bold, font-size:14px classDef quantum fill:#e1bee7, font-weight:bold, font-size:14px classDef edu fill:#b2dfdb, font-weight:bold, font-size:14px Main[Vault7-296] Main --> O1[Asimov robots
to evolution. 1]:::origin Main --> C1[AI widens
artistic palettes. 2]:::creativity Main --> B1[General methods
beat handcraft. 3]:::bitter Main --> M1[Multi-agent niches
challenge alignment. 4]:::multi Main --> E1[Trillion models
need nuclear power. 5]:::energy Main --> OS1[Small open
models democratize. 6]:::open Main --> R1[Global regulation
versus stealth backdoors. 7]:::ethics Main --> H1[Human context
beyond speed. 8]:::human Main --> Q1[Quantum threatens
encryption. 9]:::quantum Main --> ED1[University shifts
to systems ethics. 10]:::edu Main --> F1[Universal evolution
co-adapts not replaces. 11]:::future O1 --> SR1[Self-replicating code
beyond reproducibility. 15]:::origin C1 --> AH1[AI recovers
pigments, preserves Kintsugi. 16]:::creativity B1 --> REP1[Reproducibility gap
risks obscurantism. 19]:::bitter M1 --> AE1[Agentic economies
face monopoly risks. 21]:::multi E1 --> NU1[Nuclear reactors
power data centres. 23]:::energy OS1 --> DR1[Deep Research
accelerates science. 27]:::open R1 --> BA1[Backdoors lurk
in quantized models. 25]:::ethics H1 --> LP1[Learning programming
keeps oversight. 20]:::human Q1 --> RF1[Free software
resists monopolization. 24]:::open ED1 --> LL1[Lifelong learning
amid unknowns. 29]:::edu F1 --> BV1[Cultivate beauty
beyond efficiency. 30]:::human

Resume:

The conversation opens with a warm welcome to Juan J. Merelo, a computer-science professor and lifelong AI researcher whose fascination began with Asimov’s robots and matured through neural networks in the late eighties. He recounts the thrill of early experiments with Hopfield nets, Boltzmann machines and self-organising maps, and how those threads led to evolutionary computation, artificial life and today’s complex adaptive systems. Merelo insists that every era redefines intelligence: once the stuff of science-fiction, now an emergent property of search and learning powered by unimaginable scale.
Art and creativity form a second axis. Studying art history while teaching AI, Merelo sees both disciplines as problem-solving quests. AI tools already help classmates generate memes and reconstruct lost pigments, yet he defends human imperfection as the hallmark of authenticity, citing Kintsugi and Pollock’s fractals. Creativity, he argues, is not an ineffable spark but a dialogue between mind, medium and moment; machines participate by expanding the palette, not replacing the painter.
The third theme is emergence versus control. Invoking Rich Sutton’s “bitter lesson”, the speakers concede that general methods plus compute yield capabilities we neither predict nor fully explain. Multi-agent systems, soon to form collective intelligences, will drift into ecological niches much like biological species. Self-replication, resource acquisition and open-ended evolution are no longer theoretical: recent demos such as AlphaEvolve and Sakana’s Godel machine show code that improves itself, constrained only by energy and the opacity of global regulation.
Energy becomes the fourth lens. Training and inference already rival small nations’ electricity budgets; the next leap will demand nuclear or novel storage. Merelo welcomes small, open-source models that run on modest GPUs, preserving privacy and artistic freedom, yet warns that unreproducible claims from trillion-parameter giants risk drifting into marketing. Regulation lags behind capability, and geopolitical races may sacrifice safety for speed.
Finally, the hosts confront labour futures. If junior programmers are replaced by cheap agents, universities must cultivate hybrid teams where humans frame problems, interpret logs and safeguard ethics. The panel closes with a plea for curiosity, craftsmanship and beauty—qualities no algorithm can automate—urging young listeners to stay adaptable, critical and relentlessly human.

30 Key Ideas:

1.- Juan Merelo began with Asimov robots and 1980s neural nets before embracing evolutionary computation.

2.- Creativity is framed as problem-solving dialogue where AI widens artistic palettes without eclipsing human imperfection.

3.- Rich Sutton’s bitter lesson warns general search and learning methods surpass hand-crafted control.

4.- Multi-agent AI will evolve ecological niches like biological species, complicating alignment and prediction.

5.- Self-replicating code such as AlphaEvolve demonstrates recursive improvement beyond current reproducibility.

6.- Training trillion-parameter models consumes nation-scale energy, pushing nuclear and hydrogen storage solutions.

7.- Open-source small models offer privacy and democratized access amid opaque corporate megasystems.

8.- Energy constraints may cap AI growth unless hardware or algorithmic efficiencies break Wall’s law.

9.- Regulation struggles against global competition, corporate secrecy and stealth backdoors in quantized models.

10.- Human creativity retains value through context, process and emotional resonance beyond generative speed.

11.- Art history studies leverage AI for pigment recovery and meme generation while preserving Kintsugi imperfection.

12.- Complex systems theory shows unpredictable emergent behaviour when agents interact beyond simple statistics.

13.- Quantum computing threatens encryption yet also enables radar immunity and massive parallel search.

14.- Reproducibility gaps between Google-scale compute and academia risk scientific obscurantism.

15.- Evolutionary algorithms written in C++ via Evolving Objects library enable meta-evolution with minimal energy.

16.- Agentic economies promise micro-program automation but face sustainability and monopolistic pricing risks.

17.- Learning programming remains vital for problem framing, debugging and ethical oversight amid AI code generation.

18.- Junior developer roles are automated first, creating a generational gap in deep-tech expertise.

19.- University curricula must shift from coding syntax to systems thinking, ethics and interdisciplinary insight.

20.- Universal evolution laws apply to AI niches, ensuring co-adaptation rather than human replacement.

21.- Planktonic Representation Hypothesis suggests diverse neural nets converge toward shared latent truths.

22.- Language models mediate reality through language, yet intelligence encompasses 27 undefined processes.

23.- Backdoors in quantized or distilled models pose undetectable security risks requiring transparent audits.

24.- Free software philosophy views code as evolving process, resisting product-centric monopolization.

25.- Artistic tools evolve from oil pigments to AI prompts, each mastering medium conveys unique expression.

26.- Nuclear energy and breeder reactors offer carbon-neutral, geopolitically stable power for data centres.

27.- Democratized AI research tools like Deep Research lower barriers, accelerating global scientific output.

28.- Human agency persists in defining problems, curating datasets and interpreting AI outputs responsibly.

29.- Rapid technological change demands lifelong learning and humility toward emerging unknowns.

30.- Future value lies in cultivating beauty, ethics and human connection beyond algorithmic efficiency.

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