Paradigms of Intelligence (Pi) is a research team at Google led by Blaise Agüera y Arcas (VP/Fellow and CTO of Technology & Society). The team is verified on GitHub with a google.com domain and currently has 320 followers. Their stated mission is to “advance our understanding of how intelligence evolves to develop new technologies for the benefit of humanity and other sentient life.”
Pi represents a distinctive intellectual position in AI research—one that bridges computer science, evolutionary biology, physics, neuroscience, and philosophy. Their work is organized around five interconnected paradigm shifts:
The radical premise that computing existed in nature long before artificial computers. Drawing on John von Neumann’s insight that DNA functions like a Turing tape, and Wheeler’s “it from bit” hypothesis, Pi argues that life itself is inherently computational. Their landmark “Computational Substrates” paper demonstrated self-replicating programs emerging spontaneously from “primordial soups” of random code —a digital analog to abiogenesis.
Current AI infrastructure remains hobbled by classical sequential computing paradigms. Pi advocates for truly brain-like architectures with massively parallel processing and data locality , arguing that GPUs/TPUs are only stepping stones toward genuinely neural silicon.
The success of LLMs reveals something fundamental: intelligence is statistical modeling of the future (including one’s own future actions) given a growing body of knowledge, observations, and feedback from the past . This aligns with the “predictive brain hypothesis” in neuroscience.
Pi takes the controversial position that “Artificial General Intelligence” (AGI) may already be here — we just keep shifting the goalposts . They argue that individual humans are now less general than AI models in terms of breadth of capability.
Intelligence is fundamentally social and modular—from cortical columns to human societies. This frames AI development as potentially multi-agent and socially organized rather than monolithic.
Their cubff repository (159 stars) implements the “Computational Life” experiments showing how self-replicators emerge from random Brainfuck code without any explicit fitness function—just dynamic stability through self-modification . This work has been published in Physical Review and covered widely (New Scientist, Popular Science, Sean Carroll’s Mindscape podcast).
A “moonshot” initiative to scale machine-learning compute in space using solar-powered satellite constellations carrying Google TPUs . Two prototype satellites are planned for launch with Planet Labs in early 2027. The rationale: space offers up to 8x more productive solar energy and eliminates terrestrial resource constraints.
Papers on “learning-aware policy gradients” and “Embedded Universal Predictive Intelligence” explore how cooperation emerges when agents learn to model each other—connecting evolutionary symbiosis to AI alignment.
The team includes philosophers (Geoff Keeling, Winnie Street) working on questions like AI moral patiency, the attribution of confidence to LLMs, and whether AI can make trade-offs involving “stipulated pain and pleasure states.”
Blaise Agüera y Arcas’s 600-page book, published by MIT Press/Antikythera, serves as the manifesto for Pi’s worldview. It argues that prediction is fundamental not only to intelligence and the brain but to life itself . The book was included in the Financial Times’s Best Books of 2025 and praised by figures including Patricia Churchland, Terrence Sejnowski, and David Krakauer (President of Santa Fe Institute).
Pi occupies a unique intellectual niche compared to other major AI labs:
| Aspect | Pi’s Position | Contrast with Others |
|---|---|---|
| Fundamental orientation | Bottom-up, evolutionary | OpenAI/Anthropic: Engineering-focused scaling |
| View of AGI | Already here, continuously emerging | Others: Future threshold to cross |
| Key metaphor | Life/symbiosis/ecology | Others: Engineering/optimization |
| Safety framing | Through understanding collective intelligence | Others: Alignment/control problems |
| Disciplinary roots | Physics, biology, philosophy | Others: Primarily ML/engineering |
| Infrastructure vision | Space-based, neural computing | Others: Terrestrial scaling |
Pi represents a “naturalist” school of AI thinking—viewing artificial intelligence not as something fundamentally separate from biological intelligence, but as a continuation of the same computational processes that gave rise to life itself. This contrasts with both the “AI as engineering artifact” view and the “AI as existential risk” framing.
The team is notably interdisciplinary, including:
- ML researchers: João Sacramento, Alexander Meulemans, Seijin Kobayashi, Nino Scherrer
- Philosophers/ethicists: Geoff Keeling, Winnie Street
- Physicists/complexity scientists: Thomas Fischbacher, Rif A. Saurous
- Visual/generative researchers: Alexander Mordvintsev (of Neural CA fame), Eyvind Niklasson
- Infrastructure: Travis Beals (leading Project Suncatcher)
Key collaborators include researchers at Mila Montreal, ETH Zurich, and University of Chicago.
Pi represents perhaps the most philosophically ambitious AI research program inside a major tech company—one that treats the question “What is intelligence?” not as a branding exercise but as a genuine scientific and metaphysical inquiry requiring interdisciplinary synthesis. Their work suggests that understanding AI requires understanding life, evolution, and consciousness as interconnected computational phenomena.
summary and analysis of “Paradigms of Intelligence” based on the public GitHub organization, associated research publications, and wider internet context — giving you a clear picture of what it is, who’s behind it, what they’re researching, and where it sits in today’s AI landscape.
⸻
Paradigms of Intelligence (Pi) is a research team at Google focused on foundational, long-term, interdisciplinary research into the nature and evolution of intelligence — both natural and artificial. Their work is not a typical product unit, but a research organization aiming to rethink core assumptions in AI and cognition. 
The group: - • Is led by Blaise Agüera y Arcas, a well-known Google VP of Research with a long history of foundational work in machine learning, perception, and computational theory.  - • Combines philosophers, engineers, and scientists to explore what intelligence truly is, how it emerges, and how it might be built fundamentally differently from mainstream deep learning approaches.  - • Publishes across disciplines — from Nature and Communications of the ACM to ArXiv and physical sciences journals — signaling broad, collaborative scholarship. 
Mission statement: advance understanding of how intelligence evolves to benefit humanity and sentient life. 
⸻
The team’s research agenda is structured around several interrelated paradigms — each challenging dominant AI assumptions:
📌 1. Natural Computing - • Intelligence and complex behavior aren’t exclusive to silicon computers — they precede them in biology. - • The goal is a theoretical framework linking evolution, complexity, and computation. 
🧠 2. Neural Computing - • Argues that current AI processors (GPUs, TPUs) are fundamentally constrained by classical computing paradigms. - • Emphasizes massively parallel, brain-like architectures with data locality as a new substrate for future systems. 
🌐 3. Predictive Intelligence - • Intelligence is described as ongoing statistical modeling of the future based on past observations. - • Blurs traditional boundaries between training and deployment, hinting at continuously learning and interactive systems. 
👥 4. Collective Intelligence - • Intelligence is not just an isolated system — social and multi-agent dynamics may be core to scaling cognition. - • Suggests research in cooperative behaviors, modularity, and mutual modeling as paths to more capable intelligence. 
⸻
The Pi team has a significant scholarly footprint across venues: - • Peer-reviewed journals in AI, ethics, philosophy, and physical sciences. - • Preprints on ArXiv that explore learning structures, representation geometry, multi-agent systems, and cellular automata. - • Interdisciplinary pieces bridging biology, computation, and philosophical foundations of cognition. 
Some representative research topics include: - • Geometry of representations in neural nets - • Mesa optimization in Transformers - • Predictive multi-agent learning frameworks - • Broader ethical questions about AI agents - • Conceptual intersections between chemical kinetics and computation 
Their publication list reveals a broad research scope — from core computation and learning theory to ethical and societal implications.
⸻
Pi’s perspective on intelligence is distinct from many mainstream AI labs:
⚙️ Bottom-up and interdisciplinary: Intelligence is seen as a natural phenomenon evolved over billions of years — not a narrow engineering artifact.  🧬 Biology-inspired: Natural systems are viewed as powerful analogues — understanding life as computation guides new frameworks.  🔄 Continuous learning over static training: Intelligence should grow and adapt continuously, not just through discrete training runs.  🤝 Collective & social scaling: Cognition is modular and social, not solitary. 
This contrasts with the industry’s current scaling-centric view, which focuses heavily on larger models and more data. Pi’s focus is on principle, architecture, and evolution rather than scale alone.
⸻
While Pi is Google-sponsored and uniquely positioned to pursue long-term foundational questions, its research aligns with broader shifts in AI research: - • Other communities are exploring alternative paradigms (e.g., neuro-symbolic blends, algorithm discovery, agent-based learning).  - • Philosophers and cognitive scientists are integrating findings from AI to rethink cognition and intelligence. - • A growing emphasis on ethics, alignment, and societal impacts — topics reflected in Pi’s work — is influencing public and scholarly discourse.
In short, Pi’s work is part of a larger move away from “more parameters = more intelligence” toward richer architectural understanding and biologically grounded cognition.
⸻
Paradigms of Intelligence is:
✅ A Google research team focused on deep, theoretical, and interdisciplinary AI research.  ✅ Led by respected researchers like Blaise Agüera y Arcas with an unusual blend of science, philosophy, and engineering.  ✅ Organized around four key paradigms that challenge mainstream AI assumptions.  ✅ Producing academic and technical publications across venues and disciplines.  ✅ Positioned within a broader context of researchers seeking new foundations for intelligence beyond large-scale deep learning. 
⸻