This episode deconstructs the universal hierarchy of life, revealing how abstract principles of evolution and complexity transcend biology and offer a new lens for evaluating the potential of artificial life and intelligence.
The Three Cultures of Science
- Variance Culture: Focuses on diversity, deviation, and detailed observation.
- Coarse-Grained Culture: Seeks abstract, universal principles and simple theories.
- Exactitude Culture: Leverages modern computation to build high-fidelity simulations and complex models, a capability unavailable to historical figures like Galileo.
Kempes, drawing on his experience at SFI, argues that true scientific progress requires integrating these three approaches. He highlights physics' historical success with the "magic loop"—a continuous cycle where observation informs theory, which in turn generates new, testable predictions. He proposes that the biosciences, and by extension the study of AI, must adopt a similar integrated model to move beyond mere data collection or simulation and toward genuine understanding.
What Makes a Good Scientific Theory?
- Kempes defines a powerful scientific theory by its compactness and ability to make "dangerous predictions." He emphasizes that true understanding is not just about modeling existing data but about predicting unseen phenomena.
- A good theory is a form of compression, allowing vast amounts of knowledge to be transmitted efficiently through simple equations or principles, even if decoding them requires extensive training.
- The ultimate test of a theory is its ability to make what he calls a "dangerous prediction"—a forecast about something that has not yet been observed. Kempes states, “There's no data for it yet... you're making a prediction that will either be right or wrong about something that hasn't been seen yet. But that's a real notion of understanding.”
- This perspective directly challenges purely descriptive or correlational models, suggesting that for an AI system to demonstrate true understanding, it must move beyond pattern matching to generate novel, falsifiable hypotheses about the world.
Toward a Universal Theory of Life
- The conversation shifts to the ambitious goal of creating a universal theory of life—one that applies not only to biology but also to other adaptive, information-processing systems.
- Kempes argues that concepts like human culture, language, and even in silico (computer-based) life should be considered forms of life operating on unconventional substrates.
- The goal is to identify the core principles governing any living system, including its origin, its growth in complexity, and its major evolutionary transitions.
- He notes that while a complete theory is still distant, researchers are making progress on key ingredients, such as quantitative theories of agency, semantic information, and individuality. This work aims to build a framework that can explain life regardless of its material form.
The Universal Hierarchy of Life: Materials, Constraints, and Optimization
Kempes introduces a fascinating three-level hierarchy for understanding all possible life, which separates the particular from the universal. This framework is critical for investors and researchers looking to evaluate the fundamental potential of AI systems beyond their immediate implementation.
- Materials: This is the lowest, most diverse level, representing the physical substrate of life (e.g., carbon-based biochemistry). Kempes argues that this level is highly contingent and likely varies dramatically across the universe. He is "completely uncommitted to the materials," suggesting that life could be built on entirely different chemistries.
- Constraints: This middle level collapses the diversity of materials into a more unified set of physical laws. Any form of life, regardless of its composition, must obey universal principles like gravity, thermodynamics, and diffusion. These constraints shape evolution in predictable ways, leading to convergent solutions across different lineages.
- Optimization: This is the highest, most abstract level, containing principles like evolution, learning, and adaptation. Kempes references the error threshold—a concept from evolutionary theory defining the maximum mutation rate a system can sustain before losing its information—as a universal principle that applies equally to genomes, language, and potentially AI models.
This hierarchy suggests that while an AI's "material" (code, silicon) is specific, its behavior is governed by the same universal constraints and optimization dynamics as biological life.
Convergence, Transfer, and Information Phylogeny
- Convergent Evolution: Kempes uses the eye as a prime example. Eyes evolved independently multiple times because the physics of focusing light acts as a powerful "target" or attractor for evolution. This implies that certain optimal solutions for intelligence or agency may be discovered repeatedly by different AI architectures.
- Information Transfer: Unlike the slow process of genetic transfer in biology, human culture excels at rapid, horizontal information transfer. Kempes notes, “We have a higher rate of transfer, but we also have really good procedures for deciding if we want to allow the transfer or not.”
- This raises a strategic question for AI: could an "information phylogeny" emerge that transcends its material substrate? While currently speculative, this idea suggests that successful AI architectures or algorithms could behave like a "memeplex," propagating across different hardware and software ecosystems.
Simulating Life and the Embodiment Debate
- Kempes addresses the question of whether life can be simulated in a computer. He argues that in principle, with sufficient complexity and constraints, it is possible. However, the debate is hampered by the lack of a formal theory of life.
- Without a clear, compact theory, we lack the objective criteria to determine if a simulation has crossed the threshold into "life." As Kempes puts it, we can easily verify if a simulated gas obeys the second law of thermodynamics, but we cannot do the same for a simulated organism and the laws of life.
- He also touches on the philosophical debate around embodiment and materialism, acknowledging that some, like philosopher John Searle, argue that physical instantiation is essential. Kempes remains focused on identifying the abstract principles first, believing those will clarify the embodiment question later.
Phase Transitions and "Promethean Moments" in Evolution
- Kempes explains that these transitions are not random but are often triggered when an organizational model hits a hard physical limit or "evolutionary wall." For example, a single-celled organism can only get so big before diffusion becomes too slow to support its metabolism.
- To overcome this wall, a new architecture is required—a phase transition. He notes that these transitions often coincide with massive environmental shifts, like the "Snowball Earth" periods, which may have created the selective pressure for multicellularity.
- For AI researchers, this provides a model for anticipating breakthroughs. Major advances may occur when current architectures (e.g., transformers) hit fundamental scaling limits, forcing the invention of entirely new organizational principles.
Assembly Theory: A New Metric for Complexity
- Finally, Kempes introduces Assembly Theory, a novel method developed with colleagues like Lee Cronin and Sarah Walker to quantify complexity objectively.
- Assembly Theory measures the complexity of an object by calculating the shortest possible pathway to build it through the recursive use of its component parts. It provides a lower bound on the complexity of a molecule or object, independent of its origin.
- This method avoids the biases of being "Earth-centric" by not relying on prior knowledge of biochemistry. It can distinguish a truly complex, likely biotic molecule from a large but simple one (like a C60 buckyball).
- Experimentally, this theory suggests a detectable threshold between abiotic and biotic matter. For AI and crypto, Assembly Theory offers a powerful, first-principles way to measure the complexity of a system—be it a neural network or a decentralized protocol—based on its structural dependencies rather than just its size or parameter count.
Conclusion
This episode reveals that the fundamental principles of life—adaptation, information processing, and evolution—are not exclusive to biology. By applying this universal hierarchy, investors and researchers can better evaluate the long-term potential of AI systems, identifying those with the architectural depth to overcome constraints and achieve true evolutionary leaps.
