Agents Writing Scientific Workflows

In this episode of the Crazy Wisdom podcast, host Stewart Alsop sits down with Markus Buehler, the McAfee Professor of Engineering at MIT, to explore how seemingly different systems—from proteins and music to knowledge structures and AI reasoning—share underlying patterns through hierarchy, self-organization, and scale-free networks. The conversation ranges from the limits of current AI interpolation versus true discovery (using the fire-to-fusion example), to the emergence of agent swarms and their non-linear effects, to practical questions about ontologies, knowledge graphs, and whether humans will remain necessary in the creative discovery process. Markus discusses his lab's work automating scientific discovery through AI agents that can generate hypotheses, run simulations, and even retrain themselves, while Stewart shares his own experiences building applications with AI coding agents and grapples with questions about intellectual property, material science constraints, and the future of human creativity in an AI-abundant world.

Timestamps

00:00 - Introduction to Marcus Buehler's work on knowledge graphs, structural grammar across proteins, music, and AI reasoning

05:00 - Discussion of AI discovery versus interpolation, using fire and fusion as examples of fundamental versus incremental innovation

10:00 - Language models as connective glue between agents, enabling communication despite imperfect outputs and canonical averaging

15:00 - Embodiment and agency in AI systems, creating adversarial agents that challenge theories and expand world models

20:00 - Emergent properties in materials and AI, comparing dislocations in metals to behaviors in agent swarms

25:00 - Human role-playing and phase separation in society, parallels to composite materials and heterogeneity

30:00 - Physical world challenges, atom-by-atom manufacturing at MIT.nano, limitations of lithography machines

35:00 - Synthetic biology as alternative to nanotechnology, programming microorganisms for materials discovery

40:00 - Intellectual property debates, commodification of AI models, control layers more valuable than model architecture

45:00 - Automation of ontologies, agent self-testing, daughter's coding success at age 11

50:00 - Graph theory for knowledge compression, neurosymbolic approaches combining symbolic and neural methods

55:00 - Nonlinear acceleration in AI, emergence from accumulated innovations, restaurant owner embracing AI

01:00:00 - Future generations possibly rejecting AI, democratization of knowledge, social media as real-time scientific discourse

Key Insights

1. Universal Patterns Across Disciplines: Seemingly different systems in nature—proteins, music, social networks, and knowledge itself—share fundamental structural patterns including hierarchy, self-organization, and scale-free networks. This commonality allows creative thinkers to draw insights across disciplines, applying principles from one domain to solve problems in another. As an engineer and materials scientist, Buehler has leveraged these isomorphisms to advance scientific understanding by mapping the "plumbing" of different systems onto each other, revealing hidden relationships that enable extrapolation beyond what's observable in any single domain.
2. The Discovery Versus Interpolation Problem: Current AI systems, particularly large language models, excel at interpolation—recombining existing knowledge in new ways—but struggle with genuine discovery that requires fundamental rewiring of world models. Using the example of fire versus fusion, Buehler explains that an AI trained on combustion chemistry would propose bigger fires or new fuels, but couldn't conceive of fusion because that requires stepping back to more fundamental physics. True discovery demands the ability to recognize when existing theories have boundaries and to develop entirely new frameworks, something current AI architectures aren't designed to achieve due to their training objective of predicting the most likely outcome.
3. The Role of Ontologies and Knowledge Graphs: While some AI researchers argue that ontologies are unnecessary because models form internal representations, Buehler advocates for explicit knowledge graphs as essential discovery tools. External ontologies provide sharp, analytical, symbolic representations that complement the fuzzy internal representations of neural networks. They enable verification of rare connections—like obscure papers that might hold key insights—which would be averaged away in standard AI training. This neurosymbolic approach combines the generalization capabilities of neural networks with the precision of formal knowledge structures, creating more powerful discovery systems.
4. Emergent Properties and Agent Swarms: Just as materials science shows that collections of atoms exhibit properties impossible to predict from individual components, AI agent swarms demonstrate emergent behaviors beyond single models. When agents are incentivized not just to answer questions but to challenge each other adversarially, propose theories, and test hypotheses, they can spawn new copies of themselves and evolve understanding beyond their initial programming. This emergence isn't surprising from a materials science perspective—dislocations, grain boundaries, and other collective phenomena only appear at scale, fundamentally determining material behavior in ways unpredictable from studying just a few atoms.
5. The Commoditization of Intelligence: The fundamental AI models themselves are becoming commodities, as evidenced by events like the Moldbug phenomenon where people built agents using various providers interchangeably. The real value is shifting from who has the smartest model to how models are orchestrated, integrated, and deployed. This parallels historical technology adoption patterns—just as we moved past debating who makes the best electricity to focusing on applications, AI is transitioning from a horse race over model capabilities to questions of infrastructure, energy, access speed, and agent coordination at the systems level.
6. Human-AI Collaboration and Creative Control: Rather than wholesale replacement, AI enables humans to operate in an intensely creative space as orchestrators sampling from vast possibility spaces. Similar to how Buehler's 11-year-old daughter now builds sophisticated applications that would have required professional developers years ago, AI democratizes access to capabilities while humans retain the creative judgment about direction and meaning. The human role becomes curating emergence, finding rare connections, playing at the edges of knowledge, and exercising the kind of curiosity-driven exploration that AI systems lack without embodied stakes in their own survival and continuation.
7. Technology as Evolutionary Inevitability: The development of AI represents not an unnatural threat but the next stage of human evolution—an extension of our innate drive to build models of ourselves and our world. From cave paintings to partial differential equations to artificial intelligence, humans continuously create increasingly sophisticated representations and tools. Attempting to stop this technological evolution is futile; instead, the focus should be on steering it ...