Per Bak

In the AI Story

Bak's signature contribution was the sandpile model — a thought experiment in which grains of sand dropped one at a time onto a flat surface eventually reach a critical angle where the next grain might trigger anything from a single-grain shift to a system-wide avalanche. This simple model revealed a profound principle: complex systems don't need external tuning to reach criticality. They drive themselves there through their own local dynamics, grain by grain, until the pile is poised at the precise angle where small perturbations can propagate across the entire system. The mathematical signatures of this critical state — power-law distributions, long-range correlations, sensitivity to perturbation — appeared not just in sandpiles but in earthquakes, forest fires, mass extinctions, financial markets, and evolutionary dynamics.

Bak's career was marked by intellectual combativeness and unwavering conviction. He argued that self-organized criticality was not merely a property of sandpiles but a fundamental organizing principle of complex systems generally. His 1996 book How Nature Works: The Science of Self-Organized Criticality presented the framework to general audiences, making claims about the universality of power laws that irritated peers who found his extrapolations excessive. Yet Nobel laureate Philip W. Anderson defended Bak's work as having 'paradigmatic value, as the kind of generalization which will characterize the next stage of physics.' The trajectory of research since Bak's death has supported Anderson's assessment: the framework has proven increasingly applicable to domains Bak never studied.

The vindication arrived most dramatically in artificial intelligence research. In 2021, researchers demonstrated analytically that learning dynamics of neural networks are 'generically attracted towards a self-organized critical state.' In 2024, studies showed optimal deep neural network performance occurs precisely at the transition point separating stable and chaotic attractors. In March 2026, Burc Gokden demonstrated that large language models trained at self-organized criticality exhibit reasoning at inference time, while those not at criticality do not. The mechanism governing sandpiles was governing whether an AI could think — a convergence Bak would have found unsurprising but which he never lived to witness.

Bak's framework provides the most rigorous lens available for understanding the AI transition not as a product event but as a criticality event. The December 2025 threshold, the trillion-dollar SaaS Death Cross, the overnight dissolution of professional certainties — these are avalanches in a self-organized critical system, governed by power laws that make specific prediction impossible while making the class of events statistically inevitable. The grain that triggered each avalanche was unremarkable. What mattered was the global state of the pile at the moment the grain landed. Bak's physics explains why the triumphalists and elegists are both partially right and why the silent middle — those holding exhilaration and loss simultaneously — are experiencing the most accurate description of a system that is, at the critical point, genuinely contradictory.

Origin

Per Bak was born in 1948 in Denmark and trained in physics during the 1960s and 70s, a period when condensed matter physics was producing profound insights into collective phenomena. He spent formative years studying phase transitions, magnetism, and the emergence of order in complex materials. This background in many-body physics gave him the mathematical tools to recognize, when he encountered it, that certain systems organize themselves without external tuning — a radical departure from the classical assumption that order requires a designer or controller.

The breakthrough came in 1987 while Bak was at Brookhaven National Laboratory, collaborating with Chao Tang and Kurt Wiesenfeld. Their Physical Review Letters paper 'Self-Organized Criticality: An Explanation of 1/f Noise' introduced the sandpile model and proposed that a wide class of natural phenomena — previously thought to require separate explanations — could be understood as manifestations of a single universal mechanism. The paper was initially met with skepticism. Bak spent the remaining fifteen years of his life developing the framework, applying it to evolutionary biology, economics, and neuroscience, and arguing with increasing urgency that self-organized criticality was as fundamental to complexity science as thermodynamics was to physics. He died in Copenhagen in 2002, having secured a permanent place in the physics of complexity but still waiting for the broader scientific community to embrace the universality of his claims.

Key Ideas

Self-organized criticality as universal principle. Complex systems drive themselves to critical states through local dynamics alone, no external tuning required — a claim vindicated across domains from neural networks to evolutionary biology.

Power laws as signatures. The fingerprint of a critical system is a power-law distribution of event sizes, where frequency decreases as magnitude increases but extreme events remain statistically significant rather than exponentially suppressed.

Impossibility of specific prediction. At criticality, the timing and magnitude of the next avalanche are fundamentally unknowable, not due to insufficient data but due to the system's dynamics — forecasting must be replaced by structural resilience.

Sandpile as paradigm. The dropping of grains onto a pile until it reaches the critical angle is not merely an illustration but a minimal model capturing the essential dynamics of systems ranging from earthquakes to AI-driven professional disruption.

Posthumous vindication. Research in the 2020s confirmed Bak's most ambitious predictions: neural networks self-organize to criticality, optimal AI reasoning occurs at the critical point, and the mathematics of sandpiles govern whether machines can think.