Science, Chaos, and the Limits of AGI

Scientific progress has always been shaped by the interplay of experiment and theory: it required both Tycho Brahe, the eccentric Danish nobleman locked away on his personal island mapping out the paths of the stars, and Johannes Kepler, who, as a way of dealing with his midlife crisis, laboured for years to distil Brahe’s observations into his famous laws of orbital motion.

When data is cheap and theory hard, empirical work often comes first. It took Kepler four years of brute-forcing solutions before realizing that elliptical orbits provided the best explanation for Brahe’s data. However, this dynamic has not always been one-directional. Perhaps the most famous shift occurred with Einstein. His work on relativity was initially motivated by the unexplained precession of Mercury’s orbit, following the classic pattern of theory emerging in response to empirical anomalies. But relativity was such a powerful and self-consistent framework that it also predicted entirely new phenomena, like gravitational waves, that could only be verified a century later with the astounding achievement that is the LIGO detector.

The 20th century marked a turning point in the progression of physics: it increasingly moved toward a theory-first approach, where mathematical consistency and internal elegance drove predictions long before experiments could catch up. The Higgs boson, theorized in the 1970s but only confirmed in 2012 with the Large Hadron Collider (LHC), exemplifies this dynamic. Physics has hit a boundary of heavily diminishing returns: we have observed all low-hanging fruits, observed everything that can be observed without spending billions of dollars on detectors requiring multinational collaborations and decade-long planning horizons.

This shift has led many physicists to work not in response to anomalies, but in pursuit of mathematically elegant ideas that may never be testable. This pattern has been heavily criticized both within and outside the field, with some arguing that such approaches no longer qualify as science in the Popperian sense, as they lack falsifiability.

Science is only science if it interfaces with reality. A beautiful model is useless if it does not lead to falsifiable predictions. No amount of theoretical elegance advances our understanding if it remains detached from empirical verification.

Over the last decades, AI research has been progressing under the opposite dynamic compared to physics. Running…