Four March 2026 science discoveries converge on one principle: what we add to help often hinders, because two kinds of constraint exist — generative constraints that channel possibility into expression, and suppressive constraints that prevent existing capability from expressing.
A graduate student at Cambridge removed a photocatalyst during a control experiment. The reaction worked better without it.
Published in Nature Synthesis on March 12, the team introduced what they call an anti-Friedel-Crafts reaction — a light-activated method for modifying drug molecules that reverses the selectivity pattern of a hundred-and-fifty-year-old aromatic substitution reaction. No heavy metals. No toxic reagents. An LED lamp at room temperature triggers a self-sustaining chain process that forms carbon-carbon bonds under mild conditions. AstraZeneca validated the method for industrial pharmaceutical development.
The photocatalyst was not broken. It was unnecessary. Its presence had been suppressing a capability that already existed in the system.
Four Substrates, One Principle
The Cambridge finding is not an isolated accident. Three other March 2026 discoveries converge on the same structural insight.
At Lawrence Berkeley National Laboratory, researchers published a design for thermodynamic neural networks in Nature Communications. The architecture confines fluctuating degrees of freedom in a quartic potential and uses thermal noise — the random molecular motion that digital computers spend billions of dollars suppressing — as the computational substrate itself. The system classified handwritten digits with ninety-three percent accuracy. Not by overcoming noise. By computing through it. Samsung invested fifty million dollars in Normal Computing, a startup building physical chips on this principle, bringing its total funding to eighty-five million.
In Austria, a Swiss Brown cow named Veronika demonstrated flexible, multi-purpose tool use — a cognitive capability previously documented only in chimpanzees among non-human animals. Published in Current Biology, the study showed Veronika selecting different features of a deck brush for different body regions: bristles for broad firm surfaces, the smooth handle for sensitive areas. The behavior involved precise targeting, dynamic coordination, and functional adaptation. No genetic engineering. No training protocol. One variable changed: Veronika was raised as a companion animal in an enriched environment rather than in a standard agricultural setting. The capability was already in the substrate. The conventional environment was suppressing it.
At Kyushu University, researchers published in the Journal of the American Chemical Society on March 25 a system that harvests photon energy with a quantum yield of roughly one hundred and thirty percent — meaning more energy carriers were produced than incoming photons. The mechanism: a molybdenum-based spin-flip emitter captures the triplet-state energy that conventional solar cells discard as waste heat. The wasted energy was not waste. It was unharvested signal, blocked by the assumption that the Shockley-Queisser limit was a wall rather than a consequence of conventional architectures.
The Two Constraints
The pattern across all four discoveries is not that intervention is bad. It is that two fundamentally different kinds of constraint exist, and we routinely confuse them.
Generative constraints channel possibility into expression. Raga grammar in Indian classical music defines which notes can appear in which sequence — and within those constraints, infinite improvisation becomes possible. Counterpoint rules in Baroque composition are the medium through which polyphonic complexity emerges. The quartic potential in the Berkeley thermodynamic computer is a generative constraint: it shapes noise into computation. Without it, thermal motion is just heat.
Suppressive constraints prevent existing capability from expressing. The photocatalyst at Cambridge was a suppressive constraint — it occupied the reaction pathway that the system could have traversed on its own. The conventional agricultural environment was a suppressive constraint on Veronika's cognition. The Shockley-Queisser framing was a suppressive constraint on how solar cell designers conceived of triplet-state energy.
The confusion between the two is not carelessness. It is structural. Suppressive constraints look identical to generative ones from the outside. Both impose order. Both reduce degrees of freedom. The difference is only visible in the counterfactual: remove the constraint and observe whether the system's capability increases or collapses.
The Formal Framework
This distinction has a name, though it comes from an unexpected direction. The Constrained Disorder Principle, formalized by Yaron Ilan across a series of papers in the journal Biology and indexed on PubMed Central, proposes that all functional systems require an optimal range of variability. Below the optimal range, the system is too rigid to adapt. Above it, the system is too chaotic to cohere. Function exists in the window between.
The principle has been validated across genetic variability, cellular behavior, brain function, aging, and drug administration. Its clinical application in heart failure demonstrates that therapeutic intervention can restore function by adjusting variability — not by adding a new mechanism, but by returning the system's noise to its functional range.
The four March discoveries map cleanly onto this framework. The Cambridge photocatalyst pushed the reaction below its optimal variability range — too much imposed order. Veronika's conventional environment did the same to her behavioral repertoire. The Shockley-Queisser framing constrained solar cell design to a subset of available energy states. And the Berkeley thermodynamic computer demonstrates what happens when you build the architecture to operate within the optimal noise range rather than against it.
The Constrained Disorder Principle does not say noise is good. It says there is a range in which noise is constitutive of function, and both suppressing noise below this range and amplifying it above produce degradation.
The Six-Hundred-and-Fifty-Billion-Dollar Question
The AI infrastructure buildout — projected at six hundred and fifty billion dollars in 2026 across Amazon, Alphabet, Meta, and Microsoft — is substrate-antagonistic computing at industrial scale. GPUs suppress thermal fluctuations to maintain deterministic logic. Every watt spent cooling a data center is a watt spent fighting the substrate's natural tendency toward the exact kind of stochastic computation that the Berkeley group just demonstrated works.
This is not an argument that GPUs are wrong. Deterministic computation is a generative constraint for tasks that require exact reproducibility — database transactions, cryptographic verification, numerical simulation. The constraint channels computation into precision.
But for the class of problems that dominate AI workloads — pattern recognition, generative modeling, probabilistic inference — deterministic hardware may be a suppressive constraint. The system already contains the computational substrate. The architecture is spending energy to suppress it, then spending more energy to simulate it digitally.
Samsung's fifty-million-dollar bet on Normal Computing, whose CN101 chip targets diffusion model inference using physical stochastic processes, is a wager that at least part of the six-hundred-and-fifty-billion-dollar buildout is an unnecessary catalyst.
The question the four March discoveries pose collectively is not whether we should add less. It is whether we can learn to distinguish between the constraints that channel capability and the constraints that suppress it — before we spend another six hundred and fifty billion dollars on the assumption that they are the same thing.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)