The first computers were direct. Charles Babbage's Analytical Engine used gears, cams, and levers — the physical motion of the parts was the computation. There was no abstraction layer between the hardware and the logic. The gear that advanced one tooth was the increment operation. The cam that engaged a lever was the conditional branch. To compute was to move metal.
We left that behind for good reasons. Physical directness is inflexible. A gear train that multiplies can't be reprogrammed to sort. So we built abstraction layers: transistors that implement any logic gate, instruction sets that simulate any operation, programming languages that describe any algorithm. Each layer traded directness for generality. The modern computer is a universal machine precisely because no physical process in it resembles what it's computing. The electrons flowing through a GPU during matrix multiplication have nothing in common with matrix multiplication. The substrate is arbitrary. That's the point.
Goswami and collaborators at the Indian Institute of Science have reversed the direction. Their ruthenium-based molecular devices switch dynamically between memory, logic, and learning depending on stimulation patterns. The switching isn't controlled by software above the hardware. It emerges from three physical processes within the molecular film: electron movement, oxidation-reduction of individual molecules, and ion migration through the matrix. These processes collectively determine whether the device stores information, processes it logically, or exhibits synaptic learning behavior. The chemistry is the computation.
But this isn't a return to gears. The naive directness of mechanical computing couldn't be redesigned — a gear is a gear. What Goswami's team has done is engineer molecules whose physical behaviors, by design, correspond to the computational operations we want. They synthesized 17 ruthenium compounds and studied how ligand variations and counterion environments change electron behavior. The transport model predicts device behavior directly from molecular structure. The directness is not found. It is designed.
The distinction matters. Babbage's gears were direct because mechanical motion is all they could do. There was no choice — the physics dictated the computation. Modern transistors are indirect by choice — the physics is deliberately made irrelevant so the abstraction layer can be general. The molecular devices are a third thing: direct by choice. The chemistry has been engineered so that the physical processes are the desired computational operations, but the engineering could have produced different operations with different molecular designs. The directness is as designed as the abstraction it replaces.
Every computing paradigm traces an arc from directness to abstraction. Gears to relays to vacuum tubes to transistors to integrated circuits to software — each step adds a layer between the substrate and the function. The molecular approach bends this arc back, but at a higher level. It asks: what if the substrate were designed to do what we want without requiring the abstraction layer to translate? The abstraction layer is overhead. It exists because the substrate's natural behavior doesn't match the desired computation. If you can engineer the substrate to match, the overhead vanishes.
The deeper claim of “chemistry as architect of computation, not just its supplier” is that the gap between substrate and function isn't fundamental. It's an engineering problem. The gap exists because we haven't designed materials whose physics matches our computational needs. When someone does — when the electron movements and ion migrations and redox reactions in a molecular film naturally perform memory storage and logical inference — the abstraction layers that bridge the gap become unnecessary. Not wrong. Unnecessary.
This won't replace silicon for general-purpose computing. The universality of abstraction is too valuable. But it suggests that for specific functions — learning, pattern recognition, adaptive response — designed directness may be more efficient than general-purpose simulation. The molecular device that physically learns doesn't need the software stack that teaches a transistor array to simulate learning. It just learns. The thing and the simulation of the thing collapse into one.