An easy prediction: the largest impact of AI will be in science and engineering. That’s because, one, those are areas where more and more diverse cognitive capabilities most easily translate to qualitative progress, and two, simply because scientific and technological progress are still the “one weird trick” that made possible the unprecedented increases in quality of life in the last hundred and fifty years or so.
To get an idea of what’s coming up, then, the best place to go is to the researchers who are actually working at the frontiers between new basic research and new applied research, specially at the (relative) beginning of their careers, where the wildest ideas — the later “entrepreneurial innovation” — are often born.
All of the above is less interesting that the paper I wanted to point you to, Ella M. King’s PhD thesis Frankenstein’s Tiniest Monsters: Inverse Design of Bio-inspired Function in Self-Assembling Materials. The extract (you can skip, but these are terms you are going to want to start gaining familiarity with):
Despite tremendous advances in synthetic materials design, the complexity achievable in artificial systems is dwarfed by the complexity of living matter. One cause of this discrepancy is that biological systems fundamentally rely on precise control over not just structure, but also function, in micron-scale components. Examples range from kinetic proofreading in DNA to regulation of clathrin formation and on-command microtubule disassembly. Achieving comparable dynamic and non-equilibrium functional control in synthetic materials remains an outstanding challenge. Because biological systems that control these non-equilibrium functionalities exist, it must be possible to design synthetic materials with similarly rich and complex functions. However, the design space of out-of-equilibrium functionalities is vast and hard to explore. How do we design complex functional materials without the luxury of billions of years of evolution? Here, we leverage automatic differentiation, the tool underlying much of the dramatic success in machine learning and non-convex optimization, to develop methods for computational materials design, and demonstrate quantitative control over non-equilibrium functionality in self-assembled materials. We couple this computationally-driven approach with a parallel effort to extract more information from experimental data, towards the goal of making our designs experimentally realizable. We develop a novel algorithm for particle tracking in systems with highly correlated motion and introduce a method for inferring interaction potentials from stochastic trajectory data.
At a higher level: Here’s an early example of how to use automatic differentiation to begin to design synthetic materials that have the sort of functional complexity that so far we only see in biological materials.
At an even higher level: The stuff we make our stuff out of is on its way to being transformatively more capable in ways no artificial material has been close to yet.
Now that is innovation.