Building To The Future
I recently gave a talk about quantum computing at a local library ("Quantum Computing for The Curious Beginner") and one thing that stood out is how big the gap still is between what we'll want to do once we have quantum computers, and how much more serious engineering work needs to happen before we get there.
For example, Peter Shor developed his famous algorithm for finding the prime factors of an integer in 1994. Only in 2024, 30 years later, was it used to factor a not-ridiculously-small number (1,031,167) on a real quantum computer. Likewise, algorithms for molecular simulation, which are important in drug discovery and materials research, are still ahead of the available hardware—though that particular gap is closing fast.
It's fun to build to a future and dream about the cool things we'll be able to do once technology in the real world catches up. But there's a danger of losing sight of two things:
What can already be done, right now, with some clever engineering
Whether the imagined future will ever come to pass
Number 1 is a question of opportunity cost. Time spent dreaming up what you could do with a multi-million-qubit quantum computer is time not spent figuring out what to do with the devices we have available today.
Number 2 is even more dangerous. What if we never get to a quantum computer of that size? The overwhelming consensus is that, at this point, it's down to engineering, not fundamental physics, but who knows?
A version of that has been going on with all things AI. Futurists have been dreaming about self-improving hyperintelligent AI that will usher in a utopia by giving us answers to everything, including how to stop and reverse aging, or upload our brains into the matrix.
But even at a saner level, it is becoming clear that executives were projecting too much of their dreams (of replacing all customer support work with bots, for example) into the future where AI would actually be capable of that.
Given the speed at which the field evolves, it makes sense to build for next year's AI. But maybe not for next decade's AI.
