I’m not a runner, but let’s start with running. The New York Times published a few days ago an article about wavelights the controversy around them.
It starts with pacers – sets of people who run in front of a runner at a fixed pre-arranged pace. This helps runners optimize their speed (too fast and you get tired before the finish line, too slow and… well, you’re too slow), which has been one of the factors helping break records. This was originally controversial. But it helps performance, and the social and business negotiations that are always behind the determination of what’s allowed and what’s not in any field converged towards acceptance.
Wavelights are the hardware-only version of this: LED lights along a running track programmed to move at a speed optimized to help break a record. In the simplest case, the lights move at the pace of the record you want to break, and the intended strategy is to keep up with them during most of the run and push forward near the last moment. It’s not the only speed profile that breaks a record, but if you can follow it it does break the record.
So how does this relate to AI?
Ask yourself this: how do wavelights or pacers make the runner faster? It’s not because it improves their physiology or biomechanics. Rather, they provide strategic guidance — how fast to run at any given time — that helps optimize results. That the strategy is fixed right now it’s an algorithmic detail; it’s easy to imagine variations where physiological sensors and runner-specific models adjust wavelight speed in real time to get the best possible time from the runner. This is something that untrained humans are very bad at, and even trained humans (e.g. runners) can’t do as well as the external systems, specially if they are further augmented by physiological information and models. Interoceptive and time senses — how your body is doing, and how much time has passed — can be very much improved by training, but sensors and clocks are better, which is why watches, pacers, and wavelights, without adding anything to the physiology or biomechanics of the runner, improves their performance.
In ways that matters in the context of running, wavelights are AIs that make runners smarter, and therefore more effective. This isn’t intelligence in the sense of more complex algorithms, but rather simple algorithms with more precise inputs.
Conceptually speaking, the argument works with wavelights as they exist (or even wristwatches) but we can spruce it up by thinking about the combination of sensors and customized physiological profiles. This would not just offload the decision of how fast to run, but also most likely improve it — another speed-up based not on changing physiology but on offloading what’s ultimately a strategic process.
The two takeaways you want to take for general applications are:
- More things that you would think are dynamic strategic problems. If you can improve running by offloading a (barely conscious) decision process, you can improve pretty much everything, even the processes you think are too straightforward or “human.”
- Understanding what you want to optimize, what you need to control, and what information is directly relevant to this is much — much, much, much — more important than algorithmic sophistication, whether in the form of complex neural networks, Large Language Models, or anything else.