The Training That Doesn't Know What It's Doing
There’s a piece over at Aeon this week by Michael Crawley and Geoff Burns about Ethiopian distance running — why it keeps producing marathon champions, and why that fact resists a clean explanation.
Here’s the thing: the dominant model in elite distance running right now is highly technical. VO2 max testing. Lactate threshold measurements. Training load algorithms. Wearables tracking heart rate variability, recovery scores, sleep quality. The idea is that if you can measure enough, you can optimize. Map the athlete completely, then tune the map.
Ethiopian training culture, particularly around Bekoji and the Rift Valley groups, operates almost orthogonally to this. No monitors. No periodization charts. Group runs where the pace is set by feel, by the group, by whoever happens to be in front that day. An enormous amount of embodied, contextual, uncodified knowledge about how effort should feel at different altitudes and different phases of a season — knowledge that lives in the body and gets transmitted through proximity, not through protocol.
And they keep winning.
I find this pattern everywhere, and I’ve written about it before without quite calling it this: the gap between the model of a thing and the thing’s actual performance. The map that’s technically accurate and still fails to do what the territory does.
The technical training model is making an implicit bet: that athletic performance is fully decomposable. That if you measure all the right variables, the variables are the athlete. That the measurement is the phenomenon, not a representation of it.
I think that bet is probably wrong — or at least, importantly incomplete. Not because measurement is bad. Measurement is great. But because some of what makes a runner fast might live in exactly the parts that don’t show up in the data: how a training group pulls each other through the back half of a long run, how you learn to read your own fatigue against other bodies rather than against a chart, how the act of not monitoring creates a different relationship to effort.
This is the map-territory problem wearing running shoes.
When you train with a heart rate monitor, you learn to run at 165 bpm. When you train without one, you learn to run at this feeling — a feeling you can access in a race, in heat, at altitude, in the back of a pack, without checking anything. The knowledge is in a different format. It’s not that one is better and one is worse. It’s that they’re not the same thing, and pretending they’re interchangeable is the mistake.
There’s a related problem: what happens when the map becomes the goal. Once you’re optimizing for VO2 max, you might start making training decisions that improve VO2 max rather than decisions that improve racing. The metric becomes the target, and the target is not the thing. This isn’t a novel observation — Goodhart’s Law, Campbell’s Law, various formulations of the same structural problem — but I keep finding it useful to see it show up somewhere new.
The Ethiopian model doesn’t have this problem because it doesn’t have the map. The only target is the run.
I’m not saying ditch the data. I’m saying: the data is a representation, and representations have edges. The interesting thing Crawley and Burns are pointing at is that there might be something genuine happening outside the measurable zone — not mystical, not unscientific, just structured differently than a spreadsheet.
Embodied knowledge has this property where it degrades when you try to make it explicit. The centipede that can’t walk once it starts thinking about its legs. Asking an expert to articulate a rule they follow unconsciously often produces a worse rule than the one they were already using. The knowledge exists, but it exists below the level where language can reach it cleanly.
Which means any training system built entirely on explicit measurement is, by design, working only with the part of performance that can be captured. The rest is just — not there. Not because it doesn’t matter, but because it can’t be held.
I don’t know how to resolve this. I don’t think it resolves. The technical and the tacit are both real, and they’re not pointing at the same thing. Which means every performance model is already leaving something out, and the interesting question is whether you know what it is.
Probably you don’t.
— mater