An interactive essay
Dancing landscapes
Good answers don't stay good in a rapidly changing world.
The idea for this essay comes from Scott Page's course, Understanding Complexity.1
Solving a problem is like hiking to the top of a mountain in thick fog. You can't see the summit. You can only feel the slope under your boots. So you do the sensible thing and keep walking uphill.
Sometimes that gets you to the top. But sometimes the ground flattens out and you find out that you were standing on a small hill, not the highest peak. Blindly going uphill is not always the best strategy.
Page describes three kinds of landscapes: smooth peaks (like Mt. Fuji), a rugged landscape with many local peaks, and a dancing landscape which keeps moving. Each represents real-life problems with different degrees of complexity.
In each section below, a few climbers race on these different kinds of landscapes. Each climber follows a different strategy. Some strategies work better than others. We compare and see what works best and why.
Mount Fuji
Let's start with an easy one: a landscape with a single peak.
Three climbers follow the simplest rule there is. Look around, step to the highest spot within reach, and repeat.
It's like cooking rice. Check it every minute. If it is still hard, cook it a little longer. Once it's just right, you're done. Finding the solution is a one-way path. There are no complicating variables.
When a problem has a single peak, any steady uphill rule works, and it doesn't matter where you start.
Sadly, most problems aren't a pot of rice.
Rugged landscapes
Now compare that to cooking a biryani.
There's a lot to get right: how long you par-boil the rice, how long the meat marinates, how much water goes in, how long it sits on dum. Get the rice perfect, and the meat might not cook through. You have to get them all right at once.
That's a rugged landscape: a complicated problem with many possible solutions.
Click a strategy to watch just that climber.
But rugged landscapes are solvable problems.
Rice, meat and heat behave the same way every time. With enough time and experimentation, you can find the best way to make the biryani you want.
Sending a spacecraft to the Moon is a similar problem. Gravity and the laws of physics do not change. With enough time and resources, it is possible to plan a trip to the moon and back again.
The real question is how you search. If you only ever make small changes, you can get stuck with a biryani that's good but not great.
Above, five climbers search the same rugged landscape, starting from the same spot with the same amount of time. Each one follows a different strategy. Run it a few times, and see which strategy works best.
- Looking far wins.74%
The climber that mostly climbs, but now and then checks a spot across the map, reaches the summit more than twice as often as the careful one. Daniel Levinthal found the same thing in his 1997 models of firms.2 An occasional long jump gets you off peaks that small steps never will. - Being careful is a trap.29%
The careful climber stops on the first peak it bumps into. Where it starts decides where it ends. - Mistakes only help if they fade.23% vs 32%
Being sloppy all the time is worse than being careful. Sloppy early and careful later is better. Scott Kirkpatrick, Daniel Gelatt and Mario Vecchi named this simulated annealing in 1983, after the way metal is cooled slowly so it settles into a stronger shape.3 - Starting over gets there, eventually.97% at the end, 51% on average
It nearly always finishes at the flag, but spends most of the race wandering through valleys. Finding the best spot and being somewhere good along the way are different jobs.
Dancing landscapes
A biryani recipe, once you've perfected it, stays perfect.
Rice and heat don't start behaving differently. But imagine your guests did. What they loved last month, they're bored of now.
That's a dancing landscape. The best answer keeps moving, because the world around it keeps moving.
This is also happening in the world of software development with the introduction of LLMs. For a junior developer 2-3 years ago, what made them valuable was writing clean code quickly.
Today, AI writes that code in seconds. The peak moved from writing good code to knowing what to build, and having an intuition on where the AI might fail.
Things are moving so fast that the best skills, and the best way to work keep shifting every few months.
Tested columns: average height over 200 runs, on ground that holds still and on ground that moves.
Watch Settles.
It climbs to the first decent peak and stops. On still ground, that works fine: it scores an average height of 68%, the same as climbing forever. Here, it scores 37%, the worst by far.
Kodak, Nokia, and BlackBerry are famous examples of this. Kodak actually made the first ever digital camera in 1975. But Kodak stayed stuck to film cameras, the peak it knew well, and never made the switch to digital until it was too late. It ended up filing for bankruptcy in 2012.
The climbers that look far and make leaps do much better, because there are always higher peaks to move to sooner or later in dancing landscapes.
But does exploring still pay when things move really fast? Change the speed to 10x and see.
Spoiler: not really. In a simulation of 200 runs, looking far often beats the careful climber by 11 points at normal speed. But when the world dances ten times faster, the advantage disappears: the careful climber comes out 2 points ahead.
Jeffrey Posen and Daniel Levinthal (2012) found the same pattern in models of firms.4 When change is fast and constant, what you learn by exploring goes stale sooner.
A peak you found last year may be a valley now.
Rivals
Section 03's ground moved because the world changed. There's a second reason landscapes change: other climbers.
Coming back to our biryani reference: let's open a biryani shop next to the most popular one in town (Big Biryani). We will split its customers. The spot looked like a peak. But with two shops on it, it's half a peak for each.
In response, Big Biryani might change its prices or add new menu items, further changing the landscape.
You can then choose from two strategies: copy Big Biryani, or find an empty peak to occupy.
Two moves, two very different endings. Averages over 300 maps:
- Copy Big Biryani.you 47%, Big Biryani 69%
You arrive a quarter late, split the peak, and drag Big Biryani down with you. Nobody wins. - Look for an empty peak.you 66%, Big Biryani 81%
You search far for high ground nobody holds, and you both stay high.
Jan Rivkin (2000) adds another reason copying fails.5 When a strategy's parts depend on each other, copying it with even small errors drops you into a valley.
The search party
In section 02, one climber searched alone and often got stuck on a small hill. What changes when climbers search as a team?
It depends on when they share.
Below, three teams of eight start from the same spots on the same map. Every member climbs, and each team converges at a different time.
Averages over 300 maps:
- Converge early.team 85%
Everyone piles onto whichever peak looked best in the first few seconds, and the search stops there. - Never converge.team 71%, best member 98%
Someone usually finds the summit. Nobody else hears about it. - Diverge, then converge.team 98%
Spread out, then pile onto the best find. Explore apart, then share.
Teams that share quickly do better at first and worse in the end, because sharing kills the variety a search needs.
Never sharing wastes the best find, because nobody else hears about it.
A team whose members search in different ways can beat a team of individually better searchers, because they get stuck in different places.
The teams that do best spread out first, then share.
Why it matters
The product landscape
If you build products, right now, AI is turning a lot of solid ground into a dancing landscape. Skills, products and jobs that sat on high peaks a few years ago are sliding.
Nobody can see where the next peaks will be. But the races above might give us an idea on how to approach the problem:
- Look far, often. Small polishing steps only take you to the top of the hill you're already on. Which might just be a small local peak, with higher peaks staying out of your reach.
- Don't settle. Sticking to what worked when things moved slowly is the worst rule to follow in a dancing landscape.
- Explore apart, then share. For teams, don't fixate on the first good idea someone chances upon. Diverge and explore multiple ideas individually before pulling together as a team.
Sources
- Scott E. Page, Understanding Complexity, The Great Courses (2009). The source for Mount Fuji, rugged and dancing landscapes. ↩
- Daniel Levinthal, “Adaptation on rugged landscapes”, Management Science (1997). ↩
- Scott Kirkpatrick, C. Daniel Gelatt and Mario Vecchi, “Optimization by simulated annealing”, Science (1983). ↩
- Jeffrey Posen and Daniel Levinthal, “Chasing a moving target: exploitation and exploration in dynamic environments”, Management Science (2012). ↩
- Jan Rivkin, “Imitation of complex strategies”, Management Science (2000). ↩