Level-k thinking, applied to Lowball

Why you can't just pick 1, and how to figure out where on the reasoning ladder to stop.

You sit down at Lowball, look at the pick widget, and think: 1 is the lowest possible number. If I pick 1 and nobody else does, I win. Twenty seconds later you reconsider: but everyone else just had the same thought. They'll all pick 1. So 1 will collide and we'll all be out. I should pick 2. Five seconds after that: well, they're thinking that too. I should pick 3. And so on, up the ladder.

This is level-k thinking, the cognitive-science framework for games where your best move depends on what you think other people will think. Lowball is a near-perfect example of one.

The ladder

Behavioral economists model players as climbing a ladder of reasoning, where each rung represents one level of "I think that they think that..."

• Level 0: picks at random, or picks something obvious without thinking. Likely picks 1 or a personally lucky number.
• Level 1: thinks "level-0 players will pick 1, so I'll pick 2 to dodge them."
• Level 2: thinks "level-1 players are picking 2 to dodge the level-0s. I'll pick 3."
• Level 3: thinks "level-2 players are at 3, level-1 at 2, level-0 at 1. They all collide between 1 and 3. I'll pick 4 or higher."

The empirical literature on level-k reasoning (Stahl & Wilson 1995, Nagel 1995, Camerer's Behavioral Game Theory) finds that real humans cluster mostly at level 1 or level 2. Very few players climb to level 3 or beyond before settling on a pick. There are usually a few level-0 noise pickers too.

What that means for tonight's round

If most of your opponents are at level 1 or 2, they will cluster in a small range — usually 2 to maybe 8. That zone will be heavy with collisions. The winning number will tend to live just above that cluster: somewhere a level-3 or higher reasoner would land.

The pragmatic rule: if you think 1 is too low (everyone will pick it), the question is how many more levels up you need to climb to get past the crowd. Empirically, picking 1 above the modal cluster works often, but picking 2 above works almost as often and is safer.

When the ladder breaks down

Two situations where pure level-k reasoning misleads you:

1. Anti-correlation in the crowd. If the player population reads strategy guides (like this one), they're now reasoning about other strategists reasoning about them. The cluster can shift higher on its own, leaving 1 and 2 unexpectedly winnable on some days. Pay attention to recent winning numbers on the Stats panel — if they've been creeping up, the crowd is climbing higher and the low end is opening up.
2. Salient numbers. Some numbers attract way more picks than their level-k model would predict because they're memorable. 7 is universally over-picked. So is 13. So are round numbers like 10, 25, 100. These are the worst picks in absolute terms because crowds converge on them outside the normal reasoning chain. We have a whole post on this — Finding the gap.

The off-by-one trick

One concrete tactic that pulls from both of the above:

Pick a number that's one off from a salient cluster number. If 10 is going to attract a flood of picks, then 9 and 11 will both be relatively empty, because the level-k reasoners are jumping over the salient number rather than landing right next to it. Same with 24, 26, 99, 101. The off-by-one numbers are the "negative space" around the salient ones — and negative space is where uniqueness lives.

How far up should you climb?

Empirically, the median competitive Lowball pick sits between level 2 and level 4. That's a range of maybe 3 to 12 on a typical day. Picking above 12 means betting that the crowd is more strategic than usual; picking below 3 means betting they're less. Both are valid; they're just bets in different directions.

Honest summary: the only player who consistently wins this game is the player who reads the crowd that day. Yesterday's strategy is half-broken by today's strategy guide. That's the whole game.

---

More reading: Finding the gap nobody is looking at · Streak tactics