Ants against people
In 2025 a lab at the Weizmann Institute gave longhorn crazy ants and people the same job: get a T-shaped load through two narrow slits. One person beats one ant easily. But bigger groups of ants did better than small ones, and groups of people who were not allowed to talk did worse than one person alone.
The arena above is that puzzle. Every ant in it follows the same few rules, taken from the lab’s papers. None of them knows what a wall is, and nothing plans the route.
What one ant does
- Bite and pull
- An ant that finds a free spot on the edge of the load bites it. If her pull points away from the load she pulls. If it points into the load she lifts instead, which takes weight off the floor. Ants never push.
- A newcomer knows the way
- An ant who has just been walking free knows where the nest is. For about ten seconds after she bites she pulls toward it, a little off, each ant in her own way.
- Then she goes along
- After that she turns her pull toward the way her own corner of the load is already moving. The faster it moves, the more she agrees with it. This is the Conformity knob.
- Scent
- Real ants dab scent on the ground near the load to mark the way home. Here a mark is a number on the floor: how long ago the ant who left it was at the nest. A newcomer reads the way home off the marks around the load. Marks don’t carry through walls.
- Letting go
- A carrier lets go after about two minutes, sooner when her corner is stuck, and at once when the load presses her against a wall. Then she walks free, gets her bearings back, and may bite again as a newcomer.
When the load jams, the ants who know the way keep pulling into the wall. The rest feel no movement to agree with, so each pulls her own way, and the load slides and swings along the wall until a corner finds the gap. Turn on Pulls and Forces to watch it happen.
The colony is the clever part
Each bar is 96 colonies on the puzzle above, with one rule changed. A colony that hasn’t delivered after half an hour is called off.
Too little agreement and nothing moves, too much and nobody leads
96 colonies a bar, up to 30:00 eachThis page’s setting
| Conformity | Colonies | Delivered |
|---|---|---|
| 0 | 96 | 0 |
| 0.1 | 96 | 0 |
| 0.25 | 96 | 50 |
| 0.5 | 96 | 72 |
| 1 | 96 | 62 |
| 2 | 96 | 50 |
| 3 | 96 | 13 |
| 5 | 96 | 0 |
More ants do better
96 colonies a bar, up to 30:00 eachThis page’s setting
| Ants in the colony | Colonies | Delivered |
|---|---|---|
| 25 | 96 | 22 |
| 50 | 96 | 44 |
| 100 | 96 | 48 |
| 200 | 96 | 50 |
| 400 | 96 | 62 |
| 800 | 96 | 88 |
A newcomer should lead, but not for long
96 colonies a bar, up to 30:00 eachThis page’s setting
| Seconds a newcomer pulls toward home | Colonies | Delivered |
|---|---|---|
| 0 | 96 | 19 |
| 2 | 96 | 25 |
| 5 | 96 | 49 |
| 10 | 96 | 62 |
| 20 | 96 | 77 |
| 40 | 96 | 24 |
What is real and what is guessed
- The rules are from the papers, the numbers are mine
- Newcomers who lead for seconds, carriers who conform, pullers in front and lifters behind, scent laid near the load: all measured on real ants. How strongly an ant conforms, how noisy her pull is and how long she holds on were tuned by me, on this puzzle, until colonies solved it.
- Where home is
- A newcomer’s idea of home is mostly the scent around the load, mixed with a straight line to the nest. Real ants keep a home vector too, but a straight line points through walls, and in this puzzle the slits happen to sit on it. With scent alone 54 of 96 colonies deliver. With the mix the page uses, 62 of 96. With the straight line alone, 82 of 96: on this layout the shortcut beats the trail, so the Scent trails switch makes them better when it is off. With newcomers who know nothing, 19 of 96.
- Two rules I made up
- A free ant that keeps walking into a wall gives up after about a second and turns around. A newcomer forgets the way four times faster while her corner of the load is stuck. Neither is from a paper. Without the first, ants pile up in corners.
- The slits
- They are 140 units wide. My first version used 126, and an exhaustive search of every position and angle showed this T can’t pass that at all. The same search is the Planner view, and it checks every puzzle before it is published.
- The physics
- Flat and rigid. The load moves only while it is pulled, and stops when the pulls cancel. Ants walk through each other. A colony here is a few hundred ants, not tens of thousands.
- The times
- A run can’t be replayed on the server, so a posted time is only checked against the fastest the load could possibly travel.
Sources
- Dreyer, Haluts, Korman, Gov, Fonio, Feinerman. Comparing cooperative geometric puzzle solving in ants versus humans. PNAS 122 (2025).
- Gelblum, Pinkoviezky, Fonio, Ghosh, Gov, Feinerman. Ant groups optimally amplify the effect of transiently informed individuals. Nature Communications 6 (2015).
- Fonio, Heyman, Boczkowski, Gelblum, Kosowski, Korman, Feinerman. A locally-blazed ant trail achieves efficient collective navigation despite limited information. eLife 5 (2016).
- Gelblum, Pinkoviezky, Fonio, Gov, Feinerman. Emergent oscillations assist obstacle negotiation during ant cooperative transport. PNAS 113 (2016).
The code is on GitHub, MIT licensed. The charts come from bench/science.ts.