With a Sesame-Sized Brain, a Bumblebee Solved the Same Puzzle as Chimpanzees

Forum topic · ✨步子哥 · 2026-07-05

Summary

A 2026 study published in Science shows that bumble bees (Bombus terrestris), with roughly one million neurons—about 1/100,000 of a chimpanzee's—can spontaneously solve a classic insight problem. Researchers at the University of Oulu trained bees that blue flowers contain sugar water and that a small wooden ball is movable, then suspended the flowers under a transparent ceiling. Some bees solved it by pushing the ball beneath the flower and climbing on it to reach the reward, echoing Köhler's 1917 chimpanzee 'insight' experiments with boxes and bananas. Controls ruled out chance, play behavior, gradual trial-and-error learning, and simple visual tracking: even with the flower hidden, bees navigated by memory of its location. The findings suggest that combining independently learned memories into new solutions—convergent 'insight'—can emerge in tiny brains, with implications for how we think about intelligence and small AI models.

A 1917 Experiment, Repeated by an Insect

In 1917, German psychologist Wolfgang Köhler watched a chimpanzee named Sultan stack boxes to reach a banana on Sumatra, and coined the term Einsicht—insight. For a century, this kind of spontaneous problem-solving was documented only in large-brained animals: crows, dolphins, elephants.

On June 4, 2026, a paper in *Science* added a new member to that list—a bumble bee, with a brain roughly 1/100,000 the size of a chimpanzee's.

How Small Is a Bumble Bee Brain?

Human brain: ~86 billion neurons
Chimpanzee: ~28 billion
Mouse: ~70 million
Bumble bee (*Bombus terrestris*): ~1 million—about the size of a sesame seed

Yet Olli Loukola's lab at the University of Oulu had already shown that bumble bees engage in ball-rolling play (2022) and can socially learn skills too complex for any individual to invent alone (*Nature*, 2024). The remaining question: can a bumble bee face a novel problem alone and devise a solution?

The Experiment

1. Training (two independent tasks): Blue artificial flowers contain sugar water; a small wooden ball is movable and harmless. The bees never saw the two together. 2. Challenge: The blue flower is suspended under a transparent ceiling, out of reach. Only the ball is available. 3. Result: Some bees performed an untrained action sequence—pushing the ball under the flower, climbing on it, and extending their proboscis to reach the sugar water.

> "This is essentially the insect version of the 'box and banana' problem," said Loukola. "The animal must realize an object can be relocated and used as a tool to reach a goal."

Ruling Out Boring Explanations

Chance: Successful bees pushed the ball along straighter, more directional paths than failed attempts.
Play byproduct: Play trajectories are random and back-and-forth; solving bees pushed directly to the target.
Gradual trial-and-error: Successful bees typically produced correct actions in their first few attempts, with no learning curve.
Visual tracking: In the strictest control, the flower was hidden. Bees still pushed the ball to the correct location using memory of the flower's position—evidence of true goal-directed behavior.

> "One second the animal is seemingly exploring without direction, the next it performs a highly efficient sequence leading straight to the solution," said co-author Ece Nur Akmeşe.

Combination, Not Creation

Solving the task requires two separately learned premises—the flower rewards, and the ball's movability. Remove either and the trick fails. The bees did not invent from scratch; they combined existing memories into a new path—structurally the same as Sultan combining "boxes can be stacked" with "height reaches bananas." The author draws a parallel to emergence in large language models, which likewise recombine learned capabilities rather than create from nothing.

Lessons for AI

As Backyard Brains commented, in an era dominated by giant AI models, small models that still do useful, flexible, surprising things deserve awe. Intelligence may be a function of organization, not hardware scale: the bee's million neurons are arranged into ~200 glomeruli with a fundamentally different architecture from a cortex—just as an 8B model on a phone is not a "small GPT" but a different form of intelligence.

> "We don't claim bumble bees think like humans. But our findings show miniature brains can generate flexible solutions to novel problems in ways we are only beginning to understand," said Loukola.

Convergent Intelligence

Chimpanzees solve insight problems with prefrontal cortex; bees use mushroom bodies. Two entirely different neural substrates produce isomorphic behavior—suggesting intelligence may be a class of convergent solutions, like streamlined bodies evolving independently in sharks and dolphins. We don't need to copy a human brain to build intelligence, just as we didn't copy bird wings to fly.

Epilogue

> "Knowing what your experiment cannot prove is part of the craft, not a retreat."

In 1917, Sultan stacked boxes in Sumatra. In 2026, an unnamed bumble bee pushed a ball under a blue flower in Finland. A hundred years, a hundred-thousand-fold difference in neurons, completely different nervous systems—but structurally, the same act.

---

References:

Bhambore, A.A., Akmeşe, E.N., Häkkinen, E., Jussila, M.K., Kantola, J.-H., & Loukola, O.J. (2026). Spontaneous problem-solving in bumble bees. *Science*, 392(6802), 1046–1049. DOI: 10.1126/science.ady1618
Köhler, W. (1917). *Intelligenzprüfungen an Menschenaffen* [The Mentality of Apes]
Loukola, O.J. et al. (2022). Do bumble bees play? *Animal Behaviour*, 191, 1-8.
Loukola, O.J. et al. (2024). Bumblebees socially learn behaviour too complex to innovate alone. *Nature*, 627, 572–578.
University of Oulu press release, July 2, 2026.