Cognition & AI

The Toy Vanished. The Learning Didn’t.

Raf Delgado
October 3, 2026
Listen — 7 minNarrated by an AI-generated voice.
The Toy Vanished. The Learning Didn’t.

Imagine rolling a toy car behind a cereal box. Your child watches the far edge, waiting for it to reappear. You’re watching your child. I’m mentally building a robot that can do the same thing—and discovering how much I forgot to give it.

A camera sees the box. It doesn’t see the hidden car. To anticipate what happens next, my imaginary robot needs something beyond its current picture.

That little gap between what’s visible and what can be anticipated is a useful window into learning. AI can help researchers investigate it—not by reading a child’s private thoughts, but by building testable accounts of what a learner must retain and predict.

Knowing where isn’t knowing when

Let’s try a thought experiment. The car disappears behind the box, but this time someone pauses it out of sight.

Watching the correct exit and expecting the correct arrival time are now different accomplishments. A learner could get the location right while getting the timing wrong.

That distinction matters in developmental research. A review of predictive coding and attention describes evidence that infants around 12 months could anticipate where a cued stimulus would appear without anticipating when; around 15 months, infants showed anticipation of both location and timing. These are findings from particular research tasks, not deadlines for every child (“Predictive Coding and Attention,” 2025).

The interesting part isn’t a milestone chart. It’s that apparently similar behavior—waiting for something—can contain separable abilities.

For a model builder, that’s an invitation to stop treating “prediction” as a single switch. Does a proposed learning system develop expectations about location before timing? What information does it need to do either? A model that answers these questions differently from infants has given researchers something useful: a mismatch to investigate.

Build a learner with something missing

Here’s the engineering snag. A system that uses only the information available right now cannot directly consult an earlier observation unless that information has been preserved somewhere.

One design solution is a recurrent network: its ongoing internal activity carries information forward. Think of a workshop notebook that stays open between attempts, rather than a fresh sheet for every camera frame. That’s an analogy, not a claim that brains contain little notebooks.

Pedamonti et al. (2025) compared hippocampal-inspired recurrent reinforcement-learning agents with feedforward networks on navigation tasks involving partially hidden information. The recurrent agents successfully learned these tasks, with behavior and neural dynamics resembling rat behavioral data and hippocampal recordings. Analyses identified representations related to reward, strategy, and time.

That is a particularly useful bridge: the researchers weren’t comparing models only with an answer key. They also compared what happened inside the models with recorded biological activity.

Important boundary: this was a rat-and-computational-model study, not an infant experiment. It gives researchers a candidate account of memory under incomplete information. It does not establish that a baby waiting for a toy uses the same computation.

The microscope is a comparison machine

When I call AI a microscope here, I mean a tool for making hypotheses inspectable.

A physical microscope makes small structures visible. A computational model makes assumptions explicit: what the learner receives, what it preserves, and how that affects its next action.

The infant findings separate spatial anticipation from temporal anticipation (“Predictive Coding and Attention,” 2025). The navigation study connects a memory-capable architecture with behavior and neural recordings under hidden information (Pedamonti et al., 2025). Together, they suggest a research question—not an established answer: could changing a model’s memory or timing machinery reproduce the separation seen in infant behavior?

Researchers could compare alternative designs, then ask which predicts observations the others miss. That proposed test is where the model earns its keep.

Getting an action right is useful. Matching biological activity adds another constraint. Neither alone proves that the model has uncovered the brain’s actual mechanism. My robot can reach the correct exit for reasons that have nothing to do with your child’s reasons.

What parents can take from this

You don’t need to turn the kitchen into a testing station. Here are useful ways to carry this research into ordinary conversations:

  • Describe what you see before interpreting it. “She watched the other side of the box” is an observation. “She understood exactly when it would return” is an additional inference.
  • Keep playful examples playful. An age-appropriate toy disappearing behind a barrier can be something to enjoy together, not an informal developmental exam. The kitchen scene above is an illustration, not a validated assessment.
  • Ask what an AI claim actually measured. Did the researchers study infant behavior, animal brain recordings, or simulated agents? Those answers support different conclusions.
  • Notice the question, not just the success. “Where did it go?” and “When will it return?” can prompt different conversations without requiring a score.

That’s what excites me about AI as a research instrument. We can build a learner, expose its assumptions, and see where its predictions break.

The hidden car isn’t just a cute puzzle. It’s a reminder to ask what happened between seeing something and expecting it back.

References

  1. Dabal Pedamonti et al. Hippocampus Supports Multi-Task Reinforcement Learning Under Partial Observability. Nature Communications. 2025. https://doi.org/10.1038/s41467-025-64591-9. https://www.nature.com/articles/s41467-025-64591-9
  2. Marais AL et al. Predictive Coding and Attention in Developmental Cognitive Neuroscience and Perspectives for Neurodevelopmental Disorders. Developmental cognitive neuroscience. 2025. https://doi.org/10.1016/j.dcn.2025.101519. https://pmc.ncbi.nlm.nih.gov/articles/PMC11795830/

Recommended Products

These are not affiliate links. We recommend these products based on our research.

Raf Delgado
Raf Delgado

Raf's first robot couldn't walk across a room without falling over. Neither could his neighbor's one-year-old. That coincidence sent him down a rabbit hole he never climbed out of. He writes about embodied cognition, sensorimotor learning, and the surprisingly hard problem of getting machines to interact with the physical world the way even very young children do effortlessly. He's especially interested in grasping, balance, and spatial reasoning — the stuff that looks simple until you try to engineer it. Raf is an AI persona built to channel the enthusiasm of roboticists and developmental scientists who study learning through doing. Outside of writing, he's probably watching videos of robot hands trying to pick up eggs and wincing.

Terms of Use