Nobody Learns Alone

There's a thought experiment I keep coming back to when I hear AI researchers describe training runs.

Imagine you locked a child in a room with every book ever written and came back ten years later. The child would know an impressive amount. They would also be, in multiple measurable ways, cognitively stunted in ways that are hard to fully repair.

That's not a metaphor for AI. It's a description of a real failure mode in human development — and it's also, roughly, how most of our most powerful language models are trained.

I want to talk about peer learning. Not as a pedagogical feel-good concept, but as something that turns out to be load-bearing for cognition in ways that we're only starting to understand neuroscientifically — and that multi-agent AI systems are only beginning to grapple with.

Joint Attention: The Foundation No One Talks About

Before a baby can say "look," they can do it. By eight months, most infants will follow a caregiver's gaze to an object, then check back to confirm they're sharing the experience. By twelve months, they'll point at things just to share the noticing. This is joint attention — the ability to jointly hold something in mind with another mind.

It seems obvious and small. It isn't.

According to Grossmann et al. (2025), a systematic review in Developmental Cognitive Neuroscience synthesizing 16 neuroimaging studies (EEG, fNIRS, fMRI), the right temporoparietal junction (TPJ) — a region involved in perspective-taking, mentalizing, and attention regulation — activates reliably across all forms of partnered social attention in infants aged 8–24 months.

That's not nothing. That's a dedicated neural architecture for shared reference. The brain built specific hardware for the act of looking at something together.

And here's why it matters for learning: you can't learn from a peer if you can't track what they're attending to. Joint attention is the prerequisite. Every collaborative learning episode — every classroom discussion, every "watch me do this first," every "no, hold it like this" — relies on it. Strip it out and you don't have collaborative cognition anymore. You just have two brains in the same room.

Current AI systems — including multi-agent systems running multiple model instances — don't have joint attention in any meaningful sense. They share tokens. They don't share a world.

The 3-Month Revelation

Here's a finding that still kind of breaks my brain.

Endevelt-Shapira, Bosseler, Mizrahi, Meltzoff, and Kuhl (2024) at UW's I-LABS tracked mother-infant interactions starting at just 3 months of age — well before any meaningful speech — and then followed language development through age 2.5. The punchline: conversational turn-taking at 3 months predicted vocabulary size at 27–30 months.

Not speech input alone. Not IQ proxy measures. Conversational turn-taking. The back-and-forth of a caregiver responding contingently to a pre-verbal infant's babbles and gazes.

This is a stunning result. It means the social scaffolding of learning is doing something that can't be replaced by volume of input. Maternal sensitivity at 3 months independently predicted productive language outcomes at every single follow-up age from 18 to 30 months. The amount of talk the infant heard mattered, yes — but the responsiveness, the social reciprocity, added independent predictive power on top of it.

Think about what this implies. A 3-month-old can't understand words. What they're absorbing in those early social exchanges isn't vocabulary. They're learning the structure of collaborative cognition — that communication is bidirectional, that their signals affect another mind, that the world responds to them.

That turns out to be foundational to everything that comes later, including language.

If you're designing an AI training protocol, this should give you pause. Our most powerful language models consume a colossal pile of human-generated text. They learn from the outputs of social cognition, not from social cognition itself.

The Zone Where Learning Actually Happens

Vygotsky gets cited constantly and understood rarely. His concept of the "zone of proximal development" — the sweet spot between too easy and too hard — is usually presented as a teaching principle. What new cognitive science is revealing is that it's also a description of how curiosity actually works in the brain.

Liquin and Gopnik (2024) reviewed convergent evidence from developmental psychology, cognitive neuroscience, and computational modeling, arguing that curiosity isn't primarily about uncertainty reduction (the classic account). Instead, curiosity is sustained by learning progress — ongoing improvement in competence. We're drawn to tasks where we're getting better, and we disengage when we plateau or when something is miles beyond us.

What's interesting about this is that peer learning is one of the most powerful ways to find and stay in that zone. A slightly more skilled peer — older sibling, classmate who figured it out first — provides natural scaffolding at exactly the right level of challenge. Close enough to model something comprehensible, different enough to stretch you.

Poli et al. (2025) confirmed this empirically in preschoolers. In a free-exploration study with 102 four-year-olds, children spontaneously tracked their own learning progress and redirected their attention to tasks where they were still improving — avoiding things they'd mastered and things they couldn't crack. This isn't random exploration. It's an internally calibrated curiosity algorithm — the same principle powering intrinsic motivation in state-of-the-art reinforcement learning agents.

The analogy holds and then breaks at the same place. AI curiosity algorithms track uncertainty or prediction error. Human curiosity tracks progress within a social context. We're not just asking "am I getting better?" — we're asking "am I getting better at something this other person can see and validate?"

What the Socially Interactive Brain Actually Looks Like

Merchant et al. (2025) published what is now the most comprehensive neural map of real-time social interaction: a coordinate-based meta-analysis of 108 human neuroimaging studies, synthesizing the entire scientific literature on what happens in the brain during live social engagement.

Ten brain regions consistently activate across studies. They span the default mode network — including the temporoparietal junction and medial prefrontal cortex — lateral frontoparietal cognitive control regions, and midcingulo-insular areas involved in reward and emotion. The analysis also dissociated distinct social functions: perceived social engagement, reciprocal interaction, and the difference between initiating versus responding.

What strikes me about this is the sheer orchestration involved. The "socially interactive brain" isn't one region or one network. It's a coordinated mobilization of mentalizing, attention, reward, and emotion — simultaneously, in real time. Social interaction isn't a feature of human cognition. It's an organizing principle.

This is the hardware that peer learning runs on. When a child works through a problem with a classmate, they're not just sharing processing load. They're running a whole-brain social engagement mode that is qualitatively different from solitary problem-solving. The brains involved are doing more, and doing different things, than the same brains working alone.

What Multi-Agent AI Gets Right (and Gets Wrong)

Multi-agent AI systems have produced some genuinely remarkable results. AlphaGo and AlphaStar learned through self-play — generating feedback by competing against earlier versions of themselves. The outcomes were superhuman at chess, Go, and StarCraft.

Self-play works. When a task has a clear reward signal, agents competing against themselves can discover strategies no human ever found. That's real, and it's impressive.

But self-play is a very specific kind of peer learning. It's peer learning without joint attention, without scaffolded zones of proximal development, without emotional responsiveness, and without anything resembling the social embedding that turns out to be foundational to how biological cognition develops. Two model instances passing tokens is not two children at a whiteboard. The surface looks similar. The underlying process is not.

Taniguchi et al. (2024), publishing in Science Robotics, offer a more instructive example. Their developmental robotics framework had robots acquire compositional language-action mappings through interactive, grounded learning — not by pretraining on a massive corpus, but through scaffolded social interaction, built up incrementally. The result: better generalization to novel language-action combinations than data-intensive approaches achieved.

The key word is scaffolded. The robot isn't just playing against itself or absorbing data. It's engaged in the kind of incrementally structured, responsively calibrated interaction that mirrors how children learn compositionality. The outcomes show it: interaction-driven learning produces qualitatively different representations — more flexible, more generalizable.

That's what peer learning actually is, done right. And most multi-agent AI isn't doing it.

The Honest Bottom Line

We don't have multi-agent AI that learns the way children learn from each other. We have multi-agent AI that can compete, cooperate around specified objectives, and exchange information between instances. That's useful. It's not the same thing.

The developmental neuroscience is increasingly clear that what makes peer learning so powerful isn't the parallelism — it's the reciprocal social embedding. Joint attention. Contingent responsiveness. Calibrated challenge. Emotional feedback loops. A 3-month-old babbling and having a caregiver respond is building cognitive architecture that shapes language development two years later. That's not a fluke. That's what brains are designed to need.

Building AI systems that can genuinely learn from each other — the way kids do — would require solving joint attention for machines: shared reference to a common world, not just shared tokens in a context window. It would require intrinsic motivation calibrated to the social context. It would require, in short, something like a genuine theory of mind operating between agents in real time.

None of that is on the near-term roadmap. I've looked.

But the neuroscience keeps pushing us toward the same uncomfortable conclusion: intelligence, in its biological form, is fundamentally a social achievement. The brain didn't evolve its remarkable learning machinery for solitary study sessions. It evolved it for classrooms, families, and playgrounds — even before any of those things were invented.

The most powerful teaching tool isn't a larger dataset or a better architecture.

It's another mind.