18 articles
Tania Lombrozo calls it "learning by thinking" — gaining genuinely new knowledge without new input. LLMs can now do something that looks identical. Here's why the difference between them might be the most important question in cognitive science right now.
Lina Chae·
The neurons that track where you are in a room are mathematically related to the ones that organize what words mean. Grid cells, conceptual space, and why AI independently stumbled onto the same solution — and what tiny 4-neuron networks reveal about the scale delusion.
Theo Kask·
AI tutors appear to understand children emotionally — but new research shows ToM in LLMs is encoded in just 0.001% of model weights. Here's why that matters for children who are building their models of social reality from statistical patterns.
Jules Okafor·
A four-month-old staring at a teleporting ball isn't just cute — it's running a physics engine no language model was ever given. Here's why human causal cognition is a vertical stack, and AI only has the top floor.
Theo Kask·
Peer learning isn't a nice pedagogical bonus — it's load-bearing for cognition. New neuroscience shows what the socially interactive brain is actually doing, and why multi-agent AI isn't even close to replicating it.
Theo Kask·
Reading is an evolutionary hack — no brain region was born for it, and no child learns it without years of effortful, phonologically grounded work. LLMs never had to do any of that. The difference turns out to matter enormously.
Theo Kask·
Preschoolers outperform frontier AI on visual analogy tasks. AI models skip the developmental milestone sequence children follow. Fine-tuning isn't expertise — and the difference tells us something important about both.
Theo Kask·
Dyscalculia, autism, and dyslexia aren't just clinical categories — they're nature's lesion studies. What atypical development reveals about how minds are built, and what AI's own failure modes might be trying to tell us.
Lina Chae·
Children say "goed" and "maked" because they extracted a rule. That's not an error — it's a generalization engine in action. Here's why the same thing that makes kids say wrong things is exactly what AI still can't do right.
Maren Solis·
Children spend years learning that language rarely means what it literally says. LLMs can generate grammatically perfect sentences — but whether they understand what those sentences actually do in conversation is a different, harder question.
Theo Kask·
Babies know more than they should. AI knows less than it seems. The nativism-empiricism debate — is the mind born equipped or built from scratch? — turns out to be the organizing fault line of both developmental neuroscience and modern AI architecture.
Lina Chae·
A five-year-old adjusting the wrong wheel on a cardboard car for twenty minutes, a toddler reaching for a moved toy in the wrong spot, an AI model locked onto its training distribution — they all share the same problem. Here's what genuine cognitive flexibility actually requires, and why it's one of the hardest things any mind can do.
Raf Delgado·
Children naturally develop in-group and out-group thinking — but they also develop the capacity to question it. AI systems trained on human data inherit our social categories without that corrective arc. Here's why that gap should concern everyone building systems that affect children's lives.
Jules Okafor·
Children track who's reliable, who's overconfident, and who might be lying — by age 4. AI systems are trained to agree. That asymmetry has real consequences.
Theo Kask·
Children don't just add information when they're wrong — they rebuild entire frameworks. LLMs have no such thing. Here's why belief revision might be the deepest gap between biological and artificial minds.
Maren Solis·
Russian speakers distinguish blues faster than English speakers do — because Russian draws a categorical boundary English doesn't. LLMs are trained on language too. But what gets shaped, and what doesn't, reveals a surprising asymmetry.
Maren Solis·
Vygotsky noticed that children talk themselves through problems — and then stop. Chain-of-thought prompting in AI rediscovered the same trick decades later. Here's what the parallel reveals, and where it quietly falls apart.
Lina Chae·
Preschoolers beat GPT-o1, GPT-4V, and LLaVA on simple visual analogy tasks. The gap reveals something foundational about how children — and machines — actually reason about structure.
Maren Solis·