What Neanderthals Can Teach Us About AI Scaling

Big brains are not the same as deep intelligence

Neanderthals did not fail because they were stupid. Their brains were at least as large as ours, and in some cases larger. They made tools, controlled fire, buried their dead, and survived brutal environments for a very long time. If you only looked at raw hardware, the gap between them and us would seem surprisingly small.

That is what makes a strange brain organoid experiment so useful. It hints that the difference may not have been size at all. It may have been tempo.

A modern human brain takes forever to become itself. Childhood is absurdly long by animal standards. Adolescence drags on. Full maturation of some cortical systems extends well into early adulthood. From a survival perspective, this is almost comic. Human infants are helpless. Human teenagers are running a beta release with internet access. Yet that long delay seems tied to our capacity for abstraction, language, cumulative culture, and the ability to keep reworking internal models of the world for years.

The AI industry has spent the past few years betting on the opposite strategy. Train faster, scale harder, ship sooner. It has worked, up to a point. But diminishing returns are showing up everywhere: in benchmark gains, in data quality, in the cost of pretraining, in the stubborn gap between fluent output and robust understanding. The Neanderthal analogy is not proof of anything. Biology is not a product roadmap. Still, it offers a sharp way to think about a trade-off the field would rather ignore.

The gene swap that changed the clock

In 2021, a team led by Alysson Muotri at UC San Diego published a study on a gene called NOVA1, which plays a major role in neurodevelopment and RNA regulation in neurons. Modern humans carry a version that differs from the archaic variant associated with Neanderthals and Denisovans by a single amino acid. The researchers edited human stem cells to carry the archaic version, then used them to grow cortical organoids, simplified lab-grown models of early brain development.

The result was not “a Neanderthal brain in a dish.” Organoids are useful, but they are not miniature people. They lack bodies, sensory experience, vasculature, and the full developmental context that matters enormously in real brains. That caveat matters. The headlines were often more confident than the science.

Still, the differences were striking. The edited organoids developed differently in shape, neural connectivity, synaptic patterns, and electrical activity. Reports on the work often highlighted much faster maturation in some neuronal features, sometimes summarized as roughly an order of magnitude faster. Muotri’s broader interpretation was that the archaic variant pushes neural development toward earlier maturation, while the modern human variant prolongs development and plasticity.

That is the interesting part. Evolution may have changed the schedule before it changed the scale.

A chimp infant is physically and behaviorally more capable than a human infant at the same age. That looks like an advantage until you ask what a long childhood buys. Human development keeps the brain open longer. More time to wire, prune, adapt, imitate, absorb language, and be shaped by social complexity. You pay for that extended openness with vulnerability. You also get a machine that remains under construction for an unusually long time.

The Muotri study suggests that one tiny genetic tweak can shift that balance. Faster maturation may give earlier competence. Slower maturation may create the conditions for greater later complexity.

Evolution usually charges a fee

There is no free lunch in biology. Every gain is financed somewhere else.

If a nervous system matures quickly, it can support earlier survival. That matters in dangerous environments. Faster maturation means earlier motor coordination, earlier independence, earlier readiness. But the same acceleration can reduce the developmental window in which the brain remains highly plastic and reorganizable. You get to useful behavior sooner, yet you may give up some ceiling later.

Modern humans seem to have leaned hard into the opposite bargain. We stay fragile for an embarrassingly long time. In exchange, we get a brain that keeps sculpting itself through prolonged dependency, social learning, and delayed specialization. That process looks inefficient if your metric is early competence. It looks brilliant if your metric is open-ended complexity.

There is also a darker side. The same biological features that make our cognition unusually flexible may increase susceptibility to specifically human neurological disorders. Researchers have long noted that conditions such as autism spectrum disorders and Alzheimer’s disease appear tied to aspects of brain development and aging that are distinctive in humans. This does not mean complexity “causes” disease in any neat sense, and the causal chains are messy. But the broad point stands: a system optimized for prolonged plasticity and high-level integration often carries more failure modes.

That is worth dwelling on because AI discourse often treats capability gains as if they can be stacked indefinitely without changing the character of the system. Biology rarely works that way. Pushing one variable tends to deform another.

Scaling has been a speed run

Large language models improved so quickly that the whole field got addicted to the curve. Add parameters, add data, add compute, and surprising capabilities appear. Translation sharpens. Coding improves. Reasoning benchmarks jump. The pattern was strong enough to feel universal.

It was never universal. It was a regime.

Scaling works because bigger models trained on more diverse data can compress more statistical structure from language and adjacent modalities. For a while, that translates into broad capability gains. But the gains are not linear, and they are not all the same kind of gain. A model can get much better at imitation before it gets better at building durable internal representations. It can become eerily fluent before it becomes robust. It can learn to look general long before it can truly transfer knowledge into unfamiliar situations.

That distinction matters now because the easy part of scaling is ending. High-quality public text is finite. Training runs are grotesquely expensive. Benchmark progress is increasingly contaminated by saturation and leakage. More compute still helps, but it often buys polish rather than a conceptual leap. The system matures quickly in visible ways, then starts circling a plateau.

This is where the Neanderthal comparison earns its keep. The analogy is not “Neanderthals were like LLMs.” That would be silly. The useful parallel is narrower: rapid maturation can be a winning strategy in the short term while quietly constraining the path to later complexity.

Current models are astonishingly capable juveniles. They absorb huge amounts quickly, mimic adults convincingly, and impress anyone who remembers GPT-2. But they may be developing in a way that privileges early performance over deeper restructuring. Pretraining today is optimized to make a model useful fast, legible fast, benchmarkable fast. That biases the entire research stack toward visible competence.

Human brains did not become interesting by maximizing visible competence at month six.

Slower can mean richer, not weaker

When people hear “slow development,” they often imagine waste: underpowered systems taking longer to reach the same endpoint. That is not what biology suggests. Sometimes slowness is how a system avoids locking itself into shallow attractors.

A child does not learn language by ingesting a static pile of tokens once. The process is staged, embodied, interactive, and relentlessly social. Memory consolidates during sleep. Concepts are revisited across contexts. The brain alternates between plasticity and pruning, between exploration and compression. It does not simply get larger. It changes the rules by which it learns.

AI may need more of that kind of development. Not sentimentally, and not by copying neuroscience buzzwords onto a slide deck. I mean concrete design choices that trade raw training speed for richer world models later: curricula that control the order of learning, architectures with persistent memory, systems that learn through action and feedback rather than passive text prediction alone, training regimes that preserve plasticity longer instead of pushing for immediate convergence, and inference-time mechanisms that let models deliberate rather than autocomplete at industrial scale.

There is already movement in this direction. Test-time compute, tool use, synthetic environments, self-play, retrieval, memory layers, and reinforcement learning all try to give models a developmental path rather than a single giant cram session. None of these is a silver bullet. Some are glorified patches. Some may turn out to be dead ends. But the general shift is telling. The field is looking for ways to convert width into depth.

That search is partly economic and partly conceptual. The economic side is obvious: if pretraining costs keep exploding, researchers will hunt for methods that buy more capability per unit of compute. The conceptual side is more important. We may be reaching the point where “more of the same” produces systems that are increasingly polished versions of the same underlying developmental choice.

The Neanderthal organoid study points to a different intuition. A system can become competent quickly because it closes developmental options quickly. Another system can stay awkward longer because it keeps more options open.

The plateau is not the end of the story

It would be neat if this led to a clean moral. Biology says slow down, therefore AI should slow down. Real life is not that tidy.

Neanderthals were not evolutionary failures, and modern humans are not a simple upgrade path. Our own long developmental arc came with massive costs, from infant dependency to neurological fragility. In AI, longer developmental regimes may create their own problems: instability, inefficiency, harder evaluation, and systems that are less predictable because they stay plastic longer. Companies love fast feedback cycles for a reason. Slowness is expensive. It is also hard to sell.

Yet the deeper lesson still stands. Intelligence is not just a matter of how much substrate you have or how quickly you can optimize it. The schedule matters. What you hold open matters. The path to mature capability may require long periods where a system looks unimpressive by the metrics investors prefer.

That is why the Neanderthal story feels uncomfortably relevant. The industry keeps asking how to make models bloom earlier. Biology suggests that sometimes the trick is delaying the bloom so more structure can form underneath.