Context Filter, Not Context Window
Why toddlers and language models both do better with less on the table
There’s a failure mode in working with language models that every practitioner learns the hard way: you assume that more context produces better output, so you cram everything into the prompt — every document, every instruction, every edge case — and the model gets worse. It hedges, it loses the thread, it fixates on some irrelevant detail you pasted in for completeness. The fix is counterintuitive but reliable: take things out. A clean, curated context beats a maximal one. The bottleneck was never how much the model could hold. It was how much it had to ignore.
The number people fixate on is the context window — how much you can put in. The number that actually governs the output is something more like a context filter — how much the system has to suppress to find the signal. I’ve come to believe small children run on the same principle, and I have a week of evidence from my own living room that I did not enjoy gathering.
The maximal-context toddler
My three-year-old’s environment was, by any reasonable standard, well-provisioned. Toys in every category, accessible at all times — the maximal context window, lovingly assembled. The theory was abundance: more to choose from, more to engage with, more development.
After a rough stretch at home, I started to suspect the abundance itself was part of the problem. The output of all that available context was not richer play. It was noise — and an increasingly dysregulated kid. The environment held more than he could meaningfully attend to, and the result was the toddler equivalent of a model fixating on the wrong line of a bloated prompt: scattered, escalating, never settling into anything.
So we ran the experiment. We put all of it away. Every toy, boxed and out of sight. And — the harder cut — no videos.
What a filtered context produced
I expected boredom, protest, a hard adjustment. What actually happened in the days since still surprises me.
He read. He read the entire Green Eggs and Ham book, start to finish. He drew. He asked to do penmanship practice — asked. He played Khan Academy Kids games, which sit at the creation-ish, low-stimulation end of what a screen can be. The smashing-and-breaking register of play — the rough, destructive, escalating mode that fills the space when nothing else is winning — largely went quiet, because the quieter activities were finally the loudest thing in an otherwise empty room.
None of these activities are new to him. He always could have chosen them. They were simply being drowned out. When I cleared the context, the signal that had been there all along got through. Removing options didn’t deprive him of anything. It revealed what he’d actually do given a clean room to do it in.
Why less-on-the-table works for both
The mechanism is, I think, genuinely shared between the toddler and the model, and it isn’t a cute analogy, but the same constraint.
Attention is the scarce resource, not material. A language model has a finite budget for what it can actively reason over; pile in more and the budget gets spent on suppression instead of thinking. A three-year-old has a much smaller such budget and far less ability to self-suppress. Every visible toy is a small open tab demanding a fraction of an executive-function system that is, at three, barely online. Twenty open tabs on a brand-new processor is not abundance. It’s a denial-of-service attack with good intentions behind it.
Clear the table and the same system that looked scattered suddenly looks capable — because it was capable the whole time, just overcommitted. The constraint wasn’t the size of what he could hold. It was how much he had to filter out to act.
The honest caveats
A few, because a clean arc is a warning sign:
This is days old, not a doctrine. I’m reporting an early result, not a parenting philosophy with a follow-up book. Calm baselines have a way of eroding, and I fully expect to be re-running this reset on a schedule rather than declaring victory once.
The goal isn’t a barren room forever. It’s rotation — a small, curated set of materials at a time, refreshed deliberately, the way you’d assemble a tight prompt for a specific task rather than dumping your entire hard drive into the input. Curation, not deprivation.
And the screen is still the strongest thing in the room. Khan Academy Kids wins a free choice more often than the books do, because it’s engineered to. Filtering the context narrowed what the screen is here; it didn’t make the crayons more magnetic than a tablet. I’d be overselling this badly if I claimed otherwise.
The takeaway for the rest of us
If you work with these models, you already know the move: when the output degrades, your instinct to add more context is usually exactly backwards. Take things out. Curate. Trust that the signal is in there and your job is to stop burying it.
It turns out that’s also a parenting move. The next time the play in your house tips from engaged into chaotic, consider that the problem might not be too few options. It might be too many — a context window stuffed past the point where a small, still-booting mind can find the one thing it actually wants to do. Clear the table. See what gets through.
The structured, deliberately-curated version of all this — a small set of materials and one build per week, by design — is 12 Weeks of Tech Projects for Toddlers. It is, in a real sense, a context filter in book form. At buildwithyourkid.com.