Oct 1, 2025
In my parenting, I often tell my toddler that he’s being “too strong!” It’s good for times where he’s about to apply so much force that it will actually be bad, like slamming the refrigerator door shut, or throwing a ball past our hedges into the street, or squeezing his baby sister too hard.
I like “too strong” as a phrase because it characterizes his behavior in a morally neutral way. In these moments, he’s not necessarily being rough or aggressive or uncareful or disobedient. Especially with him being a boy, I want to avoid those labels.
Instead, I will treat whatever he’s doing as simply a calibration error. When he half-tries to take off his shoes and gives up, he should apply more effort. And this is just the opposite: oops, this task would work better with less force. Better luck next time!
My goal is not for him to weed out bad behavior, but for him to gain contextual control over his own strength. After all, adults are strictly stronger than toddlers, and contextual strength is a skill that we all have in order to get around the world. We size up the task at hand, make sure to apply the right level of force, and experience the task in that light. In fact, many adults have a level of peak muscular strength that doesn’t even have a safe outlet in day-to-day life. So to have a full range, we do special things like going to the gym or taking on a physically demanding job/hobby.
One day, I expect to say to my kids “too smart!” using the same principle of contextual control. “Too smart” is harder to parse. We are familiar with encouraging people with the opposite. Figure it out, be smarter.
But it’s good to be able to scale down your own peak precision, depth, certainty, creativity, generalization, knowledge-seeking, world model, honesty, and edge case handling.
Kids are being “too smart” when they correct someone for rounding the time to the quarter hour. Adults are too smart when they argue a traffic ticket by appealing to relativity. When our first kid was born, we were too smart at doctors appointments. “How long does he feed per side?” We insisted that it wasn’t really possible to generalize (and here’s why…). This time around for baby 2, we know to simplify. “About 10-15 minutes.”
There is a place for “Well, actually…” but it’s not every place.
Contextual intelligence is a core competency for educators and others who work with a range of people.
LLM research labs are currently focused on enabling higher and higher peak intelligence. This leads to a gap in contextually managing intelligence.
For many user-facing applications, LLM responses are already often too smart. Frontier LLMs use technical terms without appropriate context, they are overzealous in being helpful, and they treat questions narrowly, among other things.
The same problem rears its head in the other direction when LLMs try to hit the right level. Since there is little research effort into toning smarts down, LLMs often do it badly, coming off as smarmy, sycophantic, emoji-ridden, and patronizing. These labels are unkind, so to speak, and companies can deal with them more effectively using morally neutral language. When toning down your smarts, don’t half-ass it. Contextual intelligence is possible to do well. LLMs already politely use only one language at a time, even though they know 100 languages. We just need to extend that.
Ultimately, contextual intelligence is about 1) recognizing the needs of the situation, and 2) adapting effectively. These are rhetorical skills, which have specialized analytic frameworks, metrics of success, and improvement techniques. Giving thought to them is something that having a language person can do for you.
Image: Goltzius (1589), The Great Hercules. GPT-5 recommended it after reading a draft of this post as an allegory of “mythic overstrength.” From The Met.
