AI Can Answer Questions, But It Cannot Ask Them for You

There's a line about parenting I've always loved. The best thing you can ask your child at the end of the school day isn't "How did you do on the test?" It's "What's a good question you asked today?"

In the age of AI, that line has quietly stopped being a feel-good slogan. For parents and students alike, it has started to read like a survival question.

I've taught at university for a long time now, and the anxiety I see in my students has jumped a full level in the past two years. The newspaper coverage of "AI anxiety" actually lags reality. In the United States, the most anxious students right now aren't the ones who are struggling. They're the top students at Ivy League schools. A surprising number of them are dropping out early to chase the AI gold rush in Silicon Valley, betting that they will be the next Zuckerberg or Musk. The ones who stay often can't sleep at night, for a simple reason: AI writes faster than they do, computes better than they do, knows more than they do — so what exactly did they sign up for four years of university for?

The same anxiety isn't confined to students. Plenty of Malaysian parents are quietly asking the same thing across the kitchen table. We send our children through primary school, secondary, university, sometimes postgrad — is any of that still worth it? If AI can already do everything, what is this whole long pipeline of education actually for?

To get anywhere near an answer, I think we have to back up to a much older, and much more neglected, question. What is education really teaching us, anyway? And what does it mean to "understand" something?

It sounds abstract. But there is a surprisingly direct way into it — through AI itself.

First: how does AI "understand" anything?

How AI works under the hood is a subject I'll come back to next time. For today, just a metaphor.

You watch videos on your phone. A two-hour film, in its raw form, is an enormous amount of data — and yet your phone stores it, streams it, and plays it back without a hiccup. How?

Compression.

The thing to grasp about compression is that the skill is not "keep everything". The skill is being clever about what you throw away — throwing things away so deftly that your eyes never notice they were gone.

All compression is lossy. The craft of it isn't about remembering more. It's about knowing what doesn't matter and can be safely discarded.

AI — especially the large language models like ChatGPT and Claude — is, at its core, doing exactly this. The goal isn't to memorise every fact in the world. The goal is to find patterns general enough to apply to things the model has never seen before. Compress well enough, and even a brand-new question gets a reasonable answer.

Human learning, it turns out, is the same trick

Here's the punchline. Human learning works the same way.

I'm a physicist by trade. I sometimes joke to friends that physics is the easiest subject. People usually assume I'm humble-bragging. I'm not. Physics is "easy" in a very specific sense: it has remarkably low underlying complexity. Wrap your head around gravity, and you can already say something sensible about black holes. One equation explains why an apple falls and why galaxies rotate. That is compression in its most extreme form.

But the underlying move — finding the generalisable pattern hiding inside the mess — isn't limited to physics. It isn't even limited to science.

Take a literary example. Hemingway was a celebrated master of compressed prose: his "iceberg theory" argued that only one-eighth of a story should appear on the page, with the remaining seven-eighths left for the reader to feel underneath. A famous (and historically contested) anecdote pushes the principle to its limit. Hemingway, the story goes, once bet his friends he could write a complete story in six English words. He scribbled it on a napkin and won the bet:

"For sale: baby shoes, never worn."

Six words. No adjectives, no plot, no scene. And yet the moment you finish reading, an entire world arrives — the image, the ache, the silence of a mother who never says what happened. The literal meaning of those six words is dwarfed by what they trigger inside you. That is literary compression: an entire universe folded into one sentence.

History is the same. We don't read history books to memorise who sneezed on which Tuesday in which century. We read them to distil — to pull the patterns of human nature, power, and circumstance out of the noise, so that we can recognise them when they show up again.

So becoming an "expert" in a field, properly understood, isn't really about stockpiling knowledge. It's about mastering that field's particular way of compressing the world. Einstein mastered the compression of physical reality. Hemingway mastered the compression of language. Every discipline has its own craft — its own answer to the question: how do we fit the world inside a finite human head?

A fact that might surprise you

I've mentioned this in this column once before, but readers always find it counterintuitive. A striking share of the people running the highest-earning hedge funds on Wall Street are mathematicians and physicists by training. The single most over-represented group is, oddly enough, astrophysicists — a quiet open secret in the industry. Silicon Valley follows the same pattern. At companies like OpenAI and Anthropic, the physics influence isn't confined to the founding teams — many of their senior technical people are themselves physics PhDs.

Why?

Not because physics class taught them how to trade stocks or how to train an AI. Because the training of physics builds one very specific muscle: how to find generalisable patterns inside complex systems. Markets are complex systems. AI models are complex systems. Same muscle, different gym.

I should confess that I'm on academic leave in Malaysia at the moment, and in my spare time I've been playing around with a few macroeconomic — and, more recently, bitcoin pricing — models myself. Not to actually trade. Just out of pure curiosity about how far the same way of thinking can be stretched.

Back to Malaysia's education culture

So, having taken the scenic route, back to the original question.

The thing I find genuinely sad about Malaysian education culture — and this runs all the way from parents to schools to universities — is that we are hooked on one single yardstick when judging learning. Is it useful? Is this subject useful? Will this degree get you a job? Hidden inside that question is an assumption — that the value of education lies in its external outputs.

In the AI era, external outputs are precisely the layer that's easiest to replace. AI can already write the marketing copy, design the poster, code the basic app, clean up the spreadsheet. All those "obviously useful" skills are being commodified in front of our eyes.

What is much harder to replace is the thing sitting underneath those outputs. How a person sees a problem. How they pull a pattern out of chaos. How they decide what to ignore.

That, almost by definition, is what Malaysian education talks about least. We have poured enormous effort into training students to answer correctly, and quickly. We have poured very little into teaching them to ask well.

Which is why I keep coming back to the line I opened with. The best question for a child at the end of the day isn't "What did you score?" It is "What's a good question you asked today?"

Asking is where compression begins.

Coda: the use of the useless

Does this mean everyone should rush off and study physics? Of course not.

There are systems — human ethics, social norms, the question of how to align an AI to human values — where the toolkit of physics simply does not reach. In fact, Anthropic keeps a team of in-house philosophers on staff precisely to handle the questions physics will never teach you. If you've noticed that different AI assistants seem to have slightly different "personalities", a lot of that texture comes from teams like that one.

A society's strength lies precisely in the fact that different people compress the world differently. Physicists compress into equations. Novelists compress into stories. Philosophers compress into principles. Historians compress into patterns. A mature society learns to recognise and nurture all those different kinds of compression — instead of flattening them under one single test of "is it useful?"

More than two thousand years ago, the Daoist philosopher Zhuangzi wrote: the use of the useless turns out to be the great use of all. In the age of AI, that line is starting to sound less like ancient philosophy and more like practical advice.

If Malaysia wants to walk steadily through the AI era, the first step is not forcing every child to learn to code, or to chat fluently with an AI. The first step is simpler, and much older. We — our schools, our parents, our children — need to relearn one of the most humble skills there is:

Asking the right question.

The rest, AI can't do for you. But it also can't take that from you.