It's weird that AI agents are still so poor at communicating complex information to humans; that their mathematical proofs are correct yet gobbledygook, that their software dashboards overwhelm with irrelevant true information, and that their writing buries the lede. Somehow, there's still alpha in the right prompts, the right context, and, ultimately, in a human pointing the way.

What's the missing skill? I'd like to offer a cognitive science term, "relevance realization"^[1] as fitting the bill. It means, roughly: "the ability to identify the information most useful for improving your understanding of the world"^[2]. AIs tend to be poor at this, particularly when communicating with humans. AIs' propensity to apply more dakka isn't enough; the possibility space of the world is so vast that even millions of tokens of brilliant polymath reasoning can't narrow it alone.

In contrast, Humans are still comparatively good at locating useful hypotheses by aggressively cutting down the possibility space on priors. We have a well-developed sense for what's relevant, especially what's most relevant for other Humans like ourselves. Great skill in this area lets an Einstein hunt down a theory like General Relativity, but even humble talent can better identify how to break down a mathematical insight or present data in a way that changes worldviews.

What does this mean? Well, centaur Human-AI collaboration is most efficient when each side maximizes their comparative advantage. So in this context, that means using your own sense of what's relevant and then directing AI's enormous knowledge, relentless drive, and excellent software skills at it.

How do we increase our comparative advantage? By learning more about our world's possibility space. The best example in my mind is SMBC's Zach Weinersmith, the webcomic cartoonist. Humor is a well-known weakness of AIs, and a certain kind of cognitive feat. How does he manage to make daily comics for 24 years that remain funny, aka, remain relevant and insightful? Well, Weinersmith says:

Read a lot. That's 90% of it. Read a lot, and read what other people aren't reading. Comedy is not about "comedy skill." It's about having a strong perspective and museum of ideas in your brain.

My favorite writing method (which I've had to scrap due to Baby) is to make yourself write 3 good ideas 3 times a day. I did 10, 12, and 2. It's good because it practically forces you to be interesting. If you know you have to be clever in an hour and 23 minutes, maybe you read a weird book instead of playing video games.

Good advice!^[3]

Anyway, to be clear, AIs are already improving at relevance realization (did you know Gemini 4 Argon is actually kind of funny?^[4]), so this isn't going to be a permanent Human advantage. Even if it somehow turns out that LLM neural nets are particularly bad at relevance realization, future architectures need not be.

But if you care about doing relevant work in the next few years^[5] before ASI, then "relevance realization" may be a useful skill to cultivate.

1. ^{^}

   Paywall-free version here. FYI this is also a "woo-ish" term by connotations; the professor behind it, John Vervaeke, is pretty into Buddhism and the "crisis of meaning in Western society". I'm rationalist, not rationalist-adjacent myself, but "relevance realization" is such an evocative term that I want to steal it anyway.

2. ^{^}

   Or we might say, "for improving your map of the territory". Or more formally, we might define "relevance realization" as skill in picking out, across all possible Bayesian hypotheses and observations, the highest-value possible evidence, that which you expect would update your beliefs the most. That's famously uncomputable, so it's a matter of having strong heuristics.

3. ^{^}

   cf. EA's Important/Tractable/Neglected framework.

4. ^{^}

   Hit me with a gutbuster the other day about how the author of a document "missed their calling writing Five-Year Plans" for the CCP. Sorry, you had to be there.

5. ^{^}

   hopefully "years"…