Jun 8, 2026 · 11 min read

Not Available for Comment

The internet is awash with mental models. Anyone who’s read Taleb or Munger can rattle off tens of them — margin of safety, second-order effects, incentive-caused bias, on and on, a whole latticework you’re supposed to carry around.

I have two problems.

The first is that, as a human being, I have the working memory of a human being. If the toolkit holds 79 heuristics and 49 models, I will use exactly none of them most of the time. But if I really try, I will only grab the closest three and call the result “judgement”.

The second problem is arguably worse: I mostly can’t tell which of them are true. The ones I believe, I believe because they worked for someone who wrote a persuasive book about getting rich and famous. For any given model there is a graveyard of people who religiously used it, but unfortunately these people are usually not available for comment.

So do you throw all the models out, on the grounds that any of it might be wrong by Tuesday? No. You get more rigorous about two questions, in order: which of these are actually true, and how few of them do I actually need.

Which ones are true

Let’s start with the uncomfortable one. Most of what I believe about how the world works is a survivor’s word — pulled from someone it gloriously worked for, while the people who ran the identical playbook into a ditch don’t tend to have popular biographies or memoirs. In other words, the sample is filtered for success before it reaches me, and the sample is all there is.¹

There’s one exit, and it’s narrow. A few things are true in a way that doesn’t depend on watching anyone win. For example, compounding isn’t a tactic that happened to pay off for Buffett; it’s arithmetic, and it would hold even if every person who ever used it had died broke. Another example, fat tails, you measure across the whole population, the dead included. Incentive-caused bias you can watch in living people who won nothing, in a lab, today.

These are examples I can check myself. For everything else though, I’m trusting a survivor.

That’s the first half of the test a model has to pass to earn a place: can I verify it without a winner?

The second half is the one that actually keeps me honest — does it hold when the assumptions move? A model that only worked while families were large and populations grew isn’t a primitive, but rather a sensitive derivative of something deeper.

So, it can only be true if it not only applies to the winners, and if it holds when you lean on it. Most of the mental models touted on the internet will fail one filter or the other.

How few do I need

Fewer than most people suggest, because most of the models are some kind of derivative of another.

Consider “asymmetry” — where outcomes are very different when you consider the best and worst cases. Once you have it you don’t separately need antifragility (asymmetry that pays you for disorder), or optionality (asymmetry bought on purpose — a right with no obligation), or reversible versus irreversible decisions (asymmetry in the cost of being wrong), or margin of safety (asymmetry with a buffer).

In other words, here we have one idea with four slightly different derivatives.

A practical takeaway from the concept of “asymmetry” is Taleb’s barbell: cut what can ruin you, buy cheap exposure to what can make you, own as little as possible of the fragile middle.

Or take the 80/20 rule — a few inputs carrying the outcome because the tail of the distribution is fat. That isn’t a second insight, but rather the same one, just more business friendly.

You might notice that my approach is a form of reductionism, and reductionism can be incredibly dangerous, so let me be precise about which kind I mean. I want primitives — the word computer science uses for the irreducible building blocks that everything else is built from.

In other words, I am looking for building-block models that recombine into the rest. Some of the “rest” might not work at certain times, but the primitives are evergreen.

What I definitely don’t want is the physics fantasy, where you reduce the world to clean laws that snap back together into the whole of reality with nothing left over. Human systems are chaotic and fundamentally complex: recombine the parts and they throw off behaviour none of the parts predicted. So my reduction into the base models is simplifying my toolkit, not the world. That is, a small set of lenses I can actually carry in my human brain — held in full knowledge that the world will keep doing things none of them could have seen coming.

So I’ve run the filters across the pile, and here’s roughly what’s left — two kinds of thing, vetted two different ways. Facts about the terrain, which have to pass both filters. And the moves I make on it, which don’t go through the filters at all, because they aren’t claims about the world: an operation isn’t true or false, only valid or not, and the warrant for that is logic, not anyone’s track record.

The terrain — how the world runs:

Fat tails (power laws) — a handful of things carry the outcome, and the average is a lie.
Asymmetry (convexity) — when the best and worst cases aren’t the same size. The parent of half the models people list separately (see above), and the barbell is what you do about it.
Ruin avoidance and compounding (the absorbing barrier) — ruin is the one result you can’t average your way out of, because there’s no you left to keep playing. It’s the same instruction as never interrupt compounding, said from the other end. Some doors only open one way.
Local versus global optima — the best step available now and the best place to end up routinely point in different directions, so the move that looks best up close often isn’t. (The optimisation maths is checkable; what I hold looser is the assumption that my life has the same shaped landscape.)
Second-order effects — consequences have consequences: a promotion can be a raise and a demotion at the same time. (Strictly a discipline, not a fact — it tells me to look further down the chain, not what I’ll find there.)
The binding constraint — output is capped by whatever’s scarcest. Goldratt sold this one as a business novel; bin the novel. The thing underneath — a chain is only as strong as its weakest link — is arithmetic, and it clears the first filter cleanly.
Incentives (incentive-caused bias) — people respond to incentives, including in ways they sincerely deny. The one piece of psychology I trust, and not because it sold books: it clears the first filter because you can watch it happen in a lab today, in people who won nothing.
Iteration (the adjacent possible) — progress comes by small reversible steps, each one opening doors that weren’t reachable before.² This is the one terrain claim I keep on probation: it survives the second filter (it’s robust) but limps on the first, because the evidence is mostly survivors who iterated and won. I mostly believe it because it seems to have worked — for me, and for the people who wrote the books — which is exactly the kind of belief the rest of this essay is suspicious of. Make of that what you will. (The move hiding inside it — keep your steps small and reversible — isn’t a separate primitive; that’s just optionality, already up the list under asymmetry.)

The moves — what I do on the terrain:

There are fewer of these than you’d think, because they reduce too. Really there’s one, with one important special case.

Inversion — to crack “how do I get X”, work out what would reliably cause not-X, and avoid that. The most general move on the list; it runs on almost anything. Its warrant isn’t that it worked for Munger — you can kill a claim with one counterexample but never prove it with confirmations, so flipping the question turns a vague open search into a finite checklist.³
Via negativa — inversion aimed at risk: find what kills you, then subtract it. You know what harms you far more reliably than what helps you, and removing the bad default beats bolting on the clever thing.

Why the list stays short

This list is still too long, but at least it improves on others’. I can hold about four things at once, and maybe three when I’m tired, which is reliably when my harder decisions show up.⁴

A hundred-item latticework isn’t a useful thinking tool, and under load I won’t consider it — I’ll grab the nearest three and call it judgement, which is the local-maximum trap playing out in real time. So the small set isn’t a compromise a weak memory forced on me.

Btw, it’s also why I get suspicious when someone hands me exactly three crisp bullets. Three is the number you reach for when you want me to understand. Or when you want me to nod my head and stop thinking.⁵

How to hold them

I measure these models on two dials:

How sure am I that it’s true — high for the arithmetic, low for the psychology.⁶
And how sure am I that it applies here.

For instance, the same primitive can invert across scale: a system is antifragile because its parts are fragile and die, so “be antifragile” at the wrong level can hand you the exact opposite of what you wanted.

And since you can’t tell a global peak from a local one (without surveying a landscape too big to see) it means you should always be a little less certain than the thought leaders on LinkedIn. Even if it feels like a sure thing, with enough testing most models will break.

None of this maps the world. It’s just the handful of lenses I run a situation past — useful precisely because they’re few enough to run.

Which lands where it started. I still can’t tell you which of these are true. But that was never quite the test. The two filters don’t select for being right — they select for being wrong in a way I can walk away from. They won’t stop me making mistakes; they’re there to stop one of them being the last mistake I make. In a fat-tailed world that’s the only edge that compounds: not being right more often, but being wrong survivably.

Munger is a great example. You can’t evaluate his approach in time to use it — the only test is running his playbook for twenty years, and by then you already have your answer and can’t act on it. The thing you’d most want to verify is the thing the structure forbids you from verifying. ↩
The Apple story is illustrative — NeXT became the guts of OS X, iMovie led to iTunes, which led to the iPod, which cleared the runway for the iPhone. Each step making the next one reachable. Nice story. Survivor-selected like everything else. Thousands of companies iterated exactly as hard, also opened promising adjacent doors, and walked through them into bankruptcy. Iterating along a path is necessary and smarter than taking on moonshots from nowhere. But is adjacent iteration sufficient? ↩
Science runs on the same move — the null hypothesis. You don’t prove your theory, you try and fail to reject its opposite. The limit: inversion is strong where the ways to fail are few, enumerable, and most of the battle, and weak where success needs an irreducible positive act. You can invert your way away from an obviously bad novel; you can’t invert your way to a good one. ↩
You may have heard the number is seven, plus or minus two. I have since learned that figure is from a 1956 paper whose author opens by complaining he’s been persecuted by an integer. Later work pushes the real number down to about four, which tracks my post-children brain better. So the most-quoted statistic in popular psychology is itself a slightly-too-famous survivor. How fitting! ↩
Similar capacity limit, opposite intent. Someone explaining honestly gives you three things to remember because three fits in your head. The bad-faith pitch gives you three because a head holding three things has no room left for the objection. How many items someone hands you is a signal — read it the way you’d read what they’re holding versus what they’re selling. ↩
I’m leaning on incentives and a couple of robust biases here, not the broader catalogue. A lot of the famous psychology mostly fails to replicate — including many ideas I discovered reading books by some of the giants, such as Kahneman. But just because psychological ideas seem correct doesn’t make them correct. ↩

Not Available for Comment

Which ones are true

How few do I need

Why the list stays short

How to hold them

Footnotes

More posts