The starter stack problem.
Four mental models are named constantly. The question worth asking is not which ones to collect, but how they interact when applied together.
The framing everyone starts with.
The standard advice for someone who wants to think more clearly is to collect mental models. Read Charlie Munger. Study Farnam Street. Build a stack. That advice is correct and almost useless. It doesn't tell you which models to reach for first, how they interfere with each other, or what happens when two of them point in opposite directions on the same problem.
The four models that appear most frequently in introductory treatments are: first-principles thinking, probabilistic thinking, inversion, and second-order thinking. Each is well-sourced. Each is genuinely useful. The problem is that they are usually presented as a list, not as a system. A list you can memorize. A system you have to learn to operate.
What each model actually does.
First-principles thinking, associated most visibly with Richard Feynman's approach to physics and problem decomposition, asks you to strip a problem down to its load-bearing assumptions and reason forward from those. The move is to stop accepting inherited framings and ask what is actually, demonstrably true at the base level.
Probabilistic thinking asks the opposite question: given that you cannot know the base-level truth with certainty, what distribution of outcomes should you be holding? Daniel Kahneman's work on forecasting and cognitive bias sits behind this model. The move is to stop treating your current belief as a fact and start treating it as a probability estimate subject to revision.
Inversion, which Charlie Munger drew from the mathematician Carl Jacobi's instruction to "invert, always invert," asks you to define the goal and then reason backward from failure. Instead of asking how to succeed, you ask what would guarantee failure, then eliminate those conditions. The move is to replace vague optimism with a concrete list of things to avoid.
Second-order thinking asks what happens after the immediate consequence of a decision. Most analysis stops at first-order effects: the action produces a result. Second-order thinking asks what that result produces in turn, and what that produces after that. The move is to extend the causal chain before committing.
Why a list is the wrong container.
Presented as four separate tools, these models create a selection problem. You face a decision. You reach for a model. Which one? The question feels arbitrary, and in practice people default to whichever model they learned most recently or feel most comfortable with. That is not a thinking system. That is a cognitive comfort zone with better vocabulary.
The models become useful when you understand their relationship to each other, not just their individual definitions.
First-principles and probabilistic thinking are in productive tension. First-principles thinking is appropriate when you have access to the underlying structure of a problem and the cost of being wrong about an assumption is high. Probabilistic thinking is appropriate when the underlying structure is opaque or contested and you need to act anyway. Using first-principles thinking on a problem where the base assumptions are themselves uncertain produces confident conclusions from shaky foundations. Using probabilistic thinking on a problem where the base assumptions are knowable produces unnecessary hedging.
Inversion and second-order thinking are complementary but operate at different points in a decision. Inversion is most useful before you commit: it surfaces failure modes while you can still change course. Second-order thinking is most useful after you have identified a candidate action: it extends the consequence chain to check whether the first-order benefit survives contact with downstream effects. Running them in sequence, inversion first and second-order thinking second, produces a more complete picture than either model alone.
The architecture of a working stack.
A latticework of mental models, the term Munger used to describe his own approach, is not a collection. It is a structure where each model's application depends on what the others reveal. The models do not replace each other; they constrain each other.
A practical sequence for a novel decision looks like this. First, use first-principles thinking to identify which assumptions you are actually relying on and whether those assumptions are verifiable. Second, where assumptions are not verifiable, apply probabilistic thinking to assign rough confidence levels rather than treating uncertain beliefs as facts. Third, apply inversion to the candidate action: list the conditions that would cause it to fail, and check whether any of those conditions are already present. Fourth, apply second-order thinking to the expected outcome: trace the consequence chain at least two steps forward and check whether the downstream effects change the attractiveness of the action.
This is not a rigid procedure. It is a default order that prevents the most common errors: reasoning confidently from unexamined assumptions, ignoring failure modes, and stopping the consequence chain too early.
What the stack does not solve.
Mental models are tools for structuring thought, not substitutes for domain knowledge. A latticework of models applied to a domain you do not understand produces well-structured confusion. The models clarify the shape of a problem; they do not supply the facts needed to resolve it.
The other limit is speed. These four models, applied in sequence, require time and deliberate attention. They are not useful for decisions that must be made in seconds. They are useful for decisions where the cost of a poor choice is high enough to justify slowing down before committing. That is a smaller category than most people assume, and a more important one.