Systems | kaguc — Writing to understand systems.

The failure of good decisions

Thu, 28 May 2026 00:00:00 +0000

A pattern I keep seeing but haven’t pinned: people make good decisions, each individually defensible, and the sum is bad. Worse than that — the sum is what you would have specifically tried to avoid if you’d described it up front.

Possible names: “local optimization,” “constraint blindness.” Neither fits. Local optimization has a global frame implied; this is a thing that happens because nobody is keeping the global frame.

Coming back to this.

Systems thinking, edge case

Fri, 22 May 2026 00:00:00 +0000

The thing nobody mentions about “systems thinking” is that it can become the lens — i.e. the next universal hammer.

I caught myself this week treating a deeply specific problem (a particular failure of a particular team) as if it were an instance of a feedback loop. It wasn’t. It was just three people who hadn’t slept enough.

The discipline isn’t to use systems thinking. It’s to know when not to.

Why real learning is reconstruction, not input

Tue, 12 May 2026 00:00:00 +0000

There is a quiet assumption underneath most of how we read, listen, and watch: that comprehension is transfer. Information leaves the page, crosses the eye, arrives in the mind, and stays there. We picture knowledge as a substance — something you can pour, accumulate, or run dry of.

This is wrong in an important way. Not in the trivial sense that we forget things; in a deeper one. You cannot be given an idea. You can only be given the raw material to reconstruct one.

The model

Consider what happens when you read a paragraph that genuinely changes how you see something. You did not “receive” the new view. You did three things:

You held the sentence in working memory long enough to suspend your prior model.
You built a small, fragile reconstruction of the author’s intended structure using your own concepts as scaffolding.
You then integrated — discarded, edited, or kept — that reconstruction based on its fit with everything else you believe.

Skip any one of these and nothing happens. The sentence passes through. You feel the warmth of comprehension, mistake it for the act of learning, and move on.

Why the input model fails

The input model fails most loudly in the era of infinite content.

If learning were transfer, more bandwidth would mean more learning. Listening to twenty hours of podcasts a week should produce dramatic intellectual change. It doesn’t, and the reason is not “we forget.” The reason is that step 2 — reconstruction — requires friction. It requires the disorienting moment of holding two incompatible models in mind. Anything that smooths that friction also smooths away the learning.

The fluency of a great speaker is, in this sense, dangerous. It removes the hard step. You feel taught. You are not.

What this implies

A few things follow that are worth taking seriously:

Slow reading beats fast reading, not because of word count, but because it preserves the reconstruction step.
Writing is the highest-bandwidth learning act, because it forces you to externalize your reconstruction and discover where it breaks.
Re-reading is underrated. The second pass is the first one where reconstruction happens; the first pass was orientation.
AI summaries are an anti-pattern for learning, even when the summary is perfect. The summary skips step 2 by design.

The library is not the books. It is the index you build in your own head from having walked the shelves enough times.

The work is not to consume. The work is to reconstruct.

What the Ising model teaches about complexity

Mon, 20 Apr 2026 00:00:00 +0000

The Ising model is one of the most extraordinary objects in physics. Its definition is almost embarrassingly simple. Its behavior is almost embarrassingly profound. The gap between the two is the lesson.

The setup

Place spins $s_i \in \{-1, +1\}$ on a lattice. Each spin interacts only with its nearest neighbors, with coupling strength $J$. In an external field $h$, the total energy is

$$ H(\{s\}) \;=\; -J \sum_{\langle i,j \rangle} s_i s_j \;-\; h \sum_i s_i. $$

That is the entire model. Two terms. One says: neighbors prefer to agree. One says: the field prefers a direction.

The probability of a configuration at temperature $T$ is the Boltzmann distribution,

$$ P(\{s\}) \;=\; \frac{1}{Z}\, e^{-\beta H(\{s\})}, \qquad \beta = \frac{1}{k_B T}. $$

There is nothing else. No agents. No memory. No strategy. No structure beyond “agree with the neighbor.”

What it produces

And yet, depending on $T$ alone, you get three completely different worlds.

High $T$. Thermal noise dominates. Spins are random. Magnetization averages to zero. The system has no opinion.
Low $T$. Coupling dominates. Spins align in vast domains. The system has, in effect, chosen.
At a critical $T_c$. Neither dominates. The system is fluctuating at every scale — small clusters inside larger clusters inside still larger ones. There is no characteristic size. The correlation length $\xi$ diverges:
$$\xi(T) \sim |T - T_c|^{-\nu}.$$

That last regime is what makes the model unreasonable. A trivial local rule, held in delicate balance, produces a system that is scale-free. The distribution of cluster sizes follows a power law. The system is maximally sensitive to perturbation. Information propagates across the entire lattice.

This is emergence in its starkest form. Nothing in the local rule “knew” about clusters, about scale-invariance, about long-range correlation. It all came out of the interaction.

Why it generalizes

Once you have seen this, you see it everywhere.

Markets near a regime change look like Ising at $T_c$. Tiny news moves the whole index. Volatility clusters at every timescale.
Opinion dynamics in a network exhibit the same three phases as a function of “social temperature” — apathy, consensus, or critical fragmentation.
Neural activity in the cortex appears to sit, on average, near criticality. The conjecture is that this is what makes the brain responsive but not unstable.

The mistake is to treat each of these as a separate science. The Ising lesson is that systems of locally-interacting binary agents have a shape to their behavior, and that shape repeats. You can predict a lot, qualitatively, just by asking: where am I on the temperature axis?

The deeper point

The Ising model is not a model of anything in particular. It is a model of how simplicity becomes complexity. The transformation is not mysterious — it is a sum, an exponential, a partition function. But the consequence of that transformation, the qualitative break in behavior at $T_c$, is genuinely new.

Long-term thinking, in any system, is the discipline of asking which temperature regime you are in — and refusing to confuse one for another.

Investing is not prediction, but understanding constraints

Sun, 22 Mar 2026 00:00:00 +0000

The most common mental model of investing is forecasting. You estimate what will happen, you bet on it, the future arrives, and you are right or wrong. This model has the great virtue of being easy to explain and the great defect of being almost completely wrong.

What actually compounds

If you study people who have produced unusual results over long horizons — the ones whose track records survive thirty years rather than three — they almost universally do not describe what they did as prediction.

They describe it as understanding what cannot be different.

A constraint is a fact about a system that any future must accommodate. “A bank cannot lend money it does not have access to.” “An economy cannot consume more energy than its grid can deliver.” “A network effect cannot form without a critical density of users.” These statements are not forecasts. They are the topology inside which all forecasts have to live.

The investor who has internalized the constraint is doing something qualitatively different from the one who is forecasting within it. The forecaster is playing a game of variance. The constraint-thinker is playing a game of which-games-are-possible.

Three kinds of constraint

It helps to distinguish three layers.

Physical. The slowest and most reliable. Energy density, transmission limits, the speed of light, the cost of moving atoms. Investments that ignore these usually fail spectacularly when the physics catches up.

Structural. The shape of the system. Two-sided markets need both sides. Capital cycles take roughly seven years. Trust, once broken at scale, is rebuilt in generations, not quarters. Structural constraints are slower than narrative but faster than physics.

Behavioral. What people actually do, as opposed to what models say they should. The behavioral layer changes the fastest and is the most often mistaken for noise. It is also where most edge actually lives.

The investor’s craft is figuring out which layer is binding on the question in front of them. The forecaster’s mistake is to treat them all as the same kind of fact.

The Buffett move

The reason Berkshire’s letters reread well, decades later, is that they almost never make predictions. They almost always describe the shape of the system the bet was placed inside.

When the tide goes out, you discover who has been swimming naked.

That sentence is not a prediction. It is a constraint. Liquidity is finite; asymmetry of information is finite; eventually the shape of the system reveals itself. The prediction is the part anyone could have made up. The constraint is the part that survives.

What this implies for thinking

If you accept the constraint frame, the work changes:

Stop asking what will happen. Ask what cannot fail to be true.
Stop optimizing within the game. Ask which game you are in.
When you can’t tell, lower your conviction, not your honesty.

Forecasts age badly. The map of constraints, kept honestly, can be the same map fifty years later. That is the asset.