Laplace’s Mandate and Routing Anxiety | Jonathan Hendler

I have the sensation that I will be forced to always use the strongest, frontier model. The DSM XX will need a large class of categories for AI ailments like mine.

As long as I reach for the smartest LLM available, I tell myself I am making the best decisions. Every prompt, every half-formed thought, every tiny choice gets the full superintelligence treatment. It can feel responsible. It also can feel like the only responsible choice.

When tiny choices carry unknowable ripple effects, even the trivial ones start demanding superintelligence to navigate the chaos. Modern life turning every moment into a high-stakes simulation.

Laplace and the butterfly

Pierre-Simon Laplace, 1814. In the introduction to his Essai philosophique sur les probabilités he wrote:

We may regard the present state of the universe as the effect of its past and the cause of its future. An intellect which at a certain moment would know all forces that set nature in motion, and all positions of all items of which nature is composed, if this intellect were also vast enough to submit these data to analysis, it would embrace in a single formula the movements of the greatest bodies of the universe and those of the tiniest atom; for such an intellect nothing would be uncertain and the future, just like the past, would be present before its eyes.

That intellect became known as Laplace’s demon (Laplace himself never used the word “demon”). A perfect predictor. Complete knowledge of the present yields complete knowledge of the future. Classical mechanics as theology.

Edward Lorenz, 1961, running a weather model on a Royal McBee LGP-30. He restarted a run with initial conditions rounded from 0.506127 to 0.506. Completely different weather emerged. Sensitive dependence on initial conditions. The butterfly effect. Lorenz later asked, in a 1972 talk, whether the flap of a butterfly’s wings in Brazil could set off a tornado in Texas.

Chaos theory. Deterministic systems that are still practically unpredictable. The universe may be Laplacian in principle and still refuse to let us live that way.

Why decisions matter

Decision theory tried to give us a way through.

Von Neumann and Morgenstern (1944/1947) gave us expected utility theory: if your preferences satisfy a small set of axioms, there exists a utility function such that you should choose the option that maximizes expected utility. Rationality as calculation.

Herbert Simon looked at actual humans and said no. Bounded rationality. We do not optimize; we satisfice. We search until we find something good enough, then stop. Computational limits, attention limits, time limits. The organism is not an infinite intellect.

I think about decision density, the bits spent per perceived decision. I wrote about it years ago as a possible measure of the digital divide. Computation per capita is one thing. Computation per decision is another. What happens when the cost of “enough intelligence” for every decision drops near zero? Does the density explode?

When overthinking is overthinking

Analysis paralysis and a fear of error or of missing a superior option outweighs the value of acting in time. People stall, research forever, never decide. MRI work suggests overthinking can clog working memory and reduce creativity. The jam study (Iyengar & Lepper) is the classic: more options, less purchasing.

There is a point where additional intelligence applied to a decision produces diminishing or negative returns. Not because the model is wrong, but because the human is still the bottleneck that has to live with the output, integrate it, act on it, and then face the next decision.

The conundrum of finding optimums

Local versus global.

Gradient descent and most practical optimizers find local optima. They follow the slope downhill until the ground flattens. The global optimum may be on the other side of a ridge they will never climb. In high-dimensional, non-convex landscapes (most of life), you do not know if you are in a good valley or a mediocre one.

Superintelligence does not magically solve the problem of knowing whether a local improvement is worth the search cost for a global one. It just makes the local search extremely good and extremely fast. You can still be optimizing the wrong thing with exquisite precision.

Will it reduce or increase decision fatigue?

Maybe the only decision left becomes: always use the smartest model?

Decision fatigue (the popular name for the cumulative cost of choice) was framed by Baumeister and colleagues as a form of ego depletion; self-regulatory resources running down. Later replication work complicated the pure “muscle” model, but the phenomenology remains: after many choices, people default, avoid, or choose poorly.

If every micro-decision is escalated to the frontier model, two things can happen at once.

The cognitive load of evaluating the answer may rise. You still have to read it, trust it, decide...