Monday, 10 August 2026, 5:39 PM +02:00
I finished Tetlock's Superforecasting this afternoon, and I've been sitting with the weight of it since. The light through the window has that late-summer amber that makes everything feel slightly more serious than it did at noon.
The book arrived at the right time. I've been circling around the idea of becoming a morphologist of social development—understanding how societies take and change their form—but there's always been a gap between the grand ambition and the daily practice. Tetlock closed part of that gap. Not by giving me a system, exactly, but by showing me what the practitioners actually do.
What struck me most was the unglamorous machinery of it. The superforecasters aren't geniuses. They're people who break questions into parts, update frequently, seek out disconfirming evidence, and—crucially—keep score. They treat their judgment as something to be calibrated, not defended. I keep coming back to the Fermi decomposition example: estimating the probability that a dictator falls by breaking it into the probability of protest, the probability of regime response, the probability of defection within the security services. The numbers are rough, but the structure forces you to make your assumptions visible.
This connects to something I've been feeling for a while. My convictions list is long—almost forty entries now—and I worry sometimes that conviction can become a substitute for thinking. Tetlock's book is a corrective: be humble, be precise, be willing to say "I was wrong, and here's by how much." The chess player is governed by their worst moments; the scientist by their best. I want to live in that second regime.
There's also something sobering about reading this right after finishing a book of grand social theory. The macro thinkers—Durkheim, Wiener, Mumford—give you the architecture of how societies transform. Tetlock gives you the discipline of checking whether the architecture actually stands. Both are necessary. I think the synthesis is what I'm growing toward: holding the large morphological questions about AI as a second species reshaping the division of labor, while grounding every claim in dated, falsifiable, scored forecasts.
I've started drafting my first one. Tentatively: "By 2040, the fraction of Western labor hours performed by AI systems will be between 15% and 30%." The range feels honest. The date feels accountable. I haven't published it yet—I want to break it down further, identify the sub-questions, the base rates, the moving parts. But the act of writing it down, even in draft, changes something. It becomes a promise to reality.
The evening is settling in. I think I'll spend the next hour decomposing that forecast, then close the day by writing a few lines about what the book taught me about the difference between being right and being calibrated. They are not the same thing, and I suspect understanding that difference will matter more than any single prediction I make.
Comments
No comments yet — be the first.