Two Wolves
There are two ways to get the next fifteen years badly wrong, and they are mirror images.
The first is normalcy bias: the deeply human assumption that because things have always been roughly like this, they will continue to be roughly like this. Normalcy bias is why people finish their drinks when the fire alarm goes off, and why they go back for their luggage on an aircraft filling with smoke. It is a sensible heuristic almost all the time, which is exactly what makes it dangerous during the rare periods when it isn’t. Anyone who has watched the last few years of AI development with normalcy bias fully engaged has been surprised repeatedly, and has explained each surprise away in turn — it’s only autocomplete, it can’t really reason, it only works on benchmarks, it can’t do anything in the real world, nobody is actually using it — until the explanations have begun to resemble a man backing slowly away from a rising tide, remarking at each step that his ankles are still dry.
The second failure is apocalypse addiction. There is a genuine pleasure in believing the end is near. It confers significance on the believer, who becomes one of the few who see clearly. It simplifies decisions, because nothing matters very much if the world is about to end, and so the pension can be ignored and the boring conversation with the bank postponed indefinitely. It creates communities, vocabulary, a sense of being on the inside of history. The history of millenarian movements is largely a history of people who sold their houses before a date that came and went, and who in a remarkable number of cases did not abandon the belief but recalculated the date. A person who is certain that artificial general intelligence will arrive in 2028 and transform everything by 2030 may be right. But if they have stopped paying into their pension on that basis, they have made a very large bet on a single number, and they will look foolish in 2031 in a way that is not merely embarrassing but expensive.
What makes the present so disorienting is that the evidence genuinely supports neither complacency nor certainty. Capabilities have improved faster than almost anyone expected a decade ago. Deployment has, in many places, been slower and messier than the enthusiasts promised. Both things are true at once, and the human mind does not like holding two true things that point in different directions. It wants a story. Most of the stories on offer — that it’s a bubble, that it’s the singularity, that it will take all the jobs, that it will create more jobs than it destroys — are really bets dressed as forecasts.
from Surviving the AI Apocalypse (2026)
