In his 2015 letter to shareholders, Jeff Bezos split decisions into two kinds. Some are one-way doors: once through, you cannot come back, and they deserve slow, careful thought. Most are two-way doors: if the room on the other side is wrong, you walk back out. The common mistake, he argued, is treating the second kind like the first: deliberating for weeks over choices that could have been tried, measured and reversed in days.
The same idea has a harder, mathematical edge. When gains and losses compound, the order of events matters and a large enough loss ends the game. A bet can have a positive expected value and still ruin almost everyone who takes it too often at too large a size. The physicist Ole Peters calls this the ergodicity problem: the average over many people is not what any one person experiences over time. Survival is the condition for having any other goal at all.
Shipping a model — build the way back first
Mature ML teams spend a surprising share of their effort on undo. Shadow mode lets a new model score live traffic without acting on it, and a champion/challenger setup gives it a small slice before the whole. CI/CD for ML stops a bad model before it ships; a model registry makes going back to yesterday’s version a one-line change; packaging in containers makes sure the old version still runs; reproducibility means it can be rebuilt at all. When something does break, good incident response rolls back first and investigates second.
Sizing a position — stay in the game
The Kelly criterion gives the bet size that grows capital fastest, and shows that betting much more than that turns a winning edge into a losing strategy. Position limits cap what any one mistake can cost. Pyramiding starts small and adds only once the market has said yes. A protective put from options hedging buys a floor, paying a premium so that one outcome is ruled out in advance. Portfolio insurance tried to make the floor dynamic — and in October 1987 showed that an exit everyone plans to use at once is not reversible at all.
Training a language model — small runs before the big one
Large models are expensive one-way doors, so the field has learned to climb a ladder of cost. Prompt engineering is the cheapest change and the easiest to undo; fine-tuning costs more and changes the weights; pre-training is the one-way door. Before walking through it, labs fit scaling laws to a series of small, cheap runs and extrapolate — many small bets used to size the one big one.
The mind — why reversing is hard
The obstacles are mostly psychological. The sunk cost fallacy keeps people walking through a door they should back out of; the disposition effect reverses the wrong positions, selling winners and keeping losers; mental accounting treats recent gains as house money to bet bigger with; and FOMO turns many small entries into one large, late one. Each turns a two-way door into a one-way door from the inside.
Cheap reversals also make it safe to try many things. Trying many things brings back the warning of The Garden of Forking Paths: the more you try, the more carefully you have to judge what seems to work. Small bets are for learning fast, not for believing the first lucky one.
Last reviewed October 2026 · How this site checks its content · Report a mistake