Terence Tao Says Probability Is Only the Right Ruler When Things Repeat. What About the Decisions I Only Get to Make Once?
Terence Tao spent eighty minutes on Big Think walking through his new book, Six Math Essentials. Four things stayed with me: where probability stops applying, why doubling down is bankruptcy compressed into a small number, why traffic keeps jamming after the accident is cleared, and how a correct model that fits worse at first gets killed by the data. I checked each against my own prediction ledger and my settlement-fan calibration, and every one of them landed. Educational notes and extension, not investment advice.
Even with the eyes of Li Lou and the craft of Gongshu Zi, without compass and square you cannot make a circle or a square.
—— Mencius, “Li Lou I” (Warring States period; translation mine)
What this episode is about
Terence Tao’s interview on Big Think, published 2026-08-28, runs 84 minutes and is built around his forthcoming book, Six Math Essentials. He organizes the whole book around six ideas: numbers, algebra, geometry, probability, analysis, and dynamics. All six are thousands of years old, most of us met them as children, and mathematicians have simply developed them to an extreme degree of precision. Strip away the technical machinery, he says, and they’re all deeply intuitive; mathematics is just a language for stating them exactly.
The conversation has three parts: what each of the six ideas is; how mathematics keeps preparing, in advance, the language that science turns out to need; and how AI is changing mathematics and science. I’ve been following a few mathematicians’ talks lately to see whether I can borrow a different way of thinking. Four things stayed with me from this one, and each has to do with how decisions get made.
The key points
On probability, he traces its origin to gamblers writing to mathematician friends to ask the odds on certain dice totals: get it right and you win, get it wrong and you lose over the long run. Probability later spread to stock markets, to whether a drug will work, to any system that can’t be modeled fully from first principles. Then he adds a boundary. Probability is most effective when an event repeats, when there are thousands of trials and the odds can be measured. For something that happens once a century, probability may not be the right mathematics; the mathematics of very rare events is still being developed, and mathematicians are still looking for something better than probability there. He also says that sometimes we simply don’t know the odds.
On analysis, he uses the casino. With finite capital, no betting strategy can beat the house over the long run; that’s a theorem. Doubling your bet after every loss looks like a scheme that only needs one win to recover, but that’s because it assumes infinite money. What the strategy really does is compress the risk of loss into an event of tiny probability, and one day you’re betting millions of dollars, and you’re bankrupt. Analysis, he says, is the discipline that lets you see what those tail risks look like and reason about infinity without falling into paradox.
On dynamics, he uses the Los Angeles freeway. Each car does its own simple optimization, and the whole produces compression waves. After an accident is cleared, the obstacle is gone, but the wave takes hours to dissipate, which is why he often finds himself in a jam with no accident anywhere. He also distinguishes stable from unstable equilibrium: a hanging pendulum returns when nudged; an inverted pendulum is technically in equilibrium too, but any small push topples it. Telling which kind you’re standing on matters.
The fourth is a piece of history he brings up while discussing AI. When Copernicus proposed his heliocentric model, it was less accurate than the geocentric model of the day; it only won after Kepler replaced circles with ellipses. If Kepler and Copernicus had had AI, he says, the model that was heading toward the right answer might have been discarded because its early predictions were worse than the old one’s. Fitting the data isn’t the only criterion, and science doesn’t always get immediate feedback. He also names the overfitting risk that comes from AI moving too fast: building a complicated model unrelated to the real mechanism that fits the data extremely well and can’t extrapolate.
In the same stretch he says something I wrote down: every component getting faster doesn’t mean science as a whole accelerates. The effort might go into optimizing the wrong target, with success on paper and no real progress.
Where my thinking went
I held these four things up against two numbers of my own.
The first is my settlement fan. Every day I draw a fan of where the Taiwan index might settle, and my earliest version used the at-the-money straddle as one standard deviation and assumed a normal distribution. In August I checked it against 113 real settlements: the normal curve said 79% of days would land within one straddle, and the actual figure was 54% (as of 2026-08-23, from my own settlement records). The tails are much fatter than normal, and the fan was drawn too narrow. That’s exactly the boundary he describes: settlement happens every week, the sample is large, probability is the right tool, and I’d chosen the wrong distribution. Switching to empirical quantiles didn’t just fix the width; it revealed a skew I hadn’t known was there.
The second is my prediction ledger. Since August I’ve been writing my judgments as cards, with the criterion and expiry date fixed in advance and a machine checking the answer at expiry. Nine cards so far: one right, one wrong, seven not yet due (as of 2026-09-06). The wrong one is from August 24: five sources simultaneously said the option sellers’ buffer on the Taiwan index was exhausted, and settlement went the other way by more than 1,300 points. Taken apart afterward, four of the five sources were reading the same batch of open-interest data, and the fifth was answering a different question altogether. There was one independent piece of evidence.
Put those two numbers next to his four points and I see one shape. The settlement fan sits on the “things repeat” side: probability applies, and what was wrong was my distribution. The failed card sits on the “happens once” side: I stacked five sources into what looked like high confidence, which is doubling down in another costume, squeezing the risk into “all five sources wrong at once,” an event I assumed was tiny, when those five sources fail together.
As for the traffic jam, I thought about positions I hold. Bad news comes out, the price drops, the news is clarified, and the price keeps dropping; I used to read that as the market being irrational. His framing offers another reading: it’s a compression wave, and once the accident is cleared it takes its own time to dissipate, unconnected to the accident itself. That’s not a reason to catch falling knives. It’s a reminder that when the cause is gone and the price is still moving, ask first whether this is a delay effect, instead of rushing to find a new cause.
What this means if you’re the one deciding
The question underneath is probably: which of my decisions should I think about with probability at all. My answer is to ask first how many times this thing has happened.
Where does the habit of reaching for probability whenever there’s uncertainty come from? He gives the origin himself: seventeenth-century gamblers writing to mathematicians about dice. Back then it was right, because dice can be thrown hundreds of times a night and getting the odds right is a long-run win. Today’s situation is that we point the same ruler at once-a-century things: whether an industry gets disrupted, whether a company survives, whether a policy turns. Each of these happens once for each of us, there aren’t thousands of trials to measure odds against, and he says it plainly: probability may not be the right mathematics there.
If not probability, then what? His section on analysis points the way: look at the tail. Not “how likely is it,” but “what does the worst case look like, and can I survive it.” The error in doubling down isn’t a miscalculated probability; it’s that ruin is hidden in a tiny corner while your capital is finite. My failed card was the same. I assumed five sources failing together was near-impossible, but they fed on one dataset, and failing together wasn’t a small probability, it was one event.
The third reminder is his Kepler extension. A correct idea can perform worse than a wrong one on early data, and if you cull ideas by “how did the last few turn out,” you keep the one that fits the past and can’t extrapolate. My ledger sets long expiry dates for exactly this reason: I don’t want three days of results killing a judgment that needs three months to show.
One more thought. The line that stopped me most was his description of how solving actually feels: never a flash of insight, but eliminating methods that don’t work one by one until the path appears out of what’s left, and the feeling is always “how did I take this long to see it.” My own factor testing has run two years, and most of what’s left is negative results: moving averages, dividends, long bullish candles, relative strength, each tested and each removed. I used to count those as failures. After this episode I count them as the negative space of the problem.
Sources worth checking
- The original video: Big Think, “One of the world’s greatest mathematicians explains 6 essential concepts of math,” with Terence Tao (published 2026-08-28, 1 h 24 min 36 s; title reproduced as on the original)
- Settlement-fan calibration: my own Taiwan index settlement records, 113 settlements, normal-curve expectation 79% versus 54% observed (as of 2026-08-23)
- Prediction ledger: my own judgment cards, 9 total, 1 right, 1 wrong, 7 pending (as of 2026-09-06); the wrong card dates from 2026-08-24
- My two earlier pieces on Tao: “A Numerator Without a Denominator” (2026-08-31) and “Terence Tao’s Coin Game” (2026-09-02)
- The wireless-and-sphere-packing and compressed-sensing histories he tells are not covered here; the second part of the original has them in full
One thing to take with you
The sentence this episode left me with: probability is only the right ruler when things repeat; don’t use it to measure what happens once. It took a card that five sources agreed on and that still came up wrong for me to see I’d been measuring two different kinds of thing with the same ruler.
Here’s something I’ve tried, if you want to try it too: next time you face a decision with real uncertainty, write on a sheet of paper how many times this has happened before, counting what you’ve seen yourself plus what the market shows. If you can fill in a three-digit number, think in probabilities and go find the historical distribution. If you can’t get to ten, switch to one question, “can I survive the worst case,” and write the answer next to it. A month later, go back to the sheet and see which decisions were made with the wrong ruler.