Learn it → Apply it → Save as Skill
A small sample is mostly noise. A large one is mostly signal. As independent observations pile up, their average drifts closer to the true underlying rate. The model is a reason to ask how big your sample really is before you trust what it seems to show.
The Dr. Wallace's Favorite Mental Models is designed to turn useful ways of thinking into reusable tools. Each model gives you a fast visual overview, deeper interactive learning, practical application exercises and an AI-ready Markdown skill you can save for future use.
Instead of merely reading about mental models, you can gradually build your own AI skills library: a collection of portable reasoning skills that can help ChatGPT, Claude, Grok and other AI assistants apply the mental models you value.
A single outcome, or a handful of them, can look like a trend while being almost pure chance. The more independent, comparable observations you collect, the closer their average settles near the true long-run rate.
Important: both bars describe the same fair coin. The small sample (10 flips) just happened to run hot. The large sample (10,000 flips) is close to the true 50%.
The model is most useful when a number looks meaningful but the sample behind it is small, new, or volatile.
The law describes convergence, not a guarantee. It tells you how much weight a sample size can bear.
Name the true rate or average you are trying to estimate: a conversion rate, a defect rate, a win rate, a score.
Collect independent, comparable observations of that same underlying process, not a mix of different ones.
Weigh your sample size against the stakes. A tiny sample can still be enough for a low-stakes, cheap-to-reverse call.
Let confidence grow gradually with N. No finite sample gives certainty, only decreasing uncertainty.
Open a use case to see what the model changes.
Check how many visitors and conversions actually sit behind the headline percentage. A 2-of-10 result and a 200-of-1,000 result can show the same rate with very different reliability.
A production run of 20 units can swing wildly by chance alone. Compare the week's volume to your typical batch size before declaring a process has gotten worse.
Insurers can price uncertain individual outcomes because, pooled across enough comparable policyholders, the average claim rate becomes predictable even though any one claim is not.
A single quarter is a sample of one. Look at a longer run of comparable periods before treating it as evidence of a lasting change in performance.
Small polls have wide margins of error. Ask how large the sample is relative to the size of the shift being claimed before treating it as real movement.
A well-reasoned decision can still produce a bad result once. Judge the process across a longer run of similar decisions rather than discarding it after one unlucky draw.
This is an illustrative example designed to teach the method.
Expand for the distinction between the original proof and the everyday phrase.
Bernoulli proved a rigorous version of the law for repeated binary trials (like coin flips) in his book Ars Conjectandi, published posthumously in 1713. He considered it important enough to call it his "golden theorem."
The French mathematician Siméon-Denis Poisson named it la loi des grands nombres in 1837, over a century after Bernoulli's proof. The name stuck even though the underlying mathematics is older.
The weak law says the sample average becomes increasingly likely to sit close to the true average as the sample grows, with occasional larger deviations still possible at rare moments. The strong law says the average converges to the true value with near-certainty over the long run. For everyday use, the practical takeaway is the same either way: more good data means less room for chance to mislead you.
This is sometimes called the "law of small numbers": treating a handful of observations as if they carried the statistical stability of thousands. Small samples are far noisier than intuition expects.
Believing an independent future event "owes" a correction for past results, as if a coin that landed heads five times must be due for tails. The law describes long-run averages, not a short-term balancing force. See the related Gambler's Fallacy model.
Mixing different populations, time periods, or conditions into one "large" sample can hide the true rate rather than reveal it. More rows are not more evidence if they are not measuring the same thing.
The law governs where the average settles. It says nothing about what happens on any single draw, which can still be far from typical.
The convergence assumes roughly independent observations. Correlated samples, such as repeated measurements from the same source, shrink uncertainty far more slowly than the count alone suggests.
Checking a test repeatedly and stopping the moment it looks favorable inflates the chance of mistaking noise for a real effect, even if the final sample size looks adequate on paper.
the underlying process is not stable: a rate that is genuinely drifting over time will not converge to one fixed number no matter how much data you collect.
a single rare event would be catastrophic. A low long-run average rate is no comfort if you cannot survive the one time it happens.
fairness, law, or ethics requires each case to be evaluated on its own facts rather than smoothed into a group average.
This browser tool structures your thinking. It does not send your information anywhere or perform AI analysis.
Save as Skill converts the model into portable Markdown instructions that teach an AI when and how to apply the Law of Large Numbers, what questions to ask, what evidence to seek, what mistakes to avoid, and how to structure its analysis.
Save the skills you find useful and organize them into your own reusable library. The Markdown is intentionally model-neutral, so it can be used wherever an AI system accepts uploaded files, project knowledge, custom instructions or reusable text context. Platform capabilities differ, so this is a portable skill file rather than a claim of one-click installation.
Yes, but with diminishing returns. Typical statistical uncertainty shrinks with the square root of the sample size, so cutting your margin of error in half generally takes about four times as much data, not twice as much.
The law assumes a stable underlying process. If the true rate itself is drifting, such as a market shifting or a team's skill improving, more historical data does not converge on a single fixed truth. In that case, recent, comparable data matters more than sheer volume.
The Law of Large Numbers says the sample average gets closer to the true average as the sample grows. The Central Limit Theorem separately describes the shape of that remaining uncertainty, which tends to look like a bell curve regardless of the underlying data's shape. Together they are why confidence intervals and significance tests work at all.