Induction Works Until the World Changes
You might think a skeptic should throw out induction, since the past can't prove the future. Don't. Induction is one of the most useful tools you have. The real question is when it breaks. A short course film for INFO 7375 Computational Skepticism for AI (Northeastern University, Nik Bear Brown), built on David Hume's…
Induction Works Until the World Changes
You might think a skeptic should throw out induction, since the past can't prove the future. Don't. Induction is one of the most useful tools you have. The real question is when it breaks. A short course film for INFO 7375 Computational Skepticism for AI (Northeastern University, Nik Bear Brown), built on David Hume's An Enquiry Concerning Human Understanding (1748): the sun has risen every morning of your life, and that proves nothing about tomorrow, yet only a fool or a madman disputes experience, and custom is the great guide of human life. The film follows one real history, the U.S. Census Bureau's quarterly share of retail sales made online: ten years of small, steady steps from about 4 percent in 2010 to about 11 percent at the end of 2019, then one giant step in the spring of 2020 when COVID changed how people buy. Nothing was wrong with the math. The system that made the history had changed. A model trained on that stationary past is exactly as confident as it was the day before.
Liam, in for Bear.
Course: INFO 7375 Computational Skepticism for AI · Northeastern University · Nik Bear Brown
Chapters
0:00 The question: when does induction break? 0:17 Three terms: induction, stationary, ergodic 0:28 Hume, Enquiry (1748): will the sun rise tomorrow? 0:45 Only a fool or a madman: custom, the guide of life 0:58 Ten years of online retail share, per the Census Bureau 1:15 What held: the rules, not the number (stationary) 1:30 Ergodic: one history stands in for the whole 1:42 Spring 2020: a decade of small steps, then one giant one 1:57 Nothing wrong with the math; the system changed 2:08 A model shipped into a moving present; keep induction, check the world 2:30 Your Turn
YOUR TURN
Pick one pattern your team or your model relies on, and paste this into Claude:
Here is a pattern my team or model relies on: describe it. What process in the world produces this pattern? What would have to change in that process for the pattern to stop holding? For each change, name one signal I could watch that would tell me it had happened.
Then check two things yourself:
- Could each change really happen?
- Could you actually watch each signal, starting this week?
Sources
- David Hume, An Enquiry Concerning Human Understanding (1748), Selby-Bigge edition, Project Gutenberg 9662. Section IV: that the sun will not rise tomorrow implies no contradiction; the course of nature may change; "none but a fool or madman" disputes experience. Section V: custom, "the great guide of human life".
- U.S. Census Bureau, Quarterly Retail E-Commerce Sales: releases for 1Q2010 (CB10-64), 4Q2019 (CB20-24), 1Q2020 (CB20-71) and 2Q2020 (CB20-120, August 18, 2020); and the revised series E-Commerce Retail Sales as a Percent of Total Sales (ECOMPCTSA) via FRED. The film rounds 1Q2020 to about 12 percent and 2Q2020 to about 16 percent.
- NIST/SEMATECH e-Handbook of Statistical Methods, section 6.4.4.2, Stationarity.
- O. Peters and M. Gell-Mann, Evaluating gambles using dynamics, arXiv:1405.0585 (ergodicity: time average equals ensemble average).
- Nik Bear Brown, INFO 7375 Computational Skepticism for AI, chapter 1, The Skeptic's Toolkit (the Hume move as the course teaches it; "a system trained on a stationary past, shipped into a moving present").
Every factual claim in this video was checked against these primary sources before rendering.
@NikBearBrown
Part of the Claude & Agentic AI series. Made with the brutalist.art toolkit: Kokoro for the voice, Manim for the drawings.
▶ The full series playlist — Brutalist https://www.youtube.com/playlist?list=PLG9H-C6rp5RU
What this video covers
- 0:00 — The question: when does induction break?
- 0:17 — Three terms: induction, stationary, ergodic
- 0:28 — Hume, Enquiry (1748): will the sun rise tomorrow?
- 0:45 — Only a fool or a madman: custom, the guide of life
- 0:58 — Ten years of online retail share, per the Census Bureau
- 1:15 — What held: the rules, not the number (stationary)
- 1:30 — Ergodic: one history stands in for the whole
- 1:42 — Spring 2020: a decade of small steps, then one giant one
- 1:57 — Nothing wrong with the math; the system changed
- 2:08 — A model shipped into a moving present; keep induction, check the world
- 2:30 — Your Turn
Transcript
A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.
The question: when does induction break?
[0:00] Konnichiwa, this is Liam in for Bear with a short film for Info 7375 Computational Skepticism for AI. You might think a skeptic should throw out induction since the past can't prove the future. Don't. Induction is one of the most useful tools you have. The real question is when it breaks.
Three terms: induction, stationary, ergodic
[0:17] Three terms: induction, expecting the future to resemble the past. Stationary, a process whose statistical rules don't change over time. Ergodic, one long history speaks for the whole system.
Hume, Enquiry (1748): will the sun rise tomorrow?
[0:28] The move comes from David Hume in an Enquiry Concerning Human Understanding from 1748. The sun has risen every morning of your life. Does that prove it will rise tomorrow? Hume said no. That the sun won't rise tomorrow is no contradiction. You can picture it. The past gives no logical guarantee.
Only a fool or a madman: custom, the guide of life
[0:45] But Hume didn't say stop. Only a fool or a madman, he wrote, would dispute the authority of experience. We can't prove the future will resemble the past and we live by it anyway. He called that custom, the great guide of human life.
Ten years of online retail share, per the Census Bureau
[0:58] And most of the time induction works. For 10 years, online shopping's share of US retail sales climbed slowly and steadily from about 4% in 2010 to about 11% at the end of 2019 for the US Census Bureau. Extend the line one more quarter and you'd land close quarter after quarter.
What held: the rules, not the number (stationary)
[1:15] Notice what held, not the number, it kept rising. The rules held. Each quarter added a small step, never more than half a point. A process whose statistical rules don't change over time is called stationary. These steps were close to it and that's what induction quietly relies on.
Ergodic: one history stands in for the whole
[1:30] Ergodic is the other half. In an ergodic system, following one shop over a long stretch of time gives the same average as looking across all the shops at once. That's what lets one history stand in for the whole.
Spring 2020: a decade of small steps, then one giant one
[1:42] Then in the spring of 2020, COVID changed how people buy. Stores closed or cut back and shopping moved online. In a single quarter, online's share jumped from about 12% to about 16 for the Census Bureau. A decade of small steps, then one giant one.
Nothing wrong with the math; the system changed
[1:57] Nothing was wrong with the math. The system that made the history had changed. Hume saw that nothing forbids it. The course of nature may change, he wrote, and the past may be no rule for the future.
A model shipped into a moving present; keep induction, check the world
[2:08] A model trained on the old history makes the same bet. The chapter calls it a system trained on a stationary past shipped into a moving present. After the change, the model is exactly as confident as it was the day before. So, keep induction. Most of the time, it's the right bet. The honest rule, trust the pattern while the system that made it stays the same. When the world changes, go back and check. Your turn.
Your Turn
[2:31] Pick one pattern your team or your model relies on and paste this into Claude. Here is a pattern my team or model relies on. Describe it. What process in the world produces this pattern? What would have to change in that process for the pattern to stop holding? For each change, name one signal I could watch that would tell me it had happened. Then check two things yourself. Could each change really happen? And could you actually watch each signal starting this week? Induction works until the world changes. @nickbearbrown
