Conjecture Over Induction
Create knowledge by guessing boldly and killing errors, never by extrapolating the trend.
- Difficulty
- Moderate
- Time to result
- ~ongoing to results
- Steps
- 6
- Confidence
- 85%
The mechanism has two moves. First, creative conjecture: you guess an explanation of what is going on, and this act is genuinely creative — it brings something into existence that was not there before, which is precisely why the growth of knowledge cannot be predicted. If you could predict the invention, you would already have invented it. Second, criticism: you eliminate the guesses by experimental refutation or by direct criticism showing the explanation is bad, and keep what survives. This is the same variation-and-selection engine that runs evolution and that Edison and Tesla ran by hand. Induction runs the process backwards, cataloguing what happened and assuming it continues — which correctly predicts water heating from 20 to 30 degrees and then confidently predicts a thousand degrees, missing the boiling plateau entirely.
Origin
Naval Ravikant and Brett Hall work through David Deutsch's and Karl Popper's rejection of induction, contrasting it with Bayesian updating and illustrating it with the black swan, the boiling water experiment and the turkey.
Core principles
- 01New knowledge comes from creative guesses, never from extrapolating past observations.
- 02Explanation is the goal; prediction is what falls out of a good explanation.
- 03Error correction, not accumulation of confirmations, is what makes a process scientific.
- 04Every complex system that improves — evolution, markets, invention — runs on bold variation plus ruthless selection.
- 05Bayesian updating is powerful inside a known set of options and useless for generating new ones.
How to run it
- 1
Start from a problem, not a dataset
Name the thing that is unexplained or the conflict between what you believe and what you observe. Einstein was not extrapolating past phenomena; he was chasing specific problems in physics.
Pro tip Curiosity about one concrete problem beats general intelligence here — that was Einstein's own account of the difference.
- 2
Guess boldly
Invent candidate explanations of the mechanism, including ones that contradict the obvious reading of the observations. This step is imagination, and there is no method that produces it.
Watch out Do not restrict yourself to the theories already on the table — the set of possible explanations cannot be enumerated, which is exactly where Bayesian and Solomonoff-style schemes break down.
- 3
Ask what would break each guess
Derive the risky, narrow consequences of each conjecture — what it forbids from happening. Vague consequences mean you cannot select between guesses.
Pro tip Prefer the guess whose failure would be most obvious and most embarrassing.
- 4
Kill guesses by experiment or criticism
Run the test where you can, and where you cannot, criticise the explanation directly — many bad explanations can be discarded on their own terms without any data at all.
Pro tip Criticism is cheaper than experiment and does most of the work; save experiments for cases with two live rivals.
- 5
Hold the survivor and keep hunting for error
The remaining explanation is your best current knowledge, not the end of the process. Every theory is conjectural and each replacement explains more of the world than the last.
Watch out Treating the survivor as settled restarts the exact mistake induction makes.
- 6
Audit your forecasts for hidden extrapolation
Scan any prediction you are relying on and ask whether it is an explanation or a line drawn through past points. Pessimistic forecasts especially tend to be linear extrapolations of negative trends that ignore what creativity will add.
Pro tip The tell is a forecast that has no mechanism attached to it.
In the wild
Put a beaker of water on a constant heat source and record the temperature each minute. It climbs by roughly ten degrees a minute, a beautifully straight line. A thoroughgoing inductivist extrapolates and concludes that after two hours the water will be at a thousand degrees. What actually happens is that the temperature stalls at boiling point and stays there while the water boils away. No amount of data collection before the event could have produced that answer — only an explanation involving kinetic energy, escape velocity and latent heat tells you the plateau exists.
→ The extrapolation fails completely while the explanatory account predicts both the rise and the plateau.
Edison and Tesla did not derive their inventions from records of what had worked before; they made creative guesses and tried them out, discarding the failures. Evolution does the same thing without a mind, generating random variation and letting natural selection filter out what does not work. Both are bold guessing followed by elimination of error, which is why the same engine shows up across knowledge creation in general — the guessing supplies the novelty and the selection supplies the reliability.
→ A single variation-and-selection mechanism explains invention, biological adaptation and scientific progress alike.
A turkey is fed generously every single day and accumulates an unbroken record supporting the theory that the farmer is benevolent. The record is longest and the confidence highest on the day before Thanksgiving. Likewise, European naturalists observed white swan after white swan and concluded all swans are white, until someone reached Western Australia. In both cases the number of confirming observations was irrelevant, because the theory had no mechanism explaining why the pattern must hold.
→ Confirmation counts give no protection; only an explanation of the mechanism tells you when a pattern will break.
Common mistakes
Mistaking updating for discovery
Bayesian reasoning revises weights across options you have already listed, which is genuinely useful in medicine or a Monty Hall problem. It cannot invent the explanation that was never on the list, and treating it as a knowledge-generation method quietly caps what you can find.
Linear extrapolation of pessimistic trends
Doomsaying forecasts are usually straight lines drawn through negative data with the effect of future creativity left out. They feel intellectually serious and have an extremely poor track record.
Collecting confirmations instead of seeking refutations
Cataloguing events that agree with your theory adds no information about whether it is true. Progress comes from the test that could have killed it and did not.
Is it for you?
Best for
Researchers, founders and inventors working on genuinely open problems where the answer is not already inside the data they hold.
Not ideal for
Well-defined, finite, closed problems where the option set is already known and statistical updating is the correct and cheaper tool.
From the transcript
“science is not about cataloging a history of events that have occurred in the past and presuming they're going to occur again in the future…”
“they make bold guesses and then they weed out the things that didn't work”
“your best theories are going to be creative guesses not simple extrapolations”
From the episode
The Beginning of Infinity, Part 1