prior probability
The credence assigned to a hypothesis before incorporating new data, based on factors like simplicity and fit with prior knowledge.
As defined in 200 | Solo: The Philosophy of the Multiverse
Where it comes up · 6
- 315 | Branden Fitelson on the Logic and Use of Probability4 claims
- 200 | Solo: The Philosophy of the Multiverse3 claims
- Dwarkesh Patel Podcast - The data black hole at the center of AI1 claim
- AMA | December 20201 claim
- AMA | May 20211 claim
- AMA | October 20221 claim
Most important claims · 10 of 11Search all claims for “prior probability” →
- In a scenario where you are told a transporter is random but you always appear on pad 1, the correct approach is Bayesian: you assign prior credences to hypotheses (random, always pad 1, etc.) and update via Bayes' formula with each observation; no matter how high your prior in randomness, enough pad-1 observations shift you toward 'always pad 1,' but your credence should never reach 0 or 1 since future data could always change your mind.AMA | May 2021
- Priors are not developed with no data whatsoever; our biology, psychology, education, upbringing, and evolutionary history all count as data feeding intuitive priors. Different people starting with different priors is a feature not a bug, and rationality lies in the reasoning process from those priors; with enough data good Bayesians converge on the same answers.AMA | October 2022
- Applying Bayes' rule to cosmological scenarios is stuck on the likelihood function because the only data is 'I exist'; it is unclear whether to ask the probability that any observers exist, observers like me, or exactly me, and whether one's own existence is genuinely new evidence or old evidence already baked into the priors (the problem of old evidence), making the update ambiguous.200 | Solo: The Philosophy of the Multiverse
- We should give essentially zero credence to Boltzmann-brain-dominated cosmologies not because we observe we are not Boltzmann brains (which is what a Boltzmann brain would say) but because the belief is cognitively unstable: if you use logic and reason to conclude you are probably a random fluctuation, you have no right to trust the very reasoning steps you used, since they too would be randomly fluctuated into your brain.200 | Solo: The Philosophy of the Multiverse
- The doomsday argument fails because the assumption that we are typical observers within the population of all humans is a bad assumption; you should favor theories in which it is more likely that you specifically come into existence rather than weighting prior probabilities by the raw number of observers, which would create complications in many-worlds with differently-weighted branches.AMA | December 2020
- For RL-based training to work, the model must already have some nonzero prior probability of producing the correct solution, which is why mind-stretching amounts of human expert trajectories are needed in every skill the model should master.Dwarkesh Patel Podcast - The data black hole at the center of AI
- There are two main ways to apply typicality: the world-first approach (self-sampling assumption) assigns a prior to each cosmological scenario and then assumes you are typical within it, so a single observer in a small universe is more probably you than any one observer in a big universe; the observer-first approach (self-indication assumption) assumes you are typical across all observers in all scenarios, boosting scenarios that contain more observers.200 | Solo: The Philosophy of the Multiverse
- While individuals may legitimately differ in their prior probabilities, they are not entitled to their own likelihoods, because the likelihood information (error rates / Bayes factors) is the objective, invariant component that can be determined in the laboratory independent of any individual's beliefs.315 | Branden Fitelson on the Logic and Use of Probability
- Even a highly reliable test with a large Bayes factor will not yield a high posterior probability if the prior is extremely low—a biologically male person getting a positive pregnancy test should not conclude he is probably pregnant because his prior probability is essentially zero, demonstrating that the multiplier alone cannot determine the posterior.315 | Branden Fitelson on the Logic and Use of Probability
- To judge which of two experiments (e.g., to detect a WIMP versus to exclude an axion as dark matter) is more useful, you need only compare the likelihoods—how probable the evidence would be under each hypothesis—not the priors, because the likelihood comparison reveals whether and in which direction the evidence favors one hypothesis, even though determining the resulting probabilities would still require priors.315 | Branden Fitelson on the Logic and Use of Probability