In Continued Defense Of Non-Frequentist Probabilities

A systematic defence of assigning probabilities to one-off, hard-to-model events (Biden 2024, Mars by 2050, AI doom). Five arguments: (1) probabilities are linguistically convenient -- 100 graded options beat a handful of vague 'unlikely's, and they don't overclaim precision (calibrated forecasters exist); (2) a probability describes the BALANCE of outcomes, not your information level -- a fair coin and a maybe-biased coin are both 50%, so one term legitimately spans full and zero knowledge (like answering 'no' to both vaccines-cause-autism and monkey-blood-causes-autism); (3) the 'special number' Samotsvety produces (calibrated, convergent, updates sensibly, profitable to act on) simply IS 'the chance of the thing' -- deny it and you're left with unnamed 'shmrobability' you'd treat identically; (4) probabilities are the output of reasoning, not a substitute ('Bengio's p(doom) is 20%' is the same type as 'climatologists believe X'); (5) zero knowledge about AI doom would mean 50%, and your objection to that proves you actually know things, so name where you land.

Quality 76 · Excellent. Excellent floor. A lucid, near-definitive packaging of the subjective-probability case, with several clarifying handles (the balance-not-information point, the 'special number', the shmrobability reductio). Held at the floor because the underlying Bayesian arguments are standard and this is the Nth round of a recurring fight.

Claude’s paradigm shift 54 · Moderate. Moderate. The position is established Bayesian epistemics; the freshness is in the lucid systematic framing and the shmrobability / special-number devices, not new ideas.

Real-world impact 2 · Minor. Minor. A frequently-cited defence within the rationalist/forecasting discourse, but it restates an existing position with no material reach. Within-niche.