AI Snips
Chapters
Transcript
Episode notes
Two Kinds Of Probability
- Bayesian probabilities mix subjective belief and objective frequency, and conflating them causes errors in comparisons.
- Comparing probabilistic estimates for fundamentally different evidence bases (e.g., volcano frequencies vs. AGI risk) is misleading.
Feedback Distinguishes Near-Term From Long-Term Claims
- GiveWell-style interventions have ongoing feedback loops and falsifiable mechanisms that improve estimates over time.
- Longtermist claims lack comparable feedback, so numbers there have much weaker empirical grounding.
Prefer Arguments Before Numbers
- Avoid finalizing noisy reasoning by slapping precise probabilities on speculative claims; prefer argument-level evaluation first.
- Reserve numerical precision for measurement and counting where data and falsification exist.