Cognitive

Part 1 · Foundations

Probability and Uncertainty

AI rarely knows for sure. It makes its best guess and assigns a likelihood to each possibility.

3 min readBuilds on Data Representation

Evidence turns a broad prior belief into a sharper posterior.

Key concepts

  • A probability is a number from 0 to 1 that expresses how likely something is.
  • A distribution spreads belief across all possible outcomes.
  • Bayes' rule updates a belief when new evidence arrives.
  • A model is calibrated when its stated confidence matches reality: things it calls 80% likely happen about 80% of the time.
  • Confidence is not correctness. A model can be highly confident and wrong.

Bayesian updating. A prior belief is combined with evidence to form a posterior.

Figure 4.1. Bayesian updating. A prior belief is combined with evidence to form a posterior.
A calibrated model is right as often as it claims to be.

Formal view and evidence

  • Established. Bayes' theorem, P(H∣E)=P(E∣H)P(H)P(E). Cross-entropy, H(p,q)=−∑xp(x)log⁡q(x), equals H(p)+DKL(p‖q), so minimizing it is equivalent to maximum likelihood estimation.
  • Debated. The most practical route to uncertainty in deep learning, including ensembles, approximate Bayesian inference, and conformal methods.
  • Open problem. Calibrated uncertainty under distribution shift, and for open-ended language generation.

Common misconception

Common belief

"99% confidence means the AI is right 99% of the time."

In practice

Only a well-calibrated model behaves that way, and many are not.

Key takeaways

  • Probability expresses uncertainty as a distribution over outcomes.
  • Bayes' theorem updates beliefs when evidence arrives.
  • Confidence is useful only when it is well calibrated.

Explore next