Predictions

A prediction is the finalized computed output generated by an AI model for an input, frequently expressed as normalized probabilities across candidates or continuous score tensors.

Real-World Mental Model

“A weather forecaster stating: "There is an 82% probability of rain in Seattle tomorrow."”

INTERACTIVE ENGINE: SOFTMAX DISTRIBUTION

Live Reactive

Raw Logit z₁ (Class A)2.5

Raw Logit z₂ (Class B)1.0

Raw Logit z₃ (Class C)-0.5

Normalized Probability Distribution: Σ p = 1.00 (100%)ArgMax: Class A

Class A: p₁ = e^z₁ / Σ e^z78.6%

Class B: p₂ = e^z₂ / Σ e^z17.5%

Class C: p₃ = e^z₃ / Σ e^z3.9%

CORE PRINCIPLES

What You Must Understand

Raw model outputs (logits) are converted to probabilities using Softmax or Sigmoid functions.

Argmax selects the highest probability candidate.

Calibration measures whether a 90% confidence prediction is actually correct 90% of the time.

Reference Implementation

// Softmax to probability
const probabilities = softmax(logits);
const predictedClass = argmax(probabilities);

Production Failure Mode

Treating model confidence as a proof of factual accuracy: a hallucinating LLM can assign 99% confidence to a fabricated claim.

View Full 26-Topic Curriculum Table

Loading AI Engineering Foundations...