AI Capabilities And Limitations

A mental model of generative AI's machine-side behavior, complementing the human-side AI Fluency Framework (4Ds). Covers the two-stage training process and the behavioral fingerprints it leaves (sycophancy, verbosity, over-caution, loose confidence calibration), and the four properties that sit on a capability-to-limitation spectrum - Next Token Prediction (hallucination, fabricated specifics), Knowledge (training cutoff, staleness, uneven coverage, source amnesia), Working Memory (the fixed context window as a cliff, not gradual decline, and why it doesn't learn from your corrections), and Steerability (pattern-matched instruction-following, reasoning drift, letter-over-…