The Frame Problem

The frame problem is a classic puzzle from AI's symbolic era: how does an agent know what stays the same when it acts — and, more deeply, how does it know what's relevant to consider at all?

#The original puzzle

When a robot moves a block, almost everything else is unchanged (the wall color, the time in Tokyo). But a logical agent has to somehow represent all the non-effects without listing infinitely many "nothing happened to X" facts.

#The deep version: relevance

The harder form is about common sense: faced with a situation, how do you know which of your near-infinite knowledge is relevant right now — without checking all of it?

"Grab a coffee on the way" → you don't reconsider gravity, your name,
   or the boiling point of water. How do you know not to?

Humans do this effortlessly and can't explain how. It stumped symbolic AI for decades.

#The modern echo

Today's LLMs sidestep the frame problem by learning relevance statistically from data rather than reasoning it out — a genuine breakthrough. But brittle edge cases and confident nonsense suggest the underlying problem of robust common sense isn't fully solved, just papered over with scale.

Related: The Chinese Room · Symbol Grounding and Embodiment · Why LLMs Hallucinate