The left panel is a miniature Prolog: knowledge lives as explicit facts and rules, and answering a query means constructing a proof — every step is shown in the trace, including the dead ends it backtracks out of. It answers questions whose answers appear nowhere in the knowledge base (no grandparent fact is stored — the relation is derived), and when something is false, it says so, provably. The price: total rigidity. Misspell a name, ask about an undefined relation, and you get nothing at all.
The right panel is the statistical machine: the same knowledge written as sentences, learned by a word-level n-gram model that only tracks which words follow which. It answers in fluent language and handles questions the way an autocomplete would — by pattern. Where the pattern happens to encode the fact ("bob is the parent of …"), it is correct. Where the answer requires an inference step it has never seen written down, it confabulates — fluently, confidently, wrongly. It has no representation of who anyone's grandparent actually is.
This is the contrast behind debates about LLM "reasoning": modern models are incomparably better pattern machines than this toy, but the architecture of the question — proof versus plausibility — remains. Try the three scenarios, then ask your own questions on both sides.