Do readers need sleep? A response to "Language Models Need Sleep"

We gave a small self-reading agent, Olvasó, both stages of the SLEEP paradigm and found that consolidation is essential, while each stage carries a risk the paper does not discuss. In a toy body whose world changes with the seasons, and over 5 lives per condition: Without consolidation, a change of world is devastating. After a winter, summer was not read again at all (2–6% of warm moments). A mind that never forgets mixes the seasons; a mind that doubts its changing meanings stops trusting them. Nightly consolidation into new long-term versions fixes it. Summer was read again on the first day back as well as in the first summer (47–48%), and winter three times better. This agrees with the paper. Pruning after consolidation tripled superstition (noise read 3.7% → 10.3%; 39% in one life), because consolidated meanings are no longer doubted. Dreaming from its own memory revives interpretation, and it is dream-like: a third of dream readings belong to the wrong world, against 1% for a real trickle of input. Learning from these unfiltered dreams did little harm (39% → 36%).