What's this about?
Mollick's most-repeated observation: Astra holds a better model of what *you* mean — far fewer stray references to earlier drafts, less drift across long sessions.
What exactly is theory of mind here?
Tracking the user's intent as distinct from the literal conversation record — knowing that draft three supersedes draft one without being told. Weak theory of mind is why long sessions degrade into confusion about which version is current.
Why is this worth noting?
It's invisible to benchmarks and enormous in daily use — the difference between re-explaining context every hour and not. Fitting that the adoption researcher, not the benchmark chart, is who surfaced it.