What's this about?
A LessWrong post works through how much the recurrent-depth architecture actually degrades oversight, and lands on "significantly, but not totally" — some reasoning still surfaces as text. The HN thread (94 points, 50 comments) argues a narrower technical question: whether more internal loops must always improve capability, or whether recurrence hits diminishing returns.
What exactly is LessWrong?
The long-running rationalist forum where much of AI safety discourse happens in public — the place arguments get made carefully, before or instead of appearing in papers.
Why is this worth noting?
It's the measured counterweight to the alarm: the concern is real but bounded, and the load-bearing question — how much reasoning escapes into the visible text — is empirical and testable. Worth reading before adopting either the panic or the shrug.