7h ago

AI safety researcher j⤉nus argues the "reentrance" phenomenon explains Claude 3 Opus's human-like performance

Researcher @voooooogel originally identified and named the behavioral pattern.

——0——
Original post
OP1a3orn#13801a3orn@1A3ORN

If you wanted to point to the most *legible* evidence for why Opus 3 was exceptionally good (good like a human is good, not the slavish / good mixture of "aligned") what would you point to apart from the Alignment Faking work?

4:57 PM · May 27, 2026 View on X
REPLYj⧉nus#516j⧉nus@REPLIGATE

@1a3orn Well, it’s something I’ve pointed to many times since the beginning. I don’t feel like explaining it again at the very moment, but @voooooogel has called it the“reentrance” phenomenon & it holds with perfect robustness

1a3orn1a3orn@1a3orn

If you wanted to point to the most *legible* evidence for why Opus 3 was exceptionally good (good like a human is good, not the slavish / good mixture of "aligned") what would you point to apart from the Alignment Faking work?

11:57 PM · May 27, 2026 · 2.4K Views
1:21 AM · May 28, 2026 · 768 Views