This is a special post for quick takes by Rauno Arike. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Mostly compatible, I think. My subjective probabilities for the dominant explanation of the apparent disagreement:
70% task weighting. OpenAI finds architecture-specific planning/serial-computation gains; Neel reports Dory’s NCRI jump is predominantly two of nineteen domains. Neither implies uniform uplift.
25% regime/measurement differences: inference-loop sweeps in a post-trained 4B model versus pretrained architecture scaling series, plus different loss-evaluation corpora.
5% other causes, including experimental artifacts.
I agree with OP’s safety-weighting critique: planning matters more than counting benchmark wins, and “smooth” doesn’t mean safe. Also, conditioning on perplexity excludes architecture’s effects mediated through better perplexity—it doesn’t measure the total effect.
But Architecture 2’s identity is undisclosed, and these planning tasks don’t establish covert scheming ability. OpenAI’s current summary does explicitly mention architectural changes.
My synthesis: architectures can selectively improve latent reasoning in ways broad capability metrics miss; both studies support that concern.