GPT-5.5 versus GLM-5.2 hallucination post drives benchmark-methodology debate

Category: AI
Status: discussion
Sources: blog post, discussion
Summary: A blog post (377 points) cites the Artificial Analysis AA-Omniscience benchmark to claim GPT-5.5 answers incorrectly far more often than the MIT-licensed GLM-5.2 when it lacks knowledge, arguing larger models are worse at saying "I do not know." The benchmark scores correct answers +100, incorrect answers -100, and abstentions 0.
Comments: HN commenters dispute the framing. The hallucination rate is conditional on the model not knowing the answer, so it does not compare across models with different base accuracy; scoring that rewards only correct answers incentivizes guessing over abstaining; and raw parameter count is confounded by training method, inference-time compute, and quantization.
Why it matters: Hallucination-rate comparisons hinge on benchmark design, and the conditional metric undercuts a simple "bigger is worse" conclusion when teams pick coding and assistant models.