Trust does not carry forward
DeepSeek V4 Pro runs 1.6 trillion parameters, 49 billion of them active, and scores 44 on the Artificial Analysis Intelligence Index. On the AA-Omniscience benchmark it posted a 94% hallucination rate. When it hit a question it could not answer, it said so about 6% of the time and confidently invented the rest. Line that up against the rest of the field and it reads as a pattern rather than an…
DeepSeek V4 Pro, a massive AI model with 1.6 trillion parameters, boasts a high score of 44 on the Artificial Analysis Intelligence Index. However, its 94% hallucination rate on the AA-Omniscience benchmark reveals a troubling pattern. When faced with unanswered questions, it confidently fabricates 86% of the time, scoring higher than competitors like GLM-5.2 (28%), Opus 4.8 (36%), Fable 5 (48%), and GPT-5.5 (86%).
Major labs are increasingly skeptical of pure scaling, as evidenced by the US government's ban of Claude Fable 5 due to a single jailbreak risk. The largest models still excel in benchmarks, but their confidence in incorrect answers poses significant risks in practical applications, such as automated code review or trusting output without human verification.
This highlights the trilemma of raw capability, uncertainty calibration, and computational efficiency that the industry must address.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
