The latest models of OpenAi, O3 and O4-Mini, more frequently hallucinate than the old company AI systems, according to internal tests and third-party research. On the Personqa of Openai reference, O3 hallucinated 33%of the time – double the rate of older models O1 (16%) and O3 -Mini (14.8%). The O4-Mini still made it possible to perform, hallucinating 48% of the time. The AI non -profit AI Laboratory has discovered the O3 manufacturing processes that it claimed to use, including the current code on a MacBook Pro 2021 “outside of Chatgpt”. Stanford’s auxiliary professor Kian Katanforoosh, noted that his team had found that O3 frequently generates broken website links.
OPENAI says in its technical report that “more research is necessary” to understand why the hallucinations aggravate as the reasoning models increase.
