Shi-Botte Chatgpt cannot be trusted: he is lying in 37% of cases although it costs billions of dollars

< img src = "/uploads/blogs/29/fc/ib-fc-fcsscssc75_16240df4.jpg" Alt = "shi-bot Chatgpt can not < P > Interestingly that O3-mini from Openai, cheaper and reduced version of GPT-4O, “hallucin” in 80.3% of cases.

~ ~ < P > Using Simpleqa, the Facts Assessment tool, OpenAi acknowledged that its new large linguistic model (VMM) GPT-4.5 “hallucin”, that is, AI is fictional inventions in 37% of cases, & < P > Shi Model from a company that costs hundreds of billions of dollars, lies in more than one of the three answers it gives. Openai tries to present the problem of “lies” gpt-4.5 as something good, claiming that this chatbot does not “hallucin” as often as it will be from other companies.

< P > The graph shows how often a new Shi model is lying. It is also known that the GPT-4O, allegedly improved model of “reasoning”, “hallucin” in 61.8% of cases, which was found to be clarified by checking the facts of SIMPLEQA. It was found that O3-mini from Openai, cheaper and reduced version of GPT-4O, “hallucin” in 80.3% of cases.

< p > Of course this problem is not unique for Openai, writes media.

< p > “Currently, even the best models can generate text without” hallucinations “only in 35% of cases”, & mdash; Wenting Zhao, Doctoral Student at the University of Cornell. & mdash; “The most important conclusion from our facts testing is that we cannot yet fully trust the generation results.”

< p > In addition to distrust of a company that receives hundreds of billions of dollars of investments in products that have such problems, this is a lot of indicatives of the AI ​​industry, & mdash; What exactly sell us ? expensive, resource -intensive systems that should approach the human level of intelligence, but still cannot properly understand the basic facts, complain the authors of the material.

~ < P > Since Openai's Beenai ceases to grow in performance, the company obviously grabs a straw to restore interest in its chatbot, which was high in those days when Chatgpt first appeared. But for this we probably need to see a real breakthrough rather than even more lies, the authors summarized.