Openai launched at the end of February of this year a preview of the GPT-4.5 model, the most advanced in the company’s capabilities.
In a blog post, Sam Altman’s company has shared some of the most relevant data of this advanced model, compared to its previous versions, O1, O3-mini and 4th. In all cases, it exceeds the previous ones.
Thus, one of the values to highlight is that it presents a lower percentage in what has to do with the hallucination rate. As explained from Google Cloud on its page, it refers to the Incorrect or deceitful results that can generate these types of models.
This may be due to several factors, such as insufficient training, incorrect assumptions posed by the model to fill in the context or biases that the data causes when training the AI.
In the case of the GPT-4.5 model, OpenAI places this hallucination rate at 37.1%; It continues to be a somewhat high figure, but which is significantly reduced with respect to the others – here, lower hallucination rate implies better model performance in this area.
In comparison, the O1 model stays at 44%, followed by 4o, with 61.8%and the lightest, O3-mini, which is placed at 80.3%. In this way, GPT-4.5 allows to reduce this hallucination rate to O3-mini to more than half.
Additionally, this rate is directly related to a deeper knowledge of the world For natural language or LLM models, in addition to the precision rate. In this case, GPT-4.5 offers a 62.5%precision rate, quadruplying that of O3-Mini, 15%.
For the training of this advanced language model, OpenAI has used Microsoft Azure AI superorders.
“The result is a model that has a broader knowledge and a deeper understanding in the world, leading to the reduction of hallucinations and more reliability in a wide range of topics,” the company confirmed in its blog.
In the section of the hallucinations, several errors must be taken into account that have historically given some of the LLM. For example, as you share on its IBM page, Sidney, a Microsoft AI, admitted to falling in love with users, in addition to spying Bing employees.
A fact that is aggravated in applications as a health model, which could incorrectly identify symptoms, for example, leading to unnecessary medical interventions. At the moment, it seems that the last Chatgpt model has improved this considerably.
Know How we work in NoticiasVE.
Tags: Artificial intelligence