New Version Of Popular AI Passes Key Exams With Flying Colors, Still Produces ‘Biased’ Content
Open AI unveiled an upgraded version of ChatGPT Tuesday that is capable of passing key academic exams with ease but still has lingering issues with producing “biased” content, according to the artificial intelligence company.
GPT-4 was tested on multiple exams intended for humans including a simulated bar exam, which it passed with a score ranking “around the top 10% of test takers” compared to GPT 3.5, which ranked around the bottom 10%, according to Open AI. Like former models, GPT-4 still produces “biased” content and viewpoints, but it scores 19 percentage points higher than GPT-3.5 in factuality.
GPT-4 excelled at other exams it was tested on, scoring a 1,300 out of 1,600 on the SAT and five out of five on numerous Advanced Placement high school exams, which is a major improvement over GPT 3.5. The ChatGPT chatbot was initially released in 2022 and was a disruptive advancement in artificial intelligence, according to FOX Business.
Greg Brockman, president and co-founder of OpenAI, demonstrated some of GPT’s capabilities and limitations in a GPT-4 developer livestream on Tuesday. In the finale, he showed how GPT-4 could do taxes, pasting in 16 pages of tax code and asking the chatbot questions such as what the standard deduction and liability would be for a fictional couple.
GPT-4 also scored significantly higher in safety measures, decreasing responses to requests for prohibited content by 82%, according to OpenAI.
Previous versions of GPT were found to be less likely to detect and block hateful speech against whites, Republicans and men, according to research.
GPT-4 is currently only available to OpenAI users paying for ChatGPT Plus and as an API for developers, according to TechCrunch.
Content created by The Daily Caller News Foundation is available without charge to any eligible news publisher that can provide a large audience. For licensing opportunities of our original content, please contact firstname.lastname@example.org
openAI scored 1300 out of 1600 on the SAT, and yet it has computer speed access to all the internet. I would call that a pretty sad showing. My own interaction with this “program” was interesting, but I gave up trying to get it to correct itself after I corrected it multiple times and it apologized for making mistakes. A virtuoso advance in machine learning to be sure, but still not anywhere close to real intelligence, in my opinion.