According to OpenAI, the researchers found that the GPT-40 comes with 'medium risk' evaluating from all 4 categories (Image:X @OpenAI)

Since OpenAI launched its GPT-4o in May, it has been discussed that the GPT chatbot may carry a possibility to create unauthorised content by cloning an individual’s voice, or other chunks using the reproduced audio having copyright. The company itself assigned a team to scrutinise the weakness in the system and now declares the result of the framework.

The System Card focuses on evaluating the novel risks presented by GPT-4o’s audio capabilities as well as the guardrails we implemented to prevent the generation of harmful, biased, or copyrighted content, and to ensure the model only generates audio in one of the preset voices.

— OpenAI (@OpenAI) August 8, 2024

According to OpenAI, the researchers found that the GPT-40 comes with ‘medium risk’ evaluating from all 4 categories including cybersecurity, model autonomy, persuasion and biological threats. The research further noted that the GPT-4o doesn’t possess too many risks. Earlier in July,  OpenAI announced its new GPT-4o mini, known as the most cost-effective mini model. GPT-4o mini may support text and vision in the API but aiding more efficiency to text is on its experiment.

Also, read| Sam Altman’s OpenAI Set To RollOut ChatGPT Advanced Voice Mode Soon

As per the records, GPT-4o mini comes with a context window of 128K tokens which will hold around 16K output tokens on every single request. The company claims that the new model will meticulously affordably handle the non-English language. Even though it is called ‘mini’ the company spotlighted its capacity to compete the small rivals as well as the larger model with its potentiality.

The company stated that the GPT-4o surmounted GPT-3.5 and other models across the multimodal reasoning benchmark. The new one will be available to ChatGPT’s Free, Team and users starting in the coming week. Apart from that, the company has already attempted to train advanced language models to generate text so that the fragile models can be easily identified which is further made to ease the text for human evolution.

Also, read| OpenAI Rolls Out New GPT-4o Mini; The Most Cost-Effective Model