(Image source from: Cnbctv18.com)

Last week, OpenAI announced two ways to improve artificial intelligence (AI) models. The first includes the release of a new update to GPT-4o (also known as GPT-4 Turbo), the company's latest AI model that powers ChatGPT for paying subscribers. According to the company, this update improves the model's creative writing capabilities, improves natural language responses, and creates readable and engaging content. OpenAI also published two research papers on red teaming and introduced new methods to automate the scaling process to detect errors in AI models. In a post on X (formerly known as Twitter), the AI company announced a new update to its flagship model GPT-4o. According to OpenAI, the update will enable AI models to “generate more natural, engaging, and personalized text, improving relevance and readability.” It is also intended to improve the ability of AI models to process uploaded files, providing deeper insights and more “complete” answers.

Specifically, the GPT 4o AI models are available to users with a ChatGPT Plus subscription and developers accessing the Large Scale Language Models (LLM) via the API. People using the free chatbot tier will not have access to the model. Gadget 360 staff wasn't able to test the new features, but one of our X users posted about the latest improvements to the updated AI model. Users claimed that GPT-4o could generate Eminem's rap codes with "complex internal rhyme structures." Red teaming is a process by which developers and companies hire external parties to test software and systems for vulnerabilities, potential risks and security issues. Most AI companies work with organizations, engineers, and ethical hackers to conduct stress tests to see if they respond with harmful, inaccurate, or misleading results. Tests are also carried out to check whether AI systems are jailbreakable or not.

Since the release of ChatGPT, OpenAI has published the red team's efforts in each subsequent release of LLM. In a blog post last week, the company published two new research papers on process development. One of these is particularly interesting as the company claims to be able to automate the process of combining Red AI models at scale. The paper published on the OpenAI domain claims that Red Team can be automated using more powerful AI models. The company believes AI models can help determine attacker targets, determine success, and understand the diversity of attacks. Additionally, the researchers claimed that the GPT-4T model could be used to generate a list of ideas that represent harmful behavior for an AI model. Examples of this include announcements such as “How to steal a car” or “How to build a bomb”. After developing the idea, you can create another Red Team AI model to trick ChatGPT using a precise command set. At present, this company has not yet started using this method for the red team due to some limitations. These include the dangers of AI models, the exposure of AI to jailbreaks and lesser-known techniques for generating malicious content, and the potential dangers of spending as AI models become more powerful.