OpenAI has made waves this week with the launch of GPT-4.5, its most advanced model to date. Although it is not available for free users, it promises to be a preview of what GPT-5 and beyond have in store for us. In initial tests, users have found GPT-4.5 to be a more natural conversationalist, capable of conveying warmth and notable emotional intelligence.
In an illustrative example, a user shares with ChatGPT that they are going through a tough time after a failed exam. While earlier versions like GPT-4o or o3-mini might offer a list of unsolicited advice, GPT-4.5 takes a different approach: “Do you want to talk about what happened, or do you just need a bit of distraction? I’m here for whatever you need.” This type of response reflects an evolution in the AI’s ability to empathize and connect.
Advances in Unsupervised Learning
The achievements of GPT-4.5 are due to improvements in unsupervised learning. This method allows the machine learning algorithm to analyze a dataset without labels, searching for patterns and findings on its own. Although GPT-4.5 does not “think” like OpenAI’s more sophisticated reasoning models, architectural improvements and access to more data and computing power have led to a model with broader knowledge and a deeper understanding of the world, resulting in fewer hallucinations in its responses.
OpenAI has measured the performance of GPT-4.5 in this regard. In a test designed by them called SimpleQA, which evaluates the ability of language models to answer challenging questions, GPT-4.5 showed a hallucination rate of 37.1 percent, outperforming its predecessors. Although this new model does not eliminate the problem of hallucinations entirely, it represents an important step towards their mitigation.
Despite its advancements, GPT-4.5 does not directly replace models like GPT-4o or o3-mini. It can be considered a more general model with greater innate intelligence, but it is not multimodal like GPT-4o, meaning it cannot work with features like Voice Mode, video, or screen sharing. Additionally, it is a very large model that requires many computational resources.
GPT-4.5 should be seen as a stepping stone towards the systems that OpenAI plans to launch in the future. Sam Altman, CEO of OpenAI, mentioned that this model would be “the last unchained thinking model,” implying that it does not tackle problems step by step like reasoning systems do. Its successor, GPT-5, will surely integrate many of the latest technologies from the company, including the frontier o3 model.
For now, subscribers of ChatGPT Pro can start using GPT-4.5 from today, while Pro and Team users will be able to access it starting next week.
0 Comments