OpenAI continues to develop what has become a college student’s best friend: ChatGPT. This week, they released a preview of their newest development, GPT-4.5, claiming that it is their largest and best model for chat yet. So, let’s break down what GPT-4.5 has to offer, what improvements it has made since GPT-4.0, and where it falls short of other AI models.
What’s new?
The largest selling point of GPT-4.5 is its focus on using as many resources as possible. OpenAI tends to have a maximalist approach to machine learning, intentionally creating models that are bigger and bigger. Their goal with GPT-4.5 is to move toward something they call unsupervised learning, in which the model is given copious amounts of unlabeled information and is forced to find patterns and make sense of the resources. OpenAI hopes that through unsupervised learning AI models can better understand the nuances of human interactions, especially when it comes to chatting.Â
Right now, the GPT-4.5 demo is only available to ChatGPT Pro subscribers (for the low price of $200 a month!); these subscribers have already reported noticeable improvement in the warmth and humanness of the chatbot. OpenAI shared an example of when a user told the chatbot they failed a test. Instead of offering unsolicited advice, like previous versions of GPT often did, the chatbot says “Want to talk about what happened, or do you just need a distraction? I’m here either way.” How freaky is that??? We’re reaching a point where AI chatbots can have more empathy than some of our human friends! This is a direct result of the unsupervised learning being employed by OpenAI.
However, with larger models come greater environmental impacts. OpenAI has declined to answer questions about the specifics of its size and resources used. Any time AI is used, it draws on a lot of energy and water resources. To me, OpenAI’s refusal to give specific details about the size of GPT-4.5 tells me that they understand the environmental complications but don’t want the consumer to be acutely aware of these issues. AI users need to be aware of how their habits affect the environment, and with OpenAI continually making larger models, its users need to be aware of the size of the model’s impact.Â
How has it improved?
Users commonly agree that the biggest flaw in AI models, especially earlier versions of GPT, is constant hallucination. Hallucination refers to when AI models come up with false or misleading information and portray it as true. This problem arises when AI doesn’t have enough resources to be trained. It often takes things out of context or misinterprets patterns within the resources. By creating a larger model focused on unsupervised learning, OpenAI is decreasing the possibility of hallucinations. Of course, they’ll still happen, but the frequency at which they do will decrease greatly. This is great news for college students who abuse AI on homework and even better news for those who use AI in the workplace, especially as AI moves into sectors where mistakes are not as acceptable, like medicine and banking.
How does it compare to other AI models?
OpenAI has another software system called SimpleQA that is used as a benchmark test for new AI models. Interestingly, when OpenAI tested GPT-4.5 they found that it fell significantly behind the o3-mini model, a small, cheap reasoning model released by OpenAI in December for subjects like science and math. However, the o3-mini is a model meant to excel in those realms, so that is expected. GPT-4.5 did excel beyond o3-mini in areas like chatting, reading, and writing. This really reinforces the idea that GPT-4.5 is a fantastic chatbot. GPT-4.5 has heightened “emotional intelligence” and a deeper understanding of the world it lives in rather than it does of school subjects, thanks to unsupervised learning.Â
OpenAI’s emphasis on GPT-4.5 understanding the way humans converse gives us a glimpse into the future of AI and chatbots. GPT-5 is set to release in a matter of months and it’s clear that future models will continue to grow larger and have more refined personalities. While it is a scary thought that AI is becoming more human-like, it is also impressive to see how machines can be conditioned to accurately act like us. OpenAI and their users, alike, have high hopes for GPT-5.