OpenAI’s latest model, GPT-4.5, is the company’s most ambitious drop to date, promising improvements in anthropomorphic aspects such as intuition and emotional understanding. This research preview, available only to OpenAI’s paid subscribers, uses more computing power than its predecessor, GPT-4o, and demonstrates enhanced capabilities in tasks like writing compelling headlines.
While GPT-4o excelled at analyzing web articles on politics, GPT-4.5 showed better performance in understanding nuances and representing multiple viewpoints. The new model also improved in tasks that require a more conversational tone, making it feel less annoying and more human-like during chats.
However, the experience with GPT-4.5 was not without its limitations. Its long-form output was helpful but not surprising, and the interview questions it suggested were underwhelming. Moreover, the model struggled to provide in-depth answers when compared to a human reporter’s notes.
As OpenAI looks towards its next model, GPT-5, the company is exploring ways to recoup costs associated with training and running increasingly expensive models. One idea involves switching subscription offerings into a credit-per-use system or charging tens of thousands of dollars per month for access to premium tools.
With GPT-4.5, OpenAI has marked a new era in AGI research, focusing on the emotional and social aspects of human intelligence. While the experience was not without its flaws, it showed promise as a conversational AI that can make users feel less cloying during interactions. As the company navigates its dual goals of sustainability and superintelligence, GPT-4.5 serves as an important stepping stone towards realizing this vision.
Source: https://www.wired.com/story/gpt-4-5-openai-first-impressions