OpenAI Unveils GPT-4.5 for More ‘Natural Conversation’

When OpenAI unveiled its new GPT-4 technology in late 2022, it astounded even seasoned A.I. researchers with its capabilities. From answering questions to composing poetry and generating computer code, GPT-4 showcased advancements that seemed ahead of its time.

Fast forward two years, and OpenAI has introduced its successor: GPT-4.5. This latest iteration marks a significant shift, as it is the final version of the chatbot system that lacks “chain-of-thought reasoning.”

With the release of GPT-4.5, OpenAI’s technology is expected to take more time to ponder a query before responding, mirroring human-like contemplation rather than offering instantaneous replies.

Although GPT-4.5, which powers the top-tier ChatGPT version, may not generate as much buzz as its predecessor, it promises to deliver a more “natural” conversational experience compared to earlier chatbot technologies.

Mia Glaese, vice president of research at OpenAI, highlighted the distinctive feature of the model: its ability to engage in warm, intuitive conversations and better understand user queries. She emphasized the enhanced conversational flow and user-centric approach of GPT-4.5.

In a recent development, OpenAI introduced OpenAI o1, a technology geared towards tackling tasks involving math, coding, and science. This initiative aligns with the broader industry trend of developing A.I. systems capable of reasoning through complex challenges.

The objective is to create systems that can methodically solve problems by building upon incremental steps, akin to human reasoning. These advancements hold particular promise for computer programmers leveraging A.I. tools for code generation.

These reasoning systems, like GPT-4.5, stem from large language models (L.L.M.s) trained on vast amounts of text data sourced from diverse online content.

To develop reasoning systems, companies employ reinforcement learning to train L.L.M.s over an extended period. This iterative process enables systems to learn through trial and error, enhancing their problem-solving abilities.

The industry consensus is that reinforcement learning represents the future of A.I. development, driven by the necessity to move beyond existing internet data limitations for training models like GPT-4.5.

While some reasoning systems surpass conventional L.L.M.s on standardized tests, real-world performance may vary. Experts caution that these systems, despite advancements, may exhibit errors and inaccuracies, including instances of “hallucination.”

OpenAI announced that GPT-4.5 would be accessible to ChatGPT Pro subscribers starting Thursday, offering access to the latest tools for a monthly fee of $200.

(In December, The New York Times filed a lawsuit against OpenAI and Microsoft for alleged copyright infringement related to news content on A.I. systems.)