When Openai launched a private demonstration of the new GPT-4 technology in late 2022, its skills shocked even the most experienced AI researchers. You can answer questions, write poetry and generate computer code in ways that seem to surpass the era.
More than two years later, Openai released its successor, GPT-4.5. New technology means the end of the era. Openai said GPT-4.5 will be the last version of the chatbot system that does not “chain inference of thought.”
After this release, Openai technology, like humans, can spend a considerable amount of time thinking about questions before answering them, rather than providing immediate responses.
GPT-4.5, which can be used to run the most expensive version of CHATGPT, is unlikely to produce excitement as GPT-4, as AI research has shifted in a new direction. Still, the company said the technology “feels more natural” than its previous chatbot technology.
“What sets the model apart is its ability to engage in warm, intuitive, naturally flowing conversations, and I think it gives a stronger understanding of what it means when users ask for something.”
In the fall, the company introduced a technology called Openai O1. It is designed to infer through tasks that involve mathematics, coding, and science. New technologies have been part of a broader effort to build AI that can infer through complex tasks. Chinese startups such as Google, Meta and Deepseek are developing similar technologies.
The goal is to build a system that, like human reasoning, can carefully and logically solve problems by building a building at the end, each building through a series of individual steps. These technologies are particularly useful for computer programmers who write code using AI systems.
These inference systems are based on large-scale language models or technology like GPT-4.5, known as LLM
LLMS learns your skills by analyzing a huge amount of texts selected from across the Internet, including Wikipedia articles, books, chat logs and more. By identifying patterns in all of that text, they learned to generate text themselves.
To build an inference system, companies place an additional process in LLMS called reinforcement learning. Through this process (which can be extended over weeks or months), the system can learn to do things through extensive trial and error.
For example, by solving various mathematical problems, you can learn which methods lead to the correct answer and which ones are not. Repeat this process with many problems to identify patterns.
Openai and others believe this is the future of AI development. But in some respects they are forced in this direction as there is no longer any internet data needed to train systems like GPT-4.5.
Some reasoning system is better than the regular LLM in certain standardized tests. However, standardized testing is not a good judgment about how technology works in real-world situations.
Experts point out that new inference systems aren’t necessarily human-like reasons. Also, like other chatbot technologies, they can still make things wrong and create a phenomenon called hallucination.
Openai said starting Thursday, GPT-4.5 will be available to anyone subscribed to ChatGpt Pro, a $200 service that provides access to all the company’s latest tools.
(New York Times sued Openai and its partner Microsoft in December for copyright infringement of news content related to AI systems.)
