When GPT-5 was officially launched in August, many users were disappointed. OpenAI's long-awaited language model (the company behind ChatGPT) does not seem to have made the big breakthrough as expected.
Although the response speed is faster and the operating cost is reduced, GPT-5 is considered to lack the revolutionary mark that the technology world is waiting for.
However, OpenAI CEO Sam altman believes the situation has changed. In an interview with wired, he affirmed: It didnt seem very good at first. But it's great now.
altman believes that GPT-5 has opened a new phase as artificial intelligence begins to contribute to the advancement of scientific discoveries, something that has never happened to any model before.
In fact, the launch of the GPT-5 was once described by altman himself as a disaster. The direct introduction encountered a technical problem, the model created an incorrect data chart, forcing OpenAI to adjust response behavior to make AI more receptive and natural.
The incident has made the public skeptical, while researchers like Gary Marcus (New York University) believe that GPT-5 is overinflated when it does not meet expectations for AGI, general artificial intelligence.
In response, OpenAI President Greg Brockman affirmed that the improvements to GPT-5 do not only come from increasing data or computing power. Instead, the company has promoted enhanced learning based on human resources (RLHF). This is a method that helps the model understand and react more naturally.
Mark Chen, Research Director of OpenAI, said that the majority of users do not feel the real power of GPT-5 because this model is optimized for specialized applications such as scientific research or programming.
If you are not a Mathematician, it will be difficult for you to feel impressed when GPT-5 is in the top 5 of the Olympic Mathematics, instead of the top 200 as last year, Chen said.
Despite criticism, altman remains optimistic: What I can say confidently is that GPT-6 will be significantly better than GPT-5. and GPT-7 will also be better than GPT-6. We have quite good results with things like that.
