Brockman Claims GPT-5.4 Pro Solved New Erdős Problem, New Model to Address Writing Shortcomings

According to monitoring by Beating, Brockman revealed two specific product advancements in the Core Memory podcast. Firstly, GPT-5.4 Pro has been used by users to solve a new Erdős problem, which he described as ‘really important.’ He compared this to two years ago when OpenAI required a team of 20 people two weeks of increased computing power to train a model to win a bronze medal in the computing Olympics; now, ‘a casually trained model’ can achieve this. ‘What if we directed this capability towards drug discovery? No one is pricing this in.’ Secondly, in response to host Ashley Vance’s criticism that large models’ writing lacks ‘soul,’ Altman admitted, ‘We haven’t nailed personalization yet.’ Brockman then added, ‘We have a new model coming out in this dimension, which you can try after the podcast goes live and let us know if it has improved.’