Greg Brockman dropped two product updates on the Core Memory podcast. First, GPT-5.4 Pro has already been used by a user to solve a new Erdős problem — a notoriously hard math puzzle named after the legendary mathematician Paul Erdős. Brockman called it "really important."
He contrasted this with two years ago, when OpenAI needed a 20-person team, two weeks, and massive compute to train a model that could win a bronze medal at the International Olympiad in Informatics. Now, "a casually trained model" can do it. "What if you point this capability at drug discovery?" he asked. "Nobody is pricing that."
Second, host Ashley Vance criticized large language model writing as "soulless." Sam Altman admitted, "We haven't nailed personalization yet." Brockman then added: "We also have a new model coming on that dimension. After the podcast goes live, try it and tell us if it's better."