Noam Brown: AI 'Reasoning' Models Could Have Emerged Decades Ago

# GTC # openai

Noam Brown, a leading researcher in AI reasoning at OpenAI, recently shared insights at Nvidia's GTC conference in San Jose, suggesting that advancements in "reasoning" AI could have been achieved 20 years earlier if the right methods and algorithms had been known. He explained, "There were various reasons why this research direction was neglected," highlighting a gap in the approach that could have been filled much sooner. Reflecting on his research journey, Brown noted a crucial realization: "I noticed over the course of my research that, OK, there’s something missing. Humans spend a lot of time thinking before they act in a tough situation. Maybe this would be very useful [in AI]." This observation led him to develop AI models that mimic human-like reasoning, rather than relying solely on computational power. His work at Carnegie Mellon University, particularly with the Pluribus AI, which bested top human poker players, exemplifies this approach. Pluribus was groundbreaking because it used reasoning to solve problems, contrasting with the more common brute-force methods at the time. At OpenAI, Brown contributed to the development of o1, an AI model that uses a method known as test-time inference. This technique allows the AI to "think" before responding, enhancing its accuracy and reliability, especially in fields like mathematics and science. During the panel discussion, Brown addressed the challenge of academic research competing with the scale of experiments conducted by large AI labs like OpenAI. He acknowledged the increasing difficulty due to the growing computational demands of modern models but suggested that academics could still contribute significantly by focusing on areas that require less computational power, such as designing model architectures. He emphasized the potential for collaboration between academia and frontier labs, stating, "[T]here is an opportunity for collaboration between the frontier labs [and academia]. Certainly, the frontier labs are looking at academic publications and thinking carefully about, OK, does this make a compelling argument that, if this were scaled up further, it would be very effective. If there is that compelling argument from the paper, you know, we will investigate that in these labs." Brown's comments are particularly timely as the Trump administration has proposed significant cuts to scientific funding, a move criticized by AI experts, including Geoffrey Hinton, who argue that such cuts could jeopardize AI research efforts globally. He also pointed out the critical role of academia in improving AI benchmarking, noting, "The state of benchmarks in AI is really bad, and that doesn’t require a lot of compute to do." Current AI benchmarks often focus on obscure knowledge and fail to accurately reflect the capabilities that matter most to users, leading to confusion about AI models' true potential and progress. *Updated 4:06 p.m. PT: An earlier version of this piece implied that Brown was referring to reasoning models like o1 in his initial remarks. In fact, he was referring to his work on game-playing AI prior to his time at OpenAI. We regret the error.*

Related article

Comments (14)