AI reasoning models, designed to think through complex problems, might have been developed 20 years earlier if researchers had discovered the right methods sooner. That’s according to Noam Brown, OpenAI’s lead for AI reasoning research, who shared these insights during a panel at Nvidia’s GTC conference in San Jose.
Speaking about his journey, Brown explained, “There were several reasons this research direction was overlooked. But I noticed one thing — in tough situations, humans spend time thinking before acting. That’s something AI could benefit from too.”
Game-Playing AI Sparked Brown’s Shift Toward Reasoning Models
Brown’s interest in reasoning models began long before joining OpenAI. Back at Carnegie Mellon University, he worked on Pluribus, a groundbreaking poker AI that outplayed elite human professionals. What made Pluribus stand out was its ability to reason through situations rather than rely on brute-force tactics, a common limitation in traditional AI models.
“That work made me realize the potential of models that ‘think’ rather than just compute faster,” Brown noted.
Today, Brown is also one of the minds behind OpenAI’s experimental model, o1, which uses a technique called test-time inference. This method allows the AI to spend extra computing power thinking through a problem before answering. As a result, models like o1 offer better accuracy and reliability, especially in fields like math and science.
Academic Research Still Holds Power in Shaping AI’s Future
During the panel discussion, Brown was asked whether universities and researchers, with limited computing power, could ever compete with AI labs like OpenAI. Acknowledging the challenge, Brown admitted, “It’s harder now because models are so compute-heavy. But there are still opportunities, especially in model architecture design, which doesn’t always need vast resources.”
Brown added that academic research often sparks ideas that frontier labs like OpenAI later scale. “When an academic paper presents a strong argument that scaling the idea could work, we pay attention. And yes, we test it,” he said.
A Call for Better AI Benchmarks — Academia Can Lead the Way
Brown also emphasized that AI benchmarking is an area ripe for academic impact. “Frankly, the current state of benchmarks in AI is poor, and improving them doesn’t require huge compute,” he pointed out.
Existing AI benchmarks often assess obscure knowledge or capabilities that fail to align with real-world tasks most users care about. This disconnect creates confusion about what AI models can truly accomplish.
By focusing on creating more meaningful and practical benchmarks, Brown believes academic researchers can shape how AI progress is measured and understood.
AI Research Faces Funding Threats Amid Political Cuts
Brown’s comments come as AI research funding faces growing uncertainty. The Trump administration’s significant cuts to scientific grant-making have drawn criticism from top AI experts, including Nobel laureate Geoffrey Hinton. Many fear these cuts could hamper AI research in the U.S. and beyond.
Despite the challenges, Brown remains optimistic about collaborations between academia and major AI labs. He believes the next wave of breakthroughs could come from smarter design and better evaluation methods — not just more compute power.
