Noam Brown claims AI models could've emerged earlier

Noam Brown, an AI researcher at OpenAI, believes that advanced AI reasoning models could have emerged as early as 20 years ago. He stated this during a panel discussion at Nvidia’s GTC conference in San Jose. Brown explained that researchers simply didn’t have the right approaches or algorithms back then. He introduced OpenAI's o1 model, which uses a method called test-time inference. This technique allows the AI to take extra time to think before answering, making it more accurate in areas like math and science. Brown noted that while larger datasets for pre-training machines are still important, the focus is now also on test-time inference. During the panel, Brown discussed the challenges that academic institutions face in conducting experiments. Many lack the computing resources that AI labs have. However, he believes academics can still contribute by exploring less computing-intensive areas, such as model architecture design. Brown also suggested that there is room for collaboration between leading AI labs and academic researchers. He pointed out that AI benchmarking is a key area where academic work could make a difference. Current benchmarks often measure obscure knowledge, which can confuse people about AI capabilities. Brown highlighted that improving this area doesn't require excessive computing power and could provide clearer insights into AI performance.