Noam Brown from OpenAI revealed at the Nvidia GTC conference that reasoning AI models like o1, which utilize test-time inference, could have been developed two decades earlier if researchers had utilized the right algorithms. He explained that these reasoning models offer improved accuracy, particularly in complex fields such as mathematics and science. Brown also discussed the evolving landscape of AI research which still values pre-training methods and highlighted the ongoing need for collaboration between leading AI labs and academic institutions to foster innovation despite resource challenges.
Noam Brown of OpenAI believes reasoning AI models, like o1, could have emerged two decades earlier with the right algorithms and approaches.
Brown noted that AI reasoning models demonstrate higher accuracy and reliability than traditional models, especially in mathematics and science.
He acknowledged the growing challenge for academia in accessing computing resources, but sees opportunities for collaboration between labs and academic institutions.
Despite the rise of reasoning models, Brown stresses that pre-training is not obsolete and remains a crucial component in AI research.
Collection
[
|
...
]