EXPLORINEER
en

Archive — US — techcrunch

OpenAI research lead Noam Brown thinks AI ‘reasoning’ models could’ve arriveddecades ago

Noam Brown from OpenAI believes that AI reasoning models could have emerged decades earlier if the right approaches had been understood. He highlights the importance of test-time inference in enhancing model accuracy and notes the challenge academia faces in competing with large AI labs. Brown also criticizes the current state of AI benchmarks and calls for collaboration between research and frontier labs, especially at a time of reduced funding for scientific research.

Key points

  • Noam Brown states reasoning AI models like OpenAI's o1 could have been developed decades ago if proper approaches were known.
  • Brown emphasizes that understanding human thinking processes could enhance AI reasoning.
  • He introduces 'test-time inference' which improves accuracy in complex tasks by allowing models to reason before responding.
  • Brown explains that pre-training is not obsolete; AI labs are now balancing pre-training with test-time inference approaches.
  • He notes that academia lacks access to the resources of major AI labs but can contribute through low-compute research areas like model architecture design.
  • Brown identifies poor AI benchmarking as a significant issue and suggests that academia can improve benchmarks without requiring extensive computing.

Why it matters

  • The discussion reflects a broader trend in AI research where understanding and emulating human reasoning is essential for advancing AI capabilities.
  • Investment cuts in scientific research by the Trump administration may hinder future developments in AI, echoing concerns from notable experts like Geoffrey Hinton.
  • The challenge of AI benchmarks is part of ongoing debates about the reliability and transparency of AI systems, emphasizing the need for better assessment tools.

Noam Brown's insights underline that with better understanding and collaboration, AI reasoning models could have progressed much earlier, highlighting the urgency for academia and industry to work together, especially amid funding challenges.