推理大模型1年内就会撞墙，性能无法再扩展几个数量级 | FrontierMath团队最新研究

OpenAI's models, like o1 and o3, show a tenfold increase in required compute, indicating diminishing returns in scaling. The potential for further scalability exists, but challenges in generalization and data availability loom large.

Key Points

• OpenAI's o3 requires 10x the compute of o1 during training phases.
• Epoch AI predicts inference training may hit a wall within a year.
• Current models like DeepSeek-R1 show close performance to o1 with lower costs.
• Scaling laws suggest potential for growth, but data scarcity is a major hurdle.
• Future inference models may converge in cost and performance requirements.