DeepMind CEO Demis Hassabis 探讨大型模型的惊人有效性、十年内实现 AGI 的可能性，以及神经科学如何启发 AI 研究。
DeepMind CEO Demis Hassabis discusses the surprising effectiveness of large models, the potential for AGI within a decade, and how neuroscience inspires AI research.
要点 · TL;DR
AGI 可能在十年内实现，但仅靠扩展可能遇到瓶颈；算法创新是关键。 AGI likely within a decade, but scaling alone may hit limits; algorithmic innovation is key.
将大模型与规划和搜索结合（如 AlphaZero）对 AGI 至关重要。 Combining large models with planning and search, like AlphaZero, is crucial for AGI.
多模态数据和自我对弈能让模型扎根于真实物理世界并克服数据限制。 Multimodal data and self-play can ground models in real-world physics and overcome data limits.
核心观点 · Key points
AGI 可能在十年内实现；缩放定律出奇有效，但可能遇到瓶颈。 AGI likely within a decade; scaling laws have proven surprisingly effective but may hit limits.
将大型模型与规划搜索（如 AlphaZero）结合是实现 AGI 的关键。 Combining large models with planning and search (like AlphaZero) is key to AGI.
多模态数据（视频、音频）将使模型扎根于现实物理世界，提升理解力。 Multimodal data (video, audio) will ground models in real-world physics and improve understanding.
强化学习和合成数据（自我对弈）可以克服数据瓶颈。 Reinforcement learning and synthetic data (self-play) can overcome data bottlenecks.
安全需要更好的评估、沙盒测试和国际治理；前沿实验室必须保护权重。 Safety requires better evaluations, sandboxing, and international governance; frontier labs must secure weights.
反共识 · Contrarian takes
仅靠 Scaling 可能不够；算法创新对 AGI 同样关键。 Scaling alone may not suffice; algorithmic innovation is equally critical for AGI.
LLM 已展现出隐式扎根和抽象能力，连先驱都感到惊讶。 LLMs already show implicit grounding and abstraction, surprising even pioneers.