Demis Hassabis discusses his Nobel conjecture that patterns in nature, shaped by evolution and selection, can be efficiently discovered and modeled by classical learning algorithms, unlike random or abstract systems.
要点 · TL;DR
自然模式因进化结构而能被经典 AI 高效学习。 Natural patterns are efficiently learnable by classical AI due to evolutionary structure.
到 2030 年实现 AGI 是可能的，定义为匹配所有人类认知能力。 AGI by 2030 is plausible, defined as matching all human cognitive capabilities.
缩放定律在预训练、后训练和推理计算中仍然成立。 Scaling laws still hold for pre-training, post-training, and inference compute.
核心观点 · Key points
自然系统因进化过程而具有结构，这使得神经网络能够高效学习它们。 Natural systems have structure from evolutionary processes, making them efficiently learnable by neural networks.
到 2030 年实现 AGI 是可能的，其定义是在所有认知能力上一致地达到人类水平。 AGI by 2030 is plausible, defined as matching all human cognitive capabilities consistently.
Scaling（规模扩张）定律在预训练、后训练和推理算力上仍然成立。 Scaling laws still hold for pre-training, post-training, and inference compute.
AI 可以通过被动观察建模复杂物理（如流体动力学），挑战了具身 AI 的必要性。 AI can model complex physics like fluid dynamics from passive observation, challenging embodied AI necessity.
未来的 AI 界面将是个性化、多模态的，并由 AI 自身生成。 Future AI interfaces will be personalized, multimodal, and generated by AI itself.
反共识 · Contrarian takes
经典学习系统可能高效建模那些被认为需要量子计算机的问题。 Classical learning systems may efficiently model problems thought to require quantum computers.
像 Veo 这样的视频生成模型展现了直觉物理理解，无需具身体验。 Video generation models like Veo show intuitive physics understanding without embodiment.
科学中最难的部分是提出正确的问题，而非解决问题。 The hardest part of science is picking the right question, not solving it.
AI 进步可能需要新的突破，而不仅仅是扩展现有方法。 AI progress may require new breakthroughs, not just scaling existing methods.
聚变和太阳能带来的能源丰裕可能终结资源稀缺，焦点转向分配问题。 Energy abundance from fusion and solar could end resource scarcity, shifting focus to distribution.
本期章节 · Chapters（共 48）
自然模式与可学习性Patterns in nature and learnability
信息基础与 P vs NPInformation as fundamental and P vs NP
视频生成中的直观物理Intuitive Physics in Video Generation
AI 在电子游戏中的未来Future of AI in Video Games
90 年代游戏中的 AI 梦想Dreaming of AI in 90s games
开放世界的深度个性化Deep personalization in open worlds
超越随机生成与硬编码选择Beyond random generation and hardcoded choices
观看他人创作游戏与未来计划Watching others create games and future plans
电子游戏作为有意义体验Video games as meaningful experiences
游戏与 AI 领导力Gaming and AI leadership
AlphaEvolve 与进化启发技术AlphaEvolve and evolution-inspired techniques
搜索与新颖性发现Search and Novelty Discovery
研究品味与科学发现Research Taste and Scientific Discovery
分割假设空间与失败价值Splitting hypothesis space and the value of failure
细胞建模挑战与 AlphaFold 作用Challenges in modeling a cell and AlphaFold's role
细胞建模的时间尺度与粒度Temporal scales and granularity in cell modeling
模拟生命起源Simulating the origin of life
生命起源与大过滤器Origin of Life and Great Filters
AGI 时间线与第 37 手AGI Timeline and Move 37
AGI 定义与测试AGI Definition and Testing
AI 系统与人类程序员AI systems and human programmers
数据与计算规模扩展Data and Compute Scaling
极端丰裕与人性Radical Abundance and Human Nature
谷歌借助 Gemini 的转型Google's Turnaround with Gemini
研究文化与交付进展Research culture and shipping progress
许多用户首次接触 AIFirst encounter with AI for many users
产品直觉与 AI 产品设计Product sense and designing AI products
为未来能力设计Designing for Future Capabilities
Gemini 3 发布与版本管理Gemini 3 Release and Versioning
下游信号融入核心模型训练Integration of downstream signals into core model training
基准测试与多目标优化Benchmarks and multi-objective optimization
AI 开发中的竞争与合作Competition and collaboration in AI development
合作与科学使命Collaboration and Scientific Mission
人才争夺与薪酬Talent War and Compensation
对编程工作的影响Impact on Programming Jobs
AI 对编程与社会的影响Impact of AI on programming and society
冯·诺依曼的远见Von Neumann's foresight
科学作为连接器与 p(doom)讨论Science as a connector and p(doom) discussion
意识与人类心智Consciousness and the human mind
对人类文明的希望Hope for human civilization
结束语与感谢Closing Remarks and Gratitude
反思大卫·福斯特·华莱士演讲Reflection on David Foster Wallace's Speech
视角与智慧Perspective and Wisdom
应对网络攻击Dealing with Online Attacks
澄清 MIT 与德雷塞尔大学关系Clarifying MIT and Drexel Affiliations