Exploring the most mysterious and captivating differences between biological and artificial neural networks, including credit assignment over long time spans and the need for causal understanding.
深度网络需要更好的世界模型，而不仅仅是更多数据或参数。 Deep nets need better world models, not just more data or parameters.
主动交互和因果推理是像孩子一样学习的关键。 Active interaction and causal reasoning are key to learning like children.
核心观点 · Key points
生物神经网络擅长在极长时间跨度上进行信用分配，而当前人工神经网络难以做到。 Biological neural networks excel at credit assignment over very long time spans, which current artificial neural networks struggle with.
当前深度神经网络缺乏对世界的鲁棒和抽象理解；它们需要更好的世界模型。 Current deep neural networks lack robust and abstract understanding of the world; they need better world models.
训练目标应从被动观察转向主动智能体，通过交互和因果推理来学习。 Training objectives should shift from passive observation to active agents learning through interaction and causal reasoning.
解缠表示和分解知识对于更好的泛化和避免灾难性遗忘至关重要。 Disentangled representations and factorized knowledge are crucial for better generalization and avoiding catastrophic forgetting.
短期 AI 安全问题如偏见、就业影响和自主武器比存在风险更紧迫。 Short-term AI safety concerns like bias, job impact, and autonomous weapons are more pressing than existential risk.
反共识 · Contrarian takes
仅仅增加深度或参数不会解决神经网络中的根本表示问题。 Simply scaling up depth or parameters will not solve fundamental representational issues in neural networks.
AI 带来的存在风险非常不可能，不是紧迫问题，但值得学术研究。 Existential risk from AI is very unlikely and not a pressing concern, though worth academic study.
科学通过小步和协作进步，而非像《机械姬》中描绘的孤立天才。 Science progresses through small steps and collaboration, not isolated geniuses as depicted in movies like Ex Machina.
像 AlphaGo 这样的标志性事件被高估了；真正的进步是渐进而累积的。 Seminal events like AlphaGo are overrated; real progress is gradual and cumulative.
在理解大脑和语言运作的大框架下，语言差异是次要的。 Language differences are minor in the grand scheme of understanding how the brain and language work.
本期章节 · Chapters（共 20）
生物与人工神经网络之谜Mystery of biological vs artificial neural networks
破解信用分配难题Breaking down credit assignment
长期信用分配的架构局限Limitations of current architectures for long-term credit assignment
深度神经网络世界表征的弱点Weakest aspect of deep neural networks' world representation
架构、数据集还是训练目标？Architecture, dataset, or training objective?
像孩子一样主动交互学习Learning like children: active interaction
仅靠规模扩展足够吗？Does scaling alone suffice?
深度变革还是根本改变？Depth vs. Fundamental Changes
先验知识与常识Priors and Common-Sense Knowledge
解耦表征与机制Disentangled Representations and Mechanisms
解耦表征与泛化Disentangled Representations and Generalization
《机械姬》与 AI 安全讨论Ex Machina and AI Safety Discussion
研究的多样性Diversity in Research
AI 中的偏见与对齐Bias and Alignment in AI
人机协作与机器教学Human-Robot Collaboration and Machine Teaching
图灵测试最难的部分Hardest Part of the Turing Test
Winograd 模式与世界理解Winograd schemas and world understanding
AI 寒冬的生存与教训Surviving AI winter and lessons learned