CEO Ramine Hassani 解释 Liquid AI 如何通过生物启发式神经网络和架构搜索优化边缘设备 AI，例如可在 iPhone 上运行的 10 亿参数模型。
CEO Ramine Hassani explains how Liquid AI's biologically inspired neural networks and architecture search optimize AI for edge devices, with proof points like a 1B parameter model running on iPhones.
要点 · TL;DR
液态神经网络用更少参数实现更优效率和泛化能力。 Liquid neural networks achieve superior efficiency and generalization with fewer parameters.
闭式解使液态网络可扩展到数十亿参数。 Closed-form solution enables scaling liquid networks to billions of parameters.
结合硬件的自动架构搜索优于人工设计。 Automated architecture search with hardware-in-the-loop outperforms human design.
核心观点 · Key points
效率是 Liquid AI 研究的基石，以最小的算法格式最大化智能。 Efficiency is the cornerstone of Liquid AI's research, maximizing intelligence in the smallest algorithm format.
受生物学启发的液态神经网络以更少的参数提供更优的分布外泛化能力。 Liquid neural networks, inspired by biology, offer superior out-of-distribution generalization with fewer parameters.
液态神经网络的闭式解使其无需数值求解器即可扩展到数十亿参数。 Closed-form solution of liquid neural networks enables scaling to billions of parameters without numerical solvers.
结合硬件在环的自动化架构搜索可为特定用例和约束找到最优设计。 Automated architecture search with hardware-in-the-loop finds optimal designs for specific use cases and constraints.
注意力机制对大规模模型仍至关重要，但更简单的门控卷积足以用于小型专用模型。 Attention remains crucial for large-scale models, but simpler gated convolutions suffice for smaller, specialized models.
Liquid AI 瞄准边缘设备的设备原生基础模型，这是一个超越数据中心的大市场。 Liquid AI targets device-native foundation models for edge devices, a massive market beyond data centers.
反共识 · Contrarian takes
困惑度等代理指标常误导；在真实下游任务和硬件上评估至关重要。 Proxy metrics like perplexity often mislead; evaluating on real downstream tasks and hardware is essential.
架构设计中应去除人为偏见；系统化搜索优于专家手动调优。 Human bias in architecture design should be removed; systematic search outperforms hand-tuning by experts.
缩放定律表明，无偏架构（如 Transformer）在大规模下占优，而非有偏架构。 Scaling laws show that unstructured architectures like transformers dominate at large scale, not biased ones.
当前 AI 系统缺乏人脑使用的多样化学习算法（如强化学习、贝叶斯推理）。 Current AI systems lack the diverse learning algorithms (e.g., RL, Bayesian inference) that human brains use.
人脑的每瓦特智能远超当前 AI；需要新的学习范式，而不仅是架构。 Intelligence per watt of human brain is far beyond current AI; new learning paradigms are needed, not just architectures.
硬件制造商应投资智能层（如英伟达的 NeMo）以差异化，而不仅仅是优化内核。 Hardware makers should invest in an intelligence layer (like Nvidia's NeMo) to differentiate, not just optimize kernels.
本期章节 · Chapters（共 37）
0. 引言与嘉宾背景Introduction and Guest Background
1. 效率与真实世界机器人Efficiency and Real-World Robotics
2. 液态神经网络简介Introduction to Liquid Neural Networks
3. 应用与优势Applications and Advantages
4. 复杂性与可扩展性挑战Complexity and Scalability Challenges
5. 历史背景与闭式解Historical Context and Closed-Form Solution
6. 从生物学到液态AIFrom Biology to Liquid AI
7. 赞助商插播：AnthropicSponsor Break: Anthropic
8. 神经元vs参数：范式转变Neurons vs Parameters: A Paradigm Shift
9. 神经元作为计算单元与参数数量Neuron as Unit of Computation and Parameter Count
10. 主要瓶颈：串行到并行计算Main Bottleneck: Sequential to Parallel Computation
11. 闭式解与可扩展性限制Closed-Form Solution and Scalability Limits
12. 顺序扫描与复杂度降低Sequential Scan and Complexity Reduction
13. 缩放定律与架构偏差Scaling Laws and Architecture Bias
14. 液态神经网络的硬件需求Hardware Requirements for Liquid Neural Networks
15. 分布外泛化与持续学习Out-of-Distribution Generalization and Continual Learning
16. 液态神经网络与适应性Liquid Neural Networks and Adaptability
17. 架构搜索与客户方法Architecture Search and Customer Approach
18. 算子的统一理论Unified Theory of Operators
19. 消除人类偏见：双门控卷积Removing Human Bias: The Double Gated Convolution
20. 门控机制与输入依赖Gating mechanisms and input dependence
21. 下一代智能的效率与架构Efficiency and Architecture for Next-Gen Intelligence
22. 缩放与架构复杂性Scaling and Architecture Complexity
23. 领域特定架构：生物学与长上下文Domain-Specific Architectures: Biology and Long Context
24. 资源受限部署与液态神经网络Resource-Constrained Deployment and Liquid Neural Networks
25. 开源与硬件协作Open Source and Hardware Collaboration
26. 设备基础模型与商业应用Device Foundation Models and Commercial Applications
27. 边缘计算的巨大机遇The Massive Opportunity in Edge Compute
28. 实现大规模智能供电的难度Realizing the difficulty of powering intelligence at scale
29. 异构计算与架构搜索的挑战Challenge of heterogeneous compute and architecture search
30. 给硬件制造商的建议Advice for hardware makers
31. 硬件公司需要智能层Hardware companies need an intelligence layer
32. 本地模型效率与生态系统优化Local model efficiency and ecosystem optimization