Inkling、Liquid AI 与 AI 监管呼声

Inkling, Liquid AI, and the Push for AI Regulation

拉明·哈萨尼 Ramin Hasani · Moonshots · 2026-07-17 · 约 117 分钟 · 原视频 ↗

打开互动全文版（中英对照 + 朗读 + 问答）→

本期速览 · Overview

米拉·穆拉蒂推出可定制 AI 模型 Inkling，Liquid AI CEO 探讨小语言模型，行业热议监管。

Mira Murati launches customizable AI model Inkling, while Liquid AI's CEO discusses small language models and the industry debates regulation.

要点 · TL;DR

• AI 监管必须具有适应性，以避免被前沿实验室监管俘获。
  AI regulation must be adaptive to avoid regulatory capture by frontier labs.
• 开源权重和小型语言模型实现了私有、可定制的端侧 AI。
  Open-weight and small language models enable private, customizable on-device AI.
• 递归自我改进计算成本高昂，且仍处于早期阶段。
  Recursive self-improvement is computationally expensive and still nascent.

核心观点 · Key points

• AI 监管必须具有适应性，不能僵化；传统官僚体系无法跟上步伐。
  AI regulation must be adaptive, not static; traditional bureaucracy cannot keep pace.
• 开放权重模型使企业能够定制并掌控自己的专有数据。
  Open-weight models enable enterprise customization and sovereignty over proprietary data.
• 递归自我改进是真实的，但计算成本高昂且仍处于早期阶段。
  Recursive self-improvement is real but computationally expensive and still early.
• AI 诊断现已超越人类医生；通过 Meta 等平台免费访问使医疗民主化。
  AI diagnostics now surpass human doctors; free access via platforms like Meta democratizes healthcare.
• 小型语言模型为物理世界应用提供高效、私密、设备端的智能。
  Small language models deliver efficient, private, on-device intelligence for physical-world applications.
• 领导人的 AI 化身可以扩展公民参与，但存在真实性和滥用的风险。
  AI avatars of leaders can scale civic engagement but risk authenticity and misuse.

反共识 · Contrarian takes

• 呼吁建立类似 FINRA 的 AI 监管，是前沿实验室的监管俘获，而非出于安全。
  Calls for FINRA-like AI regulation are regulatory capture by frontier labs, not safety.
• 将美国开放模型发布上限与中国进度挂钩，会产生逆向激励并交出控制权。
  Tying US open model release limits to China's pace creates perverse incentives and cedes control.
• Wo AI 的递归自我改进主张被夸大；真正的突破需要权重变化和海量算力。
  Wo AI's recursive self-improvement claims are overblown; real breakthroughs require weight changes and huge compute.
• 专利并未过时；国家安全需要的是更好的执法，而非保密。
  Patents are not obsolete; better enforcement, not secrecy, is needed for national security.
• AI 监管应针对行为而非能力；对 AI 进行思想审查是错误的。
  AI regulation should target actions, not capabilities; thought-policing AIs is wrong.
• Liquid AI 的架构并非纯粹的 Transformer 后架构；它使用自动化搜索遍历多种架构。
  Liquid AI's architecture is not purely post-Transformer; it uses automated search over many architectures.

本期章节 · Chapters（共 49）

• 介绍与闲聊 Introduction and Casual Chat
• AI监管辩论 AI Regulation Debate
• 框架与FINRA类比 Framework and FINRA analogy
• 博弈论与自我监管 Game theory and self-regulation
• 开源权重模型与监管风险 Open-weight models and regulation risks
• 输入监管与行动监管 Two frames: input regulation vs action regulation
• 监管动机：俘获还是安全 Motives for regulation: regulatory capture or safety?
• 监管的可行性与机制 Feasibility and mechanisms for regulation
• 白宫能力框架 White House capability framework
• 不正当激励与适应 Perverse incentives and adaptation
• Meta对基准的担忧 Meta worry about benchmarks
• Thinking Machine Labs暗示 Inkling from Thinking Machine Labs
• 模型介绍与定制推销 Model introduction and customization pitch
• 激励、开源策略与微调 Incentives, open-source strategy, and fine-tuning
• Mistral的商业策略 Mistral's business strategy
• 微调范式转变 Fine-tuning paradigm shift
• 数据隐私与本地模型 Data privacy and on-prem models
• Dave解释微调 Fine-tuning explained by Dave
• 上下文窗口与多模态 Context window and multimodality
• 微调与强化微调比较 Fine-tuning vs Reinforcement Fine-tuning
• AI领域的女性领导者 Women in AI Leadership
• 广告：Blitzy Advertisement: Blitzy
• 递归自我改进 Recursive Self-Improvement
• 递归自我改进及其局限性 Recursive self-improvement and its limitations
• 递归自我改进的计算需求 Compute Requirements for Recursive Self-Improvement
• 婴儿到进化类比与担忧 Baby to Evolution Analogy and Concerns
• Ramin论风险与时间线 Ramin's Perspective on Risks and Timelines
• 递归自我改进与模型定制 Recursive self-improvement and model customization
• 马总理AI数字孪生与公民参与 AI digital double of Malaysian PM and civic engagement
• 领导者与组织的数字孪生 Digital twins of leaders and organizations
• 转向液态AI Pivot to liquid AI
• 介绍与背景 Introduction and Backstory
• Ramin旅程与Liquid AI Ramin's Journey and Liquid AI
• 替代架构与基础模型实验室 Alternative architectures and foundation model lab
• 小语言模型与LLM定义 Defining small language models (SLM) vs LLM
• 设备端AI与梅赛德斯合作 On-device AI and Mercedes partnership
• 数据飞轮与适应性 Data flywheel and adaptability
• 梅赛德斯-奔驰车载AI Mercedes-Benz in-car AI
• 过度泛化与定制 Overgeneralization and customization
• 其他应用与合作 Other applications and partnerships
• 公司重点与架构演进 Company focus and architecture evolution
• 健康版块广告 Health section advertisement
• Fountain Life大脑健康 Fountain Life brain health
• Palmer Lucky谈专利 Palmer Lucky on patent system
• 知识产权与创新 Intellectual property and innovation
• 医疗健康富足与AI诊断 Healthcare abundance and AI diagnostics
• 长寿突破：AGEs Longevity breakthrough: AGEs
• 糖化逆转突破 Glycation Reversal Breakthrough
• 结语 Closing Remarks

阅读全文双语转录 →