Flo Carllo 讨论 Lindy 新推出的 Slack AI 员工、多人 AI、记忆以及使用中国模型的政治问题。
Flo Carllo discusses Lindy's new AI employee for Slack, multiplayer AI, memory, and the politics of using Chinese models.
要点 · TL;DR
AI 员工需要丰富的共享上下文，而不仅仅是智能，才能发挥作用。 AI employees need rich shared context, not just intelligence, to be useful.
缓存对成本至关重要；高缓存率对经济可行性至关重要。 Caching is critical for cost; high cache rates are essential for economic viability.
像 DeepSeek 这样的开源模型成本效益高，足以胜任许多任务。 Open-source models like DeepSeek are cost-effective and sufficient for many tasks.
核心观点 · Key points
随着我们接近AGI，上下文比智能更重要；AI员工需要丰富的共享上下文才能有用。 Context matters more than intelligence as we approach AGI; AI employees need rich shared context to be useful.
记忆应由智能体管理，而非仅靠RAG；智能体式管理可实现自我改进和更好的检索。 Memory should be managed by an agent, not just RAG; agentic management allows self-improvement and better retrieval.
缓存对成本至关重要；维持高缓存率（85%）对经济可行性至关重要。 Caching is critical for cost; maintaining high cache rates (85%) is essential for economic viability.
像DeepSeek这样的开源模型成本效益高，足以胜任许多任务，但前沿模型仍领先。 Open-source models like DeepSeek are cost-effective and sufficient for many tasks, but frontier models still lead.
AI组织应尽可能将工作整合到单一智能体下，以降低协调成本并提高效率。 AI organizations should consolidate work under a single agent to reduce coordination costs and improve efficiency.
我们正处于半人马时代；人类与AI共同创造最佳想法，但这一阶段是暂时的。 We are in the centaur era; humans and AI co-create best ideas, but this phase is temporary.
反共识 · Contrarian takes
美国应禁止中国前沿模型，原因包括不公平的蒸馏、宣传风险以及国家安全。 Chinese frontier models should be banned in the US due to unfair distillation, propaganda risks, and national security.
微调是最后手段；大多数公司应避免微调，而依赖提示词和上下文管理。 Fine-tuning is a last resort; most companies should avoid it and rely on prompting and context management.
将AI拟人化是有益的，但AI组织不应过度拟人化。 Anthropomorphizing AI is productive, but AI organizations should not be overly anthropomorphized.
在AI原生时代，较小的团队因协调成本更低而具有优势。 Smaller teams have an advantage in the AI-native era due to lower coordination costs.
“AI员工”一词是拟物化；AI组织的原生形态将根本不同。 The term 'AI employee' is a skeuomorphism; the native form of AI organization will be fundamentally different.
记忆最终可能通过每用户LoRA存储在权重中，而不仅限于上下文文件。 Memory may eventually live in weights via per-user LoRA, not just in context files.
本期章节 · Chapters（共 55）
引言Introduction
Lindy Teammate 发布Lindy Teammate Launch
入职与社会契约Onboarding and Social Contract
上下文与入职Context and Onboarding
以会议为中心的智能体系统Meeting-Centric Agentic Systems
隐私问题与双层记忆Privacy Concerns and Two-Tier Memory
播客的净化记忆Sanitized Memory for Podcast
维护两个版本Maintaining Two Versions
赞助商插播Sponsor Break
多人 AI 的挑战Challenges in Multiplayer AI
缓存与成本管理Caching and Cost Management
定价与成本管理Pricing and Cost Management
上下文桶与子智能体Context Buckets and Sub-Agents
树平衡与 Centary 树Tree Balancing and Centary Tree
个人记忆系统Personal Memory System
Tune 与 JSON 对比Tune vs JSON for agents
拟人化模型Anthropomorphizing models
并行智能体的数据库问题Database-style problems with parallel agents
供应商致谢与沙盒Vendor Shoutouts and Sandbox
早期与自建工具Early Days and Building Own Tooling
Lindy 生活：吃狗粮与 AI 员工Life at Lindy: Dogfooding and AI Employees
用 AI 智能体管理 CIManaging CI with an AI agent
工程师角色的演变The evolving role of engineers
招聘与生产力Hiring and productivity
人类仍不可替代之处What humans are still irreplaceable for
划清界限与半人马阶段Drawing the line and the centaur phase
经济转型与翻转Economic transformation and the flip
现有企业与初创采用Incumbent vs. Startup Adoption
领先的思想者Thinkers Ahead of the Curve
单智能体与分叉Single Agent and Forking
跨组织协作与 ZK 证明Cross-Organization Collaboration and ZK Proofs