Why it’s the decade of agents (not the year), the real bottlenecks, and a 15-year intuition.
要点 · TL;DR
AI 是幽灵而非动物，缺乏持续学习和真正理解。 AI is a ghost, not an animal; it lacks continual learning and true understanding.
预训练像糟糕的进化，上下文学习是工作记忆。 Pre-training is like crappy evolution; in-context learning is working memory.
合成数据导致模型崩溃，需要人类多样性。 Synthetic data causes model collapse; human diversity is needed.
核心观点 · Key points
这将是智能体的十年，而非一年；它们需要十年克服持续学习、多模态等认知缺陷。 This will be the decade of agents, not the year; they need a decade to overcome cognitive deficits like continual learning and multimodality.
预训练像蹩脚的进化，提供模糊记忆；上下文学习则是可直接访问的工作记忆。 Pre-training is like a crappy evolution; it gives hazy recollection, while in-context learning is working memory with direct access.
当前 LLM 的强化学习很糟糕——通过吸管吸取监督——需要更好的信用分配。 Current reinforcement learning for LLMs is terrible—sucking supervision through a straw—and needs better credit assignment.
用于 RL 的 LLM 评判器可被利用；对抗样本导致奖励破解，需要更稳健的方法。 LLM judges for RL are gameable; adversarial examples cause reward hacking, requiring more robust methods.
LLM 生成的合成数据坍缩且缺乏熵；在其上训练导致模型坍缩，不像人类具有多样性。 Synthetic data from LLMs is collapsed and lacks entropy; training on it leads to model collapse, unlike human diversity.
LLM 过于擅长记忆，这干扰了学习可泛化模式；我们需要剥离记忆的认知核心。 LLMs are too good at memorization, which distracts from learning generalizable patterns; we need a cognitive core stripped of memory.
反共识 · Contrarian takes
我们建造的是幽灵，不是动物；AI 是一种不同的智能。 We're building ghosts, not animals; AI is a different kind of intelligence.
人类实际上并不使用强化学习来完成智能任务。 Humans don't really use reinforcement learning for intelligence tasks.
AI 进步与计算自动化是连续的，而非离散的爆发。 AI progress is continuous with computing automation, not a discrete explosion.
认知核心可能小到十亿参数，剥离记忆。 The cognitive core could be as small as a billion parameters, stripped of memory.
合成数据导致的模型崩溃是一个根本性问题；需要熵。 Model collapse from synthetic data is a fundamental problem; entropy is needed.
当前的编码智能体就像垃圾；自动补全才是最佳点。 Current coding agents are like slop; autocomplete is the sweet spot.
本期章节 · Chapters（共 51）
关于智能体的十年On the decade of agents
AI 的历史变迁On historical shifts in AI
早期 AI：单任务模型与智能体的误区On early AI: per-task models and the misstep of agents
动物类比与 AI 的差异On the animal analogy and why AI is different
进化 vs 预训练On evolution vs pre-training
上下文学习 vs 预训练On in-context learning vs pre-training
LLM 的工作记忆与长期记忆On working memory vs. long-term memory in LLMs
人类智能中未能复现的部分On what human intelligence we have failed to replicate
持续学习与涌现On continual learning and emergence
上下文窗口与认知架构On context windows and cognitive architecture
构建 nanoChatOn building nanoChat
从零构建 vs 使用 AI 工具On building from scratch vs. using AI tools
氛围编码与 AI 辅助编程On vibe coding and AI-assisted programming
AI 自动化 AI 研究与时间线On AI automating AI research and timelines
将架构调整集成到现有仓库On integrating architectural tweaks into existing repos
AI 代码生成的现状On the current state of AI code generation
强化学习 vs 人类学习On RL vs human learning
结果监督 vs 过程监督On outcome-based vs process-based supervision
合成数据生成与崩溃On synthetic data generation and collapse
人类与 LLM 的学习与记忆On human vs. LLM learning and memorization
模型崩溃与多样性On model collapse and diversity
认知核心的大小On the size of the cognitive core
认知核心规模On cognitive core size
未来扩展趋势On future scaling trends
硬件与软件改进On hardware and software improvements
衡量 AGI 进展On measuring progress towards AGI
自动化客服工作On automating call center work
AGI 与编程作为首个领域On AGI and coding as the first domain
LLM 的非编程应用On non-coding applications of LLMs
超级智能On superintelligence
失控风险On loss of control
AI 作为编译器 vs 替代品On AI as compiler vs replacement
AGI 作为劳动力On AGI as labor
增长与人口On growth and population
谷歌 VO 3.1On Google's VO 3.1
智能与进化On intelligence and evolution
人类智能与进化生态位On human intelligence and evolutionary niche
LLM 协作瓶颈On LLM collaboration bottlenecks
自动驾驶进展：演示到产品差距On self-driving progress and the demo-to-product gap
自动驾驶类比与 AI 部署On the self-driving car analogy for AI deployment
自动驾驶类比与 AI 部署经济学On self-driving analogy and economics of AI deployment
AI 部署的社会与法律方面On societal and legal aspects of AI deployment
教育、尤里卡与个人专注On education, Eureka, and personal focus
对教育的恐惧与愿景On fears and vision for education
技术内容教学与尤里卡On teaching technical content and Eureka