A candid take on why tool-calling architectures like MCP are unnecessary, and a personal story of how a 2015 Karpathy article led to a startup failure and eventual vindication with ChatGPT.
要点 · TL;DR
AI 模型应直接编写并运行代码,取代 MCP 和工具调用层。 AI models should directly write and run code, replacing MCP and tool-calling layers.
借助工业化模型工厂,小团队可将从训练到发布的周期压缩到数周。 Small teams can compress training-to-launch cycles to weeks with industrialized model factories.
强化学习是后训练的关键,扩展到万亿参数仍属必要。 Reinforcement learning is key to post-training; scaling to trillion parameters is still needed.
核心观点 · Key points
模型构建 90% 是工程；端到端的工业化“模型工厂”让小型团队能把从训练到发布的周期压缩到几周。 Model building is 90% engineering; an industrialized end-to-end model factory lets small teams compress training-to-launch cycles to weeks.
强化学习是后训练能力的主要驱动；持久性和验证性行为能从较小模型中解锁远多于此前的性能。 Reinforcement learning is the main driver of post-training capability; persistence and verification behavior can unlock far more from smaller models.
扩张前沿模型仍然是必要的；开放与商品化并不会消除为了竞争而进行 Scaling 的必要性。 Scaling up frontier models is still necessary; openness and commoditization do not eliminate the need to scale to compete.
网络蕴含的可学习知识远超单纯的下一个词预测所能提取；未来的预训练会更早引入强化学习和更优的课程安排。 The web holds far more learnable knowledge than plain next-token prediction extracts; future pre-training will add reinforcement learning earlier and better curricula.
应当存在许多独立的基础模型实验室；共享研究、给用户选择，比少数几家主导公司更重要。 Many independent foundation-model labs should exist; sharing research and giving users choice matters more than a handful of dominant companies.
这场竞赛以日历时间计，而非 GPU 数量；可复现性与可靠基础设施让小型团队迭代更快。 The race is measured by calendar time, not GPU count; reproducibility and reliable infrastructure let small teams iterate faster.
反共识 · Contrarian takes
MCP 和工具调用层是愚蠢的；模型应该直接编写并运行代码，而不是塞满工具调用。 MCP and tool-calling layers are stupid; models should directly write and run code in a minimal virtual machine instead.
训练运行本身并不是最贵的部分；累计研发、数据和基础设施成本才是开放权重实验室真正付出的代价。 The training run is not the expensive part; cumulative R&D, data, and infrastructure cost are what open-weight labs really pay.
大多数知识工作的最优模型规模可能接近万亿参数级，而不是再大两三个数量级。 The optimal model size for most knowledge work may be near trillion scale, not two or three orders of magnitude larger.
后训练中的墙上时钟瓶颈是强化学习时间，而非 GPU 数量，因为批量大小受限；硬件拆分能改善这一点。 RL time, not GPU count, is the wall-clock bottleneck in post-training because batch size is constrained; hardware disaggregation can help.
对开放权重模型的一刀切限制确实令人担忧，但它可能像香烟广告禁令一样固化寡头格局。 Blanket restrictions on open-weight models rightly worry incumbents, but they can entrench an oligopoly just like the cigarette-ad ban.
从巨大的未发布教师模型蒸馏未必是对的；频繁从零开始训练能带来更清晰的科学实验和洞察。 Distilling from a huge unreleased teacher is not always right; frequent from-scratch training yields clearer scientific experiments and insights.