Search Ctrl+K
Change language Switch ThemeSign In
[](https://www.bestblogs.dev/en/explore/brief "Daily Brief")[](https://www.bestblogs.dev/en/explore/newsletter "Weekly Picks")[](https://www.bestblogs.dev/en/explore/topics "Topics")[](https://www.bestblogs.dev/en/worldcup "World Cup Special")
[](https://www.bestblogs.dev/en/settings "Settings")[](https://www.bestblogs.dev/en/docs "Help Center")
Narrow Mode
88
重磅！智谱内部信“摸高计划”公开！
Zhipu founder Tang Jie released an internal letter announcing that over the next two years, the company will concentrate resources on four technical directions: long-horizon tasks, autonomous agents, self-evolution, and safety governance, aiming to push the boundaries of next-generation models. Follow·
Yesterday·1284 words (about 6 min)
AI Summary & Key Points
Summary
The article exclusively reveals Zhipu founder Tang Jie's internal letter, 'The Giant Wave is Here,' and provides a detailed interpretation of the 'Touch High' plan announced within it. The plan dictates that commercial monetization will not be the priority for the next two years; instead, resources will be concentrated on tackling four technical directions: 1) Long-horizon tasks, enabling models to complete projects spanning weeks or months; 2) Autonomous agent systems, solving challenges related to memory, continuous learning, and self-evaluation through engineering iterations to move Agents from demos to production; 3) Self-evolution, where models write code, synthesize data, and train themselves; and 4) Safety governance, with plans to invest billions in resources into mechanistic interpretability. By combining METR's capability growth data, the latest product dynamics from Anthropic and OpenAI, and Google DeepMind's ASI report, the article argues for the industry consensus and urgency regarding these four directions and includes the full text of the internal letter.
Main Points
* 1. Zhipu shifts its strategic focus to next-generation model capabilities, putting commercial monetization on hold.
The internal letter explicitly proposes concentrating resources to push the technical limits over the next two years rather than pursuing short-term revenue, which is consistent with the strategic decision to focus on Coding made in early 2025.
* 2. Long-horizon tasks are the next key milestone for model capabilities.
Citing METR data, the duration of human-expert-level tasks that models can complete is growing at a rate of doubling approximately every 90 days, evolving from 2 seconds for GPT-2 to over 5 hours for today's flagship models; models capable of multi-week tasks are expected to appear within two to three years.
* 3. The competition in autonomous agents has shifted from theoretical breakthroughs to engineering implementation and cost control.
The article points out that challenges such as memory and continuous learning are being resolved through engineering. Recent moves by Anthropic and OpenAI indicate that the industry focus has shifted to 'who can run it cheaper and who can run it unattended.'
* 4. Model self-evolution has moved from concept to practice and may drive the construction of ultra-large-scale computing clusters.
OpenAI's GPT-5.3-Codex has already participated in creating itself, and Karpathy's AutoResearch project also demonstrates the potential of AI self-improvement. The article speculates that the true purpose of overseas million-chip clusters may lie in this area.
* 5. Zhipu has elevated safety governance, particularly mechanistic interpretability, to a core strategic level.
Plans are in place to invest billions of resources to study neuron decision logic; this level of priority is rare among domestic companies and aligns with Anthropic's deep investment and sense of urgency regarding interpretability.
Sign in to highlight text and take notes as you read. Sign in now
原创 Datawhale 2026-07-11 23:09 浙江
Datawhale干货
作者：唐杰，智谱创始人，清华大学教授今天，智谱创始人唐杰发了一封内部信，标题叫《巨浪已来》。信里宣布了一个“Touch High（摸高）”计划：未来两年，不把重心放在商业变现上，集中资源投入四个技术方向，冲下一代模型。
类似的事智谱在 2025 年初做过一次：把资源收敛到 Coding 上。唐杰当时的说法是，DeepSeek R1 出来之后，Chat 范式的探索基本结束，智谱“赌”了 Coding 和 Reasoning，一种能跟 Agent 共生的能力。今天看这步走对了：从 GLM-4.5 到 GLM-5.2，智谱在多项公开测评里摸到了海外最前沿模型的能力边界；走同一条路的 Anthropic，年化收入两年半从 8700 万美金涨到了 470 亿。
这次信里的四个方向，拿这半年的AI动态对照着看，每一个都能找到别家正在做的影子，不只是智谱一家的判断。
摸高计划的四个技术方向
长程任务。模型的工作单位从“回答一个问题”变成“完成一个项目”，跨数周、数月的规划和执行。这件事有量化数据可查：评估机构 METR 一直在测一个指标——模型能以 50% 成功率完成的任务，换算成人类专家要干多久。2019 到 2025 年，这个数字每 7 个月翻一倍；2024 年之后，加速到了大约每 90 天翻一倍。GPT-2 的水平是 2 秒，去年底的旗舰模型已经是 5 个小时以上。照这个速度，能干跨数周任务的模型，大概就是两三年内的事。 自治智能体系统。信里提到，记忆、持续学习、自我评判，这三个原来被认为需要范式变革才能解决的难题，正在被逐步消解。换句话说，agent 从 demo 走向生产，卡的不再是理论突破，而是工程往前磨。海外的产品动作能印证这个判断——今年 4 月，Anthropic 把长时运行的托管智能体做成了公测产品；6 月底，又发了主打“便宜跑 agent”的 Claude Sonnet 5；几乎同时，OpenAI 把能自主拆分子任务的 GPT-5.6 开了预览。竞争点已经从“谁能做 agent”变成“谁跑得便宜、谁能无人值守”。 自我进化。模型自己写代码、自己合成数据、自己训练自己。这句话听起来像修辞，但它已经写进了发布说明：今年 2 月 OpenAI 发 GPT-5.3-Codex 时明确说，这个模型的早期版本“参与了创造它自己”——帮着调试训练、管理部署、诊断评估失败。Karpathy 做的 AutoResearch，自己跑了 700 个机器学习实验，找出 20 处训练改进。信里判断海外在建的百万芯片级算力集群，真正用途很可能就是让模型自己训练自己，这个说法放在这些事实旁边，不算激进。 安全治理。唐杰说这是四件里他最想强调的一个：计划投入百亿级资源做机械可解释性，厘清模型决策背后的神经元逻辑。这条线在海外以 Anthropic 投入最深，Dario Amodei 去年专门写过一篇《可解释性的紧迫性》，把它比作“给 AI 做核磁共振”，并给团队定了目标：2027 年之前，让可解释性能可靠地查出大部分模型问题。国内公司把这件事抬到核心引擎的优先级，还不多见。信里还引用了 Google DeepMind 6 月挂到 arXiv 的一份报告，《From AGI to ASI》，60 页，推演了从通用智能到超级智能的四条路径。里面最冷峻的论断是，即便单个模型的能力永远停在人类水平，一亿个共享同一底层大脑、零成本复制经验的实例，在群体层面就已经是超级智能。
以下是内部信全文。 智谱 内部信全文
!Image 7: 图片 一起“点**赞”三连↓**
Key Quotes
> Even if the capability of a single model remains forever at the human level, one hundred million instances sharing the same underlying brain and replicating experience at zero cost would already constitute superintelligence at the collective level.
Tags
AI Agent
LLM
Model Training & Inference
AI Safety & Alignment
AI Commercialization
Related Articles
* Just In: The Most Comprehensive Agent Harness Overview Yet, and engineering evolution, emphasizing that the engineering shell outside the model is key to moving Agents from demos to production.") * A Comprehensive Guide: The Evolution and Essence of Loop Engineering * Big News! Anthropic's Internal Skills Experience Revealed! * Redefining Skill Development: A Step-by-Step Guide & One-Stop Development Assistant Launch * Understanding Hermes in One Article: How the New Top-Tier Agent Self-Evolves from Experience * Harness Is Not the Goal, Knowledge Is the Moat — Knowledge Accumulation Practices of an AI Engineering Delivery Team * AI R&D Automation: Wiki Knowledge Base + Skill Pack * Latest! A 10,000-Word Overview of the Harness Revolution! * Latest! A 10,000-Word Review: From Prompt to Loop Evolution * OpenClaw and Hermes: A Comprehensive Review of AI Agent Architecture from Source Code
Make your daily reading actually fit you.A daily brief built from the sources you follow. Get started free HomeDiscoverWorld CupSettings