中文精读
三年之后，信念怎样接受修正：张津剑谈问题、行动与Agent责任
42章经单期精读 · 三年复盘与主体性
嘉宾：张津剑，绿洲资本合伙人｜主持：曲凯
发布：2026年1月10日｜时长：81分04秒｜整理：2026年10月18日
官方单集与节目介绍
判断方向正确，也可能仍错过重要问题
官方以智谱与MiniMax上市为三年节点，邀请第一期嘉宾张津剑再回顾AI。节目对投资结果与市场的描述属于其当时视角；上市也不能单独证明产品、回报和所有早期假设都已完成验证。
他继续相信通用模型与Agent基础设施，却承认大公司反应比预想强，自己对地域和资产类型也有偏见。更有价值的复盘，不是把三年拼成一直正确的故事，而是让坚定方向、认知修正和结果随机性同时出现。
1. Figure的遗憾，指向分类判断中的偏见
线索：02:12–12:50。 张津剑回顾2022年秋看到生成模型范式变化，随后决定重仓AI。对未投Figure的遗憾，他归因于当时觉得机器人机会应属于中国。估值起落来自个人回忆，未核融资与回报；不能用后来的价格反证当初没有风险。
他希望重来时更开放地认识不同国家、文化与资产中的人。另一个天马行空设想，是把资金转成算力，支持研究探索，把算力理解为智力访问权。这是战略想象，不是算力必然升值、个人必有配额的事实，也不是资产配置建议。
对原判断最大的修正，是大厂投入、人才与定力很强，压缩了创业公司的部分空间。AGI也须先定义：他区分全知生命与可验证环境里超过人的表现，对后者更乐观；这仍是未来判断。
编辑推论：复盘时列出“方向判断”和“排除某项目的理由”两张表。哪条理由来自证据，哪条只是熟悉的分类？胜出的例子容易记住，被排除和失败的样本也要保留，才不会只用成功证明方法。
2. 信息长成一棵树，行动仍要回到一个问题
线索：21:49–24:40、32:13–34:34。 张津剑用树干、枝叶比喻行业扩张：可以看清树，却不能数尽所有叶子。因此先决定哪些不看，再问主要矛盾在哪里，而不是穷尽所有融资消息。
他更重视诚实面对变化、定义问题和快速执行。官方导航中的“long action, short thinking”，强调行动反馈相对空想的价值；不能理解成高影响行动无需思考。想一年不动与小实验一周拿到结果，反馈条件本就不同。
融资被称为高频信号，创始人解决什么重要需求才是低频问题。曲凯补充，融资可能变成奖励函数，使人按估值比较自己。张津剑又讲一位投资朋友亲自创业后也陷入同样攀比，承认知道原则与承受压力是两回事。
编辑推论：保留用户问题、实际交付、失败和下一步，而不是只删掉融资新闻。资金确实影响生存，不能羞辱融资需求；要防止资金声量替代了本来要解决的任务。
3. Agent“主体性”先是一张身份与责任清单
线索：25:54–29:15、46:12–47:56。 Agent是能使用工具、推进任务的AI系统。张津剑从2023年的Bot六层逻辑继续追问：它是谁、代表谁、怎样付款、防欺骗、遇到争议由谁负责，协作协议怎样执行？
他以教育比强化学习、法律比代码、金融比支付，设想围绕Agent形成新基础设施。提及x402和零知识证明等方向，也只是当时关注，没有给出实现和测试。不能由“主体性”推导AI已经有法律人格、自主账户权或不受授权限制。
与人的主体性放在一起，是本期的哲学主线：人不只通过外部认可确认自己，也要选择愿意做什么。这与Agent工程不是同一层问题，可以互相启发，不能互相证明。
编辑推论：先把宏大词落到代表权、动作范围、金额、记录和撤销机制。一个模型知道自己的名字，并不等于知道何时可以替用户行动；能完成付款也不等于付款已经授权。
4. 多模态与具身，是其正在形成的研究路线图
线索：36:49–50:37。 张津剑用登山南北坡比语言模型与物理交互，认为两者可能逐渐接近。他将具身理解分为机器人产品、AI执行载体、通过机器与环境互动训练AI三个阶段，最后更看重学习系统，而不只是外壳。
这个框架让机器人是否能跑跳之外，多出“从交互学到什么”的问题。但它不是已证实的AGI必经路线。对感官、磁场和可说出信息比例的口述没有科学来源，本篇不保留为测量数字；具体团队排名与首创称号也不补成履历。
他还关注AI与科学、年轻研究者及Agent基础设施，并用人才、能源、供应链解释中国机会。资本必然流向中国、只有中美可竞争等属于个人宏观判断，不是可靠预测模型。
编辑推论：路线图可以帮助选择问题，项目验收仍看任务、数据、学习效果与部署成本。多模态丰富不自动产生智慧，年轻不自动代表科研能力，国家叙事也不能替具体团队交付。
5. 注意力与“真”，是个人选择，不是科学定理
线索：50:52–01:10:44。 张津剑谈减少应酬和惯性承诺，从个人不吃晚饭的经历，讲到更认真选择午餐。他借此问：若一年只能做一项投资，或一生只做一件事，会怎样取舍？可迁移的是有意识地选择，不是饮食健康建议。
他又从AI attention联想到人的关注点，把“真”作为个人价值标准。算法注意力、佛教觉察和人生经验是比拟，不是同一种机制；不能由此解释人类心理或模型原理。
关于运气，他提醒问题常被问给已成功的人，提问样本本身有偏。能选择方向与伙伴，后续生长仍有许多随机因素。播客既让别人预先理解机构，也让他重听自己；曲凯则讲从输出者到提问者，承认节目起量有运气，不能当稳定获客公式。
编辑推论：减少外部评判不等于忽略反馈，真实表达也不等于公开全部私人信息。把可控制的选择写清，把偶然结果留给偶然，既能反思，也能避免要求所有人接受同一人生函数。
6. 不追对手预算，是信念进入资源分配的时刻
线索：01:11:19–01:19:50。 嘉宾设想未来助理能跨时间核对言行，甚至识破伪装。这是预测，微表情不构成可靠测谎，AI也不能被许诺必然识别人的真实意图。
更具体的是他转述MiniMax创始人的预算取舍：不因竞争对手投入更多就批准追加，而是判断技术是否收敛、产品形态是否清楚、钱现阶段该投哪里。故事来自投资方视角，未独立核内部决策。
这个例子说明，信念若有经营意义，应影响当下资源分配，而非只是口号。它不意味着所有团队都该减少产品预算、只投研发，也不能保证某条路线成功。
编辑推论：每笔增量预算回答它解决什么瓶颈、何时能看到什么证据、什么情况会停止。全球创新意愿值得记录，现金、验证和组织能力仍需承接长线目标。
给三年复盘补一页“仍不知道”
以下是编辑练习。选一个持续相信的方向，分别写：保持的假设、被改掉的判断、因偏见排除的样本、无法控制的结果，以及下一阶段最小验证。
再为一条Agent任务标明代表谁、允许做什么、谁确认完成，检查愿景是否有机制。最后删去一个仅因同行在做而加的动作，保留停止条件。快速执行与保持敬畏可以共存，静待生长也不妨碍阶段核验。
回听导航与关联阅读
- 02:07：正确决定与遗憾。
- 09:44：大厂反应带来的认知修正。
- 21:13：信息过载与范围取舍。
- 25:55：Agent与人的主体性。
- 41:32：具身理解的三个阶段。
- 01:03:02：成功样本与运气。
- 01:15:24：预算取舍的例子。
相关主题：长期复盘、代理责任、选择与证据。可对照2023年Bot六层逻辑和2024年低频信号，保留观点延续与变化。
资料边界：完整阅读所供转写、旧精读与官方介绍，未逐句复听或核投资财务。估值、内部预算、科研路线、资本流向和AGI时间均为当期观点；生活实践不作健康建议，类比不作算法或科学结论，片尾音乐不作访谈论据。
English edition
Three Years Later, How Does Conviction Change? Zhang on Problems, Action and Agent Responsibility
42 Zhangjing episode analysis · Three-year review and agency
Guest: 张津剑, partner at 绿洲资本 | Host: Qu Kai
Published: January 10, 2026 | Duration: 81 minutes 4 seconds | Prepared: October 18, 2026
Official episode and introduction
Getting the direction right can coexist with missing important questions
The introduction uses the listings of 智谱 and MiniMax as a three-year milestone, inviting the podcast’s first guest to review AI again. Investment outcomes and market descriptions remain contemporary perspectives. A listing cannot alone validate products, returns or every early hypothesis.
Zhang continues believing in general models and agent infrastructure while acknowledging stronger-than-expected incumbents and his own geographic and asset-class assumptions. A useful review permits conviction, revision and randomness to coexist rather than reconstructing three years as uninterrupted correctness.
1. Regret over Figure exposes bias inside categories
Episode reference: 02:12–12:50. Zhang recalls seeing a generative-model shift in autumn 2022 and deciding to concentrate on AI. He attributes passing on Figure partly to believing robotics belonged to China. Valuations come from personal recollection, without funding or return checks. Later prices cannot prove an earlier investment risk-free.
He would have met people across more countries, cultures and asset classes. Another imaginative counterfactual converts capital into compute for research, describing compute as access to intelligence. This is a strategic scenario, not proof of inevitable appreciation or personal quotas, and supplies no asset-allocation recommendation.
His major revision concerns large companies’ spending, talent and persistence, which compressed some startup opportunities. AGI also needs definition: he distinguishes an omniscient being from outperforming people in verifiable environments, expressing more confidence in the latter. This remains a future judgment.
Editorial interpretation: Review direction separately from reasons for excluding individual opportunities. Which reasons came from evidence and which from familiar categories? Preserve excluded and failed cases alongside memorable winners rather than validating a method only through success.
2. Information becomes a tree; action still needs one problem
Episode reference: 21:49–24:40 and 32:13–34:34. Zhang compares expanding industries to trunks, branches and leaves. One can understand a tree without counting every leaf. Decide what not to examine, then identify the central constraint instead of exhausting every financing announcement.
He values honest adaptation, problem definition and execution. The official phrase “long action, short thinking” emphasizes feedback over speculation rather than eliminating thought before consequential action. Thinking for a year and producing a small experiment in a week offer different feedback conditions.
Financing is his high-frequency signal; an important customer problem is lower frequency. Qu adds that valuation comparisons can become founders’ reward function. Zhang describes an investor friend who criticized that behavior, then repeated it after starting a business, acknowledging the gap between knowing principles and experiencing pressure.
Editorial interpretation: Retain user problems, delivery, failures and next steps rather than merely deleting funding news. Capital affects survival and raising it deserves no shame. The risk is its visibility replacing the original task.
3. Agent “subjecthood” first becomes an identity and responsibility checklist
Episode reference: 25:54–29:15 and 46:12–47:56. Agents are AI systems using tools and advancing tasks. Continuing his 2023 bot discussion, Zhang asks who an agent is, whom it represents, how it pays, prevents deception, handles disputes and executes collaboration protocols.
He compares education with reinforcement learning, law with code and finance with payment, imagining supporting infrastructure. References to x402 and zero-knowledge proofs identify interests without implementations or tests. Subjecthood cannot establish existing legal personality, independent account rights or freedom from authorization.
Human agency forms the parallel philosophical theme: people should choose what they want to do beyond confirmation through others’ approval. This and agent engineering occupy different levels. They can inspire without proving each other.
Editorial interpretation: Translate the broad term into representation, permitted actions, limits, records and revocation. Knowing its name does not establish when a model may act for a person; payment capability does not establish payment permission.
4. Multimodality and embodiment form an evolving research map
Episode reference: 36:49–50:37. Zhang compares language models and physical interaction to climbing opposite sides of a mountain. He describes three understandings of embodiment: robot products, an execution vessel for AI and AI trained through machine-environment interaction. He increasingly values the learning system rather than its shell.
The framework adds a question beyond running and jumping: what is learned through interaction? It does not prove a necessary path to AGI. Claims about senses, magnetic fields and the proportion of information people can verbalize lack scientific sources and are not retained as measurements. Rankings and claims of who coined a field do not become verified biographies.
He also focuses on AI with science, younger researchers and agent infrastructure, using talent, energy and supply chains to explain Chinese opportunities. Capital inevitably moving to China or only two countries competing are personal macro judgments rather than reliable predictive models.
Editorial interpretation: A research map helps select questions; acceptance still needs tasks, data, learning outcomes and deployment costs. More modalities do not automatically produce wisdom, youth does not establish research ability and national narratives cannot replace team delivery.
5. Attention and authenticity are personal choices rather than scientific laws
Episode reference: 50:52–01:10:44. Zhang describes reducing social obligations and habitual commitments. Skipping dinner personally made him choose lunch more deliberately. He asks what changes if someone can make only one investment annually or do one thing in life. Conscious choice is transferable; dietary health advice is not.
AI attention inspires reflection on human focus, while authenticity becomes his value standard. Algorithmic attention, Buddhist awareness and lived experience are analogies rather than one mechanism. They cannot explain psychology or model architecture interchangeably.
On luck, he notes that successful people disproportionately receive the question, creating a selected sample. Direction and partners can be chosen, while subsequent growth involves randomness. Podcasts help others understand an institution and let him listen to himself. Qu describes moving from writer to questioner and credits luck in audience growth rather than claiming a repeatable acquisition formula.
Editorial interpretation: Reduced dependence on external approval need not eliminate feedback. Authenticity need not publish all private information. State controllable choices and leave accidental outcomes accidental, without imposing one life objective on everyone.
6. Refusing a rival’s budget turns conviction into resource allocation
Episode reference: 01:11:19–01:19:50. Zhang imagines assistants comparing statements over time and recognizing pretense. This is a prediction. Microexpressions do not establish reliable lie detection, and AI cannot be promised to identify inner intent inevitably.
A more concrete story concerns MiniMax’s founder declining additional budget merely because competitors spend more. The decision instead considers technical convergence, product clarity and where money matters now. This is an investor’s account without independent verification of internal decisions.
The example suggests that conviction gains operating meaning when it changes resource allocation. It does not prescribe every team reducing product budgets in favor of research or guarantee one technical route succeeds.
Editorial interpretation: Each additional budget should identify a bottleneck, expected evidence, timing and stopping conditions. Global ambition deserves recording, while cash, validation and organization must support long-term objectives.
Add a “still unknown” page to a three-year review
For an editorial exercise, choose an enduring direction and list retained assumptions, revised judgments, cases excluded through bias, uncontrollable outcomes and the smallest next validation.
For one agent task, identify representation, permissions and who confirms completion. Check whether mechanisms support the vision. Finally remove one action added only because peers do it, keeping a stopping condition. Fast execution and humility can coexist; patient growth does not prevent staged checks.
Listening guide and related reading
- 02:07: Correct decisions and regrets.
- 09:44: Incumbent reactions changing assumptions.
- 21:13: Scope under information overload.
- 25:55: Agency for people and agents.
- 41:32: Three interpretations of embodiment.
- 01:03:02: Selected success and luck.
- 01:15:24: Budget allocation.
Related themes: long-term review, delegated responsibility and evidence. Compare the 2023 six-layer bot discussion and 2024 low-frequency signals, preserving continuity and revision.
Source boundary: The complete supplied transcript, earlier note and official introduction were read without sentence-by-sentence audio or financial verification. Valuations, internal budgets, research paths, capital flows and AGI timing remain contemporary views. Personal routines are not health guidance, analogies are not algorithmic or scientific conclusions, and closing music is not interview evidence.