中文精读
个人AI的难题：记得你的经历，怎样才算理解你
42章经单期精读 · 个人模型与记忆
嘉宾：陶芳波，心识宇宙（Mindverse）创始人，身份依节目当期介绍｜主持：曲凯
发布：2024年8月31日｜时长：52分24秒｜整理：2026年10月18日
官方节目与收听入口
从会办事的AI，走向知道你要什么的AI
一个AI可以替你查资料、填表，却未必知道什么结果对你有价值。陶芳波用这个差别解释从MindOS到Me.bot的转向：先为个人建立理解需求的AI，再让它调用提供服务的工具。他同时希望AI能从日常经历中持续学习，而非始终依靠创建时写好的流程。
本期的价值在于把“懂我”拆成主体、记忆、学习与产品入口几个问题。它也有明确的交付边界：2024年访谈中提到的小几万产品用户，不等于这些人都获得了持续训练的个人模型；嘉宾随后说，这类模型当时只能维护不到100人。下文把现有使用描述、技术主张与未来设想分别保留。
1. 需求侧Bot：先明确AI代表谁
节目线索：04:26–12:26。 Bot可以理解为持续承担某种角色的AI。客服、编程、商家销售Bot提供服务；陶芳波更关注绑定一个人的需求侧Bot，由它整合需求，再与服务端沟通。区别不只在聊天界面，而在长期替谁优化。
他举个人与饮料商家各有Bot的例子：个人的AI持续对齐本人的需求，商家AI持续改善自己的服务能力，两者交换信息。每个主体拥有独立优化循环，才是划分多个Bot的理由；不必因为流程有五步就造五个角色。固定填表流程仍可作为被调用的组件。
部门或产品线是否需要单独Bot，嘉宾也认为取决于是否有独立、持续优化的需要。这个条件使“千亿Bot”从数量想象回到组织问题：谁定目标，谁给反馈，谁承担结果？编辑推论：个人Bot如果同时靠商家佣金获利，就需要解释推荐目标如何处理利益冲突。“属于我”的称呼本身不能保证目标一致。
2. 持续学习：新反馈要能改变旧办法
节目线索：14:41–18:26。 陶芳波用律师成长作类比：读法条和案例，类似模型先学习通用知识；师傅给办案流程，类似Agent获得工具与工作流；真正跟案件、收到反馈以后，律师可能改变先查什么、先问谁、怎样组合经验。
他批评的不是流程没有用，而是流程完成以后，系统仍无法利用反馈改造自己。一个用户说“这不是我需要的”，如果只保存原话，下次照旧执行，就没有完成这里要求的学习。销售每天接触不同客户，也可能提供高频反馈，但如何将反馈转成有效改进仍是工程与研究问题。
这个类比帮助定义验收，却没有证明产品已具备律师执业能力。案件输赢也受许多外部条件影响，不能简单把结局作为流程好坏的唯一标签。编辑推论：测试时应观察相似任务第二次发生时，系统选择是否改变、为什么改变，以及改变有没有误伤原来做对的部分；“新增了一条记忆”不足以回答。
3. 回忆与理解：能找到旧电影，不等于会推荐新电影
节目线索：18:59–30:54。 嘉宾借“经历的自我”和“记忆的自我”说明：原始经历不断进入，但未来判断还依赖从经历中抽出的偏好和思考方式。面对两部没看过的电影，单纯重复旧记录不够，系统需要判断用户通常喜欢什么、这次条件又是什么。
他又用《西游记》作区分：某一句原文可以回书中查找，类似RAG——先检索资料，再交给模型回答；但理解一个人，需要支持预测、推荐与决策。两者可以同时存在，嘉宾也明确说系统仍有RAG，不能把它改写成彻底抛弃检索。
陶芳波称团队比较了找细节与做推理的测试，并对自己的记忆训练方法有信心。不过本期没有公开测试集、基线、样本或消融，不能接受“只有这种方法能做好”作为普遍结论。他谈及认知科学和儿童记忆的解释，也只是本期的理论叙述，没有独立科学验证。
编辑推论：偏好推断应允许修改。一个人这周喜欢轻松电影，可能因为工作疲惫，并非永久人格；曾经不喜欢的类型，也可能在另一情境下愿意尝试。“像我”应包括对变化的敏感，而非不断强化旧标签。
4. 每个人一个模型：当时仍是小范围能力
节目线索：31:15–35:54。 陶芳波设想通用大模型提供常识与基础推理，上面再持续训练个人化模型，理想状态每天更新，使参数随个人经历变化。它与只把个人资料塞进提示词，是不同的实现主张。
但他清楚说明，当时并没有给每位用户提供这个内核，持续维护能力最多不到100人。产品在2024年四五月开始内测、有小几万用户，是另一层规模。把两句话连成“几万用户已拥有每天训练的独立模型”，就越过了原始证据。
这里还有一个容易被放大的细节：研究部门过去两年发了约200篇公众号文章，主持人误听成论文，嘉宾当场纠正。文章数量既不能变成论文数量，也不能证明技术效果。对产品决策而言，需要进一步知道每人维护成本、更新耗时、失败回退，以及训练是否稳定改善真实任务；节目没有提供这些答案。
5. “另一个自己”的入口：先有记录与联想的日常用途
节目线索：36:27–42:22。 陶芳波用HER、HIM、THEM、ME概括陪伴者、理性助理、多职业AI集合和超级自我四种定位。团队通过用户访谈倾向ME，希望AI记录、帮助，未来代表自己。这个命名体现关系假说，没有比较样本可证明所有用户都更信任“自己”。
落到当期产品，输入包括语音、文字、图片及书页；系统帮助搜索、按主题和人物组织，再把新记录与旧经历联结。例子是一次创业对话与之前另一个人的观点产生联系。延展既可在新增记忆后出现，也可每天早上九点推五条。这里可感知的价值是少丢信息、容易找回，并获得与手头问题相关的新线索。
嘉宾说自己三个月积累约1500条记忆，有些是一段长谈。这说明创始人的重度使用方式，不证明普通用户愿意同样持续输入，也没有展示推荐被采纳的长期效果。编辑推论：五条推送需要与“有几条真正相关、是否打扰、能否纠错”一起看，数量不能替代启发质量。
6. 从记录工具到超级入口，中间还隔着授权与生态
节目线索：42:31–51:59。 陶芳波把个人体验、社交与外部内容区分，认为后两类已被平台充分数字化，日记式自我对话和线下经历则可能成为新数据入口。若个人模型理解这些经历，再接电影、餐厅或社交服务，就可能长出应用生态。
主持人提出直接的阻力：主动记录门槛高。嘉宾寄望未来眼镜等设备持续、背景化采集，并预期两年内会有解决方案；这是2024年的预测，本篇没有据此认定已经实现。采集更容易，还要面对别人的谈话、敏感经历、误记与数据控制。收得更多，不自动等于判断更准。
“一颗脑子、多个设备身体”同样是未来设想。陶芳波承认公司当时没有完成生态所需能力，也认为这种敏感基础设施不应由某个私营企业独占。因此，万亿美元机会、个人模型成为下一代计算机，都应按愿景阅读。先证明一段记录怎样改善下一次行动，才有讨论入口的产品基础。
一个适合小规模测试的记忆任务
以下为编辑建议，不是Me.bot公开评测。让同一位参与者连续经历三次场景：第一次表达一个偏好并说明原因；第二次面对新选项，看AI能否给出依据；第三次明确改变条件，检查它是否更新。
| 层次 | 具体检查 | 不足以证明成功的信号 |
|---|---|---|
| 回查 | 能否找到原来的记录与时间 | 回复里提到一个熟悉词 |
| 推断 | 能否解释新选项为何相关，区分事实与猜测 | 总说“这很适合你” |
| 更新 | 新条件出现后能否纠正旧偏好，允许用户拒绝记录 | 记忆数量持续增长 |
| 代理 | 对外行动前能否说明代表谁、需要什么授权 | Bot之间成功交换信息 |
这个练习把“懂我”从抽象感受变成可讨论的行为，但不能凭几次成功推断长期人格理解。只需要偶尔找资料的人，简单检索加可编辑偏好也可能足够；是否需要持续训练，应由任务收益与成本共同回答。
回听导航与关联阅读
- 06:40：理解需求的Bot与提供供给的Bot。
- 14:19：为什么固定工作流还不够。
- 20:34：经历与抽象记忆的区别。
- 26:45：记忆训练方法的主张。
- 33:50：不到100人的个人模型能力边界。
- 38:09：当期Me.bot怎么用。
- 51:01：个人模型与设备关系的未来设想。
相关主题：AI陪伴、个人代理、长期记忆与可纠正偏好。可比较三款AI产品寻找PMF与EVE的关系记忆设计。
资料边界：依据完整提供转写、旧精读及官方当期介绍重新组织，未逐句复听。Mindverse、MindOS、Me.bot及陶芳波按官方资料校正。认知机制、模型优势、用户规模与时间表均是2024年访谈中的陈述或预测，未另作科学、技术或经营验证；“上传智能”属于愿景表达。
English edition
Personal AI’s Hard Question: When Does Remembering Your Life Become Understanding You?
42 Zhangjing episode analysis · Personal models and memory
Guest: Tao Fangbo, founder of Mindverse, as introduced in the episode | Host: Qu Kai
Published: August 31, 2024 | Duration: 52 minutes 24 seconds | Prepared: October 18, 2026
Official episode and listening link
From performing tasks to understanding what you want
An AI can retrieve information or fill out forms without knowing which result matters to you. Tao uses this distinction to explain the move from MindOS to Me.bot: first build AI that understands an individual’s needs, then let it call tools providing services. He also wants AI to keep learning from daily experience rather than depend permanently on its original workflows.
The interview separates “understanding me” into questions of identity, memory, learning and product entry points. Its delivery boundary matters. The tens of thousands of product users mentioned in the 2024 conversation were not all receiving continually trained personal models. Tao later says the team could maintain those models for fewer than 100 people. Current-use descriptions, technical claims and future proposals therefore need separate treatment.
1. Demand-side bots: Establish whom the AI represents
Episode reference: 04:26–12:26. A bot here means AI that persistently performs a role. Customer support, programming and merchant sales bots supply services. Tao focuses on a demand-side bot bound to a person, integrating needs before communicating with service providers. The distinction concerns whose interests it continually optimizes, beyond the chat interface.
His example gives an individual and a beverage business separate bots. The individual’s AI aligns with personal needs while the business’s AI improves its service capabilities; the two exchange information. Independent optimization for distinct owners justifies separate bots. A five-step process does not inherently need five personalities. A fixed form-filling workflow can remain a callable component.
Whether a department or product line needs its own bot depends, Tao says, on whether it needs independent, ongoing improvement. This condition brings the imagined abundance of bots back to organization: who sets objectives, supplies feedback and owns outcomes? Editorial inference: A personal bot funded by merchant commissions would need to explain how recommendations handle conflicting incentives. Calling it “mine” does not establish alignment.
2. Continuous learning: Feedback must change an old approach
Episode reference: 14:41–18:26. Tao uses a lawyer’s development as an analogy. Studying law and cases resembles acquiring general knowledge. A mentor’s procedural guidance resembles an agent receiving tools and workflows. Working on real cases and receiving feedback can then change what the lawyer investigates first, whom to consult and how experience is combined.
The criticism is not that procedures are useless. It is that a system may be unable to improve itself after execution. If a user says, “That is not what I needed,” saving the sentence while repeating the same approach next time does not meet this definition of learning. Sales conversations could also yield frequent feedback, but converting it into useful improvement remains a research and engineering problem.
The analogy helps define evaluation; it does not establish legal practice competence. Case outcomes also depend on external factors and cannot simply label a workflow good or bad. Editorial inference: When a similar task recurs, inspect whether the system changes its choice, why it changes and whether the adjustment damages previously correct behavior. Adding another memory alone does not answer these questions.
3. Retrieval and understanding: Finding an old film is different from recommending a new one
Episode reference: 18:59–30:54. Tao uses the “experiencing self” and “remembering self” to distinguish incoming experiences from the preferences and thought patterns abstracted from them. Choosing between two unseen films requires more than repeating a past record. The system must infer what the person tends to enjoy and what matters this time.
He makes a second distinction with Journey to the West. A particular sentence can be checked in the book, analogous to RAG: retrieving material before asking a model to answer. Understanding someone should also support prediction, recommendations and decisions. These functions can coexist. Tao explicitly says the system still uses RAG; the interview does not describe abandoning retrieval altogether.
He reports comparisons between detail-finding and reasoning tests and expresses confidence in the team’s memory-training method. No test set, baseline, sample or ablation is published in the episode, so claims that only this method succeeds cannot become universal conclusions. His explanations involving cognitive science and childhood memory are also theoretical accounts offered in the conversation, without independent scientific verification here.
Editorial inference: Inferred preferences must remain editable. Wanting light films this week may reflect exhaustion rather than a permanent trait. A previously disliked genre may suit a different occasion. Resemblance to the user should include sensitivity to change, rather than repeated reinforcement of old labels.
4. One model per person: A limited service at the time
Episode reference: 31:15–35:54. Tao proposes general models for common knowledge and basic reasoning, with personal models continually trained above them. Ideally, updates would occur daily as individual experience changes. This is a different implementation claim from placing personal facts into a prompt.
He also clearly says that this core was not provided to every user. Capacity for continual maintenance was fewer than 100 people. A beta beginning around April or May 2024 and tens of thousands of product users describe another scale. Combining them into “tens of thousands already have independently trained daily models” would exceed the evidence.
Another easily inflated detail is that the research team had produced about 200 public-account articles in two years. The host initially hears papers, and Tao corrects him. Article counts cannot become paper counts or establish technical performance. Product decisions still require maintenance costs per person, update time, recovery from failure and evidence of stable improvements on real tasks. The episode supplies none of those answers.
5. “Another me”: Begin with useful recording and connections
Episode reference: 36:27–42:22. Tao uses HER, HIM, THEM and ME for a companion, a rational assistant, a collection of occupational AIs and an enhanced self. User interviews led the team toward ME: an AI that records, helps and might eventually represent its owner. This is a relationship hypothesis, without comparative samples establishing that everyone trusts an AI framed as themselves.
The contemporary product accepts voice, text, images and book pages. It helps search and organize by topic and person, then connects new records with old experiences. One example links a current startup discussion to a previous conversation with another founder. Connections could appear after a new memory or as five suggestions at nine each morning. The tangible value is losing less information, retrieving it more easily and receiving leads relevant to a current question.
Tao says he accumulated approximately 1,500 memories over three months, including long conversations. This shows a founder’s intensive usage, not ordinary users’ willingness to keep supplying information, and provides no longitudinal evidence that recommendations were adopted. Editorial inference: Five notifications should be assessed alongside relevance, interruption and correction. Quantity cannot substitute for useful insight.
6. Between a recording tool and a major entry point lie permissions and an ecosystem
Episode reference: 42:31–51:59. Tao distinguishes personal experience, social information and external content. He argues that platforms digitized the latter two effectively, while diary-like reflection and offline experience could supply a new data foundation. A personal model understanding these experiences might connect film, restaurant and social services into an application ecosystem.
The host raises a direct obstacle: deliberate recording takes effort. Tao expects devices such as glasses to enable continuous background capture and predicts solutions within two years. This is a forecast made in 2024, not confirmation of delivery. Easier collection also raises questions about other people’s conversations, sensitive experiences, incorrect records and control of data. More collection does not automatically improve judgment.
“One brain, many device bodies” is likewise a future proposal. Tao acknowledges that the company did not then possess the necessary ecosystem capabilities and argues that sensitive infrastructure of this kind should not be monopolized by a private company. A trillion-dollar opportunity and personal models as the next personal computers are visions. Demonstrating how one record improves a subsequent action comes before a product claim about becoming a universal entry point.
A small memory exercise
This is an editorial proposal, not a published Me.bot evaluation. Give one participant three successive situations: express a preference and its reason; encounter new options and inspect the AI’s explanation; then change the conditions and check whether it updates.
| Layer | Specific check | Insufficient evidence of success |
|---|---|---|
| Retrieval | Find the original record and its date | Mentioning a familiar word |
| Inference | Explain relevance and distinguish facts from guesses | Always saying “this suits you” |
| Updating | Revise an old preference and let the user decline storage | A growing memory count |
| Representation | Explain whom it represents and required authorization before external action | Bots successfully exchanging information |
The exercise makes “understanding me” discussable through behavior, but a few successes cannot prove lasting understanding of personality. Someone who occasionally retrieves information may be adequately served by simple search and editable preferences. Whether continual training is worthwhile depends on task benefits and costs together.
Listening guide and related reading
- 06:40: Bots that understand demand versus supply services.
- 14:19: Why fixed workflows may be insufficient.
- 20:34: Experience versus abstracted memory.
- 26:45: Claims about memory training.
- 33:50: Capacity for fewer than 100 personal models.
- 38:09: Contemporary Me.bot usage.
- 51:01: The proposed relationship with devices.
Related themes: AI companionship, personal agents, long-term memory and correctable preferences. Compare three AI products seeking PMF and EVE’s relationship memory.
Source boundary: Reorganized from the complete supplied transcript, prior analysis and official contemporary introduction, without sentence-by-sentence audio checks. Mindverse, MindOS, Me.bot and Tao Fangbo follow the official metadata. Cognitive mechanisms, model advantages, user scale and timelines remain statements or forecasts from the 2024 interview, without separate scientific, technical or business verification. “Uploaded intelligence” describes an ambition.