黄东旭

If 2025 was the first year of Agents, 2026 is seeing the market shift from Token Maxing to Token Efficient, but both guests believe AI infrastructure is not in an obvious bubble. Uber burned through its full-year AI budget in 4 months, while Meta plans to limit employees’ Token consumption—signs of cooling on the surface. 张宏江’s framework is that, relative to GDP, this investment cycle remains smaller than the early infrastructure build-outs for railroads, highways and the internet; meanwhile, he estimates Token prices have fallen by roughly 10x per year over the past few years, while usage growth will outpace the decline in cost.

黄东旭 offered the hardest ROI example of the cycle: 1 person, 3 months and an extra $400-$500 a day produced db9, a cloud-native distributed database that could generate $10M in annual revenue. Last November, on tasks where it was difficult to precisely assess the gap between models, he still believed the best strategy was to use the strongest model; at the time, Opus and GPT-5.1 held a decisive lead over the second tier of Coding Agents.

OpenClaw’s significance is that it moved Agents from research into engineering, and from models into applications. 东旭 believes it turned the Agentic Loop and tool calling into a personal-assistant form that ordinary users can access more easily; configuration, stability and the ongoing resource commitment remain problems now that it is aimed at the mass market, and its historical mission may already be complete. On the capability inflection point, 东旭 cited Claude Opus 4.5 last November, while 泓君 said she started using Agents with GPT-5.4.

东旭 shifted from Token Maxing to Token Efficient under the influence of GPT-5.5, DeepSeek V4 and Fable 5. He once used Slock to have 10 top-tier Agents critique one another, consuming roughly 1 billion Tokens and costing $300-$400 in a day; Fable 5, however, can sometimes solve in one shot a complex problem that an Agent group could not resolve, even without the preceding context. DeepSeek V4 and GLM-5.2 also convinced him that cheap local open-source models can work alongside the best closed-source models.

Saving Tokens at the individual level does not mean total consumption will fall. Agentic Loops stack multiple model calls, tool inputs and outputs, and context, potentially driving Token usage more than 100x above the Chatbot era; the future will also bring multi-Agent collaboration, Agent-to-Agent Networks and higher-order Loops. 张宏江 summarized the dynamic as the Jevons paradox: as technology matures and prices fall, usage and the market expand.

The main investment opportunities sit in Agent infrastructure: data, memory and context management, observability, collaboration, Sandboxes and Agent-Native Cloud. 东旭 is most interested in infrastructure that can lift model output from 50 to 80 or 90, and believes whoever can answer “where did the Tokens go?” will have value. 张宏江 also sees enterprise vertical-data moats and FDEs as important; in their discussion, Manus and Genspark looked closer to professional-user, To B products.

张宏江 believes the singularity has arrived and, under his definition, current model capabilities already qualify as AGI; 黄东旭 is waiting for a true self-improving Gödel machine as the more concrete marker. 张宏江’s basis is that models released over the past 6 months score to the right of the human mean of 100 on both visual IQ and general IQ tests; he argues that machines surpassing human learning ability is what constitutes the singularity. The two also discussed the decline of work and the resulting existential crisis.

🔗 Original source & video: E249 | The Token Economy Turns a Corner: The Agent Evolution from OpenClaw and Hermes to In-House Local Builds

黄东旭认为，OpenClaw 标志着 AI-native 应用形态已经出现，但它只是“及格线”——“做得比 OpenClaw 差的属于上一个时代，未来只会比它更好”。 他家里不懂计算机的父母和太太都开始“养龙虾”，自己手握多个 Claude Code + Codex Pro Max 账号，峰值日耗 token 达到十亿量级，平日也有几亿，因此“很兴奋，同时也很焦虑”，总有一种“被 agent team 推着往前跑”的感觉。

Harness engineering 是继 prompt 和 context 之后的更高维度：多 agent 协同、子 agent compaction 管理、外挂记忆，本质上“更像在管一个真正的开发团队”，只是“这帮 agent 不睡觉而已”。 他看不到 harness 的边界：benchmark 只测单 agent、单任务、单环境，“模型本身很聪明了，只是我们的题设计得不好”，未来甚至可能出现“agent 的社会学”。

本期最强的判断是：LLM 无状态，各家头部模型都足够聪明，而很多价值会沉淀在记忆层，而不是大语言模型层。 “它有多了解你呢？还是得靠记忆。”他的 Mem9 插件正解决龙虾“聊着聊着就忘了”的 compaction 失忆和升级丢数据两大痛点，提供开源自部署与云端托管的永久龙虾记忆服务。

Agent 时代不是中心化大数据，而是“海量的小数据”。 “你跟小龙虾往死里聊，能聊出一本莎士比亚吗？”单个 agent 的数据量不大，但加总后可能超过淘宝、美团级的中心化数据；服务商的策略是全量保存：“永远不删数据，只要价值能 cover 存储成本。”他认为当下的存储短缺只是“产能临时短缺，会回归”，长期需求仍会持续增长。

“一虾一库”是必然趋势，而且记忆一定会在云端。 底层共享对象存储上划出一个个虚拟门户，让 agent “认为自己拥有整个数据库”；重度用户可能只有 1%，没必要为轻度用户维护专属基础设施。OpenClaw 的 local-first 是极客理念，但“牺牲了用户体验”——龙虾养死“跟家里亲人去世的感觉很像”，云端才能“马上再给你启动一只龙虾的身体，接到你的记忆上又复活了”。

商业化的核心押注是“记忆 marketplace”：下一代 agent-native application 可能不再由软件工程师编写代码，而是由无数领域专家从自己的记忆中提取高价值内容。 厨师不会写代码，但可以把与龙虾讨论做饭的记忆抽成外挂记忆，接到别人的 agent 上收费，成为“一个新的分发渠道”；基础记忆可能免费，云备份、导入导出则可以做成订阅增值服务。

Agent 基础设施将被整体重写：沙箱化的 agent 云、agent 邮箱、支付、权限和 ID 都值得重新设计，现有数据库若服务几十亿 agent，成本根本扛不住。 但 SQL 和文件系统作为模型的心智模型会长期存在。独立向量数据库“不需要单独出现”：向量检索只是检索手段之一，最终会收敛成所有数据平台的标准能力。

组织形态可能变成部落，而个人护城河在于保持手感。 他的一两人团队“过去三个月干了传统软件公司一百个人年的事”，公司 90% 的代码由 AI agent 编写，他不掌握具体代码细节，“上一个时代的软件工程已经终结了”，全职 coder 可能会很危险。但这也是软件民主化——软件不会像纺织女工时代的衣服那样变少，反而可能更多；年轻人应该学习 Unix philosophy 这类“永恒不变的东西”。

🔗 Original source & video: Silicon Valley Coordinates x PingCAP Co-founder 黄东旭 on Rebuilding Infrastructure in the Agent Era

黄东旭把 2025 年 12 月前后视为 AI Coding 跨过“singularity”的时点：他用人的 blocking 从关键瓶颈降到约 0.1%来描述这一跃迁；在 10 万行代码以内，只要结果可验证，已经可以“人类只管提需求”。 他的评分从年初不及格、局部任务 60 分、模块自治 80 分升至复杂项目约 90 分；自己每天消耗接近 billion 级 token、产出几万行代码，甚至规定项目“不允许人类提交代码”。任川则曾半开玩笑地提出更激进的看法：未来应该禁止人类写代码。

这轮 Agent 的底层主线不是 Cowork 或 Clawdbot 这样的产品壳，而是 Claude Code 确立的工具调用框架、模型长上下文能力与 context engineering 同时成熟。 黄东旭给出的关键跃迁是：在 GPT-5.1 Thinking 出现之前，SOTA 模型在相关长上下文 benchmark 上召回率约 20%—40%，GPT-5.2 接近 90%，使多轮 agentic loop 不再迅速累积幻觉；他的终局判断是“everything is coding agent”。

Skills 比 MCP 更适合现阶段的 Agent 生态，因为自然语言接口既能被模型修改，又能逐层组合，而远程、原子化的 MCP 很难继承和适配现场环境。 但 Skill 可能污染环境，因此黄东旭进一步提出 “box”：把菜谱、食材和厨房绑定成可重复、无副作用的执行单元，才能拼装出登录、绑卡、下单等复杂能力。

Clawdbot（节目中称已更名 OpenClaw）的意义更像通用 Agent 原型，而非已经建立护城河的单一产品：它把 coding agent 的确认和权限限制拿掉，再加本地控制、Skills、长期记忆与聊天入口。 黄东旭设想未来操作系统就是“聪明的编程 Agent 内核—Skills—人机交互层”；它能自行接入 Whisper、理解陌生博客代码并把经验写成 Skill，显示出粗糙但真实的自我补强迹象。

真正的瓶颈正从“能不能做”转向算力和单位经济性，企业因此可能成为第一批规模化买单者。 黄东旭愿意每天烧接近 billion 级 token，是因为“哪怕花十万美金”做出的东西能卖一百万美金；普通用户却未必能从约 200 美元成本中创造更高价值。其推演是 one-man 外包公司先重做低效 ERP 等老系统，再向小店、图书馆和家庭的长尾软件需求扩散。

组织形态可能收缩为少数 senior architect 各自带领大量 Agent，但模块之间必须严格分界。 一百个 Agent 的意义不是角色扮演，而是并行消耗算力、突破单实例每秒约十个 token 的吞吐上限；若两个 Agent 在同一模块重叠工作，就会出现“一边已走一百公里，另一边说第五公里走错了”的灾难，重写反而更便宜。

人类剩余的稀缺性不是继续和 AI 比执行，而是决定“想要什么”、建立人与人的连接，并把主体性和审美放进作品。 黄东旭对工程师的建议是“把自己变成一个更有趣的人”；对创业者则是走向两个极端——把某件事用 AI 提效 100 倍或 1,000 倍，或去做“人味儿特别重”的最后一公里。音乐、现场表演和造物心流之所以仍成立，是因为人消费的不只是结果，也是创作者本人。

🔗 Original source & video: From Clawdbot to the 2026 AI Coding Theme Boom | A Conversation with PingCAP CTO 东旭

From Day 1, PingCAP envisioned itself as the storage foundation for future AI, rather than deriving its products from database customers and monetization. In Person of Interest, the AI needed to retain everything it saw and thought, triggering 黄东旭’s pursuit of an “infinitely scalable system”; in TiKV, that became one codebase running from 3 machines to 30,000, with data splitting, migrating, and self-healing like cells. His core engineering judgment: “You can only beat complexity with simplicity; you can’t beat complexity with complexity.”

Ten years on, roughly 70% of the business is overseas, and PingCAP has shifted from software licenses to cloud services—but it did not get that model right until 2022. The early approach of deploying databases on the cloud and charging by time had fundamentally broken unit economics; after 2018, the company treated the cloud as a “faucet” and rewrote its storage kernel as pure cloud-native infrastructure. Internationalization likewise shifted from “taking a Chinese company overseas” to regional localization: each region sets its own strategy, and even with organizational redundancy, the model saves the more expensive cost of consensus.

The data-paradigm shift 黄东旭 is betting on is that data’s consumer will move from the data analyst to the large language model. His Salesforce project does not pre-build reports; it puts raw data into a database and lets the model generate dozens of SQL queries on demand, turning an analysis process that once took 3-5 days of back-and-forth into a conversation. In this vision, Salesforce, with a market value of roughly $227B, looks more like a data provider. PingCAP’s farther-out goal is to provide everyone with a free personal database and become the shared memory for all AI agents.

AI is currently more an amplifier of exceptional talent than an immediate reason to cut 50% of the workforce. 黄东旭 estimates that traditional programmers spend roughly 80% of their time on boring work, of which AI can already take over 50%-60%, effectively freeing about half their time; but complex systems still require people to choose the architecture, build the code skeleton, evaluate the result, and then have Cursor “fill in the blanks.” He explicitly rejects inexperienced people blindly relying on vibe coding: “Absolutely not,” because once the “first shot” is wrong, the agent will keep moving in the wrong direction.

In PingCAP’s logic, open source is both a belief system and a competitive strategy for building trust, contesting standards, and reaching top-tier customers. New infrastructure software does not have Oracle’s decades of accumulated credibility, so “without open source, you die”; DeepSeek’s ability to enter overseas data centers and MCP’s potential to generate network effects likewise depend on open ecosystems. 黄东旭’s historical judgment is that “the first winner is often closed source, but the second wave is bound to be open source,” while Android and iOS, or Linux and Windows, do not necessarily need a single winner.

黄东旭 temporarily treats large language models as “a new species,” not because they can already evolve autonomously, but because humans cannot understand how their formal cause produces generalization. A Transformer can be implemented in roughly 1,000 lines of code, with the goal merely of predicting the next token, yet it is like “building a table and somehow having food emerge on the tabletop”; the host’s key rebuttal is that models still depend on humans for training. The unresolved disagreement is whether advanced intelligence should be judged by task outcomes or by its ability to learn autonomously and discover new rules.

His deepest response to AI is not to accelerate further, but to remove work, property, and possessions from the definition of self. 黄东旭 sold his house and moved into an RV, planning to learn gardening, sewing, woodworking, and repair before turning 40; his answer is “to become a person.” He has also turned toward meditation, old instruments, cassettes, vinyl, and cooking by hand. His final advice is intensely concrete: “Grow a flower, cook a dish, and then seriously get to know the people around you,” because AI can code for you, but it cannot experience life for you.

🔗 Original source & video: 121: PingCAP’s 黄东旭 and His “Inner World Source Code”: Black Mirror, the Beauty of Programming, and Creative Freedom

TiDB’s core value is turning the “1,000 cups” that legacy databases must manage into a “magic cup” that expands automatically as the data grows. Founded in 2015, PingCAP spent its first 2 years mostly writing code before putting a distributed relational database behind use cases including Meituan Waimai, WeChat Pay, and bank-card transactions. 黄东旭’s view is that data keeps growing while traditional containers remain finite; TiDB’s real product value is reducing the complexity of scaling and managing multiple containers.

Open source’s moat is not publishing source code; it is using a transparent process to earn trust, attract talent, and retain control of the project’s direction. Top-tier open source does not mean throwing finished code onto GitHub. It means opening the entire build process from the first document, like turning a restaurant kitchen into a glass room. “Giving you the answer to copy is not enough”; the value lies in the thinking that keeps evolving. PingCAP has roughly 300 engineers, annual investment in the hundreds of millions of yuan, and years of accumulated historical know-how. Even a later entrant deploying 500 people at once would struggle to change the direction quickly.

Before revenue scaled, investors were buying non-consensus technical judgment and steadily rising adoption. PingCAP’s last publicly disclosed valuation was $3B, but at Series A the software could not even run and the company had no revenue. 云启’s 陈昱 reached a decision after scanning roughly 5 minutes of source code, while investors who could not read the code had to rely on lagging commercial metrics. One concrete example: a major customer had no budget but sent 4 or 5 engineers earning roughly RMB500K–1M a year to work on the project, creating a form of open-source value exchange.

Cloud services finally made it possible to commercialize open source without breaking the trust underpinning it. PingCAP was “dead set against commercialization” for at least its first 5 years. In the early days, salespeople went to large users to “collect an insurance premium,” creating information asymmetry—a practice 黄东旭 saw as “killing the hen for the eggs.” MongoDB’s growth after shifting from a software company to a services company offered a model to follow. Today, more than 70% of PingCAP’s ARR comes from cloud services; growth has stayed in the triple digits for essentially each of the past 2 years, driven mainly by internationalization and cloud.

True globalization is not “going overseas” remotely from China; it is the founder physically embedding the company, organization, and commercial relationships in the local market. 黄东旭 went to San Francisco in June 2015, just as the company began operating, took several wrong turns from 2017 onward, and moved his entire family to the US in 2022. His experience is that a company may make almost no money during its first 3 years abroad, and that the founder has to go all in personally rather than relying on a global team. English-first documentation from day one, company-wide English for internal IM and meetings in 2022, and local teams with regional CEOs and CTOs were organizational milestones produced by this “working backward” approach.

For US GTM, Chinese founders most often overestimate the language barrier and commercialization difficulty, and underestimate narrative, pricing, and long-term relationships. 黄东旭’s advice to AI founders is to live locally for at least 6 months, hire a local salesperson early, and sell the product personally as the founder. English only requires the courage to speak; before moving to the US, he spent 2 hours a day speaking with a tutor to remove the psychological barrier. The real problem is often having a valuable product but “being afraid to sell it at a high price,” or failing to win mindshare with demos, message frameworks, and stories. In the AI era, “good wine needs no bush” no longer holds.

PingCAP’s core AI bet is not another model, but providing agents with the right context, enterprise data, and a shared memory layer. 黄东旭 believes DeepSeek’s emergence at the beginning of the year showed that LLM fundamentals are now sufficient to support useful applications, with the first wave of restructuring likely to hit enterprise software. His own company still has 3 full-time employees maintaining Salesforce reports, while its data agent can already answer questions about sales activity directly. 黄东旭 first mentioned the interactions covered by NCP and A2A, then later used MCP as an example; a general-purpose shared memory layer has yet to emerge. The next 2 or 3 years will remain a “wild west,” where open standards, real usage, and network effects matter more than designing monetization first.

🔗 Original source & video: Experience, Lessons, and Methodology on AI, Open Source, Commercialization, and Globalization | A Conversation with PingCAP CTO 东旭