For the first properly recorded episode of Dine Podcast, I sat down with Tianlu Xue, co-founder and CEO of Second Enlightenment. Fifteen years ago Tianlu was the first intern I ever hired, at Zhihu. He later spent four years running product at early Xiaohongshu, then five at ByteDance. He left ByteDance last year to start his own company. His newest product, Clawly, is a cloud-based OpenClaw client built for users who have never set up a local environment. For most of our hour together, though, we were not actually talking about the product. We were talking about a question Tianlu has been turning over for months: how far ahead the frontier of AI has moved, and how little of that the ordinary user has noticed.
At Every Beginning
Tianlu’s career has placed him near the uncertain early days of three of the most significant Chinese content platforms of the last fifteen years. He was a Zhihu user before he was a Zhihu employee, back when invite codes were still being handed out one by one through a Google Form. He joined Xiaohongshu at the end of 2016, at a stage when no one on the team could have told you with any confidence what the product would eventually become. He went to ByteDance in 2020, mostly out of curiosity, wanting to see up close how Douyin had grown into the first Chinese consumer product to work globally. What he takes from that sequence, he said, is that the shape of a product at the moment it actually begins to matter almost never resembles what its builders described at the start. The people standing closest to the fire are usually the ones least able to see where it will burn.
He sees the same early uncertainty in AI agents today. The capability is already real. Public understanding of it lags far behind, and the people actually doing the work still do not know what shape it will eventually settle into.
Leaving, and Starting Over
What finally pushed Tianlu out of ByteDance in mid-2025 was the shape of the role one rung above his. Management at that scale, he found, was mostly organizational work. The product judgment and product making that had originally drawn him to the industry was being quietly crowded out by the kind of career he did not want to sign himself up for over the next decade. He gave himself an entire summer. He spent a few weeks in Italy looking at Roman ruins, did a short EIR stint at Sequoia, and came home to sit with the question of what he actually wanted to build. Somewhere in the middle of that stretch he wrote a small iOS app for himself, alone at a desk, from beginning to end, something he had not been able to do in years. The app itself did not matter commercially. What mattered was the reminder that the tools had changed enough that one curious person could now build a real, working product on their own. By the fall, he had put together his team.
Taken by Tianlu at Nagoya CastleThe Making of Clawly
In January, Tianlu’s team was building a set of productivity agents on top of the Claude Agent SDK. They had not shipped yet when the ecosystem around them moved. Late 2025 was the stretch when Opus and its surrounding tooling stepped up noticeably in capability, and the arrival of OpenClaw raised the floor high enough that almost any single-purpose agent product started to look marginal beside it. Tianlu’s team had been running OpenClaw themselves, and they knew firsthand how punishing the experience was for anyone who did not already live in a terminal. Deploys broke. Tokens had to be configured by hand. Every new capability required another round of wiring. Clawly was their shortest answer to that problem. Put an agent on a full cloud computer with enough compute to do real work, let it reach into your phone and your desktop, give it the same skill library Claude Code uses, and tuck all of that behind a chat interface.
Building Clawly has clarified something for Tianlu in particular. The foundation underneath the foundation model is thin. The model itself is a giant. Almost everything beneath it, including the tools an agent needs to reach for, the ways it takes in a user’s existing context, and the connections to the platforms where that context actually lives, is either missing or still very early. His team spent a stretch seriously considering a pivot toward infrastructure. The agent tasks most obviously worth paying for today are not consumer tasks. They are business tasks. What is blocking those tasks is not the intelligence of the model. It is access. Agents today still cannot reach most of a user’s existing data, inside or outside the major platforms, and whoever opens that gap will largely determine the shape of the next round of agent products.
The Distance
Since the beta opened, the distance Tianlu has been watching between the frontier and the user is not really about capability. It is about awareness. Last month he visited the owner of a small advertising firm in Shanghai who had found his way to Clawly through a friend. They sat down, and the first thing the man asked him was whether he had heard of Hermes. A different beta user, at a different small company, had recently demonstrated to his coworkers a spreadsheet he had built in ten minutes with Clawly. The same task used to take him an hour. The capability had been sitting in the base model for months. Nobody had ever told him where to find it.
He quoted Karpathy’s recent post to me. This is not the year of agents. It is the decade of them. Most ordinary people’s primary experience of AI is still a stale voice mode on ChatGPT, routinely mocked on TikTok as evidence that AI is stupid, while the real frontier keeps moving quietly forward somewhere else. Only then did I realize what Tianlu had been pointing at in his company’s name. It was pointing at what he hoped would be on the other side of that decade.
Mentioned in this episode
- Tianlu on X and Jike
- Japan Castles and Tianlu’s visit log on Xiaohongshu
- Clawly on the App Store and on the website
- Replit Agent, Hermes Agent, Personal Computer, Poke
- Andrej Karpathy
Music: SeeN — SunnyWild / Jillusion (HEM Records)
Dine Podcast 的第一期正式录制节目，我邀请了 Second Enlightenment 的联合创始人兼 CEO 薛天禄。十五年前，天禄是我在知乎招的第一位实习生，后来他在早期的小红书负责了四年的产品，又在字节跳动待了五年。去年他从字节离开，开始自己的公司。他的新产品 Clawly，是一款面向从未在本地配过环境的用户的云端 OpenClaw 客户端。不过那天的对话里，大部分时间其实并不在聊这款产品本身，而是在聊一个天禄最近反复在琢磨的问题：AI 的前沿已经跑出去很远，但大多数普通用户还完全没有察觉。
每一个起点
天禄的职业经历让他在过去十五年里，亲历了三家最具代表性的中国内容平台的早期阶段。他在成为知乎员工之前，就已经是知乎的用户，那时邀请码还是通过 Google 表单一份份发出去的。他在 2016 年年底加入小红书，那是一个团队里还没有任何人能够笃定地告诉你这款产品最终会长成什么样的阶段。他在 2020 年去字节，更多是出于一种好奇：想近距离看一下抖音是怎么成为第一款在全球范围内做成的中国消费级产品的。他说这段经历给他最大的启发，是一款产品真正开始产生影响力的那个瞬间，它的形态几乎从来都不是创造它的人当初所描述的样子。距离火苗最近的人，往往是最看不清它会烧向哪里的人。
他在今天的 AI agent 身上看到的是同一种早期不确定性。能力已经是真实的能力，但公共层面的认知远远滞后，并且真正在做这件事的人也还不知道它最终会沉淀成什么样的形态。
离开，再重新开始
2025 年年中，真正让天禄下决心离开字节的，是再往上一级那个岗位的形态。那个层级的管理更多是组织层面的工作，产品上的判断和构造变得越来越少，而更像是一份他不愿意在接下来十年里押上的职业。他给了自己一整个夏天。去意大利待了几个星期看罗马的遗迹，又在红杉做了一段短暂的 EIR，然后回家想了一阵自己到底想做什么。中间他抽空给自己写了一款 iOS 小应用，从头到尾一个人做，完成了过去很多年做不到的事情。这款应用商业上无关紧要，关键在于它提醒了他一件事：工具已经变化到了一种程度，一个有兴趣的个人可以独立做出一款真正可用的产品。到了秋天，他组建了自己的团队。
天禄摄于名古屋城Clawly 的诞生
今年 1 月，天禄的团队在基于 Claude Agent SDK 做一批生产力方向的 agent。他们还没上线，整个生态就已经在他们身边发生了位移。2025 年年底那段时间，Opus 和它周边的工具链在能力上明显跨上了一个台阶，而 OpenClaw 的横空出世，也到达了一个让几乎所有单点 agent 产品都显得边缘化的水位。天禄团队自己就在用 OpenClaw，亲身体会到这件事对任何一个不是天天泡在终端里的人来说都是非常痛苦的，部署脆弱、token 要手工配置、每添加一个新能力都要重新接一次线。Clawly 是他们对这个问题最简短的回答：把一个 agent 放到一台配置足够完整的云电脑上，让它能够访问你的手机和你的电脑，使用和 Claude Code 同一套 skill 生态，并把所有这些东西都藏在一个聊天界面之下。
做 Clawly 这段时间，让天禄看清楚的一件事情是，基座模型之下那一整层 agent 的底座有多薄。模型本身是一个巨人。它脚下几乎所有的东西，包括 agent 要去调用的工具，它接入用户已有上下文的方式，以及它和用户数据真正所在的那些平台之间的连接，要么还不存在，要么还非常早期。他们团队中间一度认真考虑过整体转去做 infra。今天 agent 上最值得付费的那批任务并不是消费级任务，是工作相关的任务。而今天卡住这些任务的并不是模型的智能，是访问权限。Agent 今天依然拿不到用户大部分的现有数据，无论平台内还是平台外，而谁能把这个缺口打开，谁就会在很大程度上决定下一代 agent 产品的形态。
那段距离
内测上线之后，他看到的前沿和用户之间的那段距离，其实并不在能力上，而是在认知上。他上个月去拜访了上海一家小广告公司的老板，对方是通过朋友辗转开始用 Clawly 的。他们一坐下来，老板问他的第一句话是你知道 Hermes 吗？另一位来自另一家小公司的内测用户，前不久当着同事的面演示了自己用 Clawly 花十分钟做出的一张表格，过去同一件事要花他一个小时。这项能力在基座模型里已经存在了好几个月，从来没人告诉过他应该去哪里找。
他跟我引用了 Karpathy 的那条帖子，今年不是 agent 的一年，是 agent 的十年。大多数普通人今天和 AI 最主要的交互，依然是 ChatGPT 上一个比较老的模型跑的 voice mode，在 TikTok 上反复被当成「AI 怎么这么蠢」的素材，而真正的前沿在别的地方悄悄继续往前。听到这时我才意识到，天禄给自己公司起的名字，对准的是这十年另一头他希望看到的东西。
这期节目里提到的
- 天禄的 X 和 即刻
- Japan Castles 和 天禄的打卡记录
- Clawly 的 App Store 和 官网
- Replit Agent、Hermes Agent、Personal Computer、Poke
- Andrej Karpathy
背景音乐 SeeN — SunnyWild / Jillusion (HEM Records)