在Meta从事AI开发八年后，Rohit Patel正离开该公司，TechCrunch独家获悉。这位Meta超级智能实验室前总监已创办初创公司Ndoc，将其定位为可取代Microsoft Office套件或Google Workspace的AI原生产品。
After eight years developing AI for Meta, Rohit Patel is departing the company, TechCrunch has exclusively learned. The former director at Meta Superintelligence Labs has launched Ndoc, a startup he’s pitching as an AI-native replacement for Microsoft Office suite or Google Workspace.
Patel是继首席全球事务官Joel Kaplan、AI研究员Andrew Tulloch以及AI工作产品负责人Emily Dalton Smith之后，又一位离开Meta的高调人物。
Patel is the latest in a handful of high-profile departures from Meta following the exits of Chief Global Affairs Officer Joel Kaplan, AI researcher Andrew Tulloch, and Head of Product for AI-for-Work Emily Dalton Smith.
Patel设计Ndoc的初衷不仅仅是将另一个AI工具叠加到现有工作流之上。其目标是取代人们跨Excel、Word、PowerPoint、Tableau、Replit以及“ChatGPT复制粘贴”等工具碎片化的工作方式，提供一个统一的工作空间，所有同事均可访问，即使他们未注册Ndoc账号。
Patel designed Ndoc to be more than just another AI tool layered onto an existing workflow. The goal is to replace the fragmented way people work across tools like Excel, Word, PowerPoint, Tableau, Replit, and “ChatGPT copy-paste” with one unified workspace that all colleagues can access, even if they’re not signed up for Ndoc.
Ndoc在一个布局中整合了三大组件。“舞台”标签页用于呈现关键发现的演示文稿，如幻灯片、仪表盘、模拟或微型应用。“叙述”标签页包含分析、编号引用、生成的图表，以及数字背后的“为什么”。“模型”标签页则存放底层数据集和Python脚本。
Ndoc brings three components together in one layout. The Stage tab is where presentations of key findings live, like slide decks, dashboards, simulations, or mini apps. The Narrative tab features analyses, numbered citations, generated figures, and generally the “why” behind the numbers. The Model tab is where underlying datasets and Python scripts are housed.
这一设计反映了Patel的背景。在Meta任职期间，Patel专注于强化学习——一种AI智能体通过试错学习决策并最大化奖励的机器学习类型——以及评估AI输出是否可信。Patel认为，如果AI将构建企业用于决策的电子表格和幻灯片，就应该有人能够核查其工作。例如，用户可以找到创建文档所用的Python脚本。
The design reflects Patel’s background. During his tenure at Meta, Patel focused on reinforcement learning — a type of machine learning in which AI agents learn to make decisions through trial and error and maximize rewards — and evaluating whether AI output is trustworthy. Patel believes that if AI is going to build the spreadsheets and slide decks that companies use to make decisions, someone should be able to check its work. For instance, users can find the Python scripts used to create the documents.
“即使你不懂任何 Python，也可以问 [AI 聊天]：‘嘿，这段脚本是干什么的？哪一行是我的公式？’然后你就能核实那个公式，”Patel 说。“这里没有任何幕后隐藏的东西。”在实际使用中，用户可以上传数据集，让 Ndoc 进行分析并撰写发现的叙述性报告。他们还可以添加特定于文档的上下文，包括受众、语气和记忆。“它会在你个人、那个特定文档、你的团队和你的公司之间保持记忆，”Patel 说。
“Even if you don’t know any Python, you can ask [the AI chat] to explain, ‘Hey, what does this script do? Which line has my formula?’ And you can verify that formula,” Patel said. “There’s nothing hidden behind the curtain here.”In practice, users can upload a dataset and ask Ndoc to perform an analysis and write a narrative about the findings. They can also add document-specific context, including audience, tone, and memory. “It maintains a memory across your person, that particular document, your team, and your company,” Patel said.
Patel 还表示，Ndoc 是使用开源模型的混合构建并微调而成的。因为该平台是从零开始开发的，他说运行成本比基于前沿 AI 模型构建的替代方案要低得多。
Patel also says Ndoc is built and fine-tuned using a mixture of open-source models. Because the platform was developed from scratch, he says it costs considerably less to run than alternatives built on frontier AI models.
“我以为这会是人们兴奋的点，但人们根本不在乎，”他说，指的是早期试用用户。“人们兴奋的是，他们喜欢看到所有端到端的流程都在一个地方。”Patel 更大的野心是重新发明文件本身。他说，文档可以导出为单个 “.ndoc.html” 文件，可在任何浏览器中打开，他计划将这种格式开源，以便其他开发者能构建与之兼容的应用，而不必像过去多年那样被锁定在 Microsoft 或 Google 的生态系统中。
“I thought that’s what people would be excited about, but people don’t care about that,” he said, referring to his early trial users. “What people are excited about is that they love seeing all their end-to-end in one place.”Patel’s larger ambition is to reinvent the file itself. He says documents can be exported as a single “.ndoc.html” file that can be opened in any browser, a format he plans to open source so other developers can build apps that work with it without being locked into Ndoc the way so many have been locked into Microsoft or Google for years.
他想要开源这个项目，部分原因促使他离开 Meta 独自创业。在像 Meta 这样的大公司里，要让所有人就开源格式达成一致，无疑需要 “数百人的意见，关于格式该长什么样。”
His desire to open source this project was partly why he left Meta and decided to go it alone. At a big company like Meta, getting everyone to agree on an open-source format would undoubtedly require the input of “hundreds of people and opinions on how the format should look.”
Patel 想要快速推进。“人类只能按人类的节奏前进，而……我可以按 AI 能做到的节奏前进。”对于一个只有三人团队、零融资的创始人来说，试图成为企业工作发生的平台是一场巨大的赌注，尤其是当其他 AI 实验室也在争夺同一领域时。
Patel wanted to move quickly. “Humans can only move at human pace, whereas…I can move at the pace of what AI can do.”Trying to become the platform where enterprise work happens is a huge bet for a founder with a team of three and no funding, especially when other AI labs are trying to own the same space.
OpenAI、Anthropic、Meta、SpaceXAI 和 Google 都推出了企业级人工智能工具，允许用户连接到 Google Workspace、Microsoft 365 和其他办公应用程序，以便他们的人工智能模型可以直接在这些工具中使用。上个月，OpenAI 推出了 Space，这是 ChatGPT 内的一个共享工作空间，直接与这些成熟的工具竞争。
OpenAI, Anthropic, Meta, SpaceXAI, and Google have all come out with enterprise AI tools that allow users to connect to Google Workspace, Microsoft 365, and other workplace apps so their AI models can work directly in those tools. And last month, OpenAI launched Space, a shared workspace inside ChatGPT that’s a more direct competitor to those established tools.
Patel 承认 Space 可能构成威胁，但他表示这些实验室的想法大体上是错误的。
Patel acknowledges that Space could be a threat, but says the labs largely have the wrong idea.
“公司基本上是将人工智能附加到电子表格上，然后它会为你填满单元格，”他说。“这简直让我灵魂枯竭。这就好比你刚刚发现了电力，然后你决定最好的用途是给我的耕牛装上灯，让它们耕地。不行。应该用拖拉机。”然而，人工智能实验室试图将他们的模型带到人们已经工作的地方是有原因的。改变习惯和取代整个工作流程是困难的。Patel 表示他理解这一点，因此他构建了连接器，以便客户可以导入他们已经使用的工具中的现有数据。
“Companies are basically attaching AI to spreadsheets and it fills your cells for you,” he said. “It just sort of kills my soul. This is like if you just figured out electricity, and then you decided the best use case is to put lights on my bulls that till my soil. No. Get a tractor.”There’s a reason the AI labs have tried to bring their models to where people are already working, though. It’s hard to change habits and replace entire workflows. Patel says he understands that, which is why he built connectors so customers can import existing data from the tools they already use.
“我的标准是，我不会发布任何我不用得开心的东西，”他说。“我对我的所有工作都有高度优化的 AI 工作流程。如果我开始使用这个工具，并且我一直回到我的工作流程，那就说明我没有构建一个足够好的工具……我有一个未能达到这个标准的工具坟墓。”Patel 在寻求客户支持之前，也有一个类似的衡量标准。“在有证据表明人们喜欢这款产品之前，我不想接受资金，”他说。
“My bar was that I’m not going to release anything I don’t love using,” he said. “I have very optimized AI workflows for everything I do. If I start using this tool, and I keep going back to my workflows, I haven’t built a good enough tool…I have a graveyard of tools that did not meet that bar.”Patel has a similar bar for gaining customer traction before he tries to raise funds. “I didn’t want to take money until I had some proof that people love the product,” he said.
截至周二，Ndoc 正在进行企业试点，并向加入等待名单的有限数量的客户开放了一个阿尔法版本。注册是免费的，但客户将按使用量和代币付费。他希望在未来几周内发布一个测试版，并计划于明年全面推出。
As of Tuesday, Ndoc is conducting an enterprise pilot and opening up an alpha version to a limited cnumber of customers who join the wait list. It’s free to sign up, but customers will pay for usage and tokens. He hopes to release a beta version in the coming weeks, with a full launch planned for next year.
对 Patel 来说，尽快行动比一开始就拥有完美的产品更重要。
For Patel, moving as fast as possible is more important than having the perfect product from the get-go.
他说：“执行的成本极低……所以先按一个并非最优的方案去做，之后再操心这些问题。等你制定出最佳方案时，别人早就把这件事做完了。”
“Execution is extremely cheap…so go execute on a suboptimal plan and worry about it later,” he said. “By the time you make the best plan, the execution will be done by somebody else.”