在 1024 模力工场·AI社区日主论坛上，深空之序联合创始人兼 COO 芦宇峰，以《FDE：AI 落地的核心竞争力》为题作主题分享。
值得一提的是，深空之序，是芦宇峰与前钉钉 CEO 叶军共同创办的 AI 公司，专注于跨境电商场景的 AI 业务智能体 Agent Nine。
首先，他分享了一个刺绣项目中的教训：AI 生成的图案看起来不错，老师傅一看，却发现机器需要剪线、打结 700 多次。从“生成结果”到“实际能用”，这中间的差距，正是 FDE 需要深入现场解决的问题。
在这个项目中，团队最初尝试让 AI 模仿老师傅操作软件，却发现复杂的工艺判断难以靠复制操作完成，于是转向利用标注数据和后训练，根据原图、面料和针法直接生成针迹文件。据芦宇峰介绍，部分中低难度图案已能实现分钟级产出，但具体工艺问题仍需团队驻厂，陪着老师傅实际打版、反复验证和调整。
第二个案例，是在义乌的报关项目中，团队面对的则是零散、模糊的照片和单据，通过现场收集资料、改进识别能力，将多数报关单的制作时间从至少 3 小时缩短至约 5 分钟，单份制作成本从最低约 70 元降至约 10 元。
芦宇峰强调，FDE 不能只在旁边听、拿着设备录，而要亲手做一遍业务，弄清老师傅凭经验作出的判断，以及流程中可能出现的例外。
收集到的业务资料和操作经验，一定要整理进知识库；解决问题的方法则要沉淀为可复用的 Skill，并配套覆盖真实场景和客户运行环境的测试用例，避免“开发时没问题，到了客户现场却用不了”。
在他看来，交付成效最终要看省了多少时间、降低了多少成本。而 FDE 更长远的价值，是把一次项目的经验变成产品能力，让下一个遇到同样问题的客户，不再需要团队从头驻场解决。
以下为芦宇峰讲演的完整速记，模力工场在不改变原意的前提下进行了订正和编辑。
FDE：要亲手跟老师傅做业务，再把经验解法做成 Skill
大家上午好！很高兴来到这里，和大家分享 AI+FDE。
先简单介绍一下自己。我之前很多年都在创业，也在阿里待过，在不同企业担任 CTO，前后做了 14 年。一个岗位做久了，难免会少一些新鲜感。创业这么多年，我也想尝试做业务，所以在新的创业项目中，我开始转变角色，担任深空之序联合创始人兼 COO。
现在，我要出去做业务、签单，推动具体场景落地。我负责渠道建设和生态伙伴管理，也组建了一支 FDE 团队，负责企业 Agent 的交付和落地。
今天，我想和大家分享这段时间亲身经历的几个项目，以及我们在其中遇到的问题。
我不太擅长一上来就讲很多概念，我们先看一个具体场景：刺绣。
刺绣是一个很有意思的产业，在浙江诸暨非常发达，可能占全球产业份额的 80%，甚至更高。
为什么我们会进入这个场景？因为我们讨论的是 AI 在企业中的落地，工业企业也是其中的一部分。
先看这张图：这是一只很好看的狐狸，一边是原图，另一边是刺绣图。要把一张效果图变成刺绣机器可以使用的针迹文件，其实非常复杂。这个过程需要打版师先出效果图，再使用一款昂贵的国外软件，花很长时间完成打版设计。这款软件类似工业 CAD，叫 Wilcom。
我们试用过这款软件。最初考虑的是，能不能用 Computer Use 这类能力来完成操作？后来发现做不到，因为其中涉及的判断非常复杂。
这里有哪些工艺要求？比如，不同机器有不同的配色，图案需要分成很多层，有的要先画轮廓，再做填充。这些都是刺绣的基本工艺原则。
还有一些更复杂的图案，比如山水图。刚才那张狐狸图已经算比较简单的了。如果按 1 到 10 划分难度，它最多是 4 或 5；到了需要真实还原人像的程度，就会非常复杂。
我们恰好认识诸暨一位做了多年数字化的生态伙伴，聊起这个话题时，他也在思考：AI 能为刺绣打版做些什么？
于是，我们做了一系列深入研究，发现其中有不少问题。
一方面是耗时长，可能需要 5 天，软件也很昂贵。另一方面，大量打版经验都在老师傅的脑海里，并没有形成系统的课程。教一些简单操作，比如怎么使用软件，没有问题；但如何把一张好看的图变成刺绣文件，涉及对颜色的判断、审美，以及刺绣工艺的设计，这些通常没有人教，更多依赖老师傅的经验。
所以，我们要做的，是大家前段时间经常提到的“蒸馏老师傅”。但有多少老师傅愿意被“蒸馏”？这是一个很现实的问题：我凭什么让 AI 取代自己的岗位？
这也是 FDE 在现场经常遇到的问题。你要怎么说服他，让他接受 AI 的目的不是取代他，而是提高工作效率，让他能够提前下班、早点回家，不用再用眼睛仔细盯着每一个针脚，琢磨怎样才能做得更好。
我们最初的方案，是用 Computer Use 这类能力模拟老师傅在电脑上的操作。
前期做了很多录屏，拿回来分析，想看看能不能原样复现。后来发现，根本无法通过这种方式完整复制、蒸馏整个操作过程。
于是，我们换了一个思路：为什么还要沿用原来的工作方式完成任务？为什么不能用 AI 的方式来做？
后来，我们把流程改成：上传原图，选择刺绣面料和相应针法，比如平针等，再直接生成针迹文件。
这能不能做到？我们拿了两万多套相关数据，做了大量人工标注，也做了一些后训练工作，最终实现了分钟级产出。
如果按 1 到 10 划分图案难度，10 是最难，我们认为，难度在 1 到 5 的图案已经可以做到分钟级出图。
这就是我们在诸暨工业场景中开展 FDE 现场需求调研、最终实现的效果。产出中包括刺绣图和工艺文件，里面会列出所涉及的色号、针数等信息。
当然，我们在这个过程中也遇到了很多问题。刺绣涉及很多颜色，换针时还会涉及打结和剪线。
最早，老师傅看了我们生成的那张图，发现竟然需要进行 700 多次剪线和打结。这些都是具体工艺中的问题，必须由 FDE 到现场解决。
我们的 FDE 在工厂待了很长时间，陪着老师傅亲手做过打版，再用 AI 针对这张图反复训练、验证，才得到了后来的结果。
再来看另一个 FDE 落地场景：报关。
义乌有很多外贸企业，报关是其中非常重要的一环。我们遇到的一种比较复杂的场景是“1039”，也是义乌这边很有代表性的报关场景。
企业可以请外部人员制作报关单，费用大多在 70 元到 200 元左右。根据照片的清晰度和资料的完整程度，手工制作一张报关单，大概需要半天到一天。
这些就是“1039”场景下通常收到的资料，有时只是简单拍一张照片，包括门锁、仓单、称重单等。资料不止这些，数量很多，也很杂，清晰度有时也存在问题。
这也是人工制作报关单耗时很长的原因：资料非常混乱。
有些场景已经可以通过软件在线生成资料，或者提供电子化文件，但我们接触的“1039”场景仍然比较复杂，很多环节依赖人工，提交资料的方式就是简单拍一张照片。所以，以前人工制作报关单需要花很多时间。
我们有一位 FDE 同事在义乌待了很长时间，收集报关所需的相关资料和信息，不断让 Agent 运行、测试，改进对这些数据的识别能力。
最后看效果，大部分报关单的制作时间，都可以从原来的至少 3 小时缩短到 5 分钟。原来最低一单 70 元的制作成本，用 AI 后大概是 10 元，甚至更低。
以上就是两个由 FDE 在现场开展场景调研、挖掘需求，并最终完成交付的案例。
回到这两个案例，我们认为 FDE 应该做什么？
在组建 FDE 部门时，我们会思考不同岗位的关注点。以前，研发更多关注功能和场景，产品经理关注如何整理分散的需求，形成完整的产品。FDE 关注的范围会更广，要从场景、产品到整个行业，全面了解并挖掘需求。销售则要把产品转化为行业解决方案，再把它卖出去。
客户的视角又不一样。前面这些岗位更多从软件或 AI 产品的角度出发，而客户身处具体行业和产业，关注的是：什么对自己的行业和产业最有利，什么能够真正帮助自己？
那么，FDE 进场的第一件事是什么？收集数据。
最近，我们看了不少失败案例，其中一种情况是：进场后就急着干活，先盯住一件事，想着能不能从某一个点打开局面。
我们之前也有一个项目遇到了类似问题，效果并不好。后来，我们换了方式：FDE 进场后，先收集资料，尽可能把企业的上下文信息收集起来。
这里的上下文，不只是企业已有的文档。有些小企业的文档管理并不完善，很多信息都在微信里，还有一些分散在不同员工的电脑中。这些数据都需要收集起来。资料格式可能各不相同，有些特殊格式还需要专门解析。
必要时，我们也会做数据导出，把资料收集齐全。现在 AI 的摘要能力已经很强，可以借助它拆解、整理资料，形成企业知识库，作为后续开展工作的知识来源。
收集资料之后，还有一个非常重要的过程：FDE 一定要跟着客户，跟着所要解决的问题对应的岗位，亲手把业务做一遍。千万不要只是调研，站在旁边拿录音笔或其他设备，录下客户怎么说、怎么讲。这样很难真正体会客户的难处。一定要亲自下场操作，自己走一遍流程，才能理解其中的问题。
有些地方，老师傅凭经验就能处理，甚至不觉得是问题，但对新人来说，其实非常难。
无论是报关单、刺绣，还是我们正在做的其他外贸场景，都是如此：一定要自己做一遍，了解其中可能遇到的问题和例外情况。
既然是新人，就要抱着学习的态度向老师傅请教：为什么我这样做会有问题？有没有例外？一定要把这些异常流程了解清楚，并核实处理结果。
亲手操作的过程，最后也要记录下来，写回知识库。
在“蒸馏”了一些岗位专家的经验之后，我们发现，这个过程有点像茅台的国标：工艺非常繁琐，但也非常规范。
这些过程，都可以在实施中沉淀下来。那么，应该用什么方式沉淀需求、客户场景和遇到的问题？
我们的岗位和刚才毕玄老师讲的不太一样。我们不是研发岗，而是偏业务执行的岗位。
但我们仍然强调：FDE 在交付时，要把自己运行、使用的方法一并沉淀下来。刚开始，你可能用提示词完成一项任务，但完成之后，就要尽快把它沉淀成 Skill，这样，你的能力库才会逐步完整。
我们的 FDE 团队有一个共享的 Skill 仓库，大家新写的内容都会上传，也会互相查看。遇到同样的问题时，我们会看看其他团队成员写的 Skill 能否解决，并不断尝试。
另外，还要构建交付的 Harness。在这里，可以把它理解为：FDE 必须收集足够多的测试用例，把用户场景转化为可以执行的测试。
现在，用 AI 执行测试已经没有太大难度。研发同学更多针对开发的功能、接口和界面进行测试，但基于真实业务场景的测试，一定要由 FDE 团队来做。
FDE 要收集足够多的场景用例。在研发完成交付、测试团队完成测试之后，还要用自己的 Harness 和测试用例库再跑一遍，因为这些用例来自现场真正遇到的问题。
有些 Bug 反馈给研发后，双方的理解可能不一样。研发可能从底层原理上修复了问题，但用户实际遇到的问题有没有解决，并不一定。
举个例子，我们现在很多人已经习惯使用 MacBook 开发，很久没有用过 Windows，但我们的外贸客户几乎都在用 Windows。
有时，开发完成后，即使装一台虚拟机运行，感觉没有问题，到了客户现场仍然可能出问题：客户使用的操作系统版本、防火墙，以及其他运行环境信息，都需要记录下来。开发同学不可能总去现场，这些工作就需要交给 FDE 团队完成。这是交付过程中的重要一环。
最后，还要测算业务结果。如果一件事没有带来实质性的效率提升和成本降低，它的价值就比较有限。
客户可能会问：现在人工成本这么低，请兼职、刚毕业的学生，甚至还没毕业的学生都不贵，为什么要用你的产品？
产品必须带来实质性的效率提升和成本降低，才有价值。因此，FDE 的最后一项工作，就是测算业务结果。
从实现功能，到交付具体场景，再到沉淀能力、形成 Skill 和 Harness，这就是我们刚才讲的整个过程。
前段时间，OpenAI 的一位研究员在 X 上分享过一个观点，我觉得非常好，大意是：如果下一个客户能够在没有我们帮助的情况下解决同样的问题，那么 FDE 团队就完成了使命。
我看到有篇公众号文章把它概括为“FDE 正在干掉自己的岗位”，有一点标题党的意味。可能有朋友也看到过。但我点进去后，认为它讲的不是让 FDE 失业，而是形成反馈：OpenAI 的 FDE 把与客户合作取得的成果转化为更好的产品，通过这种方式，让产品能够承担原本需要自己完成的工作。
这和一开始就让自己失业，是两个不同的概念。
最后这一页也是一个广告。我们是一家新公司，正在推动各个行业的企业 Agent 落地。目前进入的行业主要是外贸和跨境电商，这也是浙江非常繁荣的产业生态。
我们也在做银行风控，以及刚才提到的工业场景。我下周就要去拜访一些工业客户，了解 CAD 图纸生成等方面的需求。
我认为，在中国，真正能用好 AI 的人可能只有万分之一，甚至还不到。我们有责任帮助那些过去数字化基础还不够好的企业完成 AI 转型，否则，他们可能很难享受到这一波 AI 红利。
我负责渠道、生态伙伴以及 FDE 相关工作，也希望借这个机会和大家建立联系。我们有一些产品试用的机会，也欢迎大家在生态能力方面开展合作。即使在 AI 时代，即使几十个人每天并行使用多个 Agent 开发，我们也不可能完成所有功能、解决所有业务场景的问题。
所以，无论是具有行业背景的 FDE 个人，还是已经在某些行业积累了成熟 Skill、数字化能力的商业公司，我们都希望后续能够建立联系，一起合作解决这些问题。
我的分享就到这里，谢谢大家。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
At the main forum of the 1024 Moot Ai Community Day, the Co-founder of Deep Space Orders and CoO Zhou Zhou Feng shared the theme “FDE: A core competitiveness for landing”.
It is worth mentioning that the deep order of space was created by the AI Company, co-founded with the CEO Ye Army, which focuses on the cross-border electricity business scene of AI AI agent Agent Nine.
First of all, he shared the lessons of an embroidery project: AI created a pattern that looked good, and the teacher looked at it and found that the machine needed wire cutting and knotping 700 times. The gap between "generation of results" and "practical use", is exactly what FDE needs to do on the ground.
In this project, the team initially tried to get AI to imitate the teacher to operate the software, but found that complex process judgments were difficult to complete by replicating, and then shifted to using tagging data and post-training to directly generate needle files based on original drawings, fabrics and needles. According to Joo-Fung, some of the low- and medium-cost patterns have resulted in minute-level output, but specific process issues still require a team presence at the factory, accompanied by the teacher, who does the actual editing, validation and adjustment.
In the second case, in the customs clearance project in Yiu, the team faced scattered and vague photographs and documents that reduced the production time for most customs declarations from at least 3 hours to about 5 minutes, and the cost of single production from a minimum of about $70 to about $10, by collecting information on the ground and improving identification capabilities.
Joo-Fung stressed that the FDE could not just listen to and record the equipment, but would do it hand-in-hand to ascertain what the teacher had learned from experience and what exceptions might have been made in the process.
The collection of operational information and experience must be consolidated into the knowledge base; the solution to the problem is to settle into a reusable Skill, accompanied by test examples covering the real scene and the client's operating environment, so as not to “be developed in such a way that it is not used at the customer's site”.
In his view, delivery would ultimately depend on how much time and cost was saved. The longer-term value of the FDE is the ability to turn the experience of a project into a product, so that the next client with the same problem will no longer need a team presence to solve it.
The following is a full stenography of the lecture given by Joo-Fung, and the model workshop was revised and edited without changing its original intent.
FDE: You have to do business with your teacher, and you have to work out your experience.
Good morning, everyone! It's good to be here to share AIP.D.
Let me briefly introduce myself. I've been in business for many years, and I've been in Ari for 14 years. A post that's been in place for a long time will inevitably be less fresh. After all these years of starting a business, I've tried to do business, so in the new business project, I'm beginning to change my role as co-founder and co-founder of deep space.
Now, I'm going out to do business, sign up, and push for concrete scenes. I'm in charge of channel construction and eco-partner management, and I've formed a FDE team to deliver and land the company Agent.
Today, I would like to share with you a few items that we have experienced personally during this period, as well as the problems we have encountered in them.
I'm not very good at coming up with a lot of ideas. Let's look at a concrete scene: embroidery.
Embroidery is an interesting industry, which is well developed in Zhejiang, and may account for 80 per cent or even more of the global industry.
Why did we get into this scene? Because we're talking about AI landing in a business, and industrial enterprises are part of it.
Look at this picture: this is a very good fox, on the one hand, and on the other, embroidery. It's very complicated to turn an effect map into a needle file for embroidery machines. This process requires a copymaster to produce an impact map and a costly foreign software that takes a long time to complete the layout. This software is like industry CAD, called Wilcom.
We tried this software. First thought was, can we do this with a power like Computer Use? It was later found that it was not possible to do so because of the complexity of the judgements involved.
What are the process requirements here? For example, different machines have different colours, and the pattern needs to be divided into layers, some with contours to fill. These are the basic procedural principles of embroidery.
There are more complex patterns, like mountain water maps. It's easier than that. If it is difficult to divide by 1 to 10, it is at most 4 or 5; to the extent that it is necessary to restore a true human image, it is very complicated.
We happen to know and talk about this with an eco-partner who's been digitizing for many years, and he's thinking, "What can AI do for embroidery?"
So we did a series of in-depth studies and found a number of problems.
On the one hand, it's time-consuming, it could take five days, and software is expensive. On the other hand, a great deal of printed experience is in the teacher's mind and there is no systematic curriculum. There is no problem teaching simple operations, such as how to use software; however, how to turn a good-looking picture into embroidery documents involving colour judgement, aesthetics, and embroidery design, which are often unteached and rely more on teacher experience.
So, what we're going to do is we're going to talk about "Still Teacher Masters" that you've been talking about all along. But how many teachers are willing to be distilled? It's a very realistic question: how can I get AI to take my place?
And that's the problem that FDE's getting on the ground. How do you convince him that the purpose of his acceptance of AI is not to replace him, but to improve the efficiency of his work, to enable him to leave work early, to return home early, without having to stare closely at each of his needles and figure out what to do better?
Our original plan was to simulate the computer operation of teachers using skills like Computer Use.
A lot of videos were made in the run-up to get them back for analysis and to see if they could be reproduced. It was later found that it was simply impossible to replicate and distill the entire operation in this way.
So we changed our thinking: why do we have to do this the way we used to do it? Why can't it be done in the way of AI?
Later, we changed the process to: upload the original map, choose embroidery fabric and the corresponding needle method, such as needles, and directly generate needle files.
Can you do that? We have collected more than 20,000 sets of relevant data, made a large number of manual markers, and conducted some post-training work, which culminated in a minute-level output.
If one to 10 sizes are difficult, 10 is the hardest, and we think that one to five sizes are ready for minutes.
This is what we've done with the FDE field needs study in industry. The output includes embroidery maps and process documents, which include information on colour numbers, needle numbers, etc.
Of course, we also have many problems in the process. Embroidery involves many colours, and needle exchange involves knots and wire cutting.
First, the teacher looked at the picture that we created and found out that 700 times more cuts and knots were needed. These are process-specific problems that must be solved by FDE on site.
Our FDE stayed in the factory for a long time, making copies with her teacher, and using AI to cross-train and validate this picture, to get results.
And look at another FDE landing scene: customs clearance.
Customs clearance is a very important part of the large number of foreign-trade enterprises in Yiu. One of the more complex scenarios that we have encountered is “1039”, which is also a very representative closure scene on the part of the EU.
Enterprises can ask outsiders to produce customs declarations, mostly at a cost of between $70 and $200. Depending on the clarity of the photographs and the completeness of the information, it takes about half a day to produce a customs declaration manually.
These are the information usually received in the “1039” scenario, sometimes simply taking a photograph, including locks, warehouse receipts, weight orders, etc. The information is not limited to that, it is numerous and complex, and clarity is sometimes problematic.
This is also the reason why the manual production of customs declarations takes a long time: the information is very confusing.
While some scenarios can already generate information online through software or provide electronic files, the “1039” scene that we are exposed to is still complex, with many links relying on manual means of simply taking a photograph. As a result, it used to take a lot of time to manually produce customs declarations.
We have a FDE colleague who spent a long time in Yiu to gather the relevant information and information required for customs clearance, keep Agent running, testing and improving the ability to identify these data.
Finally, most customs declarations can be produced in less than three hours to five minutes. The minimum production cost of $70 is about $10, or even lower, after AI.
These are two cases in which FDE conducted scene research, excavated needs and eventually completed delivery.
Back to these two cases, what do we think FDE should do?
When we set up the FDE sector, we think about different jobs. In the past, research and development had focused more on functions and scenes, and product managers had focused on how to organize dispersed demand and produce complete products. The FDE focus will be broader, ranging from scenes, products to industry-wide, to a comprehensive understanding of and extraction of demand. Sales turn products into industry solutions and sell them.
Clients have different perspectives. These positions are more from the perspective of software or AI products, while clients are in specific industries and industries, concerned with what is best for their own industries and industries and can really help themselves?
So, what's the first thing that FDE did? Data collection.
Recently, we have seen a number of cases of failure, one of which is that of rushing to work after entering the field, staring at one thing and wondering if we can open the door from one point to another.
We have had similar problems with a project before, which has not worked well. Then we changed the way: when the FDE came in, we gathered information and collected as much information as possible about the enterprise.
Here's context, not just the documents that the business has. Some small businesses do not have well-managed files, much of the information is in WeChat, and others are scattered over the computers of different employees. These data need to be collected. The format of the information may vary, with some special formats requiring specialized analysis.
When necessary, we also produce data and collect them. AI ' s summary capacity is now strong for disassembling and compiling information to form an enterprise knowledge base as a source of knowledge for subsequent work.
After collecting the information, there is also a very important process: the FDE must follow its clients, follow the positions to be addressed, and do the business with its own hands. Let's not just do research, stand next to the tape pen or other equipment, and record what the client says or says. It's hard to really appreciate the difficulties of clients. It was important to go down and go through the process to understand the problems.
In some places, teachers can deal with it with experience, even if it is not a problem, but it is very difficult for new people.
Whether it be customs declarations, embroidery, or other foreign trade scenes we are doing, this is true: it must be done on its own to understand the problems and exceptions that may be encountered.
Since I'm new, I have to ask my teacher to learn: why would I be in trouble? Any exceptions? It is important to understand these anomalies clearly and verify the results of the processing.
The process of hand-to-handing is eventually documented and returned to the knowledge base.
After the experience of some of the “distilled” specialists, we found that the process was a little bit like a national sign: the process was very cumbersome, but it was also very prescriptive.
These processes can be deposited in the process. So, in what way should demand, client scenes and problems be deposited?
Our position is different from what Mr. Biggen just said. We're not research and development posts, we're business executive posts.
But we still stress that when FDE delivers, you have to sink the methods you run and use. In the beginning, you may have done a mission with prompt, but when it is done, you will have to sink it into Skill as soon as possible, so that your pool of capabilities will gradually be complete.
Our FDE team has a shared Skill warehouse, where you upload and view each other. When we get to the same point, we'll see if the other team members can solve the Skill and keep trying.
In addition, the delivery of Harness. Here, it is understood that the FDE has to collect enough examples of testing to convert user scenes into implementable tests.
Now, using AI to perform tests is no longer too difficult. Research and development students are testing more of the functions, interfaces and interfaces developed, but testing based on real business scenes must be done by the FDE team.
FDE to collect enough examples. Once the R & D has been completed, the testing team has completed the test, it will have to run again using its own Halness and the test database, as these examples come from real problems encountered at the site.
Some Bug feedback to R & D may lead to different understandings. Research and development may have recovered problems from the underlying principles, but it is not necessary that the problems actually encountered by users be resolved.
For example, many of us now are used to MacBook development, and have not used Windows for a long time, but our foreign trade clients are almost using Windows.
Sometimes, when the development is completed, even if a virtual machine is installed and it feels okay, problems can still arise at the customer site: The version of the operating system used by the client, the firewall and other operating environment information are all required for recording. Development students can't go to the scene all the time, and the work needs to be done by the FDE team. This is an important part of the delivery process.
Finally, operational results will be measured. The value of an event is more limited if it does not result in substantial efficiency gains and cost reductions.
Clients may ask, "Why use your product when it's such a low labour cost to hire part-time, newly graduated students and even those who have not yet graduated?"
For products to be valuable, they must lead to substantial efficiency gains and lower costs. Therefore, the last task of the FDE is to measure the results of operations.
That's the whole process we've been talking about, from the realization of functionality to the delivery of specific scenarios, to sedimentation, to the formation of Skill and Harness.
During the previous period, a researcher from Open AI shared an idea on X, which I found very good, to the effect that if the next client could solve the same problem without our help, the FDE team would have done its job.
I saw a WeChat Official Account article that summed it up as "FDE is killing its own job," which means something about the title party. Maybe friends have seen it. But when I went in, I thought it was not about getting FDE out of work, but about getting feedback: Open AI's FDE translated the results of working with clients into better products, in a way that allowed the product to take on the work that it had to do.
This is two different concepts from the idea of losing one ' s job at the outset.
The last page is also an advertisement. We're a new company that's promoting the landing of Agent, a business in all industries. The industries currently entering are mainly foreign trade and cross-border electric operators, which are also a very prosperous industrial ecology of the Zhejiang.
We're also doing banking, and the industrial scene we just mentioned. I'm going to visit some industrial clients next week to learn about the needs of CAD, and so on.
I think in China, people who can really use AI may be one in 10,000, not even yet. We have a responsibility to help businesses that in the past did not have a good digitized base to complete the AI transition, otherwise it might be difficult for them to enjoy this wave of AI dividends.
I'm in charge of the channel, the eco-partners and the FDE, and I'd like to make contact with you. We have some opportunities to experiment with products, and we welcome cooperation on ecological capabilities. Even in the AI era, even if dozens of people use multiple Agent developments in parallel every day, we can't do all the functionality and solve all the business scenes.
So, whether FDE individuals with an industry background or commercial companies that have built up mature Skill and digital capabilities in certain industries, we hope to be able to establish linkages and work together to solve these problems.
That's my share. Thank you.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.