Linux 基金会 AI CTO 谈开源 AI、中国崛起与智能体 AI

欢迎收听 Fuzz Verse 播客。

Welcome to Fuzz Verse podcast.

谢谢。感谢邀请。

Thank you. Thank you for having me.

对于不了解的人，你在 Linux 基金会担任 AI CTO 的职责是什么？能介绍一下你的日常工作和正在做的事情吗？

For those who don't know, what's your role as the AI CTO at the Linux Foundation? Can you walk us through your responsibilities and day-to-day activities?

我的工作很大程度上面向社区。我确保我们拥有一个健康的开源 AI 生态系统，有许多优秀的项目需要成长并获得关注，比如 LLM、PyTorch、Deep Speed、Ray 等。我还专注于教育，正在开展培训和认证工作，尤其是围绕 PyTorch。我帮助为开源 AI 设定强有力的技术方向。我之前的工作包括模型开放框架和 Open MDW 许可证。我也经常与社区交流。我们最近成立了 Agente AI 基金会，包含 MCP、Goose 和 Agents MD，围绕智能体式 AI 构建健康的生态系统。我的职责是做好社区 steward，帮助发展充满活力的开源 AI 生态系统，涵盖开放数据集、开放模型以及全栈的开源软件。在 Linux 基金会，许多工作负载依赖 Kubernetes 和 Linux，因此我的部分工作是民主化工具和教育的获取，推动开源 AI 生态系统持续发展。

A lot of my work is community focused. I ensure we have a healthy open source AI ecosystem, with great projects that need to grow and get visibility, like LLM, PyTorch, Deep Speed, Ray, and others. I also focus on education, working on training and certification, especially around PyTorch. I help set a strong technical direction for AI in open source. My prior work includes the Model Openness Framework and the Open MDW license. I also speak to the community. We recently stood up the Agente AI Foundation, which includes MCP, Goose, and Agents MD, building a healthy ecosystem around agentic AI. My job is to be a good steward, helping grow the vibrant open source AI ecosystem, which includes open datasets, open models, and open source software across the stack. At the Linux Foundation, many workloads rely on Kubernetes and Linux, so part of my role is to democratize access to tools and education, and keep the open source AI ecosystem moving forward.

在加入 Linux 基金会之前，你领导了 PyTorch 的技术方面，在此之前，你在企业技术领域工作了近 25 年，包括 IBM 和电信行业。这段经历如何帮助你应对当前 AI 热潮？

Before joining Linux Foundation, you led the tech side of PyTorch, and before that, you spent almost 25 years in enterprise tech, including IBM and telco. How does this experience help you navigate the new AI hype?

我的职业生涯大部分在工业界，中间在 Cable Labs 做过研究，然后回到工业界，现在又到了另一个非营利组织。这帮助我理解两个方面：企业对消费、采用和应用 AI 的需求，以及基金会方面发展生态系统和工具获取。这让我能够同时拥有这两种视角。理解 AI 研究者和开发者所经历的事情也很重要，这样我们才能解决他们的问题。

I spent most of my career in industry, with a sidebar at Cable Labs doing research, then back to industry, and now another nonprofit. It helps me understand both sides: the enterprise needs for consumption, adoption, and applied AI, and on the foundation side, growing the ecosystem and access to tools. It uniquely positions me to have both views. It's also important to understand what AI researchers and developers go through so we can solve their problems.

你刚从中国大陆长途旅行回来，访问了 DeepSeek、Moonshot、ZAI、Jipu、Quan 等主要 AI 实验室。最让你惊讶的是什么？

You just came back from a long trip in mainland China, visiting major AI labs like DeepSeek, Moonshot, ZAI, Jipu, Quan, and others. What surprised you the most?

过去几年我去过中国几次。从美国人的角度来看，最令人惊讶的是我们并不真正了解中国正在发生的事情。我们看到很多优秀的作品发表在开放科学和开源领域，但我们不知道他们的研究走向。令人大开眼界的是，当我与这些实验室会面时，他们对自己的工作非常开放，愿意分享创新。由于商业立场，他们不分享产品路线图，但对研究很开放。最启发我的是看到他们多么热情，多么兴奋地发表作品、发布模型，并看到它们与封闭 API 模型的基准对比。他们没有保留。

I've been to China a few times over the last couple years. The most surprising thing from an American perspective is that we don't really know what's happening in China. We see great work published in open science and open source, but we don't know where their research is going. What was eye-opening is that when I met with these labs, they are very open about their work, willing to share innovations. They don't share product roadmaps due to commercial positions, but they are open about their research. The most enlightening part was seeing how passionate they are and how excited they are to publish work, release models, and see how they benchmark against closed API models. They are not holding back.

与美国同行相比有何不同？

How is it different from US counterparts?

在美国，人们会非正式地谈论他们正在做的事情，但不一定会邀请你去实验室进行广泛披露。美国实验室之间人员流动很大，研究人员经常跳槽。

In the US, folks will talk a little off the record about what they're working on, but not necessarily invite you to the lab for broad disclosures. There's a lot of movement between US labs, with researchers moving between them.

AI 并非总是万能的。早期用 LLM 做欺诈检测时，每个人都想用 LLM，但经典的机器学习模型在识别欺诈交易或非法设备方面准确率更高。你会信任智能体用你的钱投资多少？现在人们不想让智能体碰他们的钱。我们已经看到聊天机器人或客服智能体给出了不该给的退款。这些事情必须通过代码以确定性的方式构建，而不是每次都依赖模型遵循你的意图。

AI is not always the best solution for everything. Early on with LLMs for fraud detection, everyone wanted to use an LLM, but classical ML models work better for identifying fraudulent transactions or illegal devices with higher accuracy. How much money would you trust an agent to invest on your behalf? People don't want agents touching their money. We've seen chatbots or customer service agents giving refunds they shouldn't have. These things must be built deterministically through code, not relying on the model to follow intent every time.

对。

Right.

中国开源模型的中国下载量首次超过西方模型。这说明了什么？

The share of Chinese downloads of Chinese open source models for the first time is higher than the western ones. What does it tell you?

我认为我们看到中国模型的情况是……

I think what we're seeing with the Chinese models is that there's...

而且，你知道，这没什么秘密。我的意思是，因为人员流动很大，尤其是在湾区。但在中国，实验室在地理上更分散，对吧？它们在北京、上海、深圳，还有杭州。所以，没有像美国湾区实验室那样的人员流动。所以，我认为文化略有不同。我还要说，中国的实验室文化相当粘性。比如你为 Moonshot 这样的公司工作，它有非常摇滚主题的办公室和年轻人。但在美国可能……

And you know, a lot of that is there's not a lot of secrets. I mean, because there's a lot of movement, especially in the Bay Area. But in China, the labs are more geographically dispersed, right? And so, they're in Beijing, Shanghai, Shenzhen, and also Hangzhou. And so, you don't have that same movement of people that we do in the US labs in the Bay Area. And so, I think the culture is slightly different. I would also say the lab culture in China is fairly sticky. Like you work for a company like Moonshot, which has a very rock and roll themed office and young folks. But in the US maybe it's...

Moonshot 在美国是王者，是吗？

Moonshot is King in the US, is it?

是的，没错。所以，这绝对是不同的文化，但还有一点人们没有真正认识到：很多研究人员，尤其是创始人，都受过西方教育。加州大学伯克利分校、卡内基梅隆大学、斯坦福大学。他们在美国接受教育，然后回到中国。现在更多发生的是，很多企业家和研究人员回到中国工作，而不是留在美国。所以历史上他们想在美国工作，但我认为中国在鼓励创新和创业文化方面做得很好。当然，这里也有湾区，对吧？也有很多人从事这个领域。

It is, yeah. Yeah. And so, it's definitely different culture, but there's also something that people don't really recognize: a lot of these researchers were Western educated, especially the founders. UC Berkeley, CMU, Stanford. They did their education in the US and then came back to China. Something happening more now is that a lot of entrepreneurs and researchers are coming back to China to work, and they're not staying in the US. So historically they wanted to work in the US, but I think China's done a good job of encouraging a culture of innovation and startups. And of course, there's a Bay Area here too, right? There are folks working in this space as well.

实际上很少有人意识到这一点，但如果你看世界地图，有两个巨大的 AI 中心：湾区，下一个就是离这里几分钟的深圳。当然还有杭州和北京。在你与许多中国研究人员的观察和交谈中，你怎么看西方所谓的 ChatGPT 时刻，在中国被称为 DeepSeek 时刻？那是一个斯普特尼克时刻。你对 DeepSeek 有什么看法？他们在中国仍然占据主导地位，还是现在分布更均匀了？

There's actually very few people realize that, but if you look at the world map, there are two massive epicenters of AI: the Bay Area and the next one is just a few minutes away from here in Shenzhen. But also Hangzhou and Beijing obviously. In your observation and conversation with a lot of Chinese researchers, how do you find that what they in the West called the ChatGPT moment and in China is called DeepSeek moment? That was a Sputnik moment. What's your take on DeepSeek? Do they still have a dominant positioning within China or is it more evenly distributed now?

我认为 DeepSeek 被视为尖兵、创新者，用更少的资源做更多的事。他们推出了许多伟大的创新，被包括美国在内的许多其他实验室采用。所以我仍然认为他们是以这种形象出现的。中国实验室对 DeepSeek 的成就和他们的工作深表敬意。我来自开源背景，非常欣赏他们向世界发布的内容。他们的研究论文非常详细，你可以复现他们所做的研究。这对整个 AI 的进步非常重要。我们需要确保有足够的披露，以便能够复现研究。我们看到一个我不太热衷的趋势：研究论文变短了，变成了技术报告。许多技术报告没有足够的披露来复现研究，这并不能真正推动 AI 的事业。取决于你如何看待我们如何实现 AGI，如果每个人的研究都是封闭的，那肯定无助于我们推动进展。

I think DeepSeek is viewed as the tip of the spear, the innovators, doing more with less. They have put out a lot of great innovations that have been adopted by many other labs, including in the US. So I still think they're seen through that light. There's a deep respect in the Chinese labs for what DeepSeek has done and what they're working on. I can really appreciate that coming from an open source background, looking at what they're putting out into the world. Their research papers are very detailed. You can reproduce the research they're doing. That's greatly important for the advancement of AI across the board. We want to make sure we have the disclosures so that you can go and replicate the research. We've seen a trend that I'm not extremely enthusiastic about: research papers have shrunk and become technical reports. There's not enough disclosure in many of these technical reports to be able to replicate the research, and that doesn't really advance the cause for AI. Depending on your position on how we're going to get to AGI, it certainly doesn't help us move the dial if everybody's research is closed.

就在你在中国的时候，MIT 和 Hugging Face 发布了一份新报告，其中中国开源模型的中国下载量首次高于西方模型。这告诉了你什么？我们是否处于全球开发者心目中中国开源模型正在超越西方模型的情况？

Just when you were in China, MIT and Hugging Face issued a new report where the share of Chinese downloads of Chinese open source models is for the first time higher than the Western ones. What does it tell you? Are we in a situation where developers globally, the mind share of Chinese open models is prevailing over Western ones?

我认为我们看到许多中国模型有性能更好的小型语言模型和中型语言模型，这推动了更多的下载。这些模型性能很高。Llama 系列目前没有其他额外的模型。所以 Qwen 是部署最广泛的模型，甚至在美国的企业中也是如此。我认为我们看到大量的爱好者、研究人员、开发者想要下载这些模型并在本地使用 Ollama，还有企业希望从成本优化中受益。使用这些较小或中型模型比使用 API 并支付更高的每 token 价格更有吸引力。

I think what we're seeing with many Chinese models is that there are better performing small language models and medium-sized language models, and that is driving more downloads. Those models are highly performant. The Llama family is not currently, there's no other additional Llama models. So Qwen is the most widely deployed model, even in enterprises in the US. I think we're seeing a huge number of hobbyists, researchers, developers that want to download these models and use them locally with Ollama, and then enterprises that want to benefit from cost optimization. It's more appealing to use some of these smaller or medium-sized models as opposed to using APIs and paying higher per-token price.

是的，我认为 DeepSeek 版本 4 的最新数字，某些模型比商业 API 便宜 50 倍，且能力相似，这在中国大陆引发了一场价格战。但有趣的是，你也见了 MiniMax 的创始人，他们刚刚发布了第一个闭源模型。他们有没有解释背后的逻辑？

Yeah, I think the latest numbers of DeepSeek version 4, which was 50 times cheaper for certain models of the commercial APIs with similar capabilities, kind of triggered the price war across mainland China. But interestingly, you also met the MiniMax founders, and they just released their first closed models. Did they explain the logic behind?

他们没有详细说明，但我认为这将成为中国的一个趋势。之前的看法是，任何超过一万亿参数的模型训练成本太高，需要将其放在 API 后面，因为部署更难，但也能实现商业收益。美国有一些开源模型，一些 API，现在 API 方面更多，我认为中国未来也会出现同样的模式。我认为这些实验室，有些今年早些时候已经 IPO，有些正在考虑 IPO。

They didn't go into detail about it, but I think this is going to be a trend in China. The perception was that anything above a trillion parameters was so expensive to train that you would need to move it behind an API because it would be harder to deploy, but also to realize commercial gains. Where the US has some open models, some APIs, and now much higher on the API side, I do think we're going to see that same model in China going forward. I think the labs, some of them had IPOs earlier this year and some are eyeing IPOs.

以及展示的压力……

And the pressure to demonstrate that...

是的，是的，他们需要，而且训练这些模型需要花钱。他们必须有商业，他们有投资者，他们必须实现某种收益。所以我认为他们会转向 API。我确信他们所有人最终都会拥有商业 API，并帮助补贴部分研究和训练成本。

Yes, yes, they need to, and it costs money to train these models. They have to have a commercial, they have investors, they have to realize some sort of gains. So I think they will shift to APIs. I'm sure all of them at some point will have commercial APIs and help subsidize some of the cost of research and training.

在我个人经历中，与 MiniMax 创始人的合作是这样的：我们正在为活动开发一个内部解决方案，一个基于语音的智能体，但很难找到适合香港观众的粤语模型。于是我联系了那位创始人，她立刻建群、拉工程师，问题就解决了。相比之下，我们用了美国公司的语音 AI 技术栈，活动进行到一半他们突然改了 API，我们花了很长时间才找到问题所在。

So in my personal experience with actually MiniMax founder, I was developing an in-house solution for events, a voice-based agent, and we were struggling to find a right Cantonese model for the Hong Kong audience. Then I messaged the founder, who I happen to know, and asked for help with MiniMax. She immediately set up a group, brought engineers, and we sorted it out. In contrast, we used a US company for the voice AI stack, and suddenly in the middle of the event they changed the API. It took us a long time to figure out where the problem was.

没错。

Right.

完全没有提前通知。我又在 LinkedIn 上联系那位创始人，但没得到回复。我提这件事是因为感觉美国对 AI 的需求太大，而中国有客户至上的态度，愿意以最低成本做最多事，还主动帮忙。也许这只是个例，但那次经历让我觉得或许该重新考虑。

It was like no heads up, nothing. So I messaged the founder on LinkedIn again, but got no response. The reason I mention this is that I feel in the US there's so much demand for AI stuff, while in China there's a customer-first approach, willingness to do as much as possible at minimum cost, and proactive help. Maybe it's just a wrong example, but that experience made me think maybe I should reconsider.

没错。我认为有些人和组织希望对模型及其生命周期有更多控制权。他们可能不希望模型在后台自动升级。显然，如果你用 API，对方可以升级模型、做小版本发布，可能破坏某些功能或改变 API。但如果你有更多控制权，运行自己的开源模型，就能设定自己的条款和升级路径。所以这对很多人很有吸引力。

Right. Yeah, I think some people and organizations want more control over the model and its lifecycle. They may not want the model to be upgraded in the background. Obviously, if you have an API, they can upgrade the model, do a minor release, maybe break something or change the API. Whereas if you have more control and run your own open models, you can set your own terms and upgrade path. So I think that's enticing to a lot of folks.

是的，我觉得整体态度和那种积极意义上的进取心——比如为了市场份额愿意为客户做一切——是因为他们没有像美国公司那样被过度需求宠坏。回到你的行程，你在博客里提到和 DeepSeek 的一些资深员工吃了午饭，他们从 High-Flyer 时期就在公司了。从你们的对话来看，他们的工作方式或内部文化有什么不同或有趣之处？有什么见解可以分享吗？

Yeah, I feel overall the attitude, the aggressiveness in a positive sense, like let's do everything for the customer to gain market share, because they're not spoiled by over-demand like you see in US companies. But going back to your trip, you also mentioned in your blog that you had lunch with some of the DeepSeek senior staff, who were with the company back when it was High-Flyer. From your conversations, what is their approach or how do they internally approach work that might be different or interesting? Any insights you can share?

当然。DeepSeek 给我印象最深的是他们的谦逊。他们非常谦虚。有些组织强调个人——论文第一作者、展示个人能力。但 DeepSeek 给人的感觉是非常集体主义：大家一起解决实际问题。办公室里的谦逊氛围，人们真正被解决问题所驱动，不太关心个人成长或六位数、七位数的薪水——就是和同事一起出色地完成工作，帮助解决问题，尤其是在资源受限的环境中。

Sure. What struck me most about DeepSeek and meeting folks there was the humility. They are very humble people. In some organizations, it's about the individual—being first author on a paper, showing what you can do. But the impression from DeepSeek is very communal: we're all working together to solve real problems. The humility in that office, how people are really motivated to solve problems and not as concerned about personal growth or six-figure, seven-figure salaries—it's really just doing a great job working with colleagues to help solve a problem, especially in resource-constrained environments.

你的行程还赶上了华盛顿的另一件有趣的事：白宫发布了一份关于蒸馏攻击美国模型的备忘录。你在与中国同行的对话中看到什么反应了吗？

Your trip also coincided with another interesting event back in Washington when the White House released a memo on distillation attacks on US models. Was there any reaction you saw from your Chinese counterparts in your conversations?

是的，我认为整个 AI 生态系统中，蒸馏是很多实验室都在做的事。一种观点是，我们都朝着同一个目标——AGI——努力，每个人都想利用最好的东西。很多人受数据限制，没有足够的高质量数据来改进模型。所以蒸馏被很多组织频繁使用，包括中国和美国。这不是中国独有的。这很有意思，因为我们谈论开放和进步，但必须认识到很多原始数据是有版权的。模型是在这些数据上预训练的。在理想世界里，数据是一种商品，可以用来改进模型。尤其是对于较小的模型或领域特定训练，从能产生高质量数据的教师模型进行蒸馏非常重要。你可能有内部专家，但未必拥有所有所需数据。所以蒸馏对社区是有益的。但同样，提供商有使用条款，如果你是商业用户，遵守它们是好的。

Yeah, I think across the entire AI ecosystem, distillation is something that a lot of labs do. One perspective is that we're all working towards the same goal—AGI—and everyone wants to leverage the best of everything. Many folks are data-constrained; they don't have the quality data needed to improve their models. So distillation is done quite often by many organizations, both Chinese and US. It's not unique to China. It's interesting because we talk about openness and progress, but we must recognize that much original data is copyrighted. Models are pre-trained on that data. In a utopian world, data would be a commodity you can use to improve models. Especially for smaller models or domain-specific training, distilling from a teacher model that produces high-quality data is super important. You may have SMEs in-house but not all the data needed. So distillation is beneficial to the community. But again, providers have terms of use, and it's good to abide by them if you're a commercial user.

我记得你在 Substack 文章里说，西方有一种说法是中国在复制粘贴西方的一切，另一种说法是出口管制减缓了中国科技和 AI 的发展，而你认为这两种说法都是错的。能解释一下为什么吗？

I think in your Substack article, you said there's a narrative in the West that China is copy-pasting everything the West does, and another narrative that export controls slow down Chinese tech and AI development, and you believe both are wrong. Can you walk us through why?

当然。中国实验室在创新。他们创造了很多伟大的创新，这些创新被用于学习，并实际应用到美国的开源模型和 API 中。这些创新是跨境的；这是所有从事 AI 研究的人共同分享成果的一部分。我认为这对整个领域都非常有益。

Sure. The Chinese labs are innovating. They're creating a lot of great innovations, and those innovations are used for learning and applied in actual US open models and on the API side. These innovations are cross-border; it's part of everyone working on AI research, sharing what they're doing. I think that's extremely beneficial across the board.

你知道，美国应该做这件事，欧洲、加拿大、中国，每个人都在产出伟大的开放科学，对吧？出口管制可能并没有达到预期的效果。当你因为无法获得资源而被迫创新时，你就会去创新。所以我认为中国发生的情况是，实验室们意识到，好吧，我们必须创新。我们需要从已有的东西中榨取更多价值。这确实发生了。我认为这很棒，因为我们可以将同样的原则和发现普遍应用于全球的 AI 研究。对硬件的限制性访问迫使中国开始投资自己的芯片。所以我们看到这个领域有了更多发展，出现了很多芯片制造初创公司。中国在稀土矿物方面具有独特优势，并且有能力构建一个非常垂直的产业链，从原材料到生产、制造和部署。

You know, the US should be doing it, and Europe, Canada, China, everybody producing great open science, right? The export controls probably haven't had the sort of intended effects. When you're forced to innovate because you don't have access to resources, you're going to innovate. So I think what's happened in China is that the labs have figured, okay, we need to innovate. We need to get more mileage out of what we've got. And that has happened. I think that's great because we can apply the same principles and discoveries universally across the world in AI research. The restricted access to hardware has forced China to start investing in their own silicon. So we're seeing more development in that area, a lot of new startups in chip manufacturing. China is uniquely positioned with rare earth minerals and the ability to build a very vertical stack from raw materials through production, manufacturing, and deployment.

为了继续这个问题，你提到了四个西方实验室现在正在借鉴的中国具体贡献：DeepSeek、GRPO、月之暗面在 Moan 优化器上的工作、字节跳动的 VEARL、MiniMax 的 scaling lighting 和注意力机制。我想知道，如果根本没有出口管制，这四个中的哪一个永远不会发生？

In order to continue this question, you named four specific Chinese contributions that Western labs are now building on: DeepSeek, GRPO, MoonShot's work on Moan Optimizer, ByteDance VEARL, MiniMax scaling lighting and attention. I wonder which one of these four would never happen if there were no export controls at all.

我几乎认为很难说哪些创新会与出口管制挂钩。我喜欢 Moan 优化器的例子，因为最初的工作并非源自中国，而是源自美国，有一篇论文发表了那项工作。然后它被后续的实验室在此基础上发展。正是这种迭代工作让我们能够取得今天的进步。我希望看到更多开放科学被拥抱：你创造了一些东西，别人改进了它，然后再改进。我们在 LLM 的演变中看到了这一点：Transformer 架构被修改，有了更多优化，不同风格的注意力机制，所有这些创新帮助我们获得更高性能的系统。所以我认为这非常重要。我们在技术栈的更高层也看到了同样的事情，不仅在模型 AI 研究方面，而且在框架、智能体框架、工具和脚手架等方面，这些都以开放形式发布。OpenClaw 在中国非常流行，现在还有 Hermes Agent——我想我在这次会议上交谈过的几乎每个人都在跟我谈论这个。所以这些开放创新、开源创新吸引了最多的关注，吸引了社区参与，并邀请开发者贡献。我认为我们在研究方面的相同情况也出现在开源软件开发中。

I almost think it's hard to say what innovations will be tied to export controls. I like the example of the Moan optimizer because the original work didn't originate in China; it originated in the US, with a paper published on that work. Then it was built on by subsequent labs. That kind of iterative work is why we're able to make the progress we can today. I would love to see more open science being embraced: you create something, somebody else improves upon it, and then improves upon it. We've seen this with how LLMs have evolved: the Transformer architecture has been modified, there have been more optimizations, different flavors of attention, and all these innovations help us get to higher performant systems. So I think this is very important. We're seeing the same things happen higher up in the stack as well, not just on the model AI research side, but also in frameworks, agentic frameworks, harnesses, and scaffolding released in the open. OpenClaw is hugely popular in China, and now Hermes Agent — I think almost everybody I've spoken to at the conference here is talking to me about that. So these open innovations, open-source innovations, get the most eyes, invite community participation, and invite developers to contribute. I think the same parallels we have on the research side, we also have in open-source software development.

是的，OpenClaw 在中国非常流行。我觉得它已经达到了顶峰，现在正在趋于平稳，可能还会下降。现在甚至有全新的服务教你如何把它从电脑上卸载，因为每个人都装了它。但有趣的是，你对开源 AI 模型的立场：如果听众中有一位创始人正在构建 AI，筹集了 200 万美元，有 18 个月的时间，你会推荐哪条路？基于 Claude？ChatGPT？DeepSeek？Qwen？你怎么看？

Yeah, OpenClaw is massive in China. I feel like it's reached its peak and now it's plateauing, maybe going down. There's a whole new service offered on how to remove it from your computer, now that everyone installed it. But interestingly, your stand on open-source AI models: if there is one founder in the audience building AI, raised 2 million USD, and has 18 months, which path would you recommend? Build on Claude? ChatGPT? DeepSeek? Qwen? What's your take?

嗯，有了 200 万美元，我会强烈建议他们不要创办一个 AI 实验室，因为那不会……

Well, with 2 million dollars, I would strongly encourage them not to start an AI lab, because that's not going to...

我的意思可能不是训练 AI，而是 AI 应用或任何东西。

I mean not maybe training AI, but AI application or anything.

我认为创新的机会非常多。对于开发产品的初创公司，你必须找到自己的利基市场、垂直领域。你要做金融相关的事情吗？在特定领域构建东西？提供某种通用的专业服务？我想我不能说，比如，每个人都喜欢 Claude Code——它是一个出色的工具。关于在本地基础设施、自托管基础设施中使用开放模型，或者使用像 Fireworks 或 Together 这样的推理提供商，或者直接使用付费 API——这是一个你必须根据组织需求做出的决定。如果你有一个初创公司，通常使用开放模型可以用更少的资源做更多的事情。这创造了一种非常顺畅的方式，可以保持较低的 token 支出，并将资金投入到营销和团队发展等方面。但同样，使用 API 的管理开销要低得多。所以我认为两者都是非常好的选择，取决于初创公司和你想要实现的目标。

I think there's so much opportunity for innovation. For startups developing products, you really have to find your niche, your vertical. Are you going to do something in finance? Build something in a particular domain? Provide some sort of generalized professional services? I don't think I could say like, everybody really loves Claude Code — it's an excellent tool. The decision around using open models in local infrastructure, self-hosted infrastructure, or using inference providers like Fireworks or Together, or just using paid APIs — that's a decision you have to make based on your organization's needs. If you've got a startup, generally you can do a lot more with less when using open models. That creates a very frictionless approach to maintaining low spend on tokens and investing in things like marketing and team growth. But again, the administrative overhead is much lower when using an API. So I see both as very good options, depending on the startup and what you're trying to achieve.

OpenAI 刚刚停止了他的模型微调服务。所以如果你想要非常经济高效且适用于非常具体的重复性任务，你现在实际上别无选择，只能微调开源模型，对吧？

OpenAI just discontinued their model fine-tuning services. So you practically don't have a choice now but to fine-tune an open-source one if you want something very cost-efficient and applied for a very specific repetitive task, right?

是的，我认为看到更多微调后的开放模型被共享，并拥有一个非常好的生态系统，那将非常棒。Hugging Face 和 ModelScope 都有很多这样的模型在线，但在众多开放模型中找出最适合你应用的那个有时很有挑战性。拥有一个非常值得信赖的微调模型生态系统，你知道它们的来源、最初基于哪个模型，并且它没有在最初是限制性许可的情况下以宽松许可重新发布，那将非常棒。

Yeah, and I think it would be really fantastic to see more fine-tuned open models being shared and having a very good ecosystem. Hugging Face and ModelScope both have a lot of these models online, but trying to figure out which one is the best for your application in a sea of open models is challenging sometimes. Having a very trusted ecosystem of fine-tuned models where you know their origination, what model they were originally built on, and that it hasn't been re-released under a permissive license when it was a restrictive license to begin with, would be great.

在开始围绕一个微调模型商业化产品之前，你需要掌握这些重要见解——这个模型可能没有按照你预想的方式获得许可。比如 Meta 的 Llama 模型，它的许可在应用方面限制相当严格。在你推广的开源或框架以及专门针对 AI 模型的框架视角下，你认为哪个标准目前接近你通过这个框架设定的标杆？在今天我们可以称之为真正开源模型的，是 DeepSeek、Llama、Qwen 还是 Mistral？

These are sort of important insights that you need to have before you start commercializing your product around a fine-tuned model that perhaps wasn't licensed the way that you thought it was supposed to be. And I think one of these examples like say Meta, right? With the Llama model where licensing is pretty restrictive in a way. For the applications. In the view of the open source or framework and the specifically framework in relation to AI models that you promote. Where do you see the standard who is currently close to achieve the standard that the bar that you're raising with this framework? In what we can call as true open source model today. Is it DeepSeek? Is it Llama? Is it Gwen? Is it Mistral?

是的，我认为这有两方面。首先，任何使用宽松许可的模型——目前它们都使用软件许可，比如 Apache 2.0 或 MIT——这些模型具有最大的自由度。所以 DeepSeek 系列、Mistral 系列，还有 AI2 或 MBZUA I 的 LN360 倡议推出的模型，都可以用于任何目的：研究、构建应用、围绕它创业等等。社区许可则带有限制，比如可能有使用量触发上限，你需要仔细辨别，但这类社区许可都包含一个条件，一旦触发你就得去谈判新许可，或者仅限非商业用途、仅限研究。所以阅读许可很重要，把它扔进你喜欢的智能体或 ChatGPT 里获取细节也很重要。这是第一部分。另一部分是，开源软件许可原本是为软件设计的，对吧？几年前我们在 Linux 基金会着手创建针对模型的许可。三四年前我们开始制定这个开放模型开放性框架。我们想定义什么构成开放模型，以及什么构成更高级别的开放科学——你不仅发布模型及其权重，还发布所有数据集、训练代码、配方、基准测试等所有辅助成果。我们看到 AI2 在这方面做得很好，允许社区复现他们的工作。我们向往这种开放科学，它对教育、可复现性、透明度和研究都非常有益。要能在此基础上构建业务或进行研究，最低可行产品就是模型及其权重，我们将其定义为开放模型。当时（现在仍广泛使用）的术语是“开放权重”。开放权重虽然暗示开放性，但实际上指的是限制性许可模型，即带有条件的模型。而我们所说的开放模型是指具有宽松许可的模型，使用上没有任何限制。实际上，开放科学很棒，但涉及商业利益时，人们默认选择开放模型。不是每个人都能复现万亿参数模型并拥有相应资源，所以我认为至少要有这些使用宽松许可的开放模型，让整个社区都能访问。人们可以学习如何使用这项技术，并在此基础上构建，这极其重要。

Yeah. I think it's a kind of a two-parter. So, for one any of the models that are using permissive licenses, which they're all using software licenses today. So using Apache 2.0 or MIT those models have the most permissiveness, right? So, any of the DeepSeek models, any of the Mistral as well. And anything that's coming out of like AI2 or the LN360 initiative out of MBZUA I. These are all models that can do use for any purpose, right? So, research, you can build your applications on it. You can create a startup around it, these sort of things. The community license are the ones that have the restrictions, so there may be utilization trigger limits and you just have to be discerning, but any of these kinds of community licenses have a condition in there that would trigger you to have to go negotiate a new license, right? Or non-commercial use or things like this, research only. So, it's important to read the license. It's important to just drop it into your favorite agent or into chat GPT and get the details, but that's one part of it. The other part is that we haven't been open source software licenses were designed for software, right? And a couple of years ago we embarked to create a license at the Linux Foundation that was directed towards models. And so, three, four years ago now we started working on this open model openness framework. And we wanted to identify what constitutes an open model, right? And what constitutes something that's higher than that, which would be like open science, like you're releasing not just the model and its associated weights, but you're also releasing all of the data sets, all of the training code, all of the recipes, any benchmarks, all of these other accessories that were needed to help get you to the final product, right? And we've seen a few, AI2's being a good steward of that sort of approach of open science and allowing the community to be able to replicate the work that they've done. And we kind of aspire to this open science and this is really great for education, it's great for reproducibility, transparency, and research purposes. The minimum viable product for what you need to be able to build a business on or study to a lesser extent would be just the model and its weights, right? And so we had defined that as an open model. And at the time and it's still being used widely now is this open weights term. And so open weights although it implies openness, it's really being attributed to the restrictive license models. So those models that have conditions in place. The what we call open model is those that have permissive license and so you have no restrictions on what you can do with it. In practice, open science is amazing but when there's commercial interests involved, people sort of default to open model. Not everybody's going to be able to reproduce a trillion parameter model and have the resources to do that and so I think it's important that we at least have these open models out there that are using permissive licenses and that the entire community has access to them. People can learn how to work with this technology, they can build on the technology. That's extremely important.

Linux 基金会目前哪个项目最受社区和贡献者关注？你认为最热门的项目是哪些？

Which of the current projects under Linux Foundation brings the most attention of community and contributors? Where do you see the most sort of the hottest ones?

Agent AI 基金会和 MCP 项目确实备受关注。很多人正在构建 MCP 服务器并将其集成到组织中。Linux 基金会的独特之处在于拥有这些在不同技术栈层面运作的基金会。在应用层，我们有 Agent AI 基金会及其项目，这对我们来说还很早期——我们成立该基金会才五六个月，现在已有近 200 名成员。所以 agent-ai 非常令人兴奋。其他更成熟的项目，比如 Kubernetes，对于所有 AI 工作负载和云工作负载都至关重要。显然，Linux 内核极其重要且部署广泛。还有 PyTorch 基金会，其项目涉及 AI 基础设施领域，包括如何构建模型、训练和微调，以及在生产中部署模型并提供推理。这些基金会都在这个技术栈中扮演着角色。

So there's certainly a lot of excitement around the Agent AI Foundation and MCP. There's a lot of people building MCP servers and integrating MCP into their organizations. The Linux Foundation is uniquely positioned with all of these foundations that work at different layers of the stack. And so, up in the application layer here we've got the Agent AI Foundation and the projects they're working on, and this is very early for us, right? Like we're 5 months, 6 months into establishing that foundation. We have close to 200 members involved now. And so, there's a lot of excitement about agent-ai. The other projects that are a lot more seasoned, like Kubernetes, are absolutely integral for all of the AI workloads and all the cloud workloads. And so, obviously the Linux kernel is extremely important and very widely deployed. And then we have the PyTorch Foundation, which has projects in the AI infrastructure space. So, how to build the models, training and fine-tuning and then deploying models in production and providing inference. So, the foundations all kind of play a role in this stack.

当我告诉我的工程团队我要见 Linux 的 AI CTO 时，他们的第一个问题是：AI 会被嵌入到 Linux 系统内核层面吗？你是否有这样的计划或想法？AI 能在其中扮演什么角色？

When I told my engineering team I'm meeting AI CTO of Linux, the first question they had is, is it going to be AI introducer embedded in the kernel level of the Linux system? Is this any kind of plan that you have in mind or any idea of that could be or what will be the where the AI can play, what kind of role AI can play within this fundamental?

我无法具体谈论内核开发方面，因为我不参与 Linux 内核的开发。但我想说，AI 作为一种工具正在许多地方被利用，比如帮助优化代码。在我们的一些项目中，很多人现在会问，当他们想使用 Claude Code 或其他智能体时，如何做出贡献。许多项目现在都有关于智能体是否可以贡献的政策。

I can't really speak to the kernel development aspect, because I'm not involved in developing the Linux kernel, but I would say that AI as a tool is being leveraged in many places, right? And so, to help optimize code and for some of the projects we work on, a lot of folks are asking now about, when they want to use Claude Code or another agent, how they make contributions. A lot of the projects have policies now around whether an agent can contribute or not.

我们看到很多项目被 PR 的数量搞得有点不堪重负，那些几乎没有知识或编程背景的人提交 PR，让维护者应接不暇。然后我们看到 AI 被用来缓解这个问题，帮助审查和合并 PR。所以我认为我们现在处于一个混乱的阶段，一旦我们理清头绪，就会进入规范阶段。但毫无疑问，智能体式编码和编码智能体现在在开源领域非常普遍。看到不同的核心维护者如何处理这个问题并试图应对 PR 的涌入，这很有趣。

We're seeing a lot of projects get a little bit overwhelmed by the number of PRs and how people with very little knowledge or coding background issue PRs and sort of overwhelm maintainers. Then we're seeing AI being used to help mitigate that, to help review PRs and consolidate them. So I think we're in this storming space right now, and I think we'll get to the norming stage once we figure it all out. But definitely, agentic coding and coding agents are very pervasive now in open source. It's interesting to see how different core maintainers are handling this and trying to reconcile the influx of PRs.

关于未来的发展方向，AI 最终会成为界面吗？或者即使是通过对话，如果 AI 嵌入到系统根目录，你直接对它说话就行，真的还需要我们习惯的那种带有按钮、设置、命令行的操作系统界面吗？

Regarding where things are going, is AI becoming the interface ultimately? Or even conversationally, do you really need the current interface of the OS as we used to, with buttons, settings, command lines, if you can just speak to AI directly if it's embedded into the root?

尽管我们现在有能力使用语音模型，比如语音到语音，但你通常不会看到人们像使用手机键盘（QWERTY 键盘）那样频繁使用它。所以我认为用户界面仍然有一席之地。

Even though we have the capability of using voice models now, like voice-to-voice, you generally don't see people using that as much as they use the keyboard on their phone, the QWERTY keyboard. So I think there's still a place for UIs.

他们可能被 Siri 伤透了心。

They were traumatized by Siri probably.

嗯，也许吧。

Yeah, maybe.

但还是有很多人对语音界面不感兴趣。而且，如果你在私人或公共场合，你真的想……

But there are still a lot of folks who are not interested in voice interfaces. Also, if you're in private or public spaces, do you really want to...

也许只是对话式的，不一定是语音。可以通过打字，但核心思想是你有一个界面层，正在重新定义我们看待软件引擎的方式。

Maybe just conversational, not necessarily voice. It could be through typing, but the idea is that you have a layer of interface that is redefining the way we look at the software engine.

是的，我认为人们使用 AI 与系统交互的方式确实在发生转变。这也取决于用户是谁，对吧？消费者与工程师、研究员、开发者不同。显然，技术性更强的人想要更深入的访问。我们看到通过 CLI 和直接集成 API 的智能体越来越多。随着模型变得更具智能体能力，我认为我们会开始看到智能体界面与人类界面有所不同。但目前，我们正在用程序化界面和人类界面来训练很多模型。

Yeah, I think there's definitely a shift in how people interface with systems using AI. It also depends on who it is, right? The consumer versus the engineer, the researcher, the developer. Obviously, those who are more technical want deeper access. We're seeing even more agents through CLIs and integrating against APIs directly. As the models become more agentic and more capable, I think we'll start to see a different set of interfaces for the agents versus the human interfaces. But right now, we're training a lot of the models on both the programmatic interfaces and the human interfaces.

你提到了智能体式 AI 和安全，以及现在涉及 API 访问。你还谈到了从 LLM 安全向智能体式 AI 安全的转变。你能多分享一些你对这个问题的看法吗？

You mentioned agentic AI and safety, and now that access to APIs is involved. You also talked about the shift from the safety of LLMs towards the safety of agentic AI. Can you share more on your thinking about this problem?

是的，安全显然对每个人都重要——下游用户、消费者以及构建系统的人。有一点很重要：模型安全是一个方面，但当我们基于模型构建系统（如智能体系统和其他形式）时，我们需要考虑安全。我始终鼓励工程中的安全优先和安全设计，因为你必须考虑人们将如何使用你构建的技术。我们希望负责任地构建这些技术。同时，这也为初创公司提供了围绕安全和安保构建业务的机会，设置护栏和约束来强制执行系统的安全。我们从模型中继承了很多智能体式 AI 的特性，但你也在构建程序化脚手架——智能体是被编码的——所以在这个层面上强制执行安全也非常重要。

Yeah, safety is obviously important to everybody—downstream users, consumers, and those building the systems. One important thing is that model safety is one aspect, but as we build systems on top of the models, like agentic systems and other forms, we need to take into account safety. A safety-first approach in engineering and safety by design is something I always encourage, because you have to think about how people will use the technologies you build. We want to build those responsibly. At the same time, it presents a great opportunity for startups to build around safety and security, putting in guardrails and harnesses that enforce the safety of the system. There's a lot we inherit in agentic AI from the model, but you're also building programmatic scaffolding—the agent is codified—so enforcing safety at that level is also very important.

当你考虑多个智能体的智能体集群时，你如何强制执行安全？我的意思是，当你不像只处理一个模型那样有那么多控制权时。

When you look at agent swarms of multiple agents, how do you enforce safety? I mean, when you don't really have that much control, unlike when you're working with just one model.

对于多智能体系统，安全必须内置于框架中。框架和约束有助于控制它。多智能体系统有不同的架构，比如生成子智能体或多智能体协调。这些必须内置于协议中，在协议层强制执行安全和安保，同时也在运行智能体的框架本身内部。安全必须在每个层面被考虑，就像我们在 Web 上做的那样。例如，在操作系统、实际程序、网络层和界面上都有强制执行。当你忙于创新和实验时，安全和安保常常是事后才想到的，但这需要被解决。我们不仅要考虑网络上的安全或应用程序的隐私问题，还要考虑下游影响——对社会、对环境等的影响，这一点非常重要。

For multi-agent systems, safety has to be built into the framework. The framework and the harness help contain it. There are different architectures for multi-agent systems, like spawning sub-agents or multi-agent coordination. These have to be built into the protocol, enforcing safety and security at the protocol layer, but also within the framework itself that runs the agent. Safety has to be viewed at each level, the same way we do it on the web. For instance, there's enforcement in the operating system, in the actual program, on the network layer, and on the interfaces. Often security and safety are a last thought when you're hard at innovation and experimenting, but it needs to be addressed. It's very important that we look at these considerations not just from security on the network or privacy issues with the applications, but also downstream effects—effects on society, on the environment, these sorts of things.

我想问你一个关于 LLM 的问题。大约一年前，你写道推理不是人类使用的那种推理；它有点像虚假推理。模型无法推理。现在一年过去了，我们看到推理能力显著提升，有什么变化吗？

One question I want to ask you is about LLMs. Around a year ago, you wrote that reasoning is not the kind of reasoning that humans use; it's sort of fake reasoning. The models can't reason. Has anything changed since then, now that we see reasoning capabilities significantly improved?

对我来说，没什么变化，因为它不是人类意义上的推理。当我们谈论下一个词预测时，我们是通过严格的文本来模仿人类推理。通过人类示例，我们可以复制你的推理方式，但我们通常不会用词元或书面形式进行推理；我们通过大脑推理。重要的是要认识到，我们在 LLM 中看到的推理与人类水平的推理程度不同，但 LLM 中已有的推理同样有用。

For me, nothing's changed because it's not reasoning in the same sense that humans reason. When we talk about next-token prediction, we're mimicking human reasoning through strictly text. Through human examples, we can replicate how you reason, but we don't generally reason out in tokens or written form; we reason through our brains. It's important to recognize that the reasoning we see with LLMs is not the same degree as human-level reasoning, but the reasoning we do have in LLMs is no less useful.

它不一定非得和人类大脑的推理方式相同，对吧？但同样的效果，对我们的应用有用吗？当然有用。它总是正确的吗？不是。人类总是正确的吗？也不是。所以我认为我们必须从这个角度来看：将 LLM 或聊天机器人拟人化，或者认为我们在与人类交互，这有点有害。

It doesn't have to be the same type of reasoning as the human brain, right? But the same effect, is it useful for our applications? Most certainly. Is it always right? No. Are humans always right? No. I think we have to look at this from the lens of personification or believing that we're interfacing with a human when we're interfacing with an LLM or chatbot is a little bit harmful.

为什么？

Why?

我认为我们看到过一些案例，人们相信自己是在与人类交互，或者陷入一种错觉，认为自己的想法是正确的。我们看到一些法庭案件，其中暗示某些聊天机器人鼓励人们做他们不应该做的事情。我认为重要的是要认识到，我们是在与一个系统打交道，我们基本上不是在和一个以与我们相同的方式思考或推理的东西交互。它们只是模式匹配了数据，有时只是将我们自己的想法回响给我们。所以人们必须认识到，这些是系统，不是人。

I think we've seen cases of folks that believe they're interfacing with a human or falling into a delusion that what they're thinking is the right path. We've seen some court cases where it's been implied that particular chatbots have encouraged people to do things they shouldn't be doing. I think it's important for us to realize that we're working with a system and that we're basically not interfacing with something that is thinking the same way we do or reasoning through things. They've pattern matched data and sometimes it's just going to echo back our own thoughts to us. So it's important for people to recognize that these are systems, not people.

我只想说，我非常喜欢你那个比喻，把 LLM 的推理比作鹦鹉，它们可以复制人类的短语甚至句子，却没有真正的理解。

I just want to say that I really like your metaphor of LLM reasoning as parrots who can replicate the human phrase or even a sentence without actual understanding of it.

对。是的，我不能把“随机鹦鹉”这个类比归功于自己，但确实，我们在训练 LLM 时是在做模式匹配。我们有意地，尤其是通过微调和强化学习，确保 LLM 遵循意图。所以你给它一堆上下文，但你也想表达你的意图。理想情况下，你得到的是你期望的东西。我认为过去几年我们在数据混合和训练模型的方法上有所改进，取得了更好的结果。

Right. Yeah, I can't be credited with the stochastic parrot analogy, but definitely again, we're pattern matching when we're training LLMs. We're intentionally, especially through fine-tuning and reinforcement learning, to make sure that the intent is followed with an LLM. So you throw a bunch of context at it but you also want to express your intent. Ideally you get back exactly the kinds of things you would expect. I think we've improved over the last few years in our data mix and the way we approach training models that has yielded much better results.

大型语言模型总是有局限性。你对 Yann LeCun 构建所谓世界模型的新尝试有什么看法？

There are always limitations of large language models. What's your take on Yann LeCun's new venture with building a so-called world model?

Yann LeCun 有自己的芯片架构，一种不同的看待方式，不一定是顺序训练。我认为这很好。对我来说，我对世界模型的兴趣更多与机器人技术、具身 AI 相关。如果你从这样的论点出发：要实现 AGI 或超级智能，你需要拥有所有相同的要素，比如多感官、物理知识等各个方面，那么世界模型就有意义。关于什么是世界模型，有不同的说法。有人说它是视频、图像和文本。其他人说你必须能够在其中创建 3D 空间，接收实时指令。有不同的方法。如果你是机器人专家，你关心行动，感知世界并预测结果的能力，你的机器人的下一步是什么。你在触发行动，你关心实时感官输入和激活执行器的能力。所以这取决于你的出发点。世界模型意味着不同的东西。甚至游戏玩法、生成游戏或选择你自己的冒险风格电影都可以来自世界模型。我很好奇事情会如何发展。扩散模型和 Transformer 是首选架构，除非你使用 Jeff 架构。

Yann LeCun has his chip architecture, a different way of looking at how, not necessarily sequential training. I think it's great. For me, my interest in world models is more associated with robotics, with embodied AI. If you come from the argument that to achieve AGI or superintelligence, you need to have all the same ingredients, like multi-sensory, physics-informed, all these aspects, then world models make sense. There are differing accounts of what a world model is. Some say it's video and images and text. Others say you need to be able to create 3D spaces, take real-time instructions. There are different approaches. If you are a roboticist, you care about action, the ability to perceive the world but also predict outcomes, what's the next move for your robot. You're triggering action and you care about real-time sensory input and being able to activate actuators. So it depends on where you're coming from. A world model means something different. This idea of even gameplay or generating games or choose your own adventure-style movies can all come out of world models. I'm very curious to see how things evolve. Diffusion models and transformers are sort of the architectures of choice unless you're using the Jeff architecture.

你提到了机器人技术。你还见过 Unitree 和其他中国机器人公司的创始人。你还暗示这个行业部分正在进入泡沫，但没有透露是什么样的泡沫。你能分享你的观察吗？

You mentioned robotics. You also met the founder of Unitree and other robotics companies in China. You also hinted that this industry is partially going into a bubble without revealing what kind of bubble. Can you share your observations?

全球有很多机器人公司，但很多在中国。大概有超过 150 家人形机器人公司。对于一个仍然很小但正在增长的市场，竞争非常激烈。能够解决机器人领域真正问题的公司，比如人类手部灵巧性问题以及自主性部分。目前很多机器人系统是远程操作或精心编排的，预先编程。自主性问题尚未解决。很多人正在研究这个问题。我确实看到的是，中国机器人初创公司非常擅长机械、物理、制造执行器、整合所有组件。美国机器人公司更专注于大脑，即背后的 AI，机器人背后的多模型系统。可能公司太多了，过度饱和，这在任何人们蜂拥而入的行业都很常见。中国机器人公司是北京方面的优先事项，所以他们投资于发展这个生态系统。像任何市场一样，会有赢家和输家。我们可能会在未来几年看到谁崛起，谁无法在顶级水平竞争。

There are a lot of robotics companies worldwide, but a lot of them in China. There's probably over 150 humanoid robotics companies. There's a lot of competition for a market that is still small but growing. The robotics companies that are able to solve real problems in robotics, like the human hand dexterity problem and the autonomy piece. A lot of robotic systems right now are teleoperated or well orchestrated, programmed ahead of time. The autonomy thing hasn't been solved. Many people are working on that. What I do see is that Chinese robotic startups are really good at the mechanics, the physics, building the actuators, bringing all the components together. US robotics companies are more focused on the brain, the AI behind it, the multi-model systems behind robotics. It's possible that there's going to be too many, oversaturation, which generally happens in any industry where folks rush. Chinese robotics companies are a priority for Beijing, so they've invested in growing that ecosystem. There will be winners and losers like in any market. We'll probably see in the coming years who rises up and who can't compete at the top level.

作为一位在企业技术领域深耕多年的人，听听你的意见很有趣。现在你关注应用，并与这些前沿 AI 公司合作，从你的观察来看，企业在尝试应用 AI 时最常见的错误是什么？他们看待和战略性地将 AI 实施到工作流程中的方式，或者他们如何构建商业模式或内部自动化。你怎么看？

One thing is interesting to hear your opinion as someone who spent years in enterprise tech. Now that you're looking at applications and working within this frontier with AI companies, from your observation, what is the most common mistake that enterprises make when they try to apply AI? The way they look at and strategically find the way to implement AI into workflows, or looking at how they can build business models or automate internally. What's your take?

我认为 AI 的采用没有挑战。

I don't think there's a challenge with adoption of AI.

我认为企业成功部署 AI 存在一个关键点。导致糟糕结果的因素有很多，其中之一就是“我知道怎么做得更好”的心态。你选择自建而非采购，但内部又没有相应的人才。你花大量时间实验，消耗资源和资本，试图创造一些新颖的东西，而这些核心组件可能早已存在于生态系统中。我们在开源领域想做的就是提供这些基础组件，让组织可以在此基础上构建，而不必完全复制。现在大家都在快速推进，尤其是在智能体式 AI 方面，有时自己动手做很有吸引力，而不是使用可能已经构建了优秀框架的合作伙伴。即使那个框架不是开源的，他们也可以提供专业服务并指导你。所以我认为寻找合作伙伴很重要，不要认为这是一段必须独自完成的旅程。显然，有些顾问会乐意收取费用为你提供指导，但我看到最多的情况是“嘿，我们能做得更好”或“我们能做出来”，结果却达不到预期的效果。

I think there's a dot is successful deployment of AI in enterprises. And that often there's a multitude of factors that lead to poor outcomes. One of those is the 'I know how I can build this better' mentality. Instead of buying, you're building, and you don't have the talent in house. You're spending a lot of time experimenting, burning through cycles, burning through capital to try to create something novel which may already have the core components out in the ecosystem. One of the things we want to do in open source is provide those foundational components so that organizations can build on them instead of having to replicate them entirely. And when we're all kind of running right now, especially on the agentic AI side, it's attractive sometimes to try to build it yourself instead of using a partner that may have already built a great framework. Even if that framework isn't open source, they could provide professional services around that and help guide you. So I think it's important to look at partners. Don't think of it as a journey you have to take on your own. Obviously, there are consultants that will happily take your money to provide you some guidance, but what I've seen the most is the 'hey, we can build this better' or 'we can build this' and then it doesn't perform at the expected outcomes.

你有没有具体的行业或用例例子，看到这种方法失败？

Do you have any specific examples of industries or use cases where you see this kind of approach failing?

是的。不点名具体组织，几年前是聊天机器人。“我们如何在基础设施上构建更好的聊天机器人？哦，我们有数据。那我们去预训练一个模型吧。”这不是个好主意。市面上有很多优秀的预训练模型。你可以基于它们进行微调。但我们看到……

Yeah. Without calling out any particular organization, a few years ago it was the chatbots. 'How can we build a better chatbot on our infrastructure? Oh, we have the data. So let's go pre-train a model.' Well, that's not a good idea. There are a lot of great pre-trained models out there. You can work from that base and then fine-tune it yourself. But we were seeing...

所以，如果你不是……训练模型通常是个坏主意。

So training models is generally a bad idea if you're not...

如果你不是实验室，就不要专注于那个领域。我们看到很多公司为不同领域和行业构建自己的模型。他们发布了一个模型后就停止了。我知道很多组织尝试预训练自己的模型，但都失败了。他们消耗了大量资本，最终一无所获。所以我认为应该利用现有的构建模块。开源领域有很多好东西。我们当然希望看到更多。我们希望看到更多开放且许可宽松的模型。所以我认为拥有一个充满活力的项目生态系统非常重要，这些项目不仅解决开发 AI 的问题，也解决应用 AI 的问题。

If you're not a lab, then don't focus on that area. We saw a lot of companies building their own models for different domains and industries. They put one model out and then stopped. There are a lot of organizations I'm aware of that tried to pre-train their own and fell short. They burned through a lot of capital to come up empty-handed. So I think build on the building blocks that are out there. There's a lot of great stuff in the open. We definitely want to see more of that. We want to see more models that are open and permissively licensed. So I think it's very important that we have this really vibrant ecosystem of projects that solve the different problems not just in developing AI, but also in applying it.

我知道的一个例子是 Revolut，他们发布了自己的模型，不是他们声称的语言模型，而是基于他们的数字数据训练的。他们在内部使用。但我想他们可能没有太多动力从现有项目中选择，加上银行和隐私的敏感性可能促使他们走那条路。我们拭目以待。目前只是第一次发布，看看第二次会怎样。

The one example I know of is Revolut, which released their own model, not as they claim a language model, but trained on their numbers, basically. They apply it internally. But I guess they probably didn't have much incentive to choose from existing projects, plus sensitivity around banking and privacy might have pushed them in that direction. We'll see. So far it was the first release. We'll see what the second will be.

是的。

Yeah.

除了“我们内部做得更好”的心态，你看到当今公司在接触 AI 时还有哪些管理或业务上的错误？

Besides the 'we build it better internally' mentality, what other mistakes in management or business do you see when companies approach AI nowadays?

还要在探索用例时识别低垂的果实。不要伸手去够最上层，试图拿下最复杂、风险最高的用例。你绝对需要对用例进行风险评估。确保你应用的用例是可实现的，并能帮助你实现自动化的优化。如果不能，你只是构建了展示能力但没有实际好处的东西，那肯定不值得投入。你要确保你在处理可实现的用例。你不会想花 6 到 12 个月构建某样东西却一无所获。显然，要做风险评估，因为很多人对前景非常兴奋。昨天讨论的一个话题是，你愿意让智能体代表你投资多少钱。随着时间的推移，它们变得更可靠，你会越来越信任它们，但现在人们会说“我不想让智能体碰我的钱”。我们见过聊天机器人或客服智能体提供不应给的退款、被越狱的例子。始终要回到安全、保障和隐私。确保你在这些方面有很强的姿态。确保你有强大的护栏，因为你不想泄露数据或个人信息。很多事情必须通过代码以确定性的方式构建，而不是依赖模型每次都遵循你的意图。

It's also about identifying the low-hanging fruit when you're looking at use cases to explore. Don't reach for the top shelf and try to pull down the most complex use cases with the highest risk. You definitely want to do a risk assessment on the use cases. Ensure that the use cases you're applying are achievable and help you realize optimizations in automations. If it doesn't and you've just built something that demonstrates capability but has no tangible benefits, then it certainly wasn't an endeavor worth taking. You want to make sure you're working on use cases that are achievable. You're not going to spend 6 to 12 months trying to build something and come up empty-handed. Obviously, do a risk assessment because a lot of people are very excited about the prospect. One topic that came up yesterday was around how much money would you trust your agent to invest on your behalf. As we go through time and they become more reliable, you'll trust them more, but right now people are like 'I don't want an agent to touch my money.' We've seen instances of chatbots or customer service agents providing refunds they shouldn't have, being jailbroken. Always go back to safety, security, and privacy. Make sure you have a very strong posture there. Make sure you have strong guardrails in place because you don't want to exfiltrate data or personal information. A lot of these things have to be built in a deterministic way through code, as opposed to relying on the model to follow your intent every time.

是的，我刚刚意识到，你通过提示词给模型设置的护栏不一定是最好的方式。在用户和 LLM 之间硬编码一个类似传感器的东西来处理边缘情况，可能是一种思路。

Yeah, I just realized the guardrails that you prompt the model with are not necessarily the best way to do it. Having something like a sensor hardcoded in between the user and the LLM to handle edge cases might be a way to look at that.

是的，是的。

Yeah, yeah.

关于智能体，现实地说，你看到过任何自主智能体完全取代人类日常工作的成功例子吗？

In regards to agents, realistically, do you see any successful examples where autonomous agents have fully replaced humans in day-to-day jobs?

是的，所以我认为对于智能体的自主性有不同的看法。我见过极端情况，比如完全自主的智能体独立存在于网络中，自行行动。但归根结底，你希望智能体完全符合智能体的定义：代表你工作以实现某些结果。目前我看到精力最集中的地方，大概在两个领域。

Yeah, so I would say there are different views of what autonomy is with agents. I've seen extremes like fully sovereign agents that exist in the ether on their own, going out there doing things. But at the end of the day, you want an agent to be exactly what the definition of an agent is: to work on your behalf to achieve some outcomes. Where I'm seeing probably the most energy focused right now, I guess in two spaces.

一方面是构建模块。我回到这些构建模块的概念，你有这些开源的构建模块，可以在上面继续构建，无论是像 MCP 这样的协议，还是某种通用框架，或者其他可以用来构建智能体的组件。技能，就是把所有这些整合起来，创造出一些东西。我认为这是很多人关注的焦点——构建协议、标准、规范，以及所需的框架，然后还有技能方面所有这些可复用的组件。另一个我看到的很多现象是，所有这些目前都非常实验性。在采用方面，最简单的用例收益最大。你必须穿透很多炒作，比如人们说他们刚靠自己的金融智能体赚了一万亿美元，那个智能体开始买股票，或者“我解雇了我的团队，现在我有 50 个智能体”。是啊，是啊。“我现在是单人公司了。我以前有 500 名员工，我把他们都解雇了，现在我有大约 12 个智能体。”这就是你必须穿透的那种东西。你要现实一点。有很多理想化的情绪，比如“哦，我可以做我的单人公司，让自己变得强大”。这很好。人们能够构建以前没有能力构建的东西，能够自动化很多他们做的事情，这很棒。我们通常不会看到某个智能体处理整个工作流程的所有环节。工作流程的这一步有一个智能体，那一步有一个智能体，另一步也有一个智能体。举个例子，比如报税。如果你不想整理收据并尝试制表，你可以扫描它们，或者把你的 PDF 文件通过一个智能体处理，让智能体把它们制成表格，创建一个电子表格，计算总数，然后给你所有这些部分，你在报税前再去验证。有很多细微的用例，但有一点是，你总是需要验证。特别是如果你在构建一个应用程序，你需要对它进行基准测试。通常你必须开发自己的基准，因为你需要确保它对你的用例有效。在你发布或投入生产之前，确保你有合适的基准。

One is on the building blocks. I go back to these concepts of building blocks where you have these open-source building blocks and you can go and build on them, whether that's a protocol like an MCP or some sort of general-purpose harness, or these other building components that you can use to build your agents. Skills and just taking all these things together and creating something out of it. I think that's where I'm seeing a lot of focus on people building protocols, standards, specs, the frameworks that are needed, and then all these reusable components on the skills side. The other area where I'm seeing a lot is that all of this is very experimental right now. On the adoption side, it's the simpler use cases that have the most benefits. You have to cut through a lot of the hype and people talking about how they just made a trillion dollars off their financial agent that started buying up stocks, or "I fired my team and now I have 50 agents." Yeah, yeah. "I have a one-person company now. I had 500 employees. I laid them all off and now I've got like 12 agents." That's the kind of thing you have to cut through. You want to be realistic. There's a lot of aspirational sentiment around things like, "Oh, I can do my one-person company and enable myself." And that's great. It's great that people can build something they may not have had the capability of building before, that they can automate a lot of the things they do. We don't generally see some agent handling everything through the entire workflow. There's an agent in this step of the workflow and in this step and this step. As an example, doing your taxes. If you don't want to sort through receipts and try to tabulate them, you can scan them or run your PDFs through an agent and have the agent tabulate them, create a spreadsheet, total it, and give you all these pieces, and you go and validate it before you file your taxes. There are a lot of nuanced use cases out there, but the one thing is that you always want to validate. Especially if you're building an application, you want to benchmark it. Often you have to develop your own benchmarks because you need to make sure it's working for your use cases. Before you ship something or put it into production, make sure you have the right benchmarks in place.

当谈到不同规模企业中的 AI 采用时，最好的策略是什么？应该从哪里开始？如果现在有人在经营一家企业，有很多不同的方法。有些老板说，“哦，每个人都必须使用 AI”，或者“如果你不用，我们会给你奖金”——我的意思是，从你的角度来看，第一步、第二步、第三步是什么？我们应该从哪里开始？

When it comes to AI adoption within businesses of different sizes, what would be the best strategy? Where should you start? If someone is running a business today, there are a lot of different approaches. Some bosses say, "Oh, everyone must use AI," or "If you don't, we'll give you a bonus" — I mean, what would be from your point of view the first step, second step, third step? Where should we start?

我认为需要认识到的一点是，你不需要直接与 AI 打交道就能获得收益。AI 正在融入你使用的产品的功能中。无论你是否需要直接与 ChatGPT 或 Gemini 交互，你都可以——比如 Adobe 的策略，在幕后集成，在它们的产品中加入更多生成式功能和智能体功能。这是通过向现有应用程序添加功能来帮助企业提高生产力的好方法。对于任何采用者来说，这是最省力的方式，因为他们不一定需要直接使用这项技术。

I think one of the things that's important to recognize is that you don't need to work with AI directly to realize benefits. AI is finding its way into the features of the products you use. Whether you don't necessarily need to interface with ChatGPT or Gemini, you can — like Adobe's strategy of integrating behind the scenes, putting more generative features in, putting more agentic features in their products. These are a great way to help up the productivity of businesses by adding features into their existing applications. It's the lowest lift for anyone adopting because they don't necessarily have to work with the technology directly.

如果你这里关注的是客户表面层面。

If you focus on the customer surface level here.

是的，没错。即使你是一家企业，也要与你的供应商、产品供应商沟通，说“嘿，我们想要这个功能”，让他们把它集成进去。这样你就不必自己构建来弥补，因为你是把功能开发外包给了第三方供应商。这是最省力的方式，特别是如果你不是 AI 原生组织，没有合适的技术人员。

Yeah, exactly. Even if you're an enterprise, talk to your suppliers, talk to your product suppliers and say, "Hey, we want this feature," and have them build it in. So you don't have to compensate by building something yourself, because you're offloading that feature development to your third-party supplier. That's the easiest lift, especially if you're not an AI-native organization, if you don't have skilled people in place.

所以和你的供应商谈谈，让他们引入基于 AI 构建的功能。

So talk to your vendor, ask them to introduce features built with AI.

是的。如果你有一个 100 人的公司、50 人的公司，你不需要自己去构建所有东西。你应该依靠你的供应商来帮助你。如果你是一个更大的组织，我总是说找顾问谈谈。不要以为你可以在第一天就自己学会所有东西。如果你是一家硅谷初创公司，你可能非常专注于自己构建。但如果你是为了其他目的而构建，并且你不是一家 AI 优先或智能体优先的公司，那么你同样需要利用现有的工具。这一点非常重要——这就是为什么开源如此重要——能够访问所有这些工具、所有这些组件，来开始构建你的产品或服务，或增强你自己的内部系统。如果我们谈论企业，以及与大量遗留系统和后台系统的集成，能够处理你自己的数据。现在有很多工具已经就位。在你经典的 SQL 世界里，现在你可以把它引入你的语言模型。现在有很多非常棒的产品，其中很多是商业产品。但你确实需要解决你的用例是什么。不要为一个你还没有识别出的问题去寻找解决方案。再次强调，这回到了第一性原理。

Yes. If you've got a 100-person company, 50-person company, you don't need to go out and build it all yourself. You should be leaning on your vendors to help you. If you're a larger organization, I always say talk to a consultant. Don't assume you can learn it all on your own on day one. If you're a Silicon Valley startup, you're probably very focused on building it yourself. But if you're building it for some other purpose and you're not an AI-first company or an agentic-first company, then again, you want to leverage the tools that are out there. It's really important — and this is why open source is so important — having access to all these tools, all these components to start building your product or service or enhancing your own internal systems. If we talk about the enterprise and integrating with a lot of legacy systems and back office systems, being able to work with your own data. There's now a lot of tooling in place. In your classical SQL world, now you can bring that into your language model. There are a lot of really great products out there now, many of them commercial. But you really have to solve for what your use cases are. Don't go searching for a solution for a problem you haven't first identified. Again, it falls back to first principles.

比如，哪些用例对你的组织最重要？你希望通过自动化实现哪些最大的优化？你如何实现成本节约？你知道，就是所有这些事情，对吧？所以你必须非常专注于你试图解决的问题。

Like what are the use cases that are most important to your organization? What are you going to realize the most optimizations through automation? How are you going to realize cost savings? You know, all of these things, right? And so you have to be very focused on the problem you're trying to solve.

但如果你让供应商为你做这件事，那就意味着你的所有竞争对手也能获得同样的能力。所以你就没有这个优势了。

But if you ask your vendor to do it for you, that means that all your competitors will have access to this as well. So you're not having this edge anymore.

是的。这绝对是差异化与生产力的问题，比如内部生产力是一回事。把这类事情推给供应商可能并不总是合理的，但通常当它是一个差异化因素时，那才是你该投资的地方，对吧？如果你要与竞争对手区分开来，如果你唯一的差异化是你能更快地处理理赔（比如在保险行业），那其实算不上护城河，对吧？那能帮你行动更快，也许还能提供更好的下游客户服务，但最终要找出差异化因素，并看看 AI 是否是推动它的好解决方案。AI 并不总是所有事情的最佳解决方案，对吧？生成式模型当然也不是所有事情的最佳解决方案。比如，我在 LLM 刚出现时看到的一件事是，对于欺诈检测，每个人都觉得需要用 LLM。其实不然。你完全有很好的传统或经典机器学习模型，能够以更高的准确性和可靠性识别潜在的欺诈交易、欺诈设备或网络上的非法设备等。所以，你不一定需要因为 LLM 是最新最伟大的东西就转向它。你也可以大量使用你的经典机器学习模型。

Yeah. It's definitely if it's differentiation versus productivity, like internal productivity is one thing. It doesn't always make sense perhaps to push that to the vendor, but often when it's a differentiating factor, then that's where you want to invest, right? If you're going to differentiate yourself from your competitor, if your only differentiation is that you are able to process claims faster if you're in insurance, that's not really a moat, right? Like that helps you move quicker and maybe provide downstream like better customer service, but ultimately figure out what the differentiating factors are and see if AI is a good solution for driving that out. AI isn't always the best solution for everything, right? And certainly generative models are not the best solutions for everything either, right? Like one of the things that I saw early when LLMs came out was that for fraud detection, everyone's like, I need to use an LLM. Well, no, you don't. You have perfectly good legacy or classical machine learning models that are able to identify potentially fraudulent transactions or fraudulent devices or illegal devices on your network or these kinds of things with higher accuracy and reliability. And so, you don't necessarily need to shift to an LLM because it's the latest and greatest thing. You can also use a lot of your classical machine learning models.

我认为现在很多保险公司面临一个大问题，整个保险理赔都成了欺诈，因为现在每个人都用 LLM 来申请保险理赔或信贷之类的。所以，你提到对于差异化，你会投资；对于内部优化，你最好与现有供应商合作。我同意，因为内部优化是有限的。收益上限默认是有限的。你能赚多少大概就在那里，对吧？但构建新的价值主张、改变商业模式、实现差异化，那种收益在某种程度上是无限的。

I think now a lot of insurers facing a massive problem where the whole thing was claiming insurance as corruption because everyone is using LLMs now to claim insurance or getting credit or anything like that. So, then you mentioned for differentiation, you would invest; for internal optimization, you probably should better off working with existing vendors. And I agree, because internal optimization is sort of limited. The upside is limited by default. How much you can earn is like 100% there, right? But building a new value proposition, changing your business model, differentiate. That sort of upside is unlimited in a way.

对。

Right.

所以，你也要看回报，对吧？

So, then your returns also you look at the returns, right?

是的。那正是更适合自己创新和构建，或者至少与能帮你在这个领域创新的伙伴合作的领域，对吧？而内部优化这类事情，反正每个人都会做。我们真的想把所有钱都投在那里吗？

Yeah. And that's the area that it's better to look at innovating and building something yourself or at least partnering with someone that can help you innovate in that space, right? Versus everyone's going to be doing this anyways. Like internal optimizations. Is that really where we want to invest all of our money?

每个人都在为此构建解决方案。比如现在每家大型 AI 公司都专注于这方面的投资。所以，你为什么要做？我是说，你反正会被他们超越。你没有那么多资源。

Everyone is building solutions for this. Like every big AI company now focused in their investment into that. So, why would you? I mean, you will be outcompeted by them anyways. You don't have that much resources.

对，没错，是的。

Right. Exactly, yeah.

你提到了 AI 优先的公司。如果你不是 AI 优先的公司，你可能不应该那样做。但你能建立一个 AI 优先的衍生公司吗？什么是最佳策略？是创建一个实验室或衍生公司之类的东西，从零开始构建吗？成为 AI 优先的公司意味着什么？

And you mentioned the AI first companies. If you're not an AI first company, you probably shouldn't do that. But can you build an AI first sort of a spin-off of your company? What would be the best strategy? Would it be like creating a lab or a spin-off or anything like that that you can just build from scratch? And what does it mean to be an AI first company?

是的。所以，如果你是一家成熟的公司，想成为 AI 优先的公司，实际上有两种方法，对吧？一种是在组织中嵌入技能，每个团队里都有一个非常懂 AI 的人，可能在应用开发层面，也可能在更底层。另一种是组建所谓的——我讨厌用“老虎团队”这个词，那是 2000 年代的词了——但他们有这些团队，可以与所有其他部门合作，帮助他们上手、学习并利用 AI。程度各有不同，对吧？因为我们谈论的是企业、初创公司，以及许多不同的组织，它们对 AI 有不同的需求，也有不同的内部构建能力。所以，如果你是初创公司，AI 优先绝对是正确的方向。你想在这个领域竞争。如果你不利用 AI，你就会被抛在后面。但如果你运营物流，比如航运物流，有很多 AI 应用场景。也许对你来说，最好的做法是创建一个小团队，帮助审视这些用例，确定我们需要构建什么工具，可以购买什么来帮助这些特定用例并进行优化。所以，有很多低垂的果实，比如客户服务聊天机器人，还有更微妙复杂的事情，比如那些正在取代呼叫中心的初创公司，对吧？他们使用语音智能体，并围绕这一点构建整个服务，不再需要拥有 200 人的呼叫中心。所以，那种公司就会成为 AI 优先的公司，对吧？

Yeah. So, if you're a well-established company and you want to be an AI first company, there are really two approaches, right? There's the embedded skill sets in the organization where you may have in each team somebody that's very AI savvy, either on the application development side or potentially lower in the stack. Or you have these sort of—I hate to use the word like tiger teams. That's a very 2000s word, but they have these teams in place that can work with all of your other departments to help get them onboarded, educated, and leveraging AI. And to various extents, right? Because we're talking about enterprises, startups, a lot of different organizations that have varying needs for AI and varying capabilities to actually build that capability in-house. And so, if you're a startup, AI first is definitely the way to go. You want to compete in this space. If you're not working with AI, you're going to be left behind. But if you're running logistics, let's say shipping logistics, lots of applications of AI to be used. And maybe it makes the best sense for you to create a small team that's going to help look at these use cases and what are the tools that we need to build, what can we buy to help with these particular use cases and optimize. And so, there's a lot of low-hanging fruit like customer service chatbots, more nuanced and complex things like when you talk about startups that are coming in the space of replacing call centers, right? And they're using voice agents and building an entire service around this that doesn't require a call center with 200 people in it. And so, that kind of company is going to be an AI first company, right?

说到客户服务，自从 MCP 推出后，你的智能体或 ChatGPT 可以通过 MCP 原生地代表你执行某些操作。那么问题来了，如果你可以直接连接，让你的智能体直接与你想要购买东西的公司的智能体对话，你真的还需要客户服务团队为你服务吗？你认为这会在未来两三年、四五年内发生吗？我们真的还需要网站和整体流程吗？比如我不再预订航班或酒店。如果未来几年你可以直接让你选择的语言模型提供商帮你订票，你真的还需要那些吗？

Speaking of customer service, now that after the introduction of MCP, your agents or your ChatGPT can go and make certain actions on behalf of you by MCP natively. So, the question would be then, do you really need a customer service team to service you if you can just connect directly and let your agent talk directly to the agent of the company that you want to buy something from? Do you see that happening within the next two, three, four years? Do we really need websites and overall like I don't book a flight or hotel. Do you really need that in the next few years if you can just ask your choice of LM provider, book me a ticket?

对。是的，我认为那是每个人都渴望拥有的：自己的个人智能体，能代表他们行动，了解他们的偏好，能主动代表他们行事，理解他们的意图。

Right. Yeah, I think that's what everybody aspires to have: their own personal agent that can act on their behalf, that knows their preferences, that can take the initiative on their behalf, understand their intent.

并且能够驾驭这种不确定性，无论是在商业领域还是他们交互的任何系统中，都能负责任地代表你行事。显然，人类参与仍然很重要。如果我告诉我的智能体，当某款沙发低于某个价格时帮我买下来，这是一个非常简单的用例。但可能有一些细微之处，作为人类你会意识到。例如，价格趋势实际上在下降，我从其他信息来源知道价格会继续下跌。所以虽然还没到买入阈值，但也许我想等等。你在股票、加密货币和其他事情上都能看到这种情况。所以有些事情我们可能更擅长处理。在购买决策上，关键在于你愿意在多大程度上信任你的智能体。但毫无疑问，我更希望让智能体帮我买票、安排行程，确保一切就绪且没有意外，同时识别我的偏好并加以应用。我认为这是未来几年的发展方向，我们会看到它越来越成为现实。

And be able to navigate that uncertainty, whether in the commercial space or whatever systems they're interacting with, to act responsibly on your behalf. Obviously, human in the loop is still important. If I tell my agent I want them to buy me a particular couch when it becomes available below a certain price, that's a very simple use case. But there may be nuanced things that you as a human would recognize. For example, the trend is actually going downward, and I know from other sources that the price will keep going down. So it hasn't hit the threshold to buy, but maybe I want to wait. You see this with stocks, cryptocurrencies, and other things. So there are things we may be better suited to do. When it comes to making purchases, it's about how much trust you put in your agent. But definitely, I would prefer to have an agent buy my tickets, organize my itinerary, and make sure everything is in place without surprises, recognizing my preferences and applying them. I think that's the direction we're moving in the next few years, and we'll see it materialize more and more.

现在正在涌现一些不断演进的规范和标准，它们将有助于智能体式接口的标准化。所以，当你有一个在线商店时，不再是传统的人类界面让你选择不同产品，那个界面可能更偏向文本驱动，更适合智能体交互。我们看到了很多这样的趋势。我们也在微交易、支付通道以及智能体式 AI 的所有这些系统中看到了大量创新。在智能体式 AI 基金会，我们第一天就启动了七个工作组，专注于从监管问题到安全、隐私、可靠性和可观测性等各个方面。所有这些领域都需要在智能体式 AI 中得到解决。由此，我们将看到不断演进的标准和规范，以及社区可以应用的解决方案和想法，以构建一个非常强大的生态系统，构建网络，并赋予智能体一定的责任来代表我们行事。我认为每个人都对这个想法感到兴奋，尤其是在中国，比如一人公司（OPC），你什么都不做，而你的智能体为你赚钱。当然，任何人都会对此感到兴奋。那么，我们如何实现这种经济呢？不断演进的智能体经济包括自主智能体的概念，它们代表你行事，可能通过区块链或传统商业进行。这一切都非常令人兴奋，并为创新提供了大量机会。这就是为什么我们看到这么多专注于智能体的初创公司。“智能体”这个词可能被过度使用了，但创新者有机会解决真正的问题。

There are evolving specs and standards coming out now that will help with agentic interfaces, to help standardize them. So when you have an online store, instead of a human interface where you select different products, that interface might be more text-driven and better suited for an agent to interact with. We're seeing a lot of that. We're also seeing a lot of innovation in microtransactions, payment rails, and all these systems in agentic AI. At the Agentic AI Foundation, we kicked off day one with seven different working groups focused on everything from regulatory issues to security, privacy, reliability, and observability. All these areas need to be solved for agentic AI. Out of that, we'll see evolving standards and specs, solutions and ideas that can be applied by the community to build a very strong ecosystem, build the networks, and endow agents with a certain amount of responsibility to act on our behalf. I think everybody is very excited about this idea, particularly in China, like the OPC (one-person company) where you sit back and do nothing while your agents make you money. Certainly, anyone would be excited about that. So how do we get to that kind of economy? The evolving agentic economy includes ideas of self-sovereign agents working on your behalf, potentially doing things on the blockchain or through traditional commerce. It's all very exciting and provides a lot of opportunity for innovation. That's why we're seeing so many startups that are agentic-focused. The word 'agent' is probably overused, but there's an opportunity here for innovators to solve real problems.

有趣。那么说到你自己，说到 Linux，我不确定你是否自己写代码，但你的工程师们，他们被允许使用 AI 吗？到什么程度？内部是如何治理的？他们通常用 AI 做什么？

Interesting. And when it comes to yourself, when it comes to Linux, and I'm not sure if you're coding yourself, but your engineers, are they allowed to use AI? To which extent? What's the governance internally? And what are the use cases they normally use AI for?

Linux 基金会下有很多项目，每个项目都有自己的 AI 使用政策，但我想说大多数项目都使用编码智能体来增强能力。我自己也写代码，和很多人一样，我首选的是 Claude Code。我承认它是一个非常有用的工具。我能直接发布它生成的代码吗？不行，我肯定需要把各个部分拼接起来，并运行我的单元测试。但它在完成所有单元测试和系统生产化所需的各个部分方面变得越来越熟练。不同项目对 AI 有不同的政策，而且 AI 在开源社区中被广泛使用。许多以前不会编码的人现在也能创建 PR，因为他们遇到了代码库并试图修复或优化某些东西。这导致维护者被大量 PR 淹没，需要审查和整理。在另一端，我们有人使用 AI 来审查 PR、合并变更并验证它们是否可行。所以我想说每个人都在某种程度上使用 AI，因为你几乎必须用 AI 来对抗 AI 的影响。这非常有趣。

We have a lot of projects in the Linux Foundation, many projects. Each of those projects has its own policies around AI use, but I'd say the majority use coding agents to augment their abilities. I myself code as well, and my go-to is Claude Code, like many people. I appreciate that it's a very useful tool. Can I ship the code it produces? No, I definitely need to tack pieces together and run my unit tests. But it's becoming a lot more proficient at doing all the unit testing and pieces needed to productionalize systems. Different projects have different policies on AI, and AI is being used extensively in the community for open source projects. Many people who couldn't code before have created a PR because they ran into the codebase and tried to fix or optimize something. That has inundated maintainers with PRs to review and sort out. On the tail end, we have folks implementing AI to review PRs, consolidate changes, and validate whether they're viable. So I'd say everybody is using AI to some degree because you almost have to use AI to counter the effects of AI. It's very interesting.

我认为工程团队现在有点难办，因为业务人员带着 Claude Code 来找他们说：“看，它对我有用，为什么对你们没用？”我觉得共识是工程团队和业务团队之间的距离正在缩短，因为业务团队实际上可以编写代码，算是产品的 MVP 或原型。他们可以应用业务逻辑，并比通过工程团队进行每次迭代更快地查看和调整。同时，工程团队可以通过看到业务逻辑的实现而受益匪浅，尽管在这种情况下代码质量可能并不重要。

I think engineering teams are having a hard time now that business folks come to them with Claude Code and say, 'Look, it works for me, why doesn't it work for you guys?' I feel like the consensus is that the distance between engineering and business teams is shortening because business teams can actually write code, sort of MVPs or prototypes of the product. They can apply business logic and see and adjust much faster than going through every iteration with the engineering team. Meanwhile, the engineering team can benefit tremendously by seeing the business logic implemented, albeit maybe the code quality doesn't really matter in that case.

是的，是的。软件开发生命周期正在被压缩，对吧？所以从需求到概念验证现在变得更容易了。但生成代码并不意味着它是最优的。生成代码并不意味着它是安全的。所以开发者仍然有很多事情要做。

Yeah, yeah. The software development life cycle is getting truncated, right? So being able to go from requirements to POC is a little easier now. But generating code doesn't mean it's optimized. Generating code doesn't mean it's secure. So there are still a lot of things that developers need to do.

是的，你需要生产规模、安全性以及所有这些东西，如果你刚开始，可能无法用你的云代码做到。最后一个问题：如果你今天是一位创始人，有 100 万美元，你现在会投资哪里？你会在这个领域关注什么？

Yeah, you need the production scale, the security, and all that stuff that you probably won't be able to do with your Claude Code if you're just starting. If the last question: if you're a founder today, you have 1 million US dollars, where would you invest right now? What would be your focus in this space?

这是一个相当开放的问题，因为人们必须弄清楚他们将在市场上如何差异化，对吧？

So that's a pretty wide-open question because people have to figure out where they're going to differentiate themselves in the market, right?

你认为机会在哪个细分领域或行业，或者你现在看到哪些低垂的果实？

Where do you see opportunity in maybe what segment or industry, or where do you see the low-hanging fruit right now?

因为我从开源的角度出发，显然倾向于解决通用问题的通用系统。我非常看好智能体。它们目前不擅长长期任务。多智能体系统以及安全方面还有很多问题需要解决。我希望看到更多专注于智能体系统安全和隐私的公司。确实有一些初创公司，但针对通用应用的还不多。我看到在高度监管的行业，比如金融，人们为非常具体的系统搭建脚手架和护栏。但如何更通用地应用呢？我们如何像处理确定性系统那样创建通用的安全和安全系统？如何将这个应用到随机性领域？很高兴看到有人在这些想法上努力，尤其是提高可靠性。在智能体式 AI 中，LLM 的失败模式（如幻觉）被继承到了智能体中。我们能够通过智能体添加更多确定性层，创建更紧密的上下文，并变得更加专注。但与此同时，你不能通过模型解决所有问题。你必须采取系统构建的方法。你要构建一个系统，而不是完全专注于模型。我认为这就是我们扩展当今 LLM 能力的方式：通过围绕它们构建系统，而不一定专注于模型本身。

Because I favor from an open-source lens, I obviously favor general-purpose systems that solve problems. I'm very bullish on agents. They aren't great at long-horizon tasks right now. There are a lot of other problems to be solved in multi-agent systems and safety and security. I would like to see what I'm not seeing enough of: companies that are focused on security and privacy for agentic systems. There are startups out there, but for general-purpose applications, not seeing as much. So I see folks in heavily regulated industries, like finance, creating scaffolding and guardrails for very specific systems. But how do you apply that more generally? How can we create general-purpose safety and security systems like we do with deterministic systems? How do we apply that in this stochastic space? It'd be great to see people working on ideas there, especially around improving reliability. In agentic AI, the failure modes of LLMs are inherited, like hallucinations. These are inherited into agents. We have the ability with agents to add more deterministic layers, create tighter context, and become more focused. But at the same time, you're not going to solve everything through the model. You have to take a system-building approach. You're going to build a system, not focus entirely on the model. I think that's how we scale the capabilities of today's LLMs: by building systems around them, not necessarily focusing on the model itself.

太棒了。谢谢你，Matt。很高兴你能来。

Fantastic. Thank you, Matt. It was great to have you.

非常感谢。谢谢。

Appreciate it. Thanks.

谢谢。

Thank you.