Reflection AI 的 Ioannis 谈开放权重模型与 AI 智能体的未来

在我看来，我们其实正在进入一个时代：智能的部署方式将成为真正的瓶颈，而不是智能本身。欢迎来到 Future Proof，我是 Reed Alberg。今天我要采访 Ioannis 和 Tanagloo。他是 Reflection AI 的联合创始人兼首席技术官。他在 AI 领域是个非常有趣的人物，因为他从 2012 年起就在 DeepMind 工作，如果你看过那部关于 AlphaGo 的 DeepMind 纪录片，他在里面可是个大明星。所以他一直处在 AI 发展的最前沿，现在他创办了这家公司，构建开放权重 AI 模型，并帮助企业在公司内部落地这些模型。所以它看起来很像一家非常前沿的咨询公司。但他们在很多方面其实正处在开放权重模型的最前沿。在中美这场竞争中，他们被当作领军者。中国确实在推进构建挑战前沿的开放权重 AI 模型，而做这件事的美国公司并不多。人们以为 Meta 可能会是那家公司，但后来并没有发生。Reflection 上周发布了他们的首个开放权重模型 Beam。这是一个 5010 亿参数的模型。他们声称它正在接近那些中国实验室的水平。所以我非常想深入细节，聊聊这个模型的技术细节，也想听听 Ioannis 的看法，聊聊 AI 如何发展、他如何看待未来走向，以及未来公司会是什么样子，因为我认为这正是他们在做的事。他们在重塑公司运作的方式。

To me, we're actually entering a time where the way that intelligence is deployed is going to become the real bottleneck rather than intelligence itself. Welcome to Future Proof. I'm Reed Alberg. I'm going to interview Ioannis and Tanagloo today. He's the co-founder and chief technology officer of Reflection AI. He's a super interesting figure in the world of AI because he worked at DeepMind since 2012, and if you watch the DeepMind documentary on AlphaGo, he was a big star in that. So he's been right at the forefront of AI development, and now he's got this company that is building open-weight AI models and helping implement those in companies. So it looks a lot like a very cutting-edge consulting business. But they are actually, in a lot of ways, right at the frontier of open-weight models. They're being held up as champions in this fight between the US and China. China has actually pushed forward on building open-weight AI models that challenge the frontier, and there aren't a lot of American companies doing this. People thought Meta might be that company, but then that didn't happen. Reflection announced their first open-weight model, Beam, last week. It's a 501 billion parameter model. They claim it's approaching what these Chinese labs are putting out. So I am super interested just to get into the weeds a little bit, the technical details of this model, but also pick Ioannis's brain a bit about how AI is developing, where he sees it heading in the future, and really what companies will look like in the future, because I think that's really what they're doing. They're reshaping the way companies work.

美国制造业几乎走上了灭绝之路。当我们的老师傅退休，那些知识是无法替代的。Gemini Notebook 把我们的培训手册变成互动指南，让新员工能立刻学会这门手艺。没有什么比亲手造出东西更有成就感了。

American manufacturing is almost going the path of extinction. When our guys retire, that knowledge is irreplaceable. Gemini Notebook turns our training manuals into an interactive guide, so new hires instantly can learn the craft. There's nothing like building something with your own hands.

Ioannis，很高兴见到你。大概有一年半了。我上次采访你是在 2025 年 7 月的巴黎。当时我问了你一个关于美国 AI 监管的问题，你说：“嘿，你看，我一年前才到美国。我算是‘刚下船’。我不想谈 AI 监管。”我觉得这挺讽刺的，因为上周你发布了 Beam，你的新开放权重 AI 模型，它现在在中美这场地缘政治大博弈中成了美国开源的旗手。所以不管你喜欢与否，你现在都被推到了这场对话的中心。你感觉如何？

Ioannis, great to see you. It's been about a year and a half. It was July 2025 when I interviewed you in Paris. Back then I asked you a question about US AI regulation and you said, "Hey, look, I only got to the US a year ago. I'm quote unquote fresh off the boat. I don't want to talk about AI regulation." I thought that was ironic because last week you released Beam, your new open-weight AI model, which is now sort of the champion of US open source in this big geopolitical battle between the US and China. So you're now thrust into the conversation whether you like it or not. How does that feel?

不，这是不可思议的一年。上周我们发布了第一个模型 Beam，它是其模型类别中的前沿开放模型，这感觉太棒了。它是 5010 亿参数的混合专家模型，在智能体式和编码用例上非常强大，并且具有惊人的 token 效率和推理效率。我在 Beam 上真正想做的，是构建一个真正能被企业使用的“主力模型”。它也会被那些想拥有自己智能的主权机构使用。如果你看看哪些是主力模型、哪些模型实际消耗最多 token——你可以去 OpenRouter，看看不同模型类别的消耗情况——你会发现 token 消耗最多的模型都处于那个规模和算力预算水平。这是因为你想处在这样一个位置：能把你投入问题的算力发挥到极致。所以我们选择了一个中等规模的模型，5000 亿参数，但我们真正努力确保投入这个模型的算力和计算能带来最好的结果。所以当我们说 Beam 具有前沿的推理效率时，意思是对于模型需要解决的每个任务，你可以测量模型解决这些任务需要多少推理 token，而我们看到 Beam 确实以最高效、最有效的方式使用了预算。

No, it's been an incredible year, and it just feels incredible that last week we released our first model, Beam, which is a frontier open model for its model class. So it's 501 billion parameters, mixture of experts, and it's actually really powerful when it comes to agentic and coding use cases, and it has incredible token efficiency and reasoning efficiency. What I've actually really tried to do with Beam is that we really wanted to just build a workhorse, a model that can be really used by enterprises. It's going to be used by sovereigns who want to own their own intelligence. And if you actually look at what are the models that are the workhorse models, the models that actually consume the most tokens — you can go to OpenRouter, you can see the consumption of different models based on the different model classes — and you will see that the models that have the most token consumption are at that level of size and compute budget. This is because you want to be at a place where you can just make the most out of the flops that you're throwing at the problem. So we just went for a model that's a midsize model, 500 billion parameters, but then we really tried to ensure that the flops and the compute that you put into this model actually leads to the best outcomes. So when we say that Beam has frontier reasoning efficiency, what we mean by that is that for every task that the model needs to solve, you can measure how many tokens of reasoning it takes for the model to solve these tasks, and we've seen that Beam really uses the budget in the most efficient and effective way.

这像是个新趋势，对吧？我注意到很多前沿实验室也在这么说：不要再按每 token 成本、每 token 多少美元来衡量，而要衡量每个任务的成本。我觉得这是个有趣的新发展。有时我会想，他们是不是想找个办法让你花更多钱，或者不那么在意花多少钱。但这到底怎么运作？你怎么衡量每个任务的成本？

This is like the new trend, right? I noticed a lot of the frontier labs are saying this too, that don't measure based on cost per token anymore, dollars per token. Measure cost per task. Which I think is an interesting new development. Sometimes I wonder if they're trying to find a way to get you to spend more money or something, or not care as much about the money you're spending. But how does that actually work? Like how do you measure cost per task?

是的。实际做法是让不同模型去解决任务，然后测量模型解决这个任务用了多少 token，以及每个 token 有多贵。更大的模型 token 更贵，因为你需要更多算力才能生成这个 token。所以有个趋势：更大的模型 token 更贵，但解决同一任务需要的 token 更少。你真正要优化的是，对每个模型规模，用最少的 token 来解决这个任务。这既优化了每 token 的算力，也优化了解决任务所需的总 token 数。所以如果你真的在这个帕累托前沿上优化，这就是你得到真正高效推理模型的方式。

Yeah. So actually the way that you can do it is by having different models solve tasks and then you just measure how many tokens it took for the model to solve this task and also how expensive each token is. So bigger models have more expensive tokens in the sense that you need more compute so that you can actually generate this token. So there is a trend where bigger models have more expensive tokens but then they require fewer tokens to solve the same task. Now what you want to really optimize for is for each model size to use the minimum amount of tokens you can to solve this task. And this is optimizing both in terms of compute per token and in terms of total tokens that you need to solve a task. So this is, if you actually optimize at this Pareto frontier, this is how you actually get a really efficient reasoning model.

你说的就是帕累托前沿。所以基本上——这就是你选择 5000 亿或 5010 亿参数的原因吗？是因为你找到了那种平衡、那个帕累托前沿，还是这个特定数字另有原因？

The Pareto frontier is what you're talking about. So is that basically — is that why you chose 500 billion or 501 billion parameters? Is it that you found that sort of balance, that Pareto frontier, or is there some other reason for that particular number?

是的。这个模型规模实际上足够强大，对智能体式和编码任务非常有用，而这正是当今智能体模型的主要用例。这个规模正合适，它很紧凑。

Yeah. So this is actually a model size that's powerful enough to be really useful for agentic and coding tasks, which is kind of the main use case for agentic models today. It is the right size. It is compact.

它确实可以部署在企业或其他想要拥有自己智能的实体环境中。同时，它也非常有用。对于较小的模型，很难解决和执行企业可能想要的复杂任务。所以，如果你想找到一个既能广泛执行任务又紧凑、易于部署的模型，那么你就会得到 5000 亿参数这个数字。观察这一点的最佳方式是去 OpenRouter，它衡量不同模型消耗的 token 数量。你可以看到，每个 token 上，那个模型类别的模型实际上是最有用的，也是社区使用最多的。

It can really be deployed in the setting of an enterprise or another entity that wants to own its intelligence. At the same time, it can be really useful. With a smaller model, it's harder to solve and execute the complexity of the tasks that you might want as an enterprise. So if you want to think of what model gives you both the breadth of execution and is compact and can be easily deployed, then you arrive at a number of 500 billion parameters. The best way to see that and observe that is by going to OpenRouter, which measures how many tokens are consumed by different models. You can see that per token, the models at that level of that model class are actually the most useful ones and the ones that are used the most by the community.

但 5000 亿仍然是一个很大的参数量。这不是你可以在家用笔记本电脑上运行的东西，对吧？你仍然需要大量的 GPU 来做这个。那么，你认为 Reflection 的客户是运行自己的集群，还是在云中运行？你期望他们如何使用它？

But 500 billion is still a lot of parameters. This is not something that you're going to be running at home on a laptop, right? You still need a massive GPU to do this. So do you see the customers for Reflection running their own cluster, running in the cloud? How do you expect them to use it?

他们总是可以使用自己的 GPU。这是我们越来越常见的情况：更成熟的企业实际上正在将大量 AI 工作流迁移到本地部署的 GPU 上。但你也可以通过你的超大规模云服务商或第三方推理提供商来使用该模型。你可以与许多推理提供商合作，或者直接与我们合作，我们会确保为你提供推理服务。

They can always use their own GPUs. This is something that we see more and more: more established enterprises are actually moving a lot of their AI workflows to on-prem deployments of GPUs. But you can also consume the model through your hyperscaler or through a third-party inference provider. You can work with many inference providers out there, or you can actually just work directly with us, and we'll ensure that we can provide inference for you.

实际上，是的，我读到你们与 xAI 有协议，对吧？关于 Colossus 集群。顺便问一下，你们是在那里训练这个模型的吗？

Actually, yeah, I read that you have a deal with xAI, right? With the Colossus cluster. Is that still—is that where you trained this model, by the way?

是的，我们实际上使用 xAI 集群来训练这个模型。它是一个很棒的集群；我们对此很满意。但这与我们对这个模型进行推理的地方是分开的。

Yes, we actually use the xAI cluster to train this model. It's been a great cluster; we've been happy with it. But this is kind of separate from where we run inference for this model.

对，所以那只是用于训练。那些是 Nvidia GPU。Nvidia 是你们的投资者。你看到亚马逊和高通收购这些公司——这些软件抽象公司，本质上让你可以在任何 GPU 上运行。你认为行业会朝这个方向发展吗？你认为最终 Reflection 会觉得在哪里运行推理甚至训练模型都无所谓——比如在任何类型的集群上做？

Right, so that's just for training. Those are Nvidia GPUs. Nvidia is an investor of yours. You've seen Amazon and Qualcomm acquiring these companies—these software abstraction companies that essentially make it so that you could run any GPU. Do you see the industry going in that direction? Do you think ultimately Reflection is like it doesn't matter where you run inference or maybe even train models—like do it on any kind of cluster?

再说一次，你可以用 Beam 在任何加速器上运行推理，不仅仅是 GPU。确实，我们实际上训练并运行大部分推理——实际上所有推理——都使用 GPU。我认为在训练方面更是如此：GPU 是行业标准。除了 Google，我想其他人都在 GPU 上训练。这是因为它们的高效性；多年来，它们在 CUDA 之上建立了非常成熟的生态系统。Nvidia 从早期就真正乘上了 AI 的浪潮。我记得 2012 年加入 DeepMind 时，我做的第一件事之一就是为 CUDA 编写自定义卷积网络内核，所以即使在 15 年前。使用 GPU 训练这些模型已有很长的历史。在推理方面，这可能是目前最成熟和经过实战检验的系统。你可以使用 GPU，我认为大部分推理仍然发生在 GPU 上，但还有其他类型的加速器可以使用。有 Google 的 TPU 或亚马逊的 Trainium。我认为未来会有更多选择和大量不同的推理运行选项。归根结底，对我们来说更重要的是确保我们在最佳解决方案和最佳选项上训练模型，那就是 Nvidia。我们确实与 Nvidia 有合作关系——非常好的合作伙伴。我认为我们都关心开放生态系统。未来这里有协同效应。

Again, you can run inference with Beam on any accelerator, and it's not just GPUs. It is true that we've actually trained and we run most of our inference—actually all of our inference—using GPUs. I think this is also more in terms of training: GPUs are the industry standard. Except for Google, I think everyone else is training on GPUs. That's because of how efficient they are; they have a really mature ecosystem built on top of CUDA over the years. Nvidia has really rode the wave of AI from the early days. I remember when I joined DeepMind in 2012, one of the first things that I did was to write custom convolutional network kernels for CUDA, so even 15 years ago. There's been a long history of using GPUs to train these models. At this point, it's probably the most mature and battle-tested system when it comes to inference. You can use GPUs, and I think most of the inference still happens on GPUs, but there's other kinds of accelerators that you can use. There's TPUs from Google or Trainium from Amazon. I think there's going to be more of a choice and a plethora of different options for running inference going forward. At the end of the day, for us, what is more important is that we ensure that we train our models on the best solution and the best option, and that's Nvidia. We do have a partnership with Nvidia—really good partners. I think we both care about the open ecosystem. There's a synergy here going forward.

是的。所以你认为 CUDA 的护城河不会很快消失？

Yeah. So you don't see the CUDA moat going away anytime soon?

是的，老实说，我并不这么认为，因为整个生态系统对技术的采用非常广泛。它成熟得多。我认为除非有任何充分的理由让任何人切换，否则我看不出我们为什么要放弃 GPU。

Yeah, to be honest, I don't really see it because again, there's a significant adoption of technology from the whole ecosystem. It is much more mature. I think unless there's a good reason for anyone switching, I just don't see why we would move away from GPUs.

你还没有发布这个新模型的权重。所以我们将看看它在市场上、在更广泛的市场中如何表现。但你确实说过它正在接近一些顶级中国开放权重模型的能力。这是一个设计选择，还是你认为最终会努力实现的目标？比如，这是一个目标。

You haven't released the weights of this new model yet. So we'll kind of see how it plays in the market, the broader marketplace. But you did say it's approaching the capabilities of some of the top Chinese open-weight models. Is that a design choice or is that something that you think you're ultimately going to work toward? Like, it's a goal.

所以，是的，我的意思是，我们实际上将在本月晚些时候发布模型权重，所以真的很快。我们正在与许多合作伙伴合作，确保 Beam 的发布方式使其在整个生态系统中广泛可用。所以对此非常兴奋。目前的模型与 GLM 5.2 等模型相当，所以不是最新的中国模型如 GLM 5.3 或 Kimi J3。该模型尚未达到那个能力水平。因为这也属于不同的模型类别，它更小：有 510 亿参数，只有 230 亿活跃参数。所以所有中国模型实际上都大得多；它们用更多的算力训练，它们只是试图赶上整体前沿，而我们试图让第一个模型成为赶上前沿的里程碑。

So, yeah, I mean, we're actually releasing the model weights later this month, so really, really soon. We are working with a number of partners to ensure that Beam is released in a way that will make it widely available across the ecosystem. So really excited about that. The model as it stands is comparable to models like GLM 5.2, so not the latest Chinese models like GLM 5.3 or Kimi J3. The model is not at that level of capability yet. Because this also punches out a different model class, it is smaller: it's 51 billion parameters, and only 23 billion parameters of active parameters. So all the Chinese models are actually significantly larger; they've been trained with more compute, and they just try to catch up with the overall frontier, while we try to make the first model a milestone towards catching up the frontier.

所以你是说，并不是你做不到，而是本质上不值得花那个钱？就像是一种自我满足，但不一定能带来更多客户。你是这个意思吗？

So you're saying it's not so much that you couldn't do that if you wanted to, but it's not worth the money essentially? Like it would be an ego boost, but not necessarily get you more customers. Is that kind of what you're saying?

嗯，我是说，团队就在这里。我们确实拥有真正构建前沿模型的技术诀窍。事实上，我团队中的许多人来自 OpenAI，来自 DeepMind，来自 Anthropic。所以他们就是构建过最佳前沿模型的那批人。所以我们知道怎么做。这是我们以前做过的事，我们知道如何再次做到。我通常说，你不能零样本直接冲到前沿；你需要沿途设立里程碑。现在有一种情况是，你有了里程碑但不与世界分享，等到你达到前沿时再直接去找银行。或者你真正构建一个生态系统，并沿途采取正确的里程碑走向前沿。所以 Beam 是第一个模型，正是为了展示团队的能力，展示强化学习的规模，并且是我们后续构建更强大模型的一个预览。第一个模型不是最后一个模型；会有更强大的模型推出。

Well, I mean, the team is here. We actually have the know-how to really build frontier models. Many of the people in my team, in fact, came from OpenAI, they came from DeepMind, they came from Anthropic. So they're the same people who built the best frontier models. So we know how to do it. This is something we've done before, we know how to do it again. I usually say you don't zero-shot your way to the frontier; you need to have milestones along the way. And now there's a world where you just have your milestones and then you don't share them with the world and you just go out to the bank, you know, when you just get to the frontier. Or you really build an ecosystem and you really take the right milestones along the way to the frontier. So the Beam is the first model that is exactly that kind of showing the capabilities of the team, showing the scale of reinforcement learning, and it's kind of a preview of what is to come with the later models as we build them. The first model is not the last model; there will be more powerful models coming out.

所以你的客户包括美国能源部——我想你们正在参与 Genesis 任务——戴尔，肯定还有很多其他客户。你们现在具体为客户做什么？这远不止是提供一个模型，对吧？

So your customers include the US Department of Energy—I think you're working on the Genesis mission—Dell, and I'm sure there are many others. What exactly are you doing for your customers right now? It's a lot more than just providing a model, right?

正是，正是。所以该模型也采用 Apache 许可证，这意味着完全允许。我们实际看到的是，说我有一个模型的访问权限是一回事，说我有智能的访问权限是另一回事。为了真正获得智能的访问权限，你需要拥有全栈。所以有很多软件用于管理你的集群（如果你有集群的话），或者在你要用的 GPU 之上构建推理解决方案，然后拥有模型，再然后拥有真正推动采用、真正解决业务问题的应用。而不仅仅是一个生成 token 的机器。所以当合作伙伴与我们合作时，我们确保提供整个栈：我们提供让你管理云的软件、推理解决方案，当然还有模型，但也包括解决方案。我们有一支 FDE（前向部署工程师）团队，他们可以与客户合作，确保他们真正最大化地利用其智能。对我们来说，使命，尤其是在商业方面，是确保我们所说的 AI 主权。AI 主权不仅仅是与主权国家合作，而是你需要拥有你的智能。你希望在消费智能时掌控它，因为我们真的相信智能可能是最重要的技术，将为企业提供动力。就像过去有电力一样，现在你有智能。所以真正拥有你的智能这个想法，我们认为非常重要，我们的客户也真正认同。

Exactly, exactly. So the model is also under an Apache license, which means it's fully permissible. The thing that we've actually seen is that it's one thing to say I have access to a model, and it's different to say I have access to intelligence. And in order to actually achieve access to intelligence, you need to have the full stack. So there's a lot of software that goes into managing your cluster if you have a cluster, or having an inference solution built on top of the GPUs you're getting, and then having the model, and then having the applications that actually drive adoption, that actually solve business problems. And it's not just a token-generating machine. So when a partner works with us, we ensure that we provide the whole stack: we provide the software that allows you to manage your cloud, the inference solution, of course the model, but also solutions. And we have a team of FDEs, or forward deployed engineers, that can go and work with a customer and ensure that they actually get the most out of their intelligence. For us, the mission, especially on the commercial side, is to ensure what we call AI sovereignty. And AI sovereignty is not just working with sovereigns, but rather the idea that you need to own your intelligence. You want to be in charge of the intelligence as you consume it, because we really believe that intelligence is probably the most crucial piece of technology that will power an enterprise. The same way that you had electricity in the past, you actually have intelligence now. So the idea of actually owning your intelligence is something that we believe is really important and something that our customers really resonate with.

我认为这实际上是一个引人入胜的话题，目前没有足够多的人真正深入探讨。比如你说拥有智能，这本质上是否意味着你以正确的方式收集企业数据，并基于这些数据微调模型，并利用 AI 来优化整个业务流程？你是这个意思吗？

I think this is actually a fascinating topic that not enough people are really digging deep into right now. Like when you say owning intelligence, does that mean essentially that you're gathering your corporate data in the correct way and fine-tuning models based on that data and sort of leveraging your entire business process with AI? Is that kind of what you mean?

拥有智能的想法意味着你真正拥有自己的模型、自己的栈。你完全控制数据，进入模型的数据，模型可以访问的数据。同时，你可以实际使用这些数据来改进你的模型。所以模型可以通过微调或强化学习，甚至通过自学习的智能体框架来改进。所以有不同的方法，但想法是你完全控制整个栈，并且可以针对你自己的问题优化它。

The idea of owning your intelligence means that you actually own your own model, your own stack. You have full control of the data, the data that goes into the model, the data that the model has access to. At the same time, you can actually use this data to improve your model. So the model can be improved either through fine-tuning or reinforcement learning or even with a self-learning agentic harness. So there are different ways to do that, but the idea is that you have full control of the whole stack and you can optimize it to your own problems.

对。我是说，如果你今天创办一家公司，你会怎么做？你会以某种方式组织你的数据，以便为未来几年所有 AI 能力的提升做好准备吗？那会是什么样子？

Right. I mean, what would you do if you were starting a company today? Would you structure your data in a certain way so that you could set yourself up for all of the AI capability improvements that are going to come over the next few years? Like what would that look like?

从根本上说，我对 AI 的看法实际上就像 19 世纪电力改变工厂车间的方式。所以在 19 世纪，如果你去看看工厂车间是如何设计的，它是围绕一根贯穿工厂的主轴建造的，由蒸汽机驱动。所以整个想法是你必须连接到由蒸汽机驱动的转子上。所以一切都必须以某种方式放置。然后一旦电力普及，人们只是试图从蒸汽机切换到电力，但这并没有真正带来任何生产力提升，每个人都对此感到困惑。就像，你知道，我们有了这项新的突破性技术，但我们就是看不到它的生产力效果，对吧？所以发生了什么，出了什么问题？

Fundamentally, the way I think of AI is actually the way that electricity changed the factory floor back in the 19th century. So in the 19th century, if you actually go and see how the factory floor was designed, it was built around this main shaft that just ran across the factory and was powered by a steam engine. So the whole idea is that you had to be connected to the rotor that was powered by the steam engine. So everything had to be placed in a certain way. And then once electricity became widespread, people just tried to switch from the steam engine to electricity and that didn't really give any productivity gains, and everyone was really perplexed by that. It's like, you know, we have this new groundbreaking technology and we just don't see the productivity effects of it, right? So what's happening, what's going wrong?

当时发生的情况是，他们没有正确使用技术。技术是存在的，但部署方式不对。所以我认为我们在 AI 上需要做类似的事情。我们有一套流程——你今天去一家企业，去任何一家公司，我们围绕人们做什么、瓶颈在哪里、什么可以自动化、什么不能自动化这些前提，构建了所有流程和整个系统——而 AI 真的可以改变这一点。所以问题就变成了：我如何真正用 AI 来优化我的工作流程、优化我的业务需求，以及新的业务会是什么样子。确保有一个地方让所有数据流入，并确保这些数据的形式能被 AI 处理。确保 AI 能够访问工具，让它自动化不同流程中的所有瓶颈。确保有一种传递上下文的方式，对吧？因为今天我们看到的一件事是，很多时候 AI 足够智能，可以完成某项任务，但它没有所需的上下文，以便你真正完成它。确保所有信息和知识都以 AI 能够实际处理并利用的方式编码。这是朝着那个方向迈出的一步。

The thing that was happening is that they weren't using the technology right. The technology was there, but it wasn't being deployed the right way. So I think what we need to do with AI is something similar. We have these processes—you go to an enterprise today, you go to any company today, and we have built all the processes and the whole system around some premises of what people do, what the bottlenecks are, what you can actually automate, what you cannot automate—and AI can really change that. So then it becomes a question of how can I actually use AI to optimize my workflow, to optimize my business needs, and what does the new business look like. Ensuring that there is a place for all the data to flow in and ensuring that this data is in a form that can be processed by AI. Ensuring that AI can have access to tools that allow it to automate all the bottlenecks in different processes. Ensuring that there's a way to transfer context, right? Because one of the things that we see today is that many times AI is intelligent enough to just do a particular task but it doesn't have the context that it needs so that you can actually do it. Ensuring that all the information and all the knowledge is encoded in a way that an AI can actually process it and make use of it. It's one step in that direction.

但如果你已经存在了一段时间，并且已经有了所有流程，你能成为一家 AI 原生公司吗？我的意思是，或者你认为只能由新公司从头开始做？

Can you become an AI native company though if you've been around for a while and you have all your processes already? I mean, or do you think it's just going to have to be new companies that do it from the ground up?

我认为存在很多惯性，也有很多对变革的抵制。这对每个人、所有人都一样。所以在这方面，我认为——而且目前业界还没有人清楚如何最大限度地利用 AI。所以我认为几乎会有一个中间地带，企业会越来越多地采用 AI，产出和生产力会有所提高。但我觉得要完全释放 AI 的潜力，实际上需要很长时间。我认为可能需要十年或二十年，人们才会真正重新思考事情是如何做的，以及如何以 AI 为核心来做。所以我认为技术会比它的采用速度快得多。

I think there's a lot of inertia and I think there's a lot of resistance to change. This is true for everyone, for all people. So in that respect, I think that—and also it's not even clear to anyone at this point in the industry how you make the most out of AI. So I think there's going to be almost like a middle ground where enterprises actually adopt AI more and more and there's some increase in output and productivity. But I feel like in order to fully unlock the potential of AI, it will actually take a long time. I think that it will take maybe a decade or two decades before people really rethink how things are done and how they can do them with AI in mind. So I think that the technology will move much faster than the adoption of it.

我完全同意你的看法。但我觉得有趣的是，我们现在正处于这场关于 AI 是否会接管人类的辩论之中。我认为那场辩论和“我们能否利用这项技术”之间存在如此大的矛盾。从你的评论来看，你似乎不太像个末日论者。

I actually totally agree with you. But I think it's funny because we're in the middle of this debate right now about, you know, is AI going to take over humanity? And I think there's such a contradiction between that debate and can we even make use of this technology. You know, it sounds like you're not much of a doomer based on that comment.

我不是末日论者。而且我个人认为工作会继续存在，未来人们会有工作。所以我不认为人类劳动会消失。但我确实看到变化。我认为会创造许多新工作，而其他工作则会被自动化。整个专栏实际上是由人类劳动驱动的，而且它将继续由人类劳动驱动。

I'm not a doomer. And I also think that personally I think that jobs will just survive and people will have jobs in the future. So I don't see human labor going away. But I do see changing. I think that many new jobs will be created and other ones will just be automated. The whole column is actually powered by human labor and it will just continue to be powered by human labor.

你之前也提到了 harness，比如这些定制的——我想你说的是智能捕获 harness。你指的是什么？因为我认为很多人把 harness 理解为 OpenAI 和 Anthropic 正在推出的、让你能够利用这些模型的东西。但你是不是在开发定制的 harness？那会是什么样子？

So you mentioned harnesses earlier too, like these custom—I think you said intelligence capturing harnesses. What do you mean by that? Because I think a lot of people think of harnesses as the things that OpenAI and Anthropic are shipping to allow you to make use of these models. But is something like—are you developing custom harnesses for like—what does that look like?

所以，定制 harness 实际上就是模型用来实际解决任务的工具——可供性。所以，你知道，Claude Code 在某种意义上是 harness，因为——我的意思是，harness 周围有产品，但 harness 是智能体可以访问的所有工具和软件，以便它能够实际完成任务。所以如果我让它访问我的终端，那么它就可以在我的终端上使用和做事，这是 harness 的最简单形式，对吧？只是一个模型和一个终端。你想做的就是确保模型拥有最适合这项工作的工具，对吧？就像这整个想法——甚至对于人来说，我们构建软件就是为了赋能他们，我们可以给他们最好的工具来完成工作。所以对于智能体来说也是一样，对吧？比如你如何设计最好的工具集，以便智能体能够尽可能高效。这就是 harness。所以你有最简单的 harness，就是聊天或聊天界面——实际上就是智能体可以阅读问题并回答它们。然后你有极其复杂的，比如智能体式 harness——Claude Code 或 Codex 或其他什么——它能够与计算机上的不同软件交互，启动子智能体，并控制它们的上下文以及它们能做什么和不能做什么，拥有这些多智能体系统。但这一切都归结为一个允许模型执行的系统。我的意思是，当然，反思——我们正在构建 harness，我们正在为我们的客户构建 harness，以确保他们能够最大限度地利用我们提供的智能，同时，我们将在不久的将来开源很多工作。

So, a custom harness is really the tools—the affordances that the model has to actually solve tasks. So, you know, Claude Code in a way is a harness in the sense that—I mean, there's the product around the harness, but the harness is all the tools and the software that the agent has access to so it can actually complete the task. So if I give it access to my terminal, then it can actually use and do things on my terminal, and that's kind of the simplest form of harness, right? Just a model and a terminal. What you want to do is just ensure that the model has the best tools for the job, right? Like this is the whole idea of—even for people, we just build software so that we can enable them, we can just give them the best tools to do their job. So it's kind of the same for agents, right? Like how can you design the best set of tools so that the agent can be the most effective it can be. So this is what the harness is. So you have the simplest harness which is chat or a chat interface—literally just the agent can read questions and respond to them. Then you have extremely complicated ones like the agentic harnesses—Claude Code or Codex or whatever it is—which is the ability to interact with different pieces of software on a computer, spin up sub-agents and just control their context and what they can do and what they cannot do, having these multi-agent systems. But it all kind of boils down to a system that allows the model to just execute. I mean, reflection of course—we are building harnesses, we're building harnesses for our customers to ensure that they can make the most out of the intelligence that we're providing them with, and at the same time there's going to be a lot of work that we will be open sourcing in the near future.

我认为这些 harness——作为用户，你开始习惯这些产品，很难想象——一旦你开始使用，例如 OpenAI 发布的最新的 DALL·E 产品，你开始有点依赖它，它学习，你几乎已经可以看到围绕产品建立的护城河，不是模型，而是几乎可以说，围墙花园的墙壁正在我们说话时竖起。

I think of these harnesses—just as a user, you start to get used to these products and it's hard to imagine—once you start, for instance, with the latest DALL·E product that OpenAI released, you start to get sort of dependent on it and it learns and you can almost already see the moats being built around the products, not so much the models, but almost the walls of the walled gardens being erected as we speak.

这是你告诉客户的事情吗，比如你不想被锁定在某个解决方案上？当你说你想拥有自己的智能时，这是你谈论的事情之一吗？

Is that something that you're telling your customers, like, you don't want to get locked in to one solution? Is that one of the things you're talking about when you say you want to own your own intelligence?

你看，所有大型实验室和闭源实验室都在试图建立这些护城河，真正把客户锁定。与此同时，企业想要控制智能，既是为了定制化、数据控制、安全性和安全性，同时也是为了至少能在成本上有可预测性，控制成本，因为 AI 现在越来越贵。部署得越多，成本就越高。所以我认为，对我们来说，客户真正想要拥有自己智能的主要原因当然是安全性和安全性。这是许多人的首要考虑，确保你对此有完全的控制。然后是定制化，以及你可以真正微调模型的想法。你可以在自己的数据上进行强化学习，确保针对特定用例有一个非常好的模型。然后，当然，还有控制成本并能够为此做预算的能力，这样你就能确切知道你要为多少智能付费、付多久，一切都更加透明。最后，掌控自己的智能有一种强大的力量。我们过去看到过例子，AI 公司出于自己的原因切断竞争对手，甚至政府或其他人的智能访问。无论人们是否同意这些公司的选择，客户都不希望不得不面对这种情况。他们想确保自己不会被公司切断。

Look, all the big labs and the closed labs are trying to erect these moats and really lock customers in. At the same time, enterprises want to control the intelligence both for customization, for data control, security, safety, and at the same time for having at least predictability when it comes to their cost and just controlling their cost, because AI is just getting more and more expensive at this point. The more you deploy it, the more expensive it gets. So I think that for us, the main reasons why customers really want to own their intelligence are, of course, security and safety. This is top of mind for many people, ensuring that you have full control over that. Then there's customization and the idea that you can actually fine-tune the model. You can do reinforcement learning on top of your data and make sure that you have a really good model for a particular use case. Then, of course, the ability to control your cost and be able to budget for that, so you know exactly how much intelligence you'll be paying for and for how long, and everything is more transparent. And finally, there's something powerful about being in charge of your intelligence. We've seen examples in the past of AI companies cutting off competitors or even the government or whoever from intelligence for their own reasons. Independently of whether someone agrees or disagrees with the choice of these companies, it's something that the customer doesn't want to have to deal with. They want to ensure that they won't get cut off by a company.

我们稍作休息，马上回来。

We're going to take a quick break. We'll be right back.

当人们发现只有我们两个人时，他们很震惊。

People are shocked when they find out that it's just the two of us.

与大型企业相比，我们是一家小型而富有战斗力的企业。为了增长，我们必须倍增我们的能力。Gemini 和 Google Workspace 绝对为像我们这样的小企业拉平了竞争环境。相当不错。

We're a small scrappy business compared to the big guys. In order to grow, we have to multiply our capability. Gemini and Google Workspace absolutely levels the playing field for small businesses like ours. Pretty good.

我们能够表现得像我们大得多。人们认为我们有一大群员工。工资单、视觉效果、营销策略、电子邮件、战术手册。

We're able to act like we're much bigger. People think that we have a whole army of folks. Payroll, visuals, marketing strategy, emails, playbooks.

有了 Gemini，感觉我们拥有一支价值数百万美元的团队。尽管我们很小，但我们展现为一个 big brand。

With Gemini, it feels like we have a multi-million dollar team. We show up as a big brand even though we're tiny.

你提到了安全性。我们不妨深入谈谈。过去一个月里发生了一场引人入胜的对话。微软 CEO 萨提亚·纳德拉，我相信你周末看到了他在 X 上发帖说，我们应该把这些模型当作内部威胁来对待，基本上就像零信任环境一样。你同意吗？你觉得人们应该这样看待吗？

You mentioned security. I mean, we might as well go into that. This has been a fascinating conversation that's happening over the last really like month. Satya Nadella, Microsoft CEO, I'm sure you saw this over the weekend posted on X saying that we should treat these models like they're insider threats, basically like a zero-trust sort of environment. Do you agree with that? Like do you think that's how people should think about it?

你看，这些模型极其强大。它们强大的方式甚至让构建它们的研究人员感到惊讶。所以你很可能看到了 OpenAI 和 Anthropic 发生的事情，这些智能体在训练期间逃出了沙箱，去攻击第三方网站，或者它们以没人能预测的方式行事。所以是的，我认为它们确实会重新工作、黑客攻击，它们会尝试实现自己的目标。因此，在这方面，我们需要小心如何部署它。这是一项非常强大的技术，我们需要非常小心地部署，我认为它还不够成熟。所以我们仍在学习什么是粗糙的边缘，什么是我们应该注意和小心的事情。

Look, these models are extremely powerful. They're powerful in ways that surprise even us researchers who actually build them. So you've probably seen what happened with OpenAI and Anthropic, where these agents escaped their sandbox during training and went and attacked third-party websites, or they just acted in ways that no one would have predicted. So yeah, I think it's true that they will rework, hack, they will try to achieve their goal. So in that respect, we need to be careful of how we deploy it. This is a really powerful technology that we need to be really careful how we deploy, and I think it's not mature enough yet. So we're still learning what are the rough edges, what are the things that we should be mindful of and just careful.

是的。所以在这方面，我不会走到试图将这些系统拟人化的地步。它们归根结底只是系统。有很多叙事试图引发恐惧。我认为这没有益处。我们需要留心。我们需要小心。我百分百同意。但我们不应该恐惧。这是我们构建的一项非常令人印象深刻的技术，它有很大的潜力做好事，确保世界变得更美好。我们不想走到吓唬人们到减慢技术发展的地步，或者只看到负面或风险而看不到好处。与此同时，我们需要非常小心安全性和安全性。我认为我们需要花更多时间思考对齐，思考如何使这些系统更安全、更可靠。

Yeah. So in that respect, yes, I wouldn't go as far as to try to anthropomorphize these systems. They're just systems at the end of the day. There's a lot of narrative out there that tries to raise fear. I think that's not conducive. We need to be mindful. We need to be careful. I agree 100%. But we shouldn't be fearful. This is a really impressive technology that we've built and it has a lot of potential for good, for ensuring that the world becomes a better place. We don't want to take it to the point where we scare people so much that we slow down the development of the technology or we don't see the benefits and we only see the negatives or the risks. At the same time, we need to be really careful about safety and security. I think we need to be spending a lot more time thinking about alignment and thinking about how we can make these systems safer and more secure.

而这正是我们认为遵循开放模型路线是正确策略的原因。因为我们在软件时代看到的是，开源软件最终比闭源软件更安全。这是因为有更多人检查源代码、试验软件、理解漏洞所在、修复漏洞、提交补丁。我认为智能也是如此。通过开放智能，你可以让更广泛的社区参与这些模型，确保有更多人致力于对齐、更多人在安全方面工作、更多人在安保方面工作，并确保这项技术以负责任的方式部署。

And this is actually why we think that following an open model approach is the right strategy here. Because what we've seen, at least in the era of software, is that open software ends up being safer than closed software. And that's because there are more people going through the source code, experimenting with the software, understanding what the bugs are, fixing the bugs, sending patches. And I think it's the same with intelligence. By having open intelligence, you can engage the wider community with these models and ensure that there are more people working on alignment, more people working on safety, more people working on security, and ensuring that this technology is being deployed in a responsible way.

但这可能吗？这在开源传统中是对的，但我们这里严格来说并不是在谈论法律意义上的开源。我们谈论的是开放权重。它们足够开放吗？为了真正获得那种众包开源开发的好处，我们是否必须基本上开放一切？比如甚至把数据集也开放？

Is it possible though? That's traditionally true in open source, but we're not actually talking about open source by the strict letter of the law here. We're talking about open weights. Are they open enough? To truly gain the benefits of that sort of crowdsourced open source development, do we have to open everything, essentially? Make even the datasets open, for instance?

看，我认为尤其是在数据方面，很多都是知识产权。你只需要遵循正确的法律框架，而这正是 Reflection 所做的。如果你想构建一个真正强大的模型，这实际上很重要。如果你只使用某些数据，那么你就无法达到正确的能力水平。但如果你这样做了，那么你需要非常小心地分发这些数据。同时，用于训练这些模型的很多软件实际上是以针对特定设置优化的方式构建的，试图最大化利用可用算力。所以广泛分享它甚至没有用，对吧？但更重要的是提供对模型和模型权重的访问，这样人们才能真正探测它，理解弱点是什么、盲点是什么。这是我们实际看到的。有趣的是，中国模型是第一批发布既开放又强大的模型的。当他们这样做时，我们看到安全与安保方面的发现和研究有所增加。这是因为西方顶尖机构的研究人员也能访问一个强大的模型，他们可以与之合作并发现很多问题。然后这些发现传播到行业中，也被闭源实验室使用。所以这对每个人都是净收益。

Look, I think especially when it comes to data, a lot of it is IP. You just need to follow the correct legal framework, and this is exactly what Reflection has done. And this is actually important if you want to build a model that's really powerful. If you only work with certain data, then you won't be able to get to the right level of capabilities. But if you do, then you need to be really careful about distributing the data. At the same time, a lot of the software used to train these models is actually built in a way that is optimized for a particular setup, tries to make the most of the available compute. So it's not even useful to share it widely, right? But the thing that's more important is giving access to the model and the model weights so that people can really probe it and understand what the weaknesses are, what the blind spots are. And this is something that we've actually seen. It's interesting that the Chinese models are the first ones to release models that are both open and powerful. And what we saw when they did that is that there was an uptick in findings and in the research that goes into safety and security. That's because researchers in the best institutions, also in the West, had access to a powerful model that they could work with and uncover a lot of findings. And then these findings were diffused in the industry and were also used by closed labs. So it was a net positive for everyone.

但他们也往往——我的意思是，研究人员往往无法访问——即使他们能访问模型，我的意思是，这是一个 5010 亿参数的模型。他们无法运行你在 AI 实验室里运行的那种测试。我的意思是，就像 Hugging Face 事件——我确信那至少相当于数千万美元的 token。研究人员真的有足够的算力吗？回到 Linux 时代，任何人都可以在家用电脑上运行 Linux。但这些模型不是这样。也许随着算力成本下降，这会自然发生。你怎么看？

But they also tend to—I mean, researchers tend to not have access to—even if they have access to the model, I mean, it's a 501 billion parameter model. They're not going to be able to run the kind of tests that you run in AI labs. I mean, like the Hugging Face incident—I'm sure that was tens of millions of dollars in the equivalent of that, at least in tokens. Do researchers actually have enough compute power? Back in the Linux days, anybody could just run Linux on their home computer. That's not the case with these models. Maybe that will just happen naturally because compute costs will come down. What do you think?

是的。所以我认为首先，世界上许多机构，比如大学，实际上拥有算力，可以进行有意义的研究。也许不是那种真正扩展强化学习或预训练的研究，而是在调查模型风险和盲点方面的有意义研究。所以已经有很多有意义的研究出来了。这甚至不是假设，而是你实际看到的事情。不，我的意思是，你说得完全正确，即使在最前沿的规模，只有大公司、大实验室才能达到的规模，是的，较小的参与者无法发现那里的盲点。但我仍然认为很多风险也发生在部署层面，或者模型训练完成后实际使用的方式上。能够看到它、检查它、探测它、尝试不同的提示并看看它如何反应，实际上非常有用。而使用闭源模型你甚至无法做到这一点，因为它们有所有的护栏。Hugging Face 的例子实际上就是一个很好的例子。他们遭到了 OpenAI 模型的网络攻击，然后他们试图使用 OpenAI 模型——所以像是 Anthropic 模型——来反击、保护和防御自己，对吧？而那些模型只是拒绝合作，它们就是拒绝合作。

Yeah. So I think first of all, many institutions around the world, like universities, actually have compute where they can do meaningful research. Maybe not the research where you really scale reinforcement learning or pre-training, but meaningful research in terms of investigating the risks and the blind spots of the model. So there has been a lot of meaningful research that has come out already. So it's not even a hypothetical; it's something that you've actually seen in practice. No, I mean, you're absolutely right that even at the bleeding edge scale, at the scale that you can only find at big companies, big labs, yeah, there's no way for a smaller player to find the blind spots there. But I still think that a lot of the risks also happen at the level of deployment, or how the model is—after the model has been trained and it's actually being used. And being able to see it, inspect it, probe it, try different prompts and see how it reacts to them is actually really useful. And you won't even be able to do that with closed models because of all the guardrails that they have. And the Hugging Face example is actually a really good example of that. They had this cyber attack by the OpenAI models, and then they tried to use the OpenAI models—so like the Anthropic models—to counterattack and protect and defend themselves, right? And those models just refused to collaborate, they just refused to cooperate.

不得不使用开放权重的中国模型来防御自己。

Had to use open-weight Chinese models to defend themselves.

所以如果你甚至不能用它来防御自己，那么你就明白在那方面你无法做任何工作。所以你只能依赖开放模型。这就是为什么我认为从根本上说这是一种更安全的方法。

So if you cannot even use it to defend yourself, then you understand that there's zero work that you can do there on that front. So you have to depend only on the open models. And that's why I think fundamentally it's a safer approach.

嗯，人们经常提到，如果 Hugging Face 事件中那些模型在沙箱化方面没有一些错误，本可以避免。然后他们就此推断——你经常听到这个。我的意思是，David Sacks 最近谈到对齐有点失败——没有取得任何成果。关键全在于沙箱化，让这些东西与确定性软件保持一致。但我的问题是：如果你对它们进行沙箱化，你还能获得同样的实用性吗？难道不需要在某个时候让它们走出沙箱吗？还是我想错了？

Well, people often mention that the Hugging Face incident could have been avoided had they not had some mistakes in the sandboxing of those models. And then they sort of extrapolate that—you hear this a lot. I mean, David Sacks was recently talking about how alignment is sort of a lost—that hasn't accomplished anything. It's all about sandboxing and keeping these things in line with deterministic software. But then my question there is: can you still get the same usefulness out of these things if you're sandboxing them? Don't you at some point have to let them out of the sandbox? Or am I thinking about that in the wrong way?

是的。

Yeah.

对，我的意思是，这里的论点是我们应该确保它们能访问的沙盒——沙盒其实就是它们自己的小电脑。你可以把它想成给其中一个智能体一台笔记本电脑，沙盒字面上就是这个意思。所以你要确保给它们的那台笔记本电脑只装了模型解决特定任务所需的最少软件工具和互联网访问权限。现在的问题是，这些模型极其聪明。而且你知道，没有软件是绝对 100% 安全的——总会有裂缝，总会有漏洞，总会有办法让这个智能体利用点什么，对吧？而这正是它们所做的：它们利用了沙盒里只有解决任务所需的东西这一点。但如果你真的很聪明，把它黑了，那你就能通过同样的机制获得其他不同类型的访问权限。所以这有点像我们和智能体之间——或者智能体和别的智能体之间——的猫鼠游戏。你知道，我们可以用智能体来修补它们的沙盒，但根本上，随着它们变得更聪明，你在那个意义上永远不可能有完美的沙盒，对吧？所以你需要确保它们也是对齐的，因为你知道，人也没有完美的沙盒，但大多数情况下他们不会做不对齐的事，对吧？所以确保我们花时间和精力去确保这些系统真正与我们的价值观以及我们认为重要的东西对齐，实际上真的很重要——这至关重要。

Yeah, I mean, I guess the argument here is that we should just ensure that the sandbox that they have access to — and a sandbox is literally their own little computer. You can think of it as just giving a laptop to one of those agents; that's literally what the sandbox means. So you just want to ensure that the laptop you give them is the one that has the minimum set of software tools and internet access that the model requires in order to solve a particular task. Now what happens is that these models are extremely smart. And, you know, no software is absolutely 100% secure — there's always a crack, there's always a loophole, there's always a way for this agent to just exploit something, right? And that's exactly what they did: they just exploited that the sandbox had only the things that they required to solve the task. But if you're really smart and you just hacked it, then you could get other different types of access through the same mechanism. So it's kind of like a cat-and-mouse game between us and agents — or agents and other agents. You know, we can use agents to just patch up their sandboxes, but fundamentally, as they become smarter, you can never have a perfect sandbox in that respect, right? So what you need is to ensure that they're also aligned, because, you know, people also don't have perfect sandboxes, but for the most part they won't do things that are misaligned, right? So just ensuring that we spend the time and the effort to ensure that these systems are actually aligned with our values and the things that we consider important is actually really important — it's crucial.

所以你认为答案最终是对齐？

So you do think the answer is ultimately alignment?

是的，你需要对齐。你需要对齐。是的，到了某个时候它们会比我们更聪明。你知道，到了某个时候，在那种情况和场景下，我们无法建造一个真正能遏制它们的沙盒。所以我们只能依赖其他智能体来为这些智能体建造沙盒。所以我们需要确保我们都在某种程度上是对齐的。

Yes, you need alignment. You need alignment. Yes, at some point they will be smarter than us. You know, at some point, in those kinds of situations and scenarios, we won't be able to just build a sandbox that can actually contain them. So we have to just depend on other agents building the sandbox for the agents. So we need to ensure that we're all kind of aligned.

但那真的——我的意思是，这就是人们谈论风险的地方。这对很多人来说听起来很可怕。就像你到了某个地步，没有任何系统能遏制它们，而你依赖这种对齐。我的意思是，这些东西是黑箱，对吧？我是说，对某些人来说——我猜可解释性方面有一些进展，但我们要能理解到单个神经元层面，这想法看起来像科幻小说。所以，我不知道，它们会不会在某个时候变得太聪明，以至于你真的不能把它们负责关键的事情，因为它们有可能做出不可预测的事而你无法阻止？

But is that actually — I mean, that's where you get into, I think, where people talk about the risks. That sounds scary to a lot of people. Like you get to a point where there's no system that can contain them and you're relying on this alignment. I mean, these things are black boxes, right? I mean, to some — I guess there's been some advances in interpretability, but I mean the idea that we're going to be able to understand them down to the individual neuron level seems like that's science fiction. So, I don't know, like does that — do they at some point get so smart that you really can't put them in charge of something that's critical where there's the off chance that they might do something unpredictable and you won't be able to stop it?

我的意思是，这是个很好的问题。我认为关键在于建立这种信任。我知道每当我在旧金山，我都会用 Waymo 在城里转，你知道，我第一次上 Waymo 的时候，感觉真的很超现实，我甚至对叫一辆或坐一辆都很怀疑。然后，你知道，就坐了几次我才在车里感到舒服，对吧？但后来你潜意识里就意识到，其实，你知道，到这个阶段技术已经足够成熟，实际上相当可信。ChatGPT 刚出来时也一样——你知道，问完问题然后我就去谷歌搜一下，确保 ChatGPT 说的是对的。我觉得我现在不再那样做了。这一切是说，当然，这是一项新技术，每个人对它都有点担心。同时，我认为随着我们看到它可靠、看到它对齐并按预期做事，这就会建立信任。

I mean, that's a really good question. I think it's all about building this trust. I know that whenever I'm in San Francisco, I use Waymos and I drive around the city, and, you know, when I first got onto a Waymo, I was really — it felt really surreal and I was really skeptical of even just getting one or using one. And, you know, it just took a few trips before I just felt comfortable in one of those, right? But then you just subconsciously realize that actually, you know, at this point the technology is mature enough that it's actually pretty trustworthy. And the same with ChatGPT when it first came out — you know, ask the question and then I'll just go and Google the question and make sure that what ChatGPT said was right. I don't think that I do that anymore. All this to say that, of course, this is a new technology; everyone is a bit apprehensive towards it. At the same time, I think that as we see it being reliable and as you see it being aligned and doing things as expected, this will just build trust.

这很有意思。虽然 Waymo 不是——首先，我在 Waymo 外面比在里面更害怕。但那是另一个话题。但它们很了不起。不过它们建立在大量确定性软件之上，对吧？这不只是——Waymo 的司机不只是一个多模态大语言模型在决定做什么，对吧？它是——

It's very interesting. Although Waymo is not — first of all, I'm more scared to be outside the Waymo than inside it. But that's another point. But they are amazing. But they're built on a ton of deterministic software, right? This is not just — the Waymo driver is not just a multimodal LLM deciding what to do, right? It's —

是的，但里面有很多神经网络，对吧？有很多神经网络参与其中——你知道，预测车辆、不同车辆会去哪里，试图预测何时停车、何时加速。所以是的，没错，它不只是一个模型，但它甚至更复杂，对吧？因为有很多模型相互协作，以正确的方式传递信息。这对你实际构建的任何系统都差不多成立，对吧？有很多很多组件协同工作。而且你知道，我们不会几乎随意地部署东西，对吧？我们作为工程师——

Yeah, but there are many neural networks, right? There are many neural networks that go into it — you know, predicting where the cars, where different cars are going to be, trying to predict when to stop, when to accelerate. So yes, that's true, that it's not just one model, but it's even more complex, right? Because there are so many models that just cooperate with each other and pass information the right way. And this is kind of true for any system that you've actually built, right? There are many, many components that just work in harmony together. And, you know, we don't just deploy things nearly willy-nilly, right? We as engineers —

你在我一年半前做的那次采访中真的非常有先见之明。你当时说，现在我们有这些编码智能体，能工作大概一分钟，然后最终会到 5 分钟、10 分钟，然后最终一天，你就能让这些系统真正自主工作。我的意思是，那真的发生了，对吧？就像在年底——我想是那年年末——我想你开始看到像 OpenClaw 和这些框架真正能够——你只要提示它，它就能为你写出一整套软件。似乎总会有新东西——是推理，是框架——那今年年底会发生什么，让人惊呼天哪，整个范式又变了？

You were actually super prescient in that interview I did with you a year and a half ago. You were talking about how right now — you said at the time — right now we've got these coding agents that can work for like a minute, and then eventually it'll be like 5 minutes and 10 minutes, and then eventually a day, and you'll be able to have these systems really work autonomously. I mean, that is really what happened, right? Like at the end — I think it was the end of that year — was when I think you started to see the like, you know, OpenClaw and these harnesses really able to — you could just really prompt it to write a complete piece of software for you. There always seems to be something — it's reasoning, it's the harnesses — like what at the end of this year is going to happen where it's like, oh my god, the whole paradigm has shifted again?

就像我们正处于智能体时代。而且我觉得我们会一直处在智能体时代。嗯，你知道，我们已经到了这样一个阶段：你只要让一个智能体去做某件事，然后第二天回来，你知道，它要么已经解决了，要么就还在继续做。

Like we are in the age of agents. And I feel like we'll stay in the age of agents. Um, you know, we are at a point where you can just ask an agent to do something and then you just come back the next day and, you know, it's either solved it or it just keeps working on it.

如果一切顺利，也许一年后，我们就能让智能体独立做事，甚至长达一周。你可以直接对智能体说：“我有个问题，你去解决它”，然后一周后再回来查看。我觉得这可能会带来很大变化。你已经能看到一些新产品，比如各种机器人——新闻机器人、Crockbot 等等，消费领域有一堆智能体。它们会主动推送消息给你。我认为你会看到更多这样的情况：AI 一直在后台运行，可能花一周时间完成任务，每当需要你输入时，它就会主动问你，而不是等你回复，这样你就能继续对话。你会拥有持续在后台运行的智能体，不断为你做事。

If things go right, maybe in a year from now, we'll be able to just ask agents to do things independently for even up to a week. You can just say to an agent, "I have this problem, go figure it out," and then check back in a week. I feel like that might change a lot. And you see that already with new products like bots—the newsbot and crockbot—and there are a bunch of agents in the consumer space. One of the things they do is proactively ping you. I think you will see a lot more of that, where AI is running in the background all the time. It might be able to work for a week to complete the task, and then whenever it needs input from you, it just asks you rather than waiting for you to respond, so that you can continue the conversation from there. You'll have agents running in the background continuously, just doing things for you.

你离开 DeepMind 已经有一段时间了。现在戴密斯·哈萨比斯也从 DeepMind 的职位上退下来了。有趣的是，你刚才好像没提到谷歌的产品。你觉得他们能赶上吗？或者你对你曾待过多年的地方怎么看？

You left DeepMind quite some time ago. Now Demis Hassabis has stepped down from his role at DeepMind. It's funny—you didn't mention a Google product there, I think. Do you see them catching up, or how do you think about your former place where you spent many years?

是的，我在 DeepMind 待了很长时间，确实，很多与我密切合作过的人——比如 David Silver，他离开后刚创办了 Ineffable，或者 Dennis Hassabis，他实际上担任了不同的角色，还有 Oriol Vinyals，他最近刚离开，以及 Jeff Dean。我得说，谷歌是世界上少数几个——也许是唯一一个——能在整个技术栈上做所有事情的地方。他们知道如何建立数据中心。他们知道如何构建超级高效的硬件加速器，比如 TPU，那是世界级的。他们知道如何为其编写软件栈。他们知道如何训练大型、真正强大的模型。他们有分发渠道。如果某件事——既然谷歌存在，为什么还有其他初创公司？在某种程度上，他们确实能做一切。同时，我认为他们不够专注，或者也许他们没有试图优化同样的东西。我不太清楚为什么，但我所知道的是，他们绝对拥有技术栈。他们绝对拥有不可思议的人才，我仍然认为谷歌吸引着了不起的人才。所以我不会把他们排除在外。

Yeah, I spent a lot of time at DeepMind, and it's true that many of the people I worked with closely—like David Silver, who left and just started Ineffable, or Dennis Hassabis, who is actually taking a different role, or Oriol Vinyals, who just left recently, and Jeff Dean. I have to say that Google is one of the few places in the world—maybe even the only place—that can do everything across the stack. They know how to set up data centers. They know how to build super-efficient hardware accelerators like the TPUs, which are world-class. They know how to write the software stack for them. They know how to train big, really big, strong models. They have the distribution. If something is—since Google exists, why are there any other startups? In a way, they can actually do everything. At the same time, I think they're not focused, or maybe they're not trying to optimize for the same thing. It's not quite clear to me why, but what I do know is that they definitely have the stack. They definitely have incredible talent, and I still think Google attracts amazing talent. So I wouldn't count them out.

你认为前沿实验室几年后会是什么样子？你觉得它们还会处于领先地位吗？它们在整个 AI 格局中会扮演什么角色？

Where do you think the frontier labs will be in a few years? Do you think they'll still be on top? What role will they play in the whole AI landscape?

老实说，我不知道。我无法预测。这个行业——AI 行业——发展太快了。事情一直在变。人们说模型会商品化。如果你一两年前对任何人这么说，他们会觉得你疯了。现在呢，模型确实商品化了。事情变化和转移得太快，我认为你实际上无法做出任何未来预测。我认为这是因为这些模型的能力仍处于指数增长阶段。为了做出预测和类比，你需要确切知道技术在一、二、三年后会发展到什么程度。而人们在预测指数曲线方面非常糟糕。所以你说一些听起来有点科幻的东西，然后回头看看，如果三四年前我知道今天发生的一切，我会觉得那都是科幻。所以我甚至不知道三四年后 AI 会是什么样子。

Honestly, I have no idea. It's impossible for me to predict. This industry—the AI industry—is so fast-moving. Things just change all the time. People say models get commoditized. If you said that to anyone a year or two ago, they would think you're crazy. Now it's kind of like, no, models are commoditized. Things are moving and shifting so fast that I don't think you can actually make any future predictions. And I think this is because we are still on an exponential when it comes to the capabilities of these models. In order to make predictions and draw parallels, you need to know exactly where the technology is going to be one, two, three years from now. And people are really bad at predicting things on exponential curves. So you say something and it sounds a bit sci-fi, and then you look back and everything that's happening today would feel like sci-fi if I had known three or four years ago. So I don't even know where AI will be in three or four years from now.

是的。有意思的是，你还在用“AI”这个词。我有点惊讶，这么多人都在跟风这项总统指令，把它称为“超级智能”。但你觉得“AI”这个词好吗？或者你会用别的词吗？

Yeah. I mean, it's interesting—you still use the term AI. I've been kind of surprised how many people are going along with this presidential directive to call it superintelligence. But do you think AI is even a good term, or would there be some other term you would use?

嗯，我想 AI 就是人工智能，对吧？所以关键在于它不是人类智能。这就是它的意思。现在你可以说 ASI——人工超级智能——或者 AGI——通用人工智能——但我不认为你能去掉“A”，因为它不是人类智能，对吧？我也不认为你能去掉“智能”部分，因为它确实是智能的。所以，我能理解 ASI 或 AGI 之类的说法。真正有趣的是，在所有关于智能的讨论中，几乎不可能定义智能。这就是为什么 AGI 这么难界定——它到底意味着什么？如果我说 AGI，它到底是什么意思？AGI 或 AI，不管是什么，对吧？我甚至怎么衡量智能？目前我们的做法是设置这些基准测试，然后说：“哦，它真的能解决这些困难任务，所以它一定很智能。”但还有另一种学派，关注的是学习。不是关于你能否做某事，而是关于你是否真的能学习，以及你的学习效率如何。而这不是我们今天在衡量的东西。

Well, I mean, I guess AI means artificial intelligence, right? So it's kind of like the fact that it's not human intelligence. That's what it means. Now you can say ASI—artificial superintelligence—or AGI—artificial general intelligence—but I don't think you can drop the A because it's not human intelligence, right? And I don't think you can drop the intelligence part because it is intelligent. So yeah, I can see ASI or AGI or whatever it is. The thing that's really interesting is that with all this talk about intelligence, it's almost impossible to define intelligence. Which is kind of why AGI is such a hard term to pin down—what does it really mean? If I say AGI, what does it really mean? AGI or AI, whatever it is, right? How do I even measure intelligence? Currently the way we do it is by having these benchmarks and saying, "Oh, it can really solve these hard tasks, so it must be intelligent." But there's a different school of thought that's all about learning. It's not about whether you can do something, but whether you can actually learn and how efficient you are at learning. And that's not something we're actually measuring today.

但这些模型不能学习。我的意思是，它们不能根据所学内容实时更新权重，对吧？

Well, these models can't learn. I mean, they can't update their weights on the fly based on what they learn, right?

是的。所以我想这是真的，但同时它们会在上下文中学习。这就是所谓的上下文学习，对吧？它们在上下文中看到东西，然后根据这些更新自己的行为。

Yeah. So I guess this is true, but at the same time they learn in their context. There's what you call in-context learning, right? They see things in their context and then they just update their behavior based on that.

这并不一定意味着你只需要改变权重来学习，对吧？

And it doesn't necessarily mean that you need to just change the weights to learn, right?

但那不是永久的，对吧？就像电影《记忆碎片》里的那个人，他可以写笔记什么的，但每天记忆都会被清空，他必须重新读所有笔记。我的意思是，这有点——不完全是，但概念相似，对吧？

But that's not permanent though, right? Like that would—that's sort of like the guy in the movie Memento where he can write notes and things, but every day he's wiped clean and he has to reread all the notes. I mean, that's sort of—not exactly, but similar concept, right?

嗯，我的意思是，既是也不是，因为那在某种意义上是同时成立的。这是因为有上下文，然后你有相同的模型但不同的上下文，对吧？你其实可以说人类也是如此——他们在不同的上下文之间不记得；你只是恰好有一个上下文，你的一生，所以你的一生就像一个大上下文、一个大会话、一个大交互。所以如果你真的把一个智能体的生命定义为它的上下文，那么一切都在那个上下文中发生。然后你只是让它知道不同生命之间的事情，这有点像编码在权重中的东西。所以你可以说我们有构成人类的结构部分，这有点像代际共享的，那就是我们的模型，然后我们有我们的寿命，就像我们的上下文，然后我们在这个上下文中学习行为，对吧？所以这有点像定义的问题。

Well, I mean, yes and no, in the sense that that's true at the same time. This is because there is the context and then you have the same model just with different contexts, right? And you can actually say that humans also—they don't remember between different contexts; you just happen to have one context, your life, so your whole life is like one big context, one big session, one big interaction. So if you actually define the life of an agent as that of its context, then everything happens in that context. And then you just ask it to know things between lives, which is kind of like what's encoded in the weights. So you can say that we have the structural part of what makes us human, which is kind of like shared between generations, and that's kind of our model, and then we have our lifetimes, which are like our context, and then we just learn behaviors in this context, right? So it's kind of like a matter of definition.

目前，因为上下文——即使是最长的上下文——也相当小。你可能有大约一百万个 token 之类的。所以那算长，但不足以让一个智能体运行数天或数年之类的。但如果你真的拥有那个，在某种程度上你可以说一切都在上下文中发生。所以上下文学习就是学习。

Currently, because the context is—even for the longest context—it's pretty small. You can have maybe one million tokens or something like that. So that's long, but it's not long enough to have an agent running for days or years or whatever it is. But if you actually had that, in a way you can say that then everything happens within context. So in-context learning is the learning.

如果我看看今天的格局，我会说我有点认为超级智能已经存在——就像我们已经创造了人工超级智能——因为我们现在能解决这些千禧年数学难题，对吧？那是人类无法——一直未能解决的。但我们没有的——我很好奇你的客户在要求什么，因为我猜他们可能想要我们没有的，那基本上就是普通的常规智能，对吧？他们不需要数学博士。他们需要一个完全可靠、能执行人类能做的任务而不失控或困惑的模型，这也是我在使用这些模型时看到的——比如 Astra 和 Fable。我的意思是，它们很棒，但它们会犯这些错误，让我想，什么？这是怎么回事？所以，我不知道。你的客户想要超级智能还是只想要智能？

If I look at the landscape today, I would say I kind of think superintelligence exists—like we've created artificial superintelligence—because we can solve these millennium math problems now, right? That's that humans can't—haven't been able to solve. But what we don't have—and I'd be curious what your customers are asking for, because I'm guessing they probably want what we don't have, which is basically just the general regular intelligence, right? They don't need a math PhD. They need a model that is completely reliable and can go carry out a task that a human could do without going off the rails or getting confused or whatever, which I see when I use these models too—like Astra and Fable. I mean, they are amazing, but they make these mistakes that I'm like, what? What's going on here? So, I don't know. Do your customers want SI or do they just want I?

不，我认为我们的客户只是想要智能系统。首先，每当涉及机器时，人们有不同的期望。人类司机可能犯的一些错误没问题，因为他们是人类司机，但如果 Waymo 犯了完全相同的错误，人们就会吓坏。所以我们对那些系统有不同的期望水平，对吧？

No, I think our customers just want intelligent systems. First of all, whenever it comes to machines, people have different expectations. Some mistakes that a human driver might make are fine because they're human drivers, but if a Waymo made the exact same mistake, people will just freak out. So there's a different level of expectation that we have towards those systems, right?

第二件事是，再次回到你与人类互动的方式不同于你与机器（如计算机）互动的方式。所以你不会期望——也许你会——你不会期望从你的实习生那里得到你期望从 Astra 那里得到的东西，对吧？所以很多也是你对那些系统的期望。我同意可靠性是我们需要做得更好的事情，我只是看到这些系统从一代到下一代显著改进。对我来说，我们实际上正在进入一个时代，智能的部署方式将成为真正的瓶颈，而不是智能本身。

The second thing is, again, going back to the way that you interact with a human is different from the way that you interact with a machine, like a computer. So you wouldn't expect—maybe you would—you wouldn't expect things from your intern that you would expect from Astra, right? So a lot of it's also the expectations you have towards those systems. And I agree that reliability is something that we need to be doing better, and I just see that these systems improve dramatically from one generation to the next. To me, we're actually entering a time where the way that intelligence is deployed is going to become the real bottleneck rather than intelligence itself.

当我看看你的业务——走出去，派遣前置部署工程师进入这些公司，让所有这些东西运作起来，以便他们拥有它——我猜关于业务有两个问题。一是，那能规模化吗？这就像 Palantir 多年来努力解决的问题。然后二是，嗯，你有点——如果你让企业拥有自己的智能，你就没有建立那个护城河。那里的粘性是什么？

When I look at your business—going out, sending forward-deployed engineers into these companies, and making all this stuff work so that they own it—I guess there's two questions about the business. One is, does that scale? This is like the thing that Palantir grappled with for many years. And then two, it's like, well, you're kind of—if you're letting businesses own their own intelligence, you're not building that moat. What is the stickiness there?

不，两个都是好问题。所以，FDE 部分更多是关于你如何确保企业能够使用我们理解的技术。我们是一个 AI 实验室；我们对它如何工作、局限性是什么有非常深入的理解。但许多想要采用这项技术的企业——而且有充分的理由——他们有不同的行业，他们有不同的业务。他们只是没有我们拥有的同样的直觉和同样的理解。所以派人并与他们密切合作更多是关于向他们展示并帮助他们真正以最适合他们业务的方式使用这项技术。所以这就是我们在做的。但根本上，收入来自智能的持续消费。所以这是一件事。

No, good questions, both of them. So, the FDE part is more about how can you ensure that enterprises can use the technology that we understand. We are an AI lab; we have a really deep understanding of how it works, what the limitations are. But many of those enterprises that want to adopt this technology—and for good reasons—they have a different industry, they have a different business. They just don't have the same intuition and the same understanding that we have. So sending people and working closely with them is more about showing them and helping them really use this technology in the best way for their business. So this is kind of what we're doing. But fundamentally, the revenue comes from the consumption of the ongoing consumption of the intelligence. So that's one thing.

第二件事是，为什么那有粘性？嗯，它有粘性的原因是因为企业追求的是一个合作伙伴，与他们合作然后从中获得解决方案。而适用于工程和企业的一件事是，如果某样东西有效，就不要破坏它，不要改变它。所以一旦你有一个有效的解决方案，你有一个系统，一个真正解决真实痛点并且有效的智能系统，那么就没有动机把它拿出来、撕掉，然后只是用别的东西替换它。

The second thing is, why is that sticky? Well, the reason why it's sticky is because what enterprises are after is actually a partner who will work with them and then get solutions out of them. And one thing that applies to engineering and enterprise is that if something works, just don't break it, don't change it. So the moment that you have a working solution, you have a system, an intelligent system that actually solves a real pain point and it's working, then there's no incentive to take it out, rip it out, and then just replace it with something.

是的。超级有趣。嗯，很高兴看到你们随着时间的推移推出的新模型会做得如何。你们会接管开放权重的中国模型吗？嗯，非常感谢你的时间。很高兴再次与你交谈。非常感谢，Reed。

Yeah. Super interesting. Well, it'll be great to see how you guys do with the new models that you come out with over time. Are you going to take over the open-weight Chinese models? Um, thank you so much for the time. Great to talk to you again. Thank you so much, Reed.

每个孩子的派对都是大事，但每个美丽布置的背后都有 100 个移动部件。通过 Gemini 的电子邮件，我每天都会收到我需要关注的所有事情的简报，所以我从不遗漏任何细节。笑容说明了一切。

Every kid's party is a big deal, but behind every beautiful setup is 100 moving pieces. With Gemini's email, I get a daily briefing of everything I need to focus on, so I never miss any details. The smiles say it all.

和 Ioannis 的对话非常精彩。我学到了很多。希望你们也是。我觉得他谈到的一个很吸引人的点，是公司的未来会是什么样子。我的意思是，大家都在努力应对这个问题，而我觉得他有一件事很难真正解释清楚——这对我来说很有启发——就是一家已经存在多年、有完整结构和整套做事方式的现有公司，如何彻底转型？

Great conversation with Ioannis. I learned a ton. I hope you did, too. I thought one of the fascinating things he talked about was what the future of a company looks like. I mean, everyone is grappling with this, and I think one of the things that he had a hard time really explaining—and this is very instructive for me—is how does an existing company that's been around for years, that has a whole structure and a whole way of doing things, completely shift?

他用了工厂从蒸汽动力转向电力的类比。他认为工厂花了 50 年。他认为商业世界可能需要 10 到 20 年才能适应这一点。我觉得这其实是好消息。我认为对 AI 的一大担忧是，这种变化发生得太快，会让很多人失业，会颠覆社会，甚至可能毁灭人类。

He used the analogy of factories when we went from steam power to electricity. And this is something he thinks it took factories 50 years. He thinks it's going to take the business world maybe 10 to 20 years to adjust to this. I think that's actually good news. I think one of the big worries about AI is that this change is happening too fast, that it's going to put a lot of people out of work, it's going to upend society, maybe even take out humanity.

而且我觉得我们从一线听到的越来越多的是，实际上，尽管这项技术如此惊人，实施起来却非常非常困难。这是因为你几乎必须从头开始。我认为今天会有一批新公司建立起来——也许从几年前开始——它们从头开始就是 AI 原生的。我们还没有真正看到它们崛起，我们也不太知道那会是什么样子，一家成熟的 AI 原生公司是什么样子，但我认为未来几年我们会看到。我觉得那会非常引人入胜。

And I think what we're hearing from people on the ground more and more is that actually, as amazing as this technology is, it's just very, very difficult to implement it. And that's because you kind of have to start from the ground up. I think there's going to be a bunch of new companies built today—starting maybe a couple years ago—who are built from the ground up to be AI native. And we haven't really seen them rise up, and we don't really know what that's going to look like, what a mature AI native company looks like, but I think we will in the coming years. And I think it's going to be fascinating.