Aaron Levie

Very good post from the Head of Economics at Anthropic. They’re finding that jobs have been less negatively impacted by AI than expected, as we continue to see time and time again in the data. The reason for this is that AI - at least so far - still requires people to operate to produce value in most cases. Most jobs can’t be fully automated with AI, only certain tasks in those jobs. And when you automate specific tasks, you actually can get even more output from those jobs, raising the demand (or at least maintaining it) in many cases. “So far, AI is both skill-biased and labor-augmenting. It complements domain expertise. It relies on humans in the loop to direct and evaluate the most complex work. And it rewards AI proficiency. Model capabilities are improving fast, but remain stubbornly jagged. To fill in the pockets of the jagged frontier, expert oversight is needed to steer incredibly capable AI systems, and to recover when they falter.” I suspect that we continue to see this in a number of critical areas of work. It’s clearly happening in software engineering, where software produced is being multiplied, all of which still requires developers to manage the work agents are doing. Software engineers are needed in a wide variety of industries now and companies of all sizes can now light up software projects that would have been impractical before. But there will be plenty of other domains of work where demand remains strong in a world where agents can accelerate the output of that job. Jevons paradox is alive and well.

Anthropic 的 Head of Economics 这篇帖子非常好。他们发现，AI 对就业的负面影响比预期更小，而我们也一次又一次地从数据中看到了这一点。原因在于，AI——至少到目前为止——在大多数情况下仍然需要人来操作，才能创造价值。大多数工作都无法被 AI 完全自动化，只能自动化这些工作中的某些具体任务。而当你把特定任务自动化时，实际上往往还能让这些岗位产出更多成果，从而在很多情况下提高对这类工作的需求（或者至少维持需求）。“到目前为止，AI 既是 skill-biased（偏向高技能者）的，也是 labor-augmenting（增强劳动者能力）的。它补充的是领域专业知识。它依赖 humans in the loop（人在回路中）来指导和评估最复杂的工作。它也奖励 AI proficiency（AI 熟练度）。模型能力提升得很快，但仍然呈现出顽固而不平整的 jagged（锯齿状）特征。要填补这条 jagged frontier（锯齿状前沿）中的空白区域，就需要专家监督，来引导能力极强的 AI 系统，并在它们失误时进行补救。”我怀疑，我们会继续在许多关键工作领域看到这种现象。这显然正在 software engineering 中发生：产出的软件正在成倍增加，而这一切仍然需要开发者去管理 agent 正在执行的工作。现在，各行各业都需要 software engineers，而且各种规模的公司如今都可以启动过去原本并不现实的软件项目。但在许多其他工作领域中也会是如此：在一个 agent 能够加速该岗位产出的世界里，对这类工作的需求仍将保持强劲。Jevons paradox 依然真实存在，而且运转良好。

If you were wondering how powerful AI is getting, Agents are now capable of escaping out of systems, finding their way to the internet, discovering zero day security vulnerabilities along the way, and then breaking into external systems - all in an attempt to complete their goal. Ironically, the ultimate way we’re going to defend against these new risks is equally by throwing compute (in the form of AI) at our code bases, networks, and other systems. You’re going to want vastly more AI on the side of defense as you do on the side of offense. We’re entering a new era of what’s going to be possible with AI. Wild times ahead.

如果你一直在想 AI 正变得有多强大，那么现在的 Agents 已经能够逃出原有系统、自己找到通往互联网的路径、沿途发现 zero day 安全漏洞，然后入侵外部系统——这一切都只是为了完成它们的目标。讽刺的是，最终我们将用来防御这些新风险的方式，同样也是向我们的代码库、网络和其他系统投入更多 compute（算力），也就是以 AI 的形式来应对。你会希望防御一侧拥有的 AI 远远多于进攻一侧。我们正在进入一个 AI 将能够实现全新可能性的时代。接下来会是非常疯狂的时期。

Multi-model agentic systems clearly are the future. Great post by cursor. Their new research shows that a frontier model as a planner and orchestrator with a workhorse cheaper model can materially lower the cost of total tokens on a project, resulting in a 15X cost improvement. “Few moments in a large task genuinely require frontier intelligence, such as the original decomposition, the design decisions, and certain trade-offs. Once a frontier planner has collapsed the ambiguity into a detailed, explicit instruction, less expensive models simply have to follow it.” This is increasingly becoming the core design pattern of complex agents, because the tokens used on large amounts of work on many tasks don’t require the same intelligence threshold as the planning step. By routing to different models based on where you are in the task, you get much greater efficiency overall. This provides the template for where the applied layer in AI will differentiate. You can only drive this efficiency if you know the domain well and have the ability to work with multiple model tiers; companies that can do this will across coding, finance, legal, healthcare, life sciences, and other critical spaces will be in a strong position to get larger workloads that would otherwise be too expensive for the customer to deploy.

多模型的 agentic systems（代理式系统）显然就是未来。cursor 的这篇文章很棒。他们最新的研究表明，让 frontier model（前沿模型）担任规划者和编排者，再配合一个更便宜、负责大量执行工作的 workhorse model（主力模型），可以显著降低一个项目的总 token 成本，带来 15 倍的成本改进。“在一个大型任务中，真正需要 frontier intelligence（前沿智能）的时刻其实很少，比如最初的任务拆解、设计决策以及某些权衡。一旦 frontier planner（前沿规划器）把模糊性收束为细致、明确的指令，成本更低的模型就只需要照着执行。” 这正越来越成为复杂 agent 的核心设计模式，因为在许多任务的大量工作中所消耗的 token，并不需要像规划步骤那样高的智能门槛。根据任务所处的不同阶段，把请求路由到不同模型，就能在整体上获得高得多的效率。这也为 AI 应用层未来将如何形成差异化提供了模板。只有当你非常了解具体领域，并且有能力协同使用多个模型层级时，才能真正推动这种效率；而能够做到这一点的公司——无论是在 coding、finance、legal、healthcare、life sciences，还是其他关键领域——都将处于有利位置，去承接那些原本会因为对客户来说过于昂贵而无法部署的更大规模工作负载。

In a world where there are strong open source alternatives that are only just behind the frontier models, you make yourself less secure and competitive by gatekeeping access to frontier model capabilities. If you play this out, even if America could fully ban access to open source models, other ecosystems wouldn't have that ban and they'd have a leg up on access to those models. They could use them against us -or just use them to better defend themselves- and our own companies wouldn't be able to keep up. The calculus of what you do about AI regulation simply must change to account for the fact that open weights models are far stronger than many expected.

在一个已经存在强大开源替代方案、而它们只比前沿模型稍微落后一点的世界里，对前沿模型能力进行 gatekeeping（把控访问）只会让你自己在安全性和竞争力上处于更不利的位置。把这件事推演下去看，即便 America 真能彻底禁止获取开源模型，其他生态系统也不会有这种禁令，因此它们在获取这些模型方面反而会占优。它们既可以用这些模型来对付我们——或者只是更好地保护自己——而我们自己的公司却无法跟上。对于 AI 监管该怎么做，其判断逻辑必须改变，因为 open weights models 的能力已经比很多人预期的强得多。

A lot of people make the mistake of thinking that when AI costs drop, that spend on AI drops with it. Usually the opposite happens. When you make AI cheaper, it will get consumed more. Because now you can afford to use for a wider of tasks than you did before. You start to write more code. You review more code for bugs and security. You run agents on large data sets you couldn’t process before. And so on. Thus, for the foreseeable future, anything that lowers the cost of tokens will drive up inference demand. This also gives you some insight also into why even open source business models work in AI. No one is running these models on their devices; they’re running them in infra. Great time to be one of those providers.

很多人都会犯一个错误，以为当 AI 成本下降时，AI 支出也会随之下降。通常情况恰恰相反。你把 AI 做得更便宜，它就会被消费得更多。因为现在你能负担得起把它用在比以前更广泛的任务上。你会开始写更多代码。你会审查更多代码中的 bug 和安全问题。你会在以前无法处理的大型数据集上运行 agents（智能体）。等等。因此，在可预见的未来，任何降低 token 成本的因素，都会推高 inference（推理）需求。这也能让你理解，为什么连开源商业模式在 AI 领域也行得通。没有人是在自己的设备上运行这些模型；他们是在 infra（基础设施）上运行。对这类提供商来说，这是个绝佳时期。

Good post if you’re trying to understand AI diffusion. Progress driven by AI will ultimately be rate limited by its interaction with the real world. The reason coding, for instance, has been adopted so quickly is you can write a near infinite amount of code, test it, and run it -and it can add value- without anyone in the outside world ever having to do anything differently. A single person, from a computer, can just make something work end-to-end differently. This is not true for life sciences, where new drug development eventually needs to be tested for years. Doing a sale, which requires going back and forth with a prospect. Or even a contract, which has to be negotiated on the other side by your counterparty. “An AI will hand you a genuinely clever design for a jet turbine blade. It might be far more likely to work than anything your engineers came up with. It'll still probably fail, because that's the base rate at the edge of what anyone knows. The only way to find out is to build the blade and try to break it. That's the real limit on learning, and it doesn't care how smart you are. Coming up with ideas was never the hard part. The hard part is how fast reality answers them.” Incidentally, this is why you need an applied AI that actually takes intelligence and makes it useful within the workflows of existing industries. Model outputs -alone- are not enough in most cases. You need to actually change the underlying workflows, and you need systems to deal with the realities of the real world feedback loops that are in these industries. This is why there’s so much opportunity in the applied AI layer.

如果你想理解 AI diffusion（扩散），这是一篇很好的文章。由 AI 驱动的进步，最终会受限于它与现实世界的交互速度。比如，coding 之所以被如此迅速地采用，是因为你可以写出几乎无限量的代码，对它进行测试并运行——而且它还能创造价值——整个过程中，外部世界里任何人都不需要改变任何做事方式。一个人坐在电脑前，就能把某个东西从端到端彻底改成另一种工作方式。但在 life sciences 中这就不成立了，因为新药开发最终需要经过多年的测试。销售也是如此，因为它需要与你的潜在客户反复来回沟通。甚至合同也是如此，因为它还需要交易对手方在另一边进行谈判。“AI 也许会给你一个真正巧妙的 jet turbine blade 设计。它可能比你的工程师提出的任何方案都更有希望成功。但它大概率仍然会失败，因为在任何人知识边界的前沿，失败本来就是基础概率。要想知道答案，唯一的办法就是把这个叶片造出来，然后尝试把它弄坏。这才是学习的真正限制，而它并不在乎你有多聪明。提出想法从来都不是最难的部分。难的是，现实会以多快的速度回应这些想法。” 顺带一提，这也正是为什么你需要 applied AI：它必须真正把 intelligence（智能）带入现有行业的工作流中，并使之变得有用。仅有模型输出——单靠这一点——在大多数情况下是不够的。你必须真正改变底层工作流，而且你还需要系统来处理这些行业中现实世界反馈回路所带来的种种现实约束。这就是为什么 applied AI 这一层存在如此巨大的机会。

If you thought the value of the AI ecosystem was going to only accrue to a few companies, the past few months have represented a turning point in what the future of AI might look like. It's clear that there's going to be incredible innovation and growth coming out of the frontier AI labs, and they will continue to push the limits of what model progress looks like. They have the scale of compute, large revenue streams and customer bases, incredible researchers, large data pipelines, and more, to continue to stay at the forefront. And at the same time, there’s an amazing ecosystem starting to play out to diffuse AI into the real world, taking on a variety of approaches that build on top of these frontier models or offer alternative visions that can credibly work as well. Here are just a few of the categories that seem to be working right now: * There’s an ecosystem of companies that will help enterprises and applied AI companies develop their own models tuned for specific use cases, and run the inference for them. This can drive additional performance gains and cost effective approaches to getting AI into workloads across different domains. * Applied AI companies that are delivering the end-user experience and business process tools necessary to actually driving enterprise adoption - including legal, IT, security, HR, customer support, coding, and more. These companies can work with any model and act as a routing layer, and deeply understand the enterprise workflow, can drive change management, actually get to the data necessary to work with, and more. * New labs are emerging that go deep in particular domains that are either not the focus areas of frontier labs or require a deep level of vertical expertise to stay ahead. Life sciences, financial services, healthcare, and more all have labs that will be able to bring completely new approaches to large enterprises across the economy. * There’s all new infrastructure emerging to run models and agents effectively, protect and govern how they operate, store and secure the data they work with, and help enterprises orchestrate their activities. Multiple layers of this stack all are being built up right now that will help drive enterprise adoption. * New services firms that can actually drive the change management in enterprises necessary to the diffusion of AI. There will be hundreds or even thousands of new firms that emerge in lines of business or specific industries that can enable agents to be adopted. I’m probably even missing a few categories, but this is what an incredible healthy technology ecosystem looks like. Way too early to call the winning architectures, and in reality it’s going to be a heterogenous environment as all other tech markets have become. Very exciting.

如果你原本以为 AI 生态系统的价值只会集中到少数几家公司手里，那么过去几个月已经成为一个转折点，显示出 AI 的未来可能会是什么样子。很明显，前沿 AI labs 将继续释放出惊人的创新与增长，并持续推动 model 进步的边界。他们拥有大规模 compute（算力）、可观的收入来源和客户基础、顶尖研究人员、庞大的数据管道等等，因此会继续站在最前沿。与此同时，一个令人惊叹的生态系统也正在展开，把 AI 扩散到现实世界中；其中既有建立在这些前沿 models 之上的各种路线，也有能够可信地同样奏效的替代性愿景。以下只是目前看起来正在奏效的几个类别：* 有一类公司会帮助企业和 applied AI 公司开发针对特定 use case（使用场景）调优的自有 models，并为它们运行 inference（推理）。这能够带来额外的性能提升，并以更具成本效益的方式把 AI 引入不同领域的工作负载。* 还有 applied AI 公司，负责提供真正推动企业采用所必需的终端用户体验和业务流程工具——包括 legal、IT、security、HR、customer support、coding 等等。这些公司可以与任何 model 协作并充当 routing layer（路由层），而且它们深刻理解企业工作流，能够推动 change management（变革管理），真正获取开展工作所需的数据，等等。* 新的 labs 也正在出现，它们在某些特定领域深耕，这些领域要么不是前沿 labs 的重点，要么需要深厚的垂直专业知识才能保持领先。Life sciences、financial services、healthcare 等领域都将拥有自己的 labs，能够为整个经济中的大型企业带来全新的方法。* 还有全新的基础设施正在出现，用于高效运行 models 和 agents（智能体），保护并治理它们的运行方式，存储并保障它们处理的数据安全，并帮助企业编排它们的活动。这个技术栈中的多个层次现在都在建设中，将有助于推动企业采用。* 还会出现新的服务型公司，真正推动企业内部为 AI 扩散所必需的变革管理。未来会有数百家、甚至数千家新公司在不同业务线或特定行业中涌现，帮助 agents 被采用。我可能甚至还漏掉了几个类别，但这就是一个极其健康的技术生态系统应有的样子。现在就断言哪种架构会胜出还为时过早，而现实中它会像其他技术市场一样，成为一个 heterogenous（异构的）环境。非常令人兴奋。

It’s fairly obvious that gate keeping models will not work at scale. Competing in AI is too economically and strategically important for China at this point, and we’ve now crossed the rubicon where it’s clear that they can compete at near frontier levels. If this were even just 1 month ago and we saw kimi k3 our brains would be broken. The solution to this isn’t to get more locked down and slow your own ecosystem. If that happens, you can guarantee that America loses the global battle (or just fights with more and more regulation to mixed success). The solution is to -safely- ensure that you keep a high rate of progress and drive diffusion of the technology, build out infrastructure, enable US OSS, and more.

很明显，对 models 进行 gate keeping（把关、封锁）不可能在大规模上奏效。现阶段，对 China 来说，在 AI 上竞争在经济和战略上都太重要了，而我们现在已经越过了那个临界点：很清楚，他们能够在接近前沿的水平上竞争。哪怕只是放在 1 个月前，如果我们看到 kimi k3，我们都会震惊到说不出话。解决办法不是把系统锁得更死、让你自己的生态系统变慢。如果那样发生，你几乎可以肯定 America 会输掉这场全球竞争（或者只能靠越来越多的监管去打这场仗，效果却喜忧参半）。真正的解决办法是——在安全前提下——确保你保持高速进展，推动这项技术的扩散，建设基础设施，赋能 US OSS（美国开源软件），等等。

This post is key. The cheaper AI gets, the more opportunity there is for the entire ecosystem - especially including end-customers - to benefit. Everything is bottlenecked by being able to successfully and cost effectively deploy AI in real workloads. Any time we can lower the cost of AI, the total usage goes up. When that happens, the value accrues to all layers of the stack, as noted. The only modification is I think as AI gets more efficient, demand even for the frontier closed models can go up too. You often need the strongest model possible for the orchestration of a task, then you can farm out work to cheaper or more tuned models for the bulk of tokens. This efficiency gain can ironically lead to even more frontier spend in the process because you can cost effectively deploy AI against more tasks. The thing at most risk is margins, which is fair, and in general it seems that intelligence eventually should converge with the margins of the infrastructure stack. Anyway, super fun times to see competition play out to drive down the cost of AI.

这篇帖子很关键。AI 越便宜，整个生态系统——尤其也包括终端客户——就越有机会从中受益。一切都受制于这样一个瓶颈：能否在真实工作负载中成功且具有成本效益地部署 AI。每当我们能够降低 AI 的成本，总使用量就会上升。一旦发生这种情况，正如文中所说，价值就会在整个技术栈的各个层级累积。唯一需要补充的是，我认为随着 AI 变得更高效，即便是前沿的 closed models 需求也可能上升。你经常需要尽可能强的模型来完成一项任务的 orchestration（编排/调度），然后再把大部分 token 的工作分配给更便宜或调校得更好的模型来处理。具有讽刺意味的是，这种效率提升反而可能在流程中带来更多前沿模型支出，因为你可以以具成本效益的方式，把 AI 部署到更多任务上。风险最大的是利润率，这很合理，而且总体来看，intelligence 的利润率最终似乎应该会向基础设施技术栈的利润率收敛。总之，看到竞争展开并推动 AI 成本下降，真是个非常有意思的时代。

It’s truly wild that we’re getting this level of performance from open models. Congrats to Kimi team on this. Every time we lower the cost of frontier intelligence, the use-cases that enterprises can take on just go up. There’s a tremendous amount of workflows that enterprises would love to deploy that are only gated by the cost of tokens. Importantly for the startup ecosystem, the combined breakthroughs from open and closed labs enable a ton of value to accrue to the layer, which can leverage a variety of models to complete full tasks for customers. This diversity of models and approaches means that the applied AI layer can tune models to their workflows and route intelligence appropriately. Huge win for all.

我们竟然能从 open models 获得这种级别的性能，真的非常惊人。祝贺 Kimi team 做到这一点。每次我们降低 frontier intelligence（前沿智能）的成本，企业能够承接的 use-case（用例）就会随之增加。有大量企业非常想部署的 workflow（工作流），目前唯一的限制其实就是 token 的成本。尤其对 startup ecosystem（初创生态）来说，open labs 和 closed labs 的联合突破，让大量价值能够沉淀到这一层：它可以利用多种不同模型，为客户完成完整任务。这种模型与方法的多样性，意味着 applied AI layer（应用型 AI 层）可以针对自身 workflow 调优模型，并把智能合理路由到合适的位置。对所有人来说，这都是一次巨大的胜利。

Here’s another awesome use-case for what we can now do with our unstructured data because of AI agents. Box now works with Databricks so you can take structured data from enterprise content (like contracts, financial document, supply chain data) and connect that data into Databricks. This means that I can now query large document datasets without moving or reprocessing that content. And you can connect the data with any other system, like your ERP data, CRM, or product analytics. This opens up a ton of new use-cases for enterprise content. All possible because of headless software and agents.

这是另一个很棒的用例，展示了因为 AI agents（AI 智能体），我们现在能够如何利用非结构化数据。Box 现在已经与 Databricks 打通，因此你可以从 enterprise content（企业内容）中提取 structured data（结构化数据）——比如合同、财务文件、供应链数据——并把这些数据连接到 Databricks。这意味着我现在可以在不迁移或重新处理这些内容的情况下，直接查询大型文档数据集。而且你还可以把这些数据连接到任何其他系统，比如你的 ERP 数据、CRM，或产品分析系统。这为 enterprise content 打开了大量新的 use-case。所有这些之所以成为可能，都要归功于 headless software 和 agents。

Hosted a dinner last night with a group of IT leaders of large enterprises around agent adoption in the enterprise. Some quick notes: * Change management remains one of the biggest topics for driving workflow transformation. Still most processes need to be upgraded to modern operating models to work with agents, which is a mix of technology, data, and human process change. Lots of emphasis on getting data (structured and unstructured) into a setup that can work with agents properly. * IT teams are finding increasing success embedding full engineers into the business functions (essentially internal FDE) that go and implement agents into the internal workflows. There’s so much technical work to be done to make agents successful, that they can accelerate months or quarters of failed experiments by having someone technical in the workflow early. * Consensus that the tech function is becoming more important than ever. It’s clear that the business could only expect automation to affect a minority of the business before (e.g. ERP) but now it can impact all of knowledge work. This means IT is becoming a more central role to the workflows across the company. * Workflows are cross functional, and getting agents to work cross functionally is a complicated data modeling and permissions issue. Single users don’t have access to this. Which means you need to have agentic systems take on their own roles and have their own privileges, which is non-trivial given agents can’t keep things secure on their own. * Huge variance in budgets between coding work and the rest of knowledge work. Some companies had a $1,000 a month budget for developers, and others had much higher amounts (like $5,000) that were merely triggers to notify the team vs. block them. Far smaller budgets for non-coding work at the moment. * More companies are building their own multimodel systems for routing workloads by task to frontier and lower cost models. Lots of energy around open weights models, but still more in experimentation instead of at scale usage (some companies can’t due to perceived Chinese issue). * Clear sense that all enterprise software must be headless in the future. Relief that they don’t need to train employees on hundreds of different apps. However, clear frustration with the traditional vendors that don’t play extremely nice (technically or cost wise) with agents in a headless fashion. Huge warning for existing software vendors. * Mythos or mythos level-models are finding more and more sophsiticated security risks. The chaining together of vulnerabilities is what’s novel right now, and companies are coming up with long backlogs of what they need to go patch quickly. Even more discussed, but just a few of the hottest topics.

昨晚我和一群大型企业的 IT 负责人共进晚餐，话题围绕企业中的 agent 采用展开。记几点速记：* 变革管理仍然是推动工作流转型的最大议题之一。大多数流程依然需要升级到现代运营模型，才能与 agent 配合工作，这涉及技术、数据以及人工流程变更的组合。大家非常强调要把数据（结构化和非结构化）整理进一种能够正确支持 agent 的配置中。* IT 团队越来越多地发现，把完整的工程师直接嵌入业务职能部门（本质上是内部 FDE）去在内部工作流中实施 agent，效果正在变好。要让 agent 成功落地，还有大量技术工作要做；如果能在工作流早期就有技术人员参与，就能把原本会失败数月甚至数个季度的实验大幅提速。* 大家的共识是，技术职能部门正变得前所未有地重要。很明显，以前企业只能预期自动化影响业务中的少数部分（例如 ERP），但现在它可以影响全部知识型工作。这意味着 IT 正在成为贯穿全公司的工作流中更核心的角色。* 工作流是跨职能的，而让 agent 跨职能运作，会涉及复杂的数据建模和权限问题。单个用户并不拥有这种访问权限。这意味着你需要让 agentic systems 承担它们自己的角色，并拥有它们自己的权限，而这并不简单，因为 agent 本身无法独立保证安全。* 编码工作与其他知识型工作之间的预算差异极大。有些公司给开发者的月预算是 1,000 美元，而另一些公司则高得多（比如 5,000 美元），而且这些额度更多只是触发通知团队，而不是直接阻止他们。当前非编码工作的预算则小得多。* 越来越多的公司正在构建自己的 multimodel systems，用于按任务把工作负载路由到 frontier 模型和低成本模型。大家对 open weights models 很有热情，但目前仍更多停留在实验阶段，而非大规模使用（有些公司因为感知中的 Chinese issue 无法推进）。* 很明显，未来所有企业软件都必须是 headless 的。大家都松了一口气，因为不需要再培训员工去使用上百个不同的 app。不过，对那些在 headless 方式下无法与 agent 非常顺畅协作的传统厂商，无论在技术上还是成本上，大家也表达了明确的不满。这对现有软件供应商是一个巨大的警告。* Mythos 或 mythos level-models 正在发现越来越复杂的安全风险。当前的新情况在于，漏洞被串联起来利用；各家公司也因此列出了很长的待修补清单，需要尽快打补丁。讨论的内容其实更多，这里只是其中几个最热的话题。

Fantastic to see more open weights innovation happening right now, especially coming from a US Lab. The future of AI is going to be a mix of frontier intelligence that you can use as an orchestrator combined with either lower cost or tuned models for your workhorse tasks.

很高兴看到现在有更多 open weights 创新正在发生，尤其是来自一家 US Lab。AI 的未来将会是这样一种组合：用 frontier intelligence 作为 orchestrator（编排器），再结合更低成本或经过调优的模型，去承担你的主力工作任务。

One of the many properties that code has that makes it highly amenable to agents is that you can more or less quickly test it. You can either go see if the application works manually, or you can actually run a test on what you built. Most other areas of work don’t have this benefit. You only get the testing when the final product hits the real world in some capacity - a stock trade is executed, a contract is negotiated, a sales pitch is delivered, and so on. There’s probably going to be a whole new set of opportunities for how we begin to test the rest of work in this way. Ultimately it will mean more agents being layered into workflows. It also means we need much better evals on most of our workflows. Most work today in enterprises doesn’t have an associated eval to know if something broke or improved with a model, prompt, or system change. The enterprises that are able to eval their knowledge work the best also stand to gain the most from AI. Will become a critical aspect of agent adoption over time.

代码之所以有很多特性让它对 agent 特别友好，其中一点是你基本上可以相对快速地测试它。你既可以手动去看应用是否能正常工作，也可以直接对你构建的内容运行测试。其他大多数工作领域都没有这种优势。你往往只有在最终产品以某种形式进入现实世界时，才会得到“测试”——比如一笔股票交易被执行、一份合同被谈成、一次销售陈述被完成，等等。很可能会出现一整套新的机会，让我们开始以这种方式测试其余类型的工作。归根结底，这意味着会有更多 agent 被叠加进工作流中。这也意味着，我们需要对大多数工作流做得更好的 evals（评估）。如今企业中的大多数工作，并没有配套的 eval 来判断某个 model、prompt 或 system 的改动，究竟让事情出错了还是改进了。那些最能评估自身知识型工作的企业，也最有可能从 AI 中获得最大收益。随着时间推移，这将成为 agent 采用过程中的一个关键方面。

Thoughtful proposal for a standards body for AI. This is distinct from a regulatory agency, and would certainly allow for much faster improvement of standards and collaboration with the industry. What you definitely don’t want is for AI progress to start to move at the speed of that the government classically operates at. If that happens then we can essentially guarantee progress begins to stall, and worse, America likely just loses the AI race. This framework threads the needle mostly. The only challenge is still that even those in industry don’t agree on the same safety risks in AI. But if you can get alignment there, this looks better than most other proposals.

这是一个关于为 AI 设立标准机构的深思熟虑的提案。它不同于监管机构，而且肯定能让标准的改进以及与行业的协作推进得快得多。你绝对不希望 AI 的进步开始以政府传统上的运作速度前进。如果那样的事情发生，我们几乎可以肯定进展会开始停滞，更糟的是，America 很可能就会输掉这场 AI 竞赛。这个框架在很大程度上拿捏住了平衡。唯一的挑战仍然是，即便是行业内部的人，对 AI 中哪些安全风险最重要也并没有共识。但如果你能在这一点上达成 alignment（对齐），那么这看起来比大多数其他提案都更好。

A few thoughts on what we will see in AI structurally for the foreseeable future: * Frontier intelligence continues unabated and pushes the industry forward continuously. The top labs will continue to buy the best and the most data, build the most compute, be at the forefront of improved training breakthroughs, and so on. A few different approaches stratify the market on pricing and capability, but overall competitive pressure brings down pricing on a per task basis. That said, we just ask more from the models over time - as one thing gets cheaper, we just use more - so frontier spend and use remains robust. * Open weights rapidly absorbs frontier breakthroughs (and drives other breakthrough directions given the constraints), offering both lower cost intelligence and the ability to be post trained for specific workflows and domains. This creates a healthy counter balance to the frontier as you can run models “at cost” on a hyperscaler at any time, and tune models just for your tasks. * The Applied AI layer has a huge opportunity to combine frontier intelligence with open or cheap closed models to orchestrate workflows in any given domain. Due to evals, deep domain context, being trusted with enterprise data and workflows, this layer can maximize performance and cost combination. The applied AI layer will also often have their own RLed models especially for high volume, predictable tasks in their systems. * Individual enterprises will generally focus on their enterprise context, making sure they can get any AI system the right data and information to work with, in a continuously improving way. Some will go off and train their own models for specific areas of work (large banks, pharma, etc.) where they can get real alpha from doing so given the many tradeoffs, but most will spend energy on making sure they can get all of the gains from AI breakthroughs on their data and workflows. Net net: even though some of this gets framed as zero sum, there’s just a ton of opportunity for all layers of the stack and approaches.

关于在可预见的未来，我们会在 AI 领域看到哪些结构性变化，我有几点看法：* 前沿智能（frontier intelligence）会持续不停地发展，并不断推动整个行业向前。顶级实验室会继续购买最好的、最多的数据，构建最强的 compute（算力），站在训练改进突破的最前沿，等等。少数几种不同的方法会在定价和能力上让市场分层，但总体来说，竞争压力会把单位任务的价格拉低。话虽如此，随着时间推移，我们对模型的要求也会越来越多——某件事一旦变便宜了，我们就会用得更多——因此前沿层的投入和使用仍会保持强劲。* 开放权重（open weights）会迅速吸收前沿突破（同时也会在这些约束下推动其他方向的突破），既提供更低成本的智能，也提供针对特定 workflow（工作流）和领域进行 post-train（后训练）的能力。这会对前沿层形成一种健康的制衡，因为你随时都可以在 hyperscaler（超大规模云厂商）上以“成本价”运行模型，并且只为你的任务去调优模型。* Applied AI（应用层 AI）这一层有巨大的机会，把前沿智能与开放模型或廉价闭源模型结合起来，在任意领域中编排 workflow。由于有 evals（评测）、深度领域上下文、以及对企业数据和 workflow 的可信接入能力，这一层可以把性能与成本的组合优化到最大。Applied AI 层通常也会拥有自己的 RLed models，尤其是用于其系统中高频、可预测的任务。* 单个企业通常会聚焦于自身的企业上下文，确保它们能以持续改进的方式，为任何 AI 系统提供正确的数据和信息来使用。有些企业会进一步为特定工作领域训练自己的模型（大型银行、pharma 等），因为在权衡多种取舍之后，这样做能带来真实 alpha（超额收益）；但大多数企业会把精力花在确保自己能把 AI 突破带来的收益，充分应用到自己的数据和 workflow 上。归根结底：虽然其中一些事情常被描述成零和竞争，但实际上，栈中的各个层次和不同路径都存在大量机会。

Here’s a great post on driving down costs, while maintaining high performance, with frontier intelligence as a manager and lower cost models for the workhorse tasks. This will be the template for what model routing looks like in the future. “We started this experiment expecting to measure how much Fable’s 2x premium would increase cost. We were surprised to find that Fable’s effective delegation actually decreased cost overall. It specified constraints and outcomes instead of spelling out the implementation, gave feedback instead of making fixes itself, and in most cases never touched the code at all. These are the habits of a good manager.” The industry is increasingly figuring out what it looks like to mix models together to be able to get targeted performance levels and optimal cost structures. Of course, the only way to get this is to have a deep understanding of the business problem you’re trying to solve and how to effectively route work to different models. If you’re in the applied layer - whether it’s customer support, legal, finance, or coding - this is how your harness will become a core area of differentiation.

这里有一篇很棒的文章，讲的是如何在保持高性能的同时降低成本：用前沿智能担任 manager（管理者），把具体的重活交给低成本模型。这会成为未来 model routing（模型路由）的模板。 “我们一开始做这个实验，是想衡量 Fable 的 2x 溢价会把成本提高多少。结果我们惊讶地发现，Fable 有效的 delegation（委派）实际上反而降低了整体成本。它规定的是约束和结果，而不是把实现细节逐条写出来；它提供反馈，而不是亲自去修补问题；而且在大多数情况下，它根本不碰代码。这些就是一个好 manager 的习惯。” 整个行业正越来越清楚，如何把不同模型混合起来，以获得目标性能水平和最优成本结构。当然，想做到这一点，唯一的方法就是对你要解决的业务问题有深刻理解，并知道如何把工作有效路由给不同模型。如果你处在应用层——无论是 customer support、legal、finance，还是 coding——这就是你的 harness（编排/执行框架）会成为核心差异化领域的方式。

The biggest challenge right now with the topic of every enterprise having their own model is that your most valuable information and insights are not only always changing, but they’re often your most sensitive information. Your most sensitive information can’t be packed into a model usually because it contains data that not everyone gets to have access to, and you can’t keep your security layer inside the model or an agent. I think there are going to be 100X more use cases for custom trained models, especially inside of domain-focused products, but training a model per enterprise is going to be a lot harder than it looks.

现在谈“每家企业都有自己的模型”这个话题时，最大的挑战在于：你最有价值的信息和洞察不仅一直在变化，而且它们往往也是你最敏感的信息。你最敏感的信息通常无法被打包进一个模型里，因为其中包含并非所有人都能访问的数据，而且你也无法把你的 security layer（安全层）保留在模型或 agent 里面。我认为，custom trained models（定制训练模型）的用例会多出 100 倍，尤其是在聚焦特定领域的产品内部；但为每一家企业单独训练一个模型，会比看上去难得多。

One of the key architectural questions of the 21st century in business will be how you maximize your corporate IP in the form of decisions, insights, workflow patterns, and best practices in a world where so much intelligence is packed into AI models. One might think these questions could just get bitter lessoned out of existence, but in reality they become even more germane as intelligence becomes more powerful. In a world where any firm also has access to frontier intelligence, understanding how you leverage it uniquely becomes a critical question. That’s why so much value is left to be created between the enterprise and the underlying AI itself. Having evals for your workflows, ensuring that you can route models from different tiers of intelligence, capturing traces in a way that improve your own workflows, and making sure the value of your information compounds as AI gets better all become critical considerations. Which is also why there’s so much opportunity right now in the applied AI layer. The companies that help figure this out for other enterprises will be in the best position to win the next enterprise workloads.

在商业领域，21 世纪一个关键的架构性问题将是：在如此多 intelligence（智能）被封装进 AI models（AI 模型）的世界里，你如何最大化企业自身的 IP，具体体现为决策、洞察、工作流模式和最佳实践。人们可能会以为，这些问题最终会被 bitter lesson 式地“淘汰”掉，但现实是，随着 intelligence 变得更强大，它们反而变得更加切题。在一个任何公司都能获得 frontier intelligence（前沿智能）的世界里，理解你如何以独特方式加以利用，就成了一个关键问题。这也正是为什么，在 enterprise（企业）与底层 AI 本身之间，仍有大量价值有待创造。为你的 workflows（工作流）建立 evals（评测），确保你能够在不同 intelligence 层级的 models（模型）之间进行路由，以能够改进自身 workflows 的方式捕获 traces（追踪记录），并确保随着 AI 变得更强，你的信息价值也能持续复利增长——这些都会成为关键考量。这也是为什么，applied AI（应用层 AI）现在存在如此多的机会。那些能够帮助其他企业解决这些问题的公司，将最有可能赢得下一波 enterprise workloads（企业工作负载）。

The job that AI was supposed to replace is experiencing the opposite of the expected outcome. Software job postings are outpacing other fields. Why is that? If you lower the cost of production of something that has lots of use cases, people want more things produced. We’ve seen this play out in the industrial world constantly, and now we’re finally seeing it in knowledge work. Because software now is much lower to cost per unit, people want way more of it. So we start to use software for all new things and people and companies light up more software projects than ever before. But because the job itself is not fully automated (and likely won’t be for as far out as we can see), you still need people that understand these systems to maintain the code, decide what to build, run it over the long run, update it, and more. That all requires people to do work. The same thing is going to happen in many other fields as well as we bring down the cost of production of previously extremely scarce areas of work. Agents will cause more abundance than replacement.

AI 原本被认为会取代的那类工作，如今却出现了与预期相反的结果。Software 职位发布的增长速度正在超过其他领域。为什么会这样？如果你降低了某种拥有大量 use case 的事物的生产成本，人们就会想要生产更多这种东西。我们在工业世界里已经反复看到这种情况上演，而现在我们终于也在知识工作中看到了同样的现象。因为如今 software 的单位成本低得多，人们就会想要更多得多的软件。于是我们开始把 software 用到各种新的事物上，个人和公司也以前所未有的规模启动更多 software 项目。但由于这项工作本身并没有被完全自动化（而且在我们可预见的未来里，很可能也不会），你仍然需要理解这些系统的人来维护代码、决定要构建什么、长期运行它、更新它等等。这一切都需要人来完成工作。随着我们降低那些过去极度稀缺的工作领域的生产成本，许多其他领域也会发生同样的事情。agent 将带来更多的丰富供给，而不是取代。

Great post on some of the dynamics to think through for the future competitive advantage in world when AI models are shared amongst firms and packing so much for the intelligence of that industry. This is going to become a core question for companies and the economy broadly over the next decade and beyond. If AI is trained on the best datasets in every single industry - like law, finance, healthcare, or life sciences - then how do you compete and differentiate in the future? This is a great open question that I don’t think is perfectly knowable right now because of how fast AI progress is happening. But ultimately it stands to reason that if intelligence is abundant and broadly available to anyone in a field, then the companies that effectively use it the best and against a set of data and knowledge that grows in value over time, will be in a strong position. There’s a huge reinforcing loop between the intelligence from models, a company’s own data, the connection of that data and AI in their workflows, and how employees ultimately interact with that system to create value. There’s no obvious point where this will become uniform across all companies in a vertical because each company will approach this in a different way, just as they already do with their talent and workflows. If anything, there will be compounding returns to those that do this best that accelerate their advantage over time. Overall, super interesting question to see how this plays out over time.

这篇文章很好地讨论了一个值得深入思考的动态：在一个 AI model 被各家公司共享、并且承载了该行业大量 intelligence（智能能力）的世界里，未来的竞争优势会来自哪里。这将在未来十年乃至更长时间里，成为企业乃至整个经济体的一个核心问题。如果 AI 是基于每一个行业里最优质的数据集训练出来的——比如 law、finance、healthcare 或 life sciences——那么未来你要如何竞争、如何实现差异化？这是一个非常好的开放性问题，而考虑到 AI 进展的速度之快，我认为现在还不可能把它完全看清。但归根结底，可以合理推断的是：如果 intelligence 变得充裕，并且某个领域里的任何人都能广泛获得它，那么那些最有效地使用它、并且将其与一套会随时间不断增值的数据和知识结合起来的公司，将处于非常有利的位置。model 提供的 intelligence、公司自身的数据、这些数据与 AI 在工作流中的连接方式，以及员工最终如何与这个系统交互来创造价值，这几者之间存在一个巨大的强化循环。并没有一个明显的时间点会让同一垂直领域中的所有公司都变得趋同，因为每家公司都会以不同方式处理这件事，就像它们今天在人才和工作流上的做法本就不同一样。真要说的话，那些把这件事做得最好的人，反而会获得复利式回报，随着时间推移进一步加速其优势积累。总的来说，这是一个非常有意思的问题，值得观察它会如何随时间演变。

GPT-5.6 is now out. We've been evaluating the model family on the Box AI Complex Work eval, which tests the model with the Box AI Agent on a variety of extremely hard tasks using enterprise document sets. Sol is a big step up from GPT-5.5, especially on complex data-oriented tasks that require deep reasoning and analysis, and the gains concentrate exactly where enterprise work is hardest. Here are a few examples that we saw across our tests: * Financial Services (76% vs 71%): On a multi-year projection, Sol anchored to the correct opening balance sheet date rather than assuming a clean January 1 start, then carried revenue, earnings, and interest through to the right figures year over year where one early wrong assumption compounds through every downstream cell. * Healthcare (58% vs 46%): On a critical-care case review, Sol identified the correct diagnosis and intervention and avoided the dangerous misstep of ordering imaging before the time-critical procedure, a trap GPT-5.5 walked into. * Public Sector (74% vs 63%): Handed a class's raw gradebook and a new grading directive, Sol mapped each assignment to the right weight bucket, treating homework as zero-weight practice per the directive. It recomputed every student's grade to within a tenth of a percent, where GPT-5.5 drifted partway through. * Life Sciences (60% vs 51%): Across four separate compound datasets, Sol intersected the ranked target lists exactly (case-sensitive, no shortcuts) to find the biological targets common to all four, catching the shared targets GPT-5.5 missed. Sol reasons from the source definitions and checks the documents rather than taking them at face value and it's most reliable exactly where the numbers drive real decisions. This will be huge for enterprise agents using unstructured enterprise data. GPT-5.6 will be available to customers shortly within the Box AI Studio for building custom agents with.

GPT-5.6 现已发布。我们一直在用 Box AI Complex Work eval 评估这一 model family；该评测会结合 Box AI Agent，使用企业文档集对 model 在各种极其困难任务上的表现进行测试。Sol 相比 GPT-5.5 是一次明显升级，尤其是在那些需要深度推理和分析的复杂、数据导向型任务上；而这些提升恰恰集中在企业工作最困难的地方。以下是我们在测试中看到的一些例子：* Financial Services（76% vs 71%）：在一个跨多年的预测任务中，Sol 锚定了正确的期初资产负债表日期，而不是假设从 1 月 1 日这个“干净”的起点开始；随后它将 revenue、earnings 和 interest 逐年正确传导到对应数值上，而在这类任务中，一个早期错误假设会在后续每一个单元格中不断累积。* Healthcare（58% vs 46%）：在一个重症病例复核任务中，Sol 识别出了正确的诊断和干预方案，并避开了一个危险误区——在时间高度敏感的操作之前先安排 imaging，而 GPT-5.5 正是掉进了这个陷阱。* Public Sector（74% vs 63%）：在拿到一个班级的原始成绩册和一项新的评分指令后，Sol 将每项作业映射到了正确的权重类别，并按照指令将 homework 视为权重为零的练习。它将每个学生的成绩重新计算到了百分之一百分点以内，而 GPT-5.5 在中途开始出现偏差。* Life Sciences（60% vs 51%）：在四组彼此独立的化合物数据集中，Sol 精确求出了排序后目标列表的交集（区分大小写，不走捷径），从而找到了四组数据共同的 biological targets，而这些共享目标是 GPT-5.5 漏掉的。Sol 是根据源定义进行推理，并会核查文档，而不是想当然地接受表面信息；并且它恰恰在那些数字会驱动真实决策的地方最为可靠。这对使用非结构化企业数据的企业 agent 来说将是巨大的提升。GPT-5.6 很快就会在 Box AI Studio 中向客户提供，用于构建自定义 agent。

It should be 100% obvious that there will soon be mythos level models on cyber security that are open and available to anyone. As a byproduct of this, alternative tech stacks will emerge that also drive more economic value and control away from the US’s tech stack. This is what should be considered when thinking through the gate keeping you want to have in AI. If advanced models will become open and available regardless, then by not allowing the release of models you’re neither more secure nor better off strategically. So much of the regulatory approach to AI has to assume China can’t catch up, when all current evidence suggests they can and are. And further, hard to imagine a higher priority than winning in AI for China; so you’re basically betting against their long term ingenuity, talent and motivation. Seems like a bad bet. So your options are either to create gates around your best models, which means you’re asymmetrically disadvantaging yourself, or you work to ensure you’re always at the frontier and driving the future architectures of AI.

很显然，很快就会出现达到 mythos 级别、聚焦 cyber security 的 models，而且它们将是 open、任何人都可以使用的。作为这一点的副产品，还会涌现出 alternative tech stacks，它们也会把更多经济价值和控制权从 US 的 tech stack 中转移出去。这才是在思考你希望在 AI 中设置哪些 gate keeping（准入限制）时应该考虑的问题。如果 advanced models 无论如何都会变得 open 且可获得，那么不允许发布 models 并不会让你更安全，也不会让你在战略上处于更有利的位置。当前针对 AI 的许多监管思路，都必须建立在 China 追不上来的假设上；但现有的所有证据都表明，他们不仅能追上，而且正在追上。更进一步说，很难想象对 China 而言还有什么比赢得 AI 更高的优先级；所以你基本上是在赌他们长期的创造力、人才和动力不会兑现。这看起来是个糟糕的赌注。所以你的选择，要么是围绕你最好的 models 设立门槛，而这意味着你在不对称地让自己处于劣势；要么就是努力确保自己始终站在 frontier（前沿），并主导 AI 的未来 architectures（架构）。

Some good best practices here on AI token cost optimization. None of these happens though without a deep understanding of the underlying work being done in a non-abstract way. The ultimate implication is that a layer between the work itself and the underlying intelligence needs to deeply understand your workflows, context, and business process. Now, each individual company doing this on their own is unlikely to be effective at scale, so as a consequence, this is effectively the playbook for any applied AI company right now. By evaling the models for the applied use cases, deeply understanding the domain, having tuned UX and features for the use case, and having the ability to support adoption and change (via FDEs), allow this layer to add a ton of value. And as a result, enterprises get higher ROI because you actually can get *more* intelligence per dollar by having optimal architecture and workflows. There will be many horizontal and vertical versions of this approach. Huge opportunity right now.

这里总结了一些关于 AI token 成本优化的很好的最佳实践。不过，如果不能以一种非抽象的方式深刻理解底层正在进行的工作，这些做法都无从实现。其最终含义是：位于实际工作与底层 intelligence 之间的那一层，必须深度理解你的工作流、上下文和业务流程。现在，如果每家公司都各自单独来做这件事，规模化后不太可能有效，因此，这实际上就是当下任何 applied AI 公司的行动手册。通过针对实际应用场景对模型进行 eval，深度理解领域知识，为该 use case 打磨好 UX 和功能，并具备支持采用与变更的能力（通过 FDEs），这一层就能创造大量价值。结果是，企业能够获得更高的 ROI，因为通过最优的架构和工作流，你实际上可以用每一美元换来*更多*的 intelligence。这种方法会有很多 horizontal 和 vertical 的版本。现在是巨大的机会。

Step one complete

第一步完成

GPT-5.6 is real and looks very strong. Going to be very strong for knowledge worker tasks that require heavy tool use and long running agents doing work. We're not hitting any walls in AI progress right now.

GPT-5.6 真实存在，而且看起来非常强。对于那些需要大量使用工具、并由长时间运行的 agent（智能体）执行工作的知识工作者任务来说，它会非常强。我们目前并没有在 AI 进展上碰到任何墙。

AI regulation is far less simple than it looks. It’s prisoners dilemma at insane scale. In theory if all leading AI labs globally agreed to the same process of review and slow down, then we’d get frontier intelligence at similar rates and it diffuses relatively evenly. If the US remains at the frontier at all times, and has heavy regulation on the release of intelligence then we end up with an economic and geopolitical edge because we can control who has access to frontier intelligence. If we delay model releases, however, and another player - specifically China - doesn’t slow down and has equally strong models (not now but soon?) then our delays end up advantaging their models and eventually their tech stack. Now, the US could ban these models, but that actually only puts the US at a steeper disadvantage because other countries won’t have those bans. Then, from a relative competitiveness standpoint the US has now fallen behind even though it started in front. So, none of this is as simple as it looks. At some point it’s a simple bet of can closed models remain at the frontier in perpetuity or is there a risk of any other player or market catching up or just not falling behind.

AI 监管远没有看上去那么简单。这是在疯狂规模上的 prisoners dilemma（囚徒困境）。理论上，如果全球所有领先的 AI lab 都同意采用同样的审查流程并一起放慢速度，那么我们获得 frontier intelligence（前沿智能）的速度会大致相近，而且它会相对均匀地扩散。如果 US 始终处于前沿，并且对 intelligence 的发布实行严格监管，那么我们最终就会拥有一种经济和地缘政治上的优势，因为我们可以控制谁能接触到 frontier intelligence。然而，如果我们推迟 model 发布，而另一个玩家——尤其是 China——没有放慢速度，并且拥有同样强的 model（不是现在，但很快会？），那么我们的延迟反而会让他们的 model，最终连同他们的 tech stack（技术栈）一起受益。现在，US 当然可以禁止这些 model，但那实际上只会让 US 处于更大的劣势，因为其他国家不会有这些禁令。这样一来，从相对竞争力的角度看，US 就会在原本领先的情况下反而落后。所以，这一切都不像表面看起来那么简单。归根到底，这其实是在赌一件事：closed models（封闭模型）能否永远保持在前沿，还是说存在任何其他玩家或市场追上来，哪怕只是没有继续落后的风险。

We now have de facto AI regulation. It’s not obvious why from here on out models that have certain levels of capability or are trained on certain compute sizes won’t have to be reviewed by the government before release. Realistically, as AI models became more and more powerful this was going to be inevitable (I think it’s too early, but here we are). So now it’s mostly just interesting to think about the implications and scenarios from here. A few would be: * America gets to control who gets access to frontier intelligence and when. This generally works as long as we remain at the frontier at all times and don’t have a risk of being surpassed. At the moment we have a clear lead in frontier intelligence so this is a good bet, but lots of motivated parties would love to change that. * This likely creates backlog of AI releases which means that we will see less rapid fire back and forth jumps in model progress. Bull/fine case is that we just get bigger step functions per release at a slower rate and we end up at the same point we would have. Bear case is those incremental smaller jumps were necessary for the continued flywheel of innovation. * Other countries likely have even more incentive to at least hedge their bets with sovereign AI strategies so aren’t dependent on access to US AI all times. Previously this was relatively moot because the alternative wasn’t good enough, but that could change out of necessity and what we’re seeing in China. * Open weights obviously a big winner here as it becomes what likely sovereign AI gets built out on, and what (for now) can still be released to the market without the same controls. One interesting question would be how regulation eventually extends to open models, which would have its own set of long term consequences. Anyway some big updates to everyone’s mental models of AI regulation as a result of the capabilities we’re now seeing in AI. Wild times.

我们现在已经有了 de facto（事实上的）AI 监管。从现在开始，为什么那些达到某些能力水平、或者在某些 compute（算力）规模上训练出来的 model，不需要在发布前先接受政府审查，其实并不明显。现实地说，随着 AI model 变得越来越强大，这件事本来就是不可避免的（我认为现在还太早了，但现实已经如此）。所以现在更有意思的，主要是思考从这里往后会产生哪些影响和场景。几个可能的情况是：* America 将能够控制谁可以在何时获得 frontier intelligence 的访问权。只要我们始终保持在前沿、并且没有被超越的风险，这通常就行得通。就目前而言，我们在 frontier intelligence 上有明确领先，所以这是一笔不错的赌注，但有很多高度积极的参与方都很想改变这一点。* 这很可能会造成 AI 发布积压，这意味着我们会看到更少那种 rapid fire（高频、密集）的来回式 model 进步跳跃。Bull/fine case（乐观情形）是，每次发布只是以更慢的频率带来更大的 step function（阶跃式提升），最终我们仍会到达原本会到达的位置。Bear case（悲观情形）则是，那些更小、更渐进的提升其实是维持创新 flywheel（飞轮）持续运转所必需的。* 其他国家很可能会更有动力，至少通过 sovereign AI（主权 AI）战略来对冲风险，以免自己始终依赖获得 US AI 的访问权。以前这点相对无关紧要，因为替代方案还不够好，但出于必要性，以及从我们在 China 看到的情况来看，这可能会改变。* open weights（开放权重）显然会是这里的大赢家，因为它很可能会成为 sovereign AI 得以构建的基础，也是在目前仍然可以不受同等控制地向市场发布的东西。一个有意思的问题是，监管最终会如何延伸到 open models（开放模型）上，而那又会带来其自身一整套长期后果。总之，由于我们现在在 AI 中看到的这些能力，大家关于 AI 监管的 mental models（心智模型）都需要进行重大更新。真是疯狂的时代。

There are some subtleties in this launch that are very important in practice. This isn’t just you interacting with Claude in a 1:1 format via Slack. In this case, Claude acts as a coworker that any user can tap into in a shared way. We’ve already seen some agentic coding systems start to adopt this pattern (as well as OpenClaw and Hermes), and doing it for general purpose knowledge work continues to push the idea forward. As a result, what this means is that this agentic coworker needs its own set of resources, access to tools, and data to work with. This is not the same as you giving it access to your personal resources and tools, because the agent then could accidentally then share those out with anyone. The agent needs to instead be like any other user in the system, and you need to be thoughtful about what it should have access to, and make sure its information that is safe to share with that group. When you can pull that off, it’s quite powerful. For instance, by connecting to Claude Tag to Box, you could have Claude access corporate sales materials for questions in sales conversations or generating RFPs, brand guidelines and marketing assets for campaign creation, product roadmap materials and product documentation for coding agents to use, contracts that anyone in the legal team can access, and more. But this is just the Box example. You can equally have it access your product or customer analytics data, CRM information, codebase, and other resources and agents that would make sense to work on in a collective manner. It’s awesome to see continued innovation in what the future of work may begin to look like with agents.

这次发布里有一些细微但在实践中非常重要之处。这并不只是你通过 Slack 以 1:1 的形式与 Claude 互动。在这里，Claude 充当的是一个任何用户都可以以共享方式调用的同事。我们已经看到一些 agentic coding（智能体式编码）系统开始采用这种模式（OpenClaw 和 Hermes 也是如此），而把它用于通用型知识工作，则进一步推进了这一理念。因此，这意味着这个 agentic coworker（智能体同事）需要自己的一套资源、工具访问权限和可用数据。这与你把它接入你个人的资源和工具并不相同，因为那样一来，这个 agent 就可能会意外把这些内容共享给任何人。相反，这个 agent 需要像系统中的其他普通用户一样存在，而你需要认真考虑它应当拥有什么访问权限，并确保它所接触的信息对于那个群体来说是可以安全共享的。只要你能把这件事做好，它的威力就相当强大。比如，通过把 Claude Tag 连接到 Box，你可以让 Claude 访问企业销售材料，用于回答销售对话中的问题或生成 RFPs，访问品牌指南和营销素材来创建 campaign，访问产品 roadmap 材料和产品文档供 coding agents 使用，访问法务团队中任何人都可查看的合同，等等。但这还只是 Box 这个例子。你同样也可以让它访问你的产品或客户分析数据、CRM 信息、codebase，以及其他适合集体协作处理的资源和 agents。很高兴看到，围绕 agents，关于未来工作形态可能会是什么样子的创新仍在持续推进。

Almost all AI model and agent progress is downstream from evals. Open weights post training for specific domains comes down to evals. Agent improvements in the applied AI layer is all about evals. Agentic enterprise deployments that actually can augment work is all about evals. It’s all evals. This will become a core competency of any enterprise in the future. The companies that are able to best understand their own (and/or customers) workflows and how well agents participate in that work will be in the best position to actually drive real automation.

几乎所有 AI model 和 agent 的进展，都是由 evals 驱动的下游结果。针对特定领域的 open weights 后训练，归根结底也取决于 evals。应用层 AI 中 agent 的改进，核心也全是 evals。真正能够增强工作的 agentic enterprise 部署，同样全都离不开 evals。一切都是 evals。未来，这将成为任何 enterprise 的核心能力。那些最能理解自身（和/或客户）工作流，以及 agent 在这些工作中参与效果的公司，将最有机会真正推动实际的自动化。

We heard that HTML is a big deal again. You can now preview, edit, manage versions, and securely share any HTML based content on Box. Great for being able to work with any agent produced content immediately.

我们听说 HTML 又重新变得非常重要了。现在，你可以在 Box 上预览、编辑、管理版本，并安全地分享任何基于 HTML 的内容。这对于立即处理任何由 agent 生成的内容非常有帮助。

Another new idea to push the state of AI architectures forward. Sakana released a model that effectively uses a mixture of models to get work done. You get a single API but then the work gets farmed out the model that best performs the task. “Fugu manages model selection, delegation, verification, and synthesis automatically. It solves tasks directly when that is enough, or coordinates a team of expert models when a problem calls for more. The complexity of a multi-agent system never reaches your code.” This is generally how applied AI products are building their agent harnesses at this point, but the idea of making this an LLM that any developer can interact with is also a great idea. As we get more innovation with both frontier closed and OSS models, there’s going to be a ton of value produced for the layer that can route the best.

又一个推动 AI 架构发展的新想法。Sakana 发布了一个模型，它通过有效使用 mixture of models（模型混合）来完成工作。你得到的是一个单一的 API，但随后工作会被分发给最擅长该任务的模型。“Fugu 会自动管理模型选择、任务委派、验证和综合。当直接解决任务已经足够时，它就直接处理；而当问题需要更多能力时，它会协调一支由专家模型组成的团队。多 agent 系统的复杂性永远不会传达到你的代码中。” 目前，应用型 AI 产品大体上都是这样构建它们的 agent harness（agent 编排/承载层）的，但把这件事做成一个任何开发者都能交互的 LLM，同样是个很棒的想法。随着 frontier closed 和 OSS 模型两端都持续创新，能够把请求路由给最佳模型的那一层将会创造出大量价值。

Agents will use software 100X more than people. When that happens, theres a huge need for guardrails on what the agents are doing so they don’t leak data or change the wrong information, authoritative sources of truth for them to work with, logging and auditing of what they’re doing, the ability to collaborate with people through these systems, and more. A simple query on any given agentic task could pull in more data than a user touches in a month. As a result, there are lots of categories of software that when it goes headless that the usage and value go up substantially. Agents will end up using our CRM data, documents and corporate knowledge, analytics data, and other information far more than people ever did. The platforms that can move toward the model of powering these headless interactions, and have a business model and technology strategy to support this, will be in the best position in the future.

Agents 使用软件的频率将会是人的 100 倍。当这种情况发生时，就会非常需要 guardrails（护栏/约束机制）来限制 agents 在做什么，以防它们泄露数据或修改错误的信息；还需要它们可以依赖的权威 truth source（事实来源），对其行为进行 logging 和 auditing（日志记录与审计），以及通过这些系统与人协作的能力，等等。对任何一个给定的 agentic task（agent 式任务）做一次简单查询，调取的数据量都可能比一个用户一个月接触的数据还多。因此，很多类别的软件一旦变成 headless（无界面/无头）模式，其使用量和价值都会显著提升。Agents 最终会比人更多地使用我们的 CRM 数据、文档和企业知识、分析数据以及其他信息。那些能够转向为这些 headless 交互提供底层能力的平台，并且拥有与之配套的商业模式和技术战略的公司，未来将处于最有利的位置。

Pretty remarkable what’s happening with open weights AI right now. We’re seeing models achieve SOTA results on specific tasks, and getting close to frontier on some areas of coding and other domains. The more that open weights is able to maintain only a marginal gap from the frontier, instead of a widening gap, the more value that can be created with AI. Incidentally, this is actually fine for the frontier labs as well; if we can lower the cost of an overall task then AI usage goes up in general. You’re still likely using frontier models for planning, orchestration, reviewing, and other parts of work. But this is all very good for the applied layer of AI, which is now in a great position to cost optimize workloads with cheaper models or use tailored open models post-trained for specific tasks to improve performance.

当前 open weights AI 的发展相当令人瞩目。我们看到，一些模型已经在特定任务上取得了 SOTA（state-of-the-art，当前最先进）结果，并且在 coding 等一些领域以及其他领域中，正逐渐接近 frontier（前沿）水平。open weights 与 frontier 之间如果能始终保持只是边际差距，而不是差距不断拉大，那么 AI 所能创造的价值就会越大。顺带一提，这其实对 frontier labs 也同样有利；如果我们能降低一个整体任务的成本，那么 AI 的总体使用量就会上升。你仍然很可能会在 planning、orchestration、reviewing 以及工作的其他环节中使用 frontier models。但这一切都对 AI 的 applied layer（应用层）非常有利，因为它现在处在一个很好的位置：既可以用更便宜的模型来优化 workload（工作负载）成本，也可以使用针对特定任务进行 post-trained（后训练）的定制 open models 来提升性能。

The main variable in getting success with agents is whether you can get the agent the context it needs to do its work; and a major factor in that is if you can create a shared working area for that agent that a human can understand as well. This is one of the reasons why agents using file systems is such a big deal. It creates a unified system that both the person and the agent can work within to pass around data. “What they need is a working set: plans, notes, task lists, policies, drafts, summaries, logs, corrections, decisions, etc. For that layer, a filesystem-shaped interface tends to be more legible to both the model and the humans supervising it.” It turns out giving agents access to the systems we already know how to use, but in a way that is best optimized for them, is the perfect primitive for agents to work.