Aaron Levie

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.

决定 agent 能否成功的主要变量，是你是否能为它提供完成工作所需的 context（上下文）；而其中一个关键因素，则是你是否能为这个 agent 创建一个人类也能理解的共享工作区。这也是为什么使用 file systems（文件系统）的 agents 如此重要的原因之一。它创造了一个统一的系统，让人和 agent 都能在其中协作并传递数据。“它们需要的是一个 working set（工作集合）：plans、notes、task lists、policies、drafts、summaries、logs、corrections、decisions 等等。对于这一层，一个类似 filesystem 的接口，往往对 model 和监督它的人类来说都更清晰易读。” 事实证明，让 agents 访问我们已经知道如何使用的系统，但以最适合它们的方式进行优化，正是 agents 开展工作的完美 primitive（基础构件）。

The fact that open weights models are being discussed credibly at this level of capability should be a huge update for many. The implications of open models getting to frontier performance ensures that you can always have sovereign AI, have the ability to post train for your specific workflows, cost optimize for various workloads, and actually afford to do much more with AI (which opens up meaningfully different applications). Huge win for the applied AI layer.

在这样的能力水平上，open weights models 被认真地纳入讨论，这一事实本身对很多人来说都应当是一次巨大的认知更新。open models 达到 frontier performance（前沿性能）所带来的影响在于：你将始终可以拥有 sovereign AI（主权 AI），能够针对自己的特定工作流进行 post train（后训练），为不同工作负载做成本优化，并且真正负担得起用 AI 做更多事情（这会打开一些明显不同的新应用）。这对 applied AI layer（应用层 AI）来说是一次巨大胜利。

This is a good update for getting access to Fable. It also gives us a view into what the future is likely going to look like with AI regulation. The government will have frameworks that are used to determine future model releases past a certain threshold of capability or compute levels. Given all the constituents involved, and the economic and societal significance of AI, this was practically an inevitability. It may seem small but the implications are massive. It will mean that each model update will go through an extensive review, testing, and feedback process. And in that processes lots of groups will weigh in on the risk of the model, and there will be lots of subjectivity on what the actual risks are or practicalities of exploiting those risks. A positive potential future here would be we still get massive model progress but they just happen in bigger jumps at once, where the labs pack in major improvements since the cost and slow down of each review stacks up. On the other side, the risk is that past a certain threshold we may not get to see the rapid back and forth of model progress that we’ve gotten used to which can have negative compounding effects. Hoping for the former outcome.

这对获得 Fable 的访问权限来说是一个不错的进展。这也让我们得以一窥未来 AI 监管很可能会是什么样子。政府会建立一些 framework（框架），用来判定未来模型发布是否越过了某个 capability（能力）或 compute（算力）水平的阈值。考虑到其中牵涉的所有利益相关方，以及 AI 在经济和社会层面的重要性，这几乎是不可避免的。这件事看起来可能不大，但其影响极其深远。这意味着此后的每一次模型更新，都将经历广泛的审查、测试和反馈流程。而且在这个过程中，很多群体都会对模型的风险发表意见，对于实际风险究竟是什么、以及利用这些风险在实践中是否可行，也会存在大量主观判断。这里一种积极的可能未来是：我们仍然能看到模型取得巨大的进步，只不过这些进步会以更大的跳跃一次性出现，因为每次审查带来的成本和放缓会不断累积，于是实验室会把重大改进打包进去。另一方面，风险在于，一旦越过某个阈值，我们可能就看不到已经习以为常的那种模型快速往复迭代了，而这可能带来负面的复合效应。希望结果会是前一种。

The past couple months we may be witnessing what the Applied AI layer will look like at scale. Despite some of the initial critique that this would just be a thin layer on the LLM, it’s turning out that actually driving agentic workflows in an enterprise is far more complex. And anywhere there’s complexity you generally gain a moat and value over time. Here are a few of the components that appear to make up the playbook based on the examples we’re collectively seeing in coding, legal, healthcare, customer support, financial services and other fields: * Build the features that bridge the gap between the intelligence and the workflow. Some workflows can be automated by simply going to a general purpose interface, but others need tuned interfaces and features tied to the work they’re augmenting or automating. They need features that are specific to capturing the kind of data that’s needed as context for the agent. And they need a variety of bespoke tools for the agent to use, and unique interfaces for the human-in-the-loop UX. Going far deeper than just presenting the output tokens is clearly critical, and the more depth there is here definitionally the more sustaining value. * Act as the model router balancing frontier intelligence with cheaper models. A natural advantage that any model neutral platform has is that it can naturally (in a business model-aligned way) leverage whatever level of intelligence is necessary for the workflows they’re automating to get done. There are plenty of scenarios where you need GPT-5.5 or Fable level capability, and also lots of workloads where a more efficient closed or open weights do the trick. Only the companies that have deep evals on specific tasks across all models, and the ability business model wise to leverage them, are in a great position. * Drive the actual implementation and change management via FDE or equivalent. A big reason the applied layer works at scale is that most enterprises need some degree of help and support with change management in implementing agents for their workflows. Data has to be cleaned up and moved to modern systems, processes have to be re-engineered and documented, workflows have to be evaled, SLAs have to get achieved, and so on. All of this is going to be unique for every type of process that gets implemented, which means the companies that have expertise in a given domain and come with all the relevant best practices will be in a strong position. * Implement domain specific GTM that creates expertise in that field. Beyond FDEs the companies that can build sales and GTM motions aligned to their domains also have a natural advantage. Most IT and line of business leaders have too many things to do in any given day; so if you’re not on their agenda, likely someone else is. Depending on the industry, there are entirely different sets of language you use, ways of working through security and compliance, regulatory controls you have to support, industry events that companies convene at, different system integrator and consulting partners you need to work with, and so on. The more generalized this gets the less you can speak the customers language, which is where the applied layer has a leg up. A final note. There remains a view that a lot of this is all mitigated by model intelligence alone, and the bitter lesson solves all of this in the limit. That’s possibly true, but enterprises need help changing *today*. And many aspects of how to bring intelligence to real world work don’t only depend on the axis of the pure capability of the model, so most of what you’re doing now to win ends up being important no matter how good the models get.

过去几个月里，我们或许正在见证大规模 Applied AI（应用层 AI）会是什么样子。尽管一开始有一些批评认为，这不过是叠在 LLM 之上的一层很薄的外壳，但事实证明，要真正把 agentic workflows（agent 驱动的工作流）推进到企业环境中，复杂得多。而凡是存在复杂性的地方，通常就会随着时间形成 moat（护城河）和价值。基于我们在 coding、legal、healthcare、customer support、financial services 及其他领域共同看到的案例，下面是一些似乎正在构成这套 playbook（方法论）的组件：* 构建能够弥合 intelligence 与 workflow 之间差距的功能。有些工作流只需要进入一个通用界面就能自动化，但另一些则需要经过调优的界面，以及与其所增强或自动化的工作紧密绑定的功能。它们需要专门用于采集 agent 所需上下文数据的功能；也需要各种定制工具供 agent 使用，以及面向 human-in-the-loop UX 的独特界面。显然，仅仅呈现输出 token 远远不够，必须做得更深；而这里做得越深，从定义上讲，可持续的价值也就越大。* 充当 model router（模型路由器），在前沿 intelligence 与更便宜的模型之间做平衡。任何 model neutral platform（模型中立平台）的一个天然优势在于，它可以很自然地——并且以符合其商业模式的方式——调用完成其所自动化工作流所需的 intelligence 水平。有很多场景确实需要 GPT-5.5 或 Fable 级别的能力，但也有大量工作负载使用更高效的 closed 或 open weights 模型就足够了。只有那些对所有模型在具体任务上做了深入 evals（评测），并且在商业模式上也有能力去灵活利用它们的公司，才真正处于有利位置。* 通过 FDE 或同类角色推动实际实施与 change management（变更管理）。应用层之所以能够大规模发挥作用，一个重要原因在于，大多数企业在为其工作流部署 agents 时，都需要某种程度的变更管理帮助与支持。数据必须被清洗并迁移到现代系统中；流程必须被重新设计并文档化；工作流必须经过 eval；SLA 必须达成，等等。所有这些对每一种被实施的流程都会是独特的，这意味着，那些在特定 domain（领域）中拥有专业能力，并且自带相关最佳实践的公司，将处于强势位置。* 落地 domain specific GTM（领域特定的 Go-To-Market），从而在该领域形成专长。除了 FDE 之外，那些能够建立与其所在领域相匹配的销售和 GTM 动作的公司，也拥有天然优势。大多数 IT 和业务条线负责人每天都有太多事情要处理；所以如果你不在他们的议程上，很可能别人就在。视行业而定，你需要使用完全不同的话语体系，采用不同的安全与合规推进方式，支持必须满足的监管控制，参加公司会聚集的行业活动，并与不同的 system integrator 和咨询伙伴合作，等等。越是做得泛化，你就越难真正说客户的语言，这正是应用层占优的地方。最后补充一点。仍然有人认为，这一切最终都会被模型 intelligence 本身所化解，bitter lesson 会在极限情况下解决所有这些问题。这或许没错，但企业需要的是今天就获得变革帮助。而且，将 intelligence 引入现实世界工作中的许多方面，并不只取决于模型纯能力这一条轴线，所以你现在为了赢而做的大部分事情，不管模型未来变得多强，最终都会依然重要。

One of the biggest questions in AI is how far behind open weights models remain from closed models at any given time. There are huge differences in market structures depending on whether open weights models remain 3 or 6 months behind, or if they fall behind by years. The answer to this will determine how the chip stack plays out, where inference can be run, what sovereign AI looks like, what happens at the applied AI layer, what the margin structure looks like in AI, how much companies can afford to spend on AI, and more. At the moment the open weights players appear to be holding up at keeping close to frontier levels of capability. Will be fun to see how this plays out.

AI 领域中最大的问题之一，是 open weights models（开放权重模型）在任意一个时间点究竟会落后 closed models（闭源模型）多远。市场结构会因此出现巨大差异：到底是 open weights models 只落后 3 个月或 6 个月，还是会落后几年。这个问题的答案将决定 chip stack（芯片栈）会如何演变、inference（推理）可以在哪里运行、sovereign AI（主权 AI）会是什么样子、applied AI（应用层 AI）会发生什么、AI 的利润率结构会是什么样、公司能负担多少 AI 支出，等等。目前看来，open weights 阵营似乎仍能把能力维持在接近 frontier（前沿）水平的位置。接下来会怎么发展，应该会很有意思。

The Cursor deal is symbolically quite significant. It was effectively the first mega success in the applied layer of AI. They firmly proved out the value proposition of having a deep domain focus, the role you play as a model router, when to lean into frontier models vs. when to train your own, and the role of applied AI GTM and distribution to make sure you’re actually taking advantage of the market opportunity. Every aspect of their business was tuned to carve out ground and keep doubling down in a highly competitive space. This is really the first at scale template for how to execute this playbook.

Cursor 这笔交易在象征意义上相当重要。它实际上是 AI 的 applied layer（应用层）中第一个真正意义上的超级成功案例。他们非常扎实地验证了这些价值主张：深耕垂直领域的价值、你作为 model router（模型路由器）所扮演的角色、何时应该依赖 frontier models（前沿模型）而不是自己训练，以及 applied AI 的 GTM（go-to-market，市场进入）和 distribution（分发）在确保你真正抓住市场机会方面所起的作用。他们业务的每一个方面都经过精细调校，以便在一个竞争极其激烈的领域中占据地盘，并持续加码。这确实是第一个大规模展示如何执行这套 playbook（打法手册）的模板。

Key post that gives a bit of insight into what the future of AI could look like. “The most interesting thing happening in AI isn't that one model is getting smarter. It's that intelligence is becoming increasingly customizable. The companies that win won't necessarily be the ones with the biggest models. They'll be the ones that turn intelligence into something uniquely their own.” The ability to combine your unique data, workflows, and a layer that can route intelligence to whatever model best performs the task is clearly the future.

这篇关键帖子让人稍微看清了 AI 的未来可能会是什么样子。“AI 里正在发生的最有趣的事情，并不是某一个 model 变得更聪明了，而是 intelligence（智能）正变得越来越可定制。最终胜出的公司，未必是那些拥有最大 model 的公司，而会是那些把 intelligence 变成自身独特能力的公司。” 能够把你独有的数据、workflows（工作流），以及一层可以把 intelligence 路由到最适合完成任务的 model 的系统结合起来，这显然就是未来。

It’s very easy to say “we need an FDA for AI” or some equivalent government agency. Well this is what that would look like. The capabilities of AI models have near infinite permutations. It’s going to be very hard have a purely objective set of metrics that can be universally applied before every model release, without extended back and forth, research, and debate between model labs, academics, and the government. Now, imagine this same process with every country that you’re doing business with, globally, for every single model release. Add in a backlog of dozens of AI model releases, and you can quickly see how in the limit this will dramatically slow down all AI progress. This is why we need to primarily focus on regulating the applied uses of AI, where the risk actually shows up.

说一句“我们需要一个 AI 版的 FDA”，或者某个类似的政府机构，是非常容易的。好，这就是那种东西实际会是什么样子。AI model 的能力几乎有无限多种排列组合。想要在每次 model 发布之前，都制定出一套纯粹客观、还能被普遍适用的指标，而不经历 model labs、学术界和政府之间反复拉扯、研究和辩论，这将会非常困难。现在，再想象一下：对于你在全球开展业务所涉及的每一个国家，在每一次 model 发布时，都要走一遍同样的流程。再加上积压的几十个 AI model 发布，你很快就能看出，如果走到极限，这会显著拖慢整个 AI 的进展。这也是为什么我们需要主要去监管 AI 的 applied uses（实际应用场景），因为风险真正出现的地方是在那里。

Open source going to win big

Open source 会大获全胜。

Great post. The companies that are able to get their unique IP, institutional knowledge, and data into a format and architecture that lets them capture all of the gains and progress in AI are going to be in the best position in the future. “the real opportunity is not in picking the best model but instead in building a learning loop on top of models where human capital and token capital compound. You can offload a task, or even a job, but you can never offload your learning. The future of the firm is the ability to compound that learning across people and AI. This requires a new architectural approach where every business is able to build agentic systems that improve over time, while still retaining control over their IP. A company should be able to switch out a “generalist” model without losing the “company veteran” expertise built into their learning system.” We’re all collectively figuring out the right architecture for the future of AI. But it’s clear that so much of the power and value will accrue to wherever can best leverage any AI system against their information. This is also why the applied AI layer will also gain so much value over the coming years.

很棒的文章。那些能够把自己独特的 IP、institutional knowledge（机构知识）和数据整理成一种格式与架构，从而抓住 AI 带来的全部收益与进展的公司，未来会处于最有利的位置。“真正的机会不在于挑选最好的 model，而在于在 model 之上构建一个 learning loop（学习闭环），让 human capital（人力资本）和 token capital（token 资本）实现复利增长。你可以外包一个 task，甚至一份工作，但你永远无法外包自己的学习。企业未来的核心，是能够让这种学习在人与 AI 之间持续复利。这需要一种新的架构方法，让每一家企业都能构建会随时间不断改进的 agentic systems（agent 系统），同时仍然保有对自身 IP 的控制。公司应该能够替换掉一个‘generalist’ model，而不会失去其学习系统中沉淀下来的‘company veteran’级专业经验。” 我们都还在共同摸索适用于 AI 未来的正确架构。但很明显，巨大的能力与价值将会积累到那些最善于让任何 AI system 结合自身信息发挥作用的地方。这也是为什么在未来几年里，applied AI 这一层也会获得如此巨大的价值。

The big winner in all of this is going to be open weights models. This is a huge win for the field, as a risk that was entirely theoretical and untested 2 days ago (that a model could be pulled back), now has a new precedent that’s been set. The game theory the US should highly consider, and the risk with regulating AI at the model layer vs. applied layer, is that other countries now have even more incentive to develop sovereign AI. If at any moment a model can be become unavailable to your country’s users or businesses, this poses very real risk on relying on technology from a particular country. As a result, it forces major countries to charter their own path on AI development, which reduces America’s leadership role in this tech stack over time. The most likely solution that other countries will rely on is open weights models, which currently is generally not coming from the US. America should be considering all of these downstream implications as it decides how and where in the stack to be regulating AI. At the same time, we should be doing a ton more OSS innovation.

这整件事里的最大赢家将会是 open weights models。这对整个领域来说是一次巨大的利好，因为一个在 2 天前还完全停留在理论层面、从未被验证过的风险（即一个 model 可能会被撤回），现在已经有了新的 precedent（先例）。美国应当高度重视其中的 game theory（博弈逻辑），以及在 model layer（模型层）而不是 applied layer（应用层）监管 AI 所带来的风险：其他国家现在会有更强的动力去发展 sovereign AI（主权 AI）。如果一个 model 在任何时刻都可能对你国家的用户或企业变得不可用，那么依赖来自某一个特定国家的技术就会构成非常现实的风险。因此，这会迫使主要国家在 AI 发展上走自己的道路，而这会随着时间推移削弱 America 在这套 tech stack（技术栈）中的领导地位。其他国家最有可能依赖的解决方案是 open weights models，而目前这类模型通常并不是来自美国。America 在决定应当如何以及在这套 stack 的哪些位置监管 AI 时，应该考虑所有这些下游影响。同时，我们也应该大幅增加 OSS 创新。

The layer that can route to the best AI model for the particular job is going to increase in value substantially. There are at least 3 big reasons: * Cost optimization: there are plenty of use cases where you need frontier intelligence for some tasks and something far cheaper for others. Even in the same task you may use frontier intelligence for planning and review of the work, but an OSS or cheaper model for the bulk of the workload. This is going to be standard across large buckets of work going forward. * Capability maximization: despite the bitter lesson and models generally getting better in the same direction, there are still lots of differences between models. Some are better at tool use, others better at coding, and others again better at certain domains of knowledge work. The ability to route between these at different times is a huge advantage. * Risk mitigation: while the Fable situation is somewhat of a black swan, it’s possible we’re heading toward a regulatory environment where governments may restrict models at different times based on their approval mechanisms or new things they discover. This means you’re going to want flexibility in being able to deploy workloads across different providers as a form of risk mitigation. Ultimately, it’s going to increasingly be a a strategic advantage for the applied AI layer that they can effectively route between models. Will be very interesting to see how this evolves.

那个能够为特定工作路由到最佳 AI model（模型）的层，其价值将会大幅提升。至少有 3 个重要原因：* 成本优化：有很多 use case（用例）里，你会需要在某些任务上使用 frontier intelligence（前沿智能），而在另一些任务上使用便宜得多的模型。即便是在同一个任务里，你也可能会用 frontier intelligence 来做规划和工作审查，但把大部分工作负载交给 OSS 或更便宜的模型。今后这会在大类工作中成为标准做法。* 能力最大化：尽管 bitter lesson（苦涩教训）依然成立，而且 models（模型）总体上是在朝同一方向变强，但模型之间仍然存在大量差异。有些更擅长 tool use（工具使用），有些更擅长 coding（编程），还有些则更擅长某些知识工作领域。在不同时间点在这些模型之间进行路由的能力，是一个巨大的优势。* 风险缓释：虽然 Fable 这次的情况某种程度上算是 black swan（黑天鹅）事件，但我们也有可能正在走向一种监管环境：政府可能会基于自己的审批机制，或基于他们新发现的情况，在不同时间限制不同模型。这意味着，你会希望自己具备灵活性，能够把工作负载部署到不同 provider（提供方）上，把这作为一种风险缓释手段。归根结底，对于 applied AI（应用层 AI）来说，能否有效地在不同模型之间进行路由，将越来越成为一种战略优势。这个方向接下来会如何演化，会非常值得关注。

Everyone thinks this is some kind of 4D chess or conspiracy. But it’s quite standard to try and jailbreak AI models, and by definition they would share that research with the government given that’s whole point. I don’t think Amazon assumed this would be the next move.

每个人都觉得这是什么 4D chess（四维棋）或者阴谋论。但尝试 jailbreak AI models（越狱 AI 模型）其实是相当标准的做法，而且按定义来说，他们既然做这项研究，本来就会把研究结果分享给政府，因为那本来就是整个事情的目的。我不认为 Amazon 预料到这会是下一步动作。

This whole Fable export control situation is actually net positive to regulation discourse. It’s an early peek into what AI regulation would end up looking like at scale when enacted at the model layer instead of the specific application of the AI. The government would have sole discretion over when a model can be released to the to public, based on a bunch of factors that they inherently control. In this case, based on the available reporting, the risk is that the model can be jailbroken to deliver increased cyber exploit capabilities. The issue is that actually you want models to be able to have those capabilities on the defense side of cyber as well, and for all intents and purposes, by Anthropic’s own response, you can execute these capabilities today in other models. So thus the whole challenge will be that you’re debating with the government, over months and months, with every model release, what these models are actually capable of and what their risks are. Inherently, there’s not only a lot of subjectivity in determining those risks, but there’s also many other factors that go into the risks being practical in the first place. The net result is that we would end up with backlog of AI releases, progress in the market inherently would dramatically slow down, and AI would start to look more like any other sclerotic industry. If this paradigm had existed 3 years ago at the start of the current AI wave, we’d likely currently be stuck on GPT-4 level intelligence at this point. This is why, wherever possible, we should be regulating the applied use of AI. We should continue to study and enforce the dangerous use of AI in cyber attacks, financial services risks, fraud, biowarfare, and other spaces. AI safety is incredibly important, but slowing down progress this early in the development of AI I suspect is net harmful.

整个 Fable export control（出口管制）事件，实际上对监管话语来说是净正面的。它让我们提前看到了一眼：如果 AI regulation（AI 监管）是在 model layer（模型层）而不是针对 AI 的具体应用来实施，那么大规模落地后会是什么样子。政府将对一个模型何时可以向公众发布拥有完全裁量权，而依据则是一系列本质上由他们掌控的因素。在这个案例里，根据现有报道，风险在于这个模型可能被 jailbreak（越狱），从而提供更强的 cyber exploit（网络利用/攻击）能力。问题在于，你其实也希望模型在 cyber（网络安全）的防御侧具备这些能力，而且就实际效果而言，按照 Anthropic 自己的回应，今天你已经可以在其他模型上执行这些能力了。所以，真正的挑战将会是：你要在每一次模型发布时，花上数月又数月与政府争论这些模型到底具备什么能力、其风险又是什么。这里面不仅在风险认定上天然存在大量主观性，而且风险之所以会在现实中成立，本身还取决于许多其他因素。最终结果就是，AI 发布会出现积压，市场进展会不可避免地显著放缓，而 AI 行业会开始看起来像其他任何僵化的行业一样。如果这种范式在 3 年前、也就是这一轮 AI 浪潮开始时就已经存在，那么我们现在很可能还卡在 GPT-4 水平的智能上。这就是为什么，只要有可能，我们就应该监管 AI 的应用使用层。我们应该继续研究并执法打击 AI 在 cyber attacks（网络攻击）、financial services risks（金融服务风险）、fraud（欺诈）、biowarfare（生物战）以及其他领域中的危险使用。AI safety（AI 安全）极其重要，但在 AI 发展这么早的阶段就放慢进展，我怀疑总体上是有害的。

This is a big turning point for AI regulation. The government is starting to deem some models too powerful for certain uses, which creates a precedent for a range of possible controls in the future. I’m in the camp that this is unnecessary and we should be primarily regulating the use of AI, as opposed to the underlying models. But, equally, there are plenty of people that actually prefer this outcome. Either way, it’s unlikely that we’re going back to a world where the government doesn’t have far more meaningful involvement in the rate of AI progress.

这对 AI 监管来说是一个重大的转折点。政府开始认定某些 model（模型）对于特定用途来说过于强大，这为未来一系列可能的管控措施开创了先例。我属于这样一种观点：这没有必要，我们主要应该监管的是 AI 的使用，而不是底层 model（模型）本身。但与此同时，也有很多人实际上更倾向于这样的结果。不管怎样，我们不太可能再回到一个政府对 AI 进展速度没有更深度、也没有更实质性介入的世界。

This is pretty freaking cool

这真他妈挺酷的。

Incredible. Congrats to @elonmusk and the entire SpaceX team on the 25 years of blood, sweat and tears to build a world-defining company. Amazing to have examples like this that push the future forward. The downstream implications of this are enormous.

难以置信。祝贺 @elonmusk 和整个 SpaceX 团队，在 25 年的 blood, sweat and tears（心血与艰辛付出）中打造出一家定义世界的公司。能有这样推动未来向前发展的榜样，实在太了不起了。这件事带来的下游影响将是巨大的。

At Box, we just surveyed 1,640 IT leaders across the US, Japan, and Europe about agentic AI adoption. Many standout findings, but a big one was that the companies that adopted AI the most are planning to grow headcount the most. Obviously lots of ways you can read that data and variables mixed in, but it’s actually quite intuitive that the companies that become most productive want to (and are able to) reinvest back into the business to keep getting the gains going. The narrative of jobs being wiped out assumes that companies will take a fixed approach to what they want to be able for work on. What’s happening in practice is it’s causing companies to want to light up more engineering projects, sell to more customers, automate more processes to give time back, and more. That all leads to more work to be done by people.

在 Box，我们刚刚就 agentic AI 的采用情况，调查了来自美国、日本和欧洲的 1,640 位 IT 领导者。结果中有很多突出的发现，但其中一个重要结论是：采用 AI 最多的公司，也计划扩大最多的员工规模。显然，这组数据可以有很多种解读方式，其中也混杂着各种变量；但实际上，这一点很符合直觉：那些生产力提升最多的公司，会希望（也有能力）把收益重新投入业务中，以持续扩大这些增益。“工作岗位会被消灭”的叙事，假设了公司对于自己希望完成的工作范围会采取一种固定不变的思路。而现实中正在发生的是，这会促使公司启动更多 engineering 项目，向更多客户销售，自动化更多流程以腾出时间，等等。这一切最终都会带来更多需要由人来完成的工作。

Lots of evidence of huge jumps in capability for Fable across coding (and related) tasks. It’s also a major jump in accuracy and success in complex knowledge work tasks. In our Box AI Complex Work Eval, we tested the model against Opus 4.8 and saw huge boosts across almost every industry. For our eval we give the Box AI Agent, using Fable, a set of hard real world knowledge work problems that deal with enterprise documents. Then score how the agent performs the tasks. The main differentiators for Fable vs Opus 4.8 is that it doesn't take shortcuts on complex reasoning, it gets multi-step calculations right, and it's significantly more consistent across runs. We saw the biggest leaps in Media & Entertainment (78% vs 61%), Technology (81% vs 73%), Financial Services (89% vs 83%), and Healthcare (66% vs 60%). Here are some specific examples: * Legal M&A due diligence: On a task reviewing NDA terms against a semiconductor company's contracting policy, Fable correctly identified that a joint-ownership clause violates exclusivity requirements while a liability cap is permitted under a Super Cap exception. Fable scored 100% vs Opus's 78%. * Healthcare: On a clinical radiology error audit across 12 reports, Fable precisely categorized each error by severity grade and correctly concluded no Grade 3 errors existed. Opus prematurely escalated a case to "major error requiring immediate departmental review" when the evidence didn't support it — Fable 63% vs Opus 41%. * Media & Entertainment: On a genre profitability projection task, Fable correctly recognized that a 20% Argentine tax deduction was already embedded in the source spreadsheet figures and didn't double-apply it. Opus applied it again on top — a compounding error across 4 genre calculations that took its score negative on the task vs Fable's 74%. * Retail analytics: On a task analyzing high-growth product articles against an investment benchmark, Fable correctly computed each article's growth rate individually and identified that only 2 of 5 exceeded the threshold. Opus confused "high growth relative to average" with "above the benchmark" — scoring 61% vs Fable's 94%. * Financial Services: On a 5-year debt facility projection, Fable correctly applied interest to opening balances and used the right capex figure. Opus applied interest to the total facility amount and computed tax from the wrong base — two compounding errors. Fable scored 83% vs Opus's 62%. * Technology: On a SaaS feature valuation requiring computation of a Feature Value Index across multiple regions, Fable applied the formula correctly and got exact values for the markets. Opus got the arithmetic wrong on multiple criteria — Fable scored 100% vs Opus's 74%. Overall, huge step change in complex analysis, work that requires analytical reasoning, and deep domain understanding. Fable will be available shortly in the Box AI Studio for customers to build agents with.

有大量证据表明，Fable 在 coding（编码）及相关任务上的能力实现了巨大跃升。在复杂知识工作任务中的准确性和成功率也有显著提升。在我们的 Box AI Complex Work Eval 中，我们将该模型与 Opus 4.8 对比测试，看到它在几乎所有行业中都有大幅提升。在这项 eval 中，我们让使用 Fable 的 Box AI Agent 处理一组与企业文档相关、具有真实世界难度的知识工作问题，然后根据 agent 完成任务的表现进行评分。与 Opus 4.8 相比，Fable 的主要差异在于：它在复杂推理上不会走捷径，能够正确完成多步骤计算，并且在多次运行中的一致性显著更高。我们看到提升最大的领域包括 Media & Entertainment（78% 对 61%）、Technology（81% 对 73%）、Financial Services（89% 对 83%）和 Healthcare（66% 对 60%）。以下是一些具体例子：* Legal M&A due diligence：在一项根据某 semiconductor company 的 contracting policy 审查 NDA 条款的任务中，Fable 正确识别出 joint-ownership clause 违反了 exclusivity requirements，而 liability cap 在 Super Cap exception 下是被允许的。Fable 得分 100%，Opus 为 78%。* Healthcare：在一项针对 12 份报告的临床 radiology 错误审计任务中，Fable 准确按严重等级对每个错误进行了分类，并正确得出不存在 Grade 3 错误的结论。Opus 在证据并不支持的情况下，过早将一个案例升级为“major error requiring immediate departmental review”——Fable 63%，Opus 41%。* Media & Entertainment：在一项类型片盈利能力预测任务中，Fable 正确认识到 20% 的 Argentine tax deduction 已经包含在源 spreadsheet 数据中，因此没有重复应用。Opus 又额外应用了一次——这个在 4 类题材计算中不断累积的错误，使其在该任务上的得分变成负分，而 Fable 为 74%。* Retail analytics：在一项将高增长产品相关文章与投资基准进行分析比较的任务中，Fable 正确分别计算了每篇文章的增长率，并识别出 5 篇中只有 2 篇超过阈值。Opus 将“相对平均值的高增长”和“高于 benchmark（基准）”混淆了——得分 61%，Fable 为 94%。* Financial Services：在一项 5 年期债务融资 projection 任务中，Fable 正确地将利息应用于期初余额，并使用了正确的 capex figure。Opus 则将利息应用于整个融资额度，并从错误的基数计算税额——这是两个会叠加放大的错误。Fable 得分 83%，Opus 为 62%。* Technology：在一项需要跨多个地区计算 Feature Value Index 的 SaaS feature valuation 任务中，Fable 正确应用了公式，并得出了各市场的精确数值。Opus 在多个指标上都算错了——Fable 得分 100%，Opus 为 74%。总体而言，这是在复杂分析、需要分析性推理的工作以及深度领域理解方面的一次巨大跃迁。Fable 很快将在 Box AI Studio 中上线，供客户用来构建 agents。

This is a critical post to read if you’re building an applied AI company right now. “An application earns its place in the untrainable corner by doing unglamorous work: arranging a company's private reality so a model can act on it, handing the model the tools to act, working with the customer to change the reality of its workforce. A company that brings the translation is tough to copy – and the translation never ends. Integration and maintenance run as long as the relationship does, won by teams that put domain-specialized engineers and tools next to the customer.” There’s still an insanely large gulf between model capabilities and what it takes to apply them to specific corporate workflows. Some of that is technology that needs to be built, a lot is access to (and formatting of) the right data to work with, and a ton more is on the change management and specific implementation work (FDEs, etc.) it takes to make AI work in any specific corporate setting. 2 things can be very true at once: frontier models and labs will continue to grow an incredible amount, and there will be a vast ecosystem of software and services companies that emerge to bring the power of these models to real enterprises. This makes room for new infrastructure provides, applied AI companies in every vertical, new versions of system integrators, and more players. Incredibly exciting time on all fronts.

如果你现在正在打造一家 applied AI 公司，这是一篇必须读的重要文章。“一个应用之所以能在不可训练的角落里占据一席之地，靠的是做那些并不光鲜的工作：整理一家公司的私有现实，让模型能够基于它采取行动；把模型执行行动所需的工具交到它手里；与客户合作，改变其劳动力体系的现实。能够完成这种翻译的公司很难被复制——而且这种翻译永远不会结束。只要合作关系还在，集成与维护就会持续进行，而胜出的是那些把领域专精的工程师和工具放在客户身边的团队。” 模型能力与将其应用到特定企业工作流所需的一切之间，仍然存在大得惊人的鸿沟。其中一部分是尚需构建的技术，很大一部分是获取正确数据并将其格式化以供使用，更多则在于让 AI 在任何特定企业环境中真正发挥作用所需的变革管理与具体实施工作（FDEs 等）。有两件事可以同时成立：frontier models 和 labs 会继续实现惊人的增长；同时，也会涌现出一个庞大的软件与服务公司生态，把这些模型的力量带给真实企业。这为新的基础设施提供商、各个垂直领域的 applied AI 公司、新版本的 system integrators 以及更多参与者留出了空间。这对所有方向而言，都是一个令人无比兴奋的时刻。

If you thought AI progress was slowing down, well here's the immediate answer to that. Huge jump in capability across the board. This is going to deliver major improvement in agents across almost all knowledge work categories.

如果你以为 AI 的进展正在放缓，那么这就是对此最直接的回答。整体能力出现了巨大跃升。这将给几乎所有知识工作类别中的 agents 带来重大改进。

Great post. So much about model performance is a function of how much compute you’re doing at inference time. This means compute-normalized benchmarks is the only logical path forward. And yet, the challenge is it’s a lot harder than it seems given it’s subjective how much compute to apply, which means models behave differently at different thresholds (simplistically, model X’s min thinking may beat model Y’s min thinking, but be reversed at high), and there are a near infinite set of thresholds you could choose to set. But either way, moving more in this direction would be great for better understanding AI progress.

很棒的一篇文章。模型性能在很大程度上取决于你在 inference time 投入了多少 compute。这意味着，按 compute 归一化的 benchmarks 才是唯一合乎逻辑的前进路径。不过，难点在于，这件事比看上去要难得多，因为该投入多少 compute 本身就带有主观性，这意味着模型在不同阈值下会表现不同（简单说，model X 的低思考量版本可能胜过 model Y 的低思考量版本，但在高阈值下结果可能反过来），而你几乎可以设定出无限多个不同的阈值。无论如何，朝这个方向多推进一些，都会非常有助于我们更好地理解 AI 的进展。

There’s no amount of intelligence that can get packed into AI models that replaces the need for context. For any sufficiently general purpose AI, you will always have to guide it in the direction you want as it has an infinite range of directions it can go in. As long as the same model is used by a lawyer, an engineer, a financial analyst, or a healthcare professional, and as long as you’re trying to do anything uniquely differentiated or specific, then instructions, domain context, and proprietary data will always need to get into the context window for the model to be useful. This is partly why AI automation doesn’t come for free, and why there’s still a wide spectrum of who’s getting the largest gains from AI and who’s not. You have to put in real work, and you get real value on the other end. This is one of the advantages that applied AI will also have in the market. Any layer of abstraction above just the raw intelligence that can meaningfully get you off to the races faster will likely continue to be valuable.

无论往 AI model 里塞进多少智能，都无法取代对 context（上下文）的需求。对于任何足够通用的 AI，你始终都必须把它引导到你想要的方向，因为它可能前往的方向实际上是无限的。只要同一个 model 被 lawyer、engineer、financial analyst 或 healthcare professional 共同使用，并且你想做的事情具有任何独特差异化或具体性，那么 instruction（指令）、domain context（领域上下文）和 proprietary data（专有数据）就始终需要进入它的 context window（上下文窗口），这个 model 才会真正有用。这也是为什么 AI automation（自动化）并不是免费获得的部分原因之一，也说明了为什么在“谁从 AI 中获得最大收益、谁没有”这件事上，至今仍存在很大的光谱差异。你必须投入真正的工作，而另一端你也会获得真正的价值。这也是 applied AI（应用型 AI）在市场中的优势之一。任何建立在纯粹原始智能之上的 abstraction layer（抽象层），只要能切实帮助你更快起步并进入状态，就很可能会持续具备价值。

The numbers may be a bit extreme here, but unquestionably use-cases have to stratify in the next year or two between model families. We’ll see a split between frontier intelligence for high end tasks and work, and much cheaper models for high volume workloads that can sufficiently be peeled off to cheaper models. Frontier will still be far bigger than today because the use-cases will demand it, but the low-end will get quite a bit larger as well. The big update here is that the layer that can efficiently route the workload to the right model will then become increasingly valuable since that becomes one of the new hard problems in AI agents. Agent orchestration that can cost optimize while still performing the task successfully will be in a strong position.

这里的数字也许有点极端，但毫无疑问，在未来一两年里，各类 use-case（使用场景）必须在不同的 model family（模型家族）之间分层。我们会看到一种分化：一边是用于高端任务与工作的 frontier intelligence（前沿智能），另一边则是便宜得多的模型，用于那些量很大、且能被充分剥离给更便宜模型处理的 workload（工作负载）。Frontier 的规模仍会比今天大得多，因为 use-case 会需要它；但低端市场也会扩大不少。这里最大的变化是：能够高效地把 workload 路由到正确模型的那一层，将变得越来越有价值，因为这会成为 AI agent（智能体）中的新难题之一。能够在成功完成任务的同时优化成本的 agent orchestration（agent 编排），将处于非常有利的位置。

This is what the market got wrong about AI eating enterprise software. Building good software in the past was very hard. Yes, AI has made that a bit easier, though it’s still hard to build something that’s got good taste, differentiated, high quality, secure, and so on. But nevertheless, that’s only one component of building a platform that enterprises rely on. The plurality of costs in most enterprise software companies is actually on GTM, because at scale most enterprise software categories are tough to break into and need a heavy amount of consultative selling and support for implementation and integration of solutions. AI hasn’t reduced the need for that, and in many cases requires it even more now, as landscapes get even more busy and complicated for buyers to navigate through. If you make one thing cheaper and more abundant (development of software) then the new problem of discoverability and market differentiation (GTM) becomes the hardest part.

这正是市场在“AI 会吞掉 enterprise software（企业软件）”这件事上判断错的地方。过去，构建优秀的软件非常困难。没错，AI 确实让这件事稍微容易了一些，尽管要做出有品味、差异化、高质量、安全等等的东西，仍然很难。但无论如何，那只是构建一个企业所依赖的平台的其中一个组成部分。大多数 enterprise software 公司的成本大头，其实是在 GTM（go-to-market，市场进入/销售转化）上，因为一旦到了规模化阶段，多数 enterprise software 品类都很难打入，必须依赖大量顾问式销售，以及为解决方案的实施与集成提供支持。AI 并没有减少这方面的需求，而且在很多情况下，现在反而更需要它，因为买家需要穿越的市场格局变得更加拥挤和复杂了。如果你让一件事变得更便宜、更充足（software 的开发），那么新的难题——可发现性和市场差异化（GTM）——就会成为最难的部分。

Box now has a markdown editor on the web. Full CLI support. Commenting. Full version history. Box Drive also lets you connect to any desktop client as a mounted drive, so you instantly work with all your files in Claude Cowork, Codex, Obsidian, Cursor, or any other app.

Box 现在在 web 上有了一个 markdown editor（Markdown 编辑器）。完整的 CLI 支持。评论功能。完整的版本历史。Box Drive 还允许你把它作为挂载盘连接到任何 desktop client，这样你就能立即在 Claude Cowork、Codex、Obsidian、Cursor 或任何其他 app 中处理你的所有文件。

Token costs are becoming one of the hottest topics for any enterprise I talk with right now. It’s very bullish for AI in general because it means these systems are being used at a scale that wasn’t contemplated before. It also gives way to another form of differentiation that will emerge for the applied AI layer, which is model routing. As tokens take on a significant amount of the cost of any given workflow, then companies will inevitably want to ensure that their dollars go into the most efficient use of tokens for the particular job at hand. Frontier intelligence will always be relevant at the high end of tasks, like coding, legal and financial analysis, healthcare, and more. And dollars spent here will only go up over time. But, equally, you can peel off individual tasks to lower cost models (whether they’re from open weights vendors or the major labs) and deliver a more efficient end outcome. To do this effectively, the applied AI layer needs to understand the workflows in their domain better than anyone else, and be able to mix and match models to different jobs. If you’re doing document extraction, you need to know which models perform better or worse for any given document type. If you’re legal analysis, you want to know which models perform various types of tasks best. And so on. This will become one of the bigger differentiation points over time. The companies with the best evals, the best ability to route the workloads, and those that have business models directly aligned to customers financial goals, will be in a great position.

Token 成本正迅速成为我现在与几乎所有企业交流时最热门的话题之一。从整体上看，这对 AI 非常 bullish（看涨），因为这意味着这些系统正以此前未曾设想过的规模被使用。这也为 applied AI（应用层 AI）带来了另一种将会出现的差异化形式，也就是 model routing（模型路由）。随着 token 在任何给定 workflow（工作流）的成本中占据相当大的比重，企业必然会希望确保自己的每一美元，都被用于针对当前具体任务最有效率的 token 使用方式。对于编码、法律与金融分析、医疗健康等高端任务，frontier intelligence（前沿智能）始终都会具有相关性。而花在这些场景上的资金只会随着时间推移而不断增加。但与此同时，你也可以把其中的单个任务剥离出来，交给成本更低的模型（无论它们来自 open weights vendors，还是 major labs），从而交付一个整体上效率更高的最终结果。要想高效做到这一点，applied AI 层必须比任何其他人都更理解其所在领域的 workflow，并且能够针对不同任务灵活组合不同模型。如果你在做 document extraction（文档提取），你就需要知道，对于任何给定的文档类型，哪些模型表现更好，哪些更差。如果你在做法律分析，你就会想知道，哪些模型最擅长完成不同类型的任务。其他领域也是如此。随着时间推移，这将成为更重要的差异化因素之一。那些拥有最佳 evals（评测）、最强 workload（工作负载）路由能力，以及其商业模式与客户财务目标直接对齐的公司，将处于非常有利的位置。

Coding is basically the pinnacle of what you could reasonably automate with AI, and yet we still need human engineers to oversee agents for them to be effective. The AI models are trained on an incredible amount of sophisticated code. The users are highly technical and can use the latest tools quickly. The work is “verifiable” because you can test an app. The outcomes are often removed from the quality of the code (you can have sloppy code but the app can still work). And the context for the agent is often already digitized and sitting in the codebase. That’s an incredible amount of benefits that AI coding agents get to work with. Some of those apply to knowledge work, but most don’t in areas where the work needs to be fully reviewed to be useful, or where data isn’t as abundantly digitized. This makes the job for agents in knowledge work more complicated. So if with all of that, engineers still remain in very high demand, the risks are going to be less than what’s perceived for other areas of knowledge work. Agents will let people do far more than they did before, but the people don’t go away.

编程基本上已经是你可以 reasonably（相当合理地）用 AI 自动化的顶峰了，但即便如此，我们仍然需要人类工程师来监督 agent，agent 才能真正有效。AI 模型是在数量惊人的复杂代码上训练出来的。用户本身也高度技术化，能够快速使用最新工具。这项工作是“verifiable”（可验证的），因为你可以测试一个 app。最终结果往往与代码质量并不完全绑定（代码可以写得很 sloppy，但 app 仍然能运行）。而且，agent 所需的上下文常常早已数字化，并且就存在于 codebase 中。这些都是 AI coding agent 能够利用的巨大优势。其中有些优势也适用于知识工作，但在大多数领域并不成立，尤其是那些必须经过完整审查才有用的工作，或者数据还没有被如此充分数字化的领域。这就让 agent 在知识工作中的任务变得更加复杂。所以，如果在具备所有这些有利条件的情况下，工程师依然处于非常高的需求之中，那么其他知识工作领域面临的风险就会比人们想象的小。agent 会让人们做到比以前多得多的事，但人并不会消失。

Good thought provoking post from Anthropic. I think this paragraph points to the key element of the optimistic scenario of AI: “There has been an explosion of new ideas, initiatives, tools, and simulations, as a result of Anthropic employees working with highly capable models—far more than we have the capacity to pursue. The rate at which organizations can spot and fix these bottlenecks may be a skill that improves over time, and it may become the most important skill for any organization.” AI lowers the barrier dramatically to allowing us to do more. As a result of that, we have far more ideas than we can pursue, and for the ones that we want to pursue we’re ultimately limited by our ability to go take on the surrounding work to execute those ideas. There’s almost no amount of AI progress that can happen where that goes away. AI is going to let us build much more software, launch more marketing campaigns, research more drugs, and so on. All of this work, even when augmented by agents, still ultimately requires people to manage.

Anthropic 这篇帖子很发人深省。我认为，这一段指出了 AI 乐观情景中的关键要素：“随着 Anthropic 员工与能力极强的 model（模型）协作，新的想法、倡议、工具和模拟大量涌现——多到远远超出我们能够推进的容量。组织识别并修复这些瓶颈的速度，可能会成为一种会随着时间推移而提升的能力，并且可能成为任何组织最重要的能力。”AI 大幅降低了让我们做更多事情的门槛。结果就是，我们拥有的想法远多于能够真正推进的数量；而对于那些我们确实想推进的想法，我们最终仍然受限于自己是否有能力去承担并完成其周边所需的工作，从而把这些想法执行出来。几乎无论 AI 取得多大进展，这一点都不会消失。AI 将使我们能够构建更多软件、发起更多营销活动、研究更多药物，等等。所有这些工作，即使有 agent（智能体）增强，最终仍然需要由人来管理。

The jobs data coming out continues to suggest the opposite of what a lot of people had thought would happen. Just take engineering, as the prime example of the area with greatest AI impact (and perceived risk). Most companies now have far more software projects than ever before because of AI, and effectively only engineers are going to be the ones doing that work. You can get by for a while by being non-technical building software, but eventually someone has to understand what the thing is that got built, has to maintain it, has to fix security issues that come up, upgrade the systems beneath it, and so on. That’s all jobs. Now apply that to a number of other job functions. AI is going to cause companies to hire more in sales because agents can let them process more leads and do more customer research. AI will cause an explosion of new marketing roles because of how much more efficient it is to launch campaigns and target. The list goes on. AI is going to have the opposite effect that lots of people thought on jobs.

持续公布的就业数据表明，情况恰恰与很多人原先以为会发生的相反。就拿 engineering 来说，它是 AI 影响最大（也是大家感知中风险最高）的领域中的典型例子。现在，大多数公司因为 AI 而拥有比以往多得多的软件项目，而实际上，真正会去完成这些工作的只会是 engineers。你也许可以在一段时间里以非技术身份去构建软件，但最终总得有人理解被构建出来的东西到底是什么、负责维护它、修复出现的安全问题、升级其底层系统，等等。这些全都是工作岗位。再把这个逻辑应用到许多其他职能上。AI 会让公司在 sales 上招聘更多人，因为 agents（智能体）能让他们处理更多 leads（销售线索），并做更多客户研究。AI 还会催生大量新的 marketing 岗位，因为它让发起 campaign（营销活动）和做 target（定向投放）变得高效得多。这样的例子还可以一直列下去。AI 对就业的影响，将会与很多人原先所想的完全相反。

Even with employer caps, the spend on AI tokens dramatically exceeds any other historical spend on software. Typically, companies maybe would spend on the order of $10-50 for a software license per month per employee, but now will pay hundreds or thousands on tokens to augment their productivity. This shows you how big the TAM for intelligence is in the enterprise. The markets for AI are going to dramatically expand the size of the traditional software markets over time.

即便存在 employer caps（雇主支出上限），企业在 AI tokens（令牌）上的花费，仍然显著超过历史上任何其他软件支出。通常来说，公司过去给每位员工支付的软件 license（许可证）月费，大概也就是 10 到 50 美元这个量级；但现在，它们会为 tokens 花费数百甚至数千美元，以提升员工生产力。这说明，在企业市场中，intelligence（智能）的 TAM（总可服务市场）有多么庞大。随着时间推移，AI 市场将显著扩大传统软件市场的规模。

As token budgets take on a larger part of operating expenses over time, model routing is the inevitable conclusion. This is also one of the biggest areas of differentiation for the applied AI layer over time. By understanding the different work patterns in your domain, and having strong evals for that domain, you’ll be able to cost/performance optimize effectively. We’re still likely at the point where most use-cases will need frontier performance for the foreseeable future; but soon you will be able to peel off individual use-cases and send them to lower cost models once the quality is sufficient for the task. Enterprises individually trying to figure this out themselves at scale will likely not be possible, so the products that can intelligently route these workflows to the right tier of model will be in a strong position to aggregate more demand.

随着 token 预算在运营支出中长期占据越来越大的比例，model routing（模型路由）将成为不可避免的结论。这也将是 applied AI layer（应用型 AI 层）长期最重要的差异化领域之一。通过理解你所在领域中的不同工作模式，并为该领域建立强有力的 evals（评估），你就能够有效地在成本与性能之间做优化。我们目前很可能仍处在这样一个阶段：在可预见的未来，大多数 use-case（用例）仍然需要 frontier performance（前沿性能）；但很快，一旦质量足以胜任任务，你就可以把单独的 use-case 剥离出来，发送给成本更低的模型。各家 enterprise（企业）如果想各自独立地、大规模地把这件事摸索清楚，可能并不现实，因此，能够智能地将这些工作流路由到正确模型层级的产品，将处于能够聚合更多需求的有利位置。

This is effectively the #1 problem for AI agents in the enterprise. As we go from agentic coding (where a large amount of context is in the code base, and users are technical enough to get the rest to the agent easily) to a world of knowledge work agents, the context problem becomes much more acute. We see this every day with customers at Box. For existing digital knowledge, it’s often fragmented across legacy systems or environments that don’t play nice with agents, and have access controls that don’t map to the real work that needs to be done, which become a huge hurdle for getting agents the context they need. This has to all get moved to modern, secure cloud environments. But also, companies often haven’t captured and digitized some of the critical context that agents need to work with. Decisions, processes, and workflows often live in people’s heads and tribal knowledge that need to get turned into unstructured data for agents. This is actually one of the biggest points of leverage for applied AI companies, because they can work to specialize in getting agents exactly the information and domain expertise they need. But it’s also one of the reasons why FDEs and new system integrator plays will also work so well right now. The companies that figure this out will be able to get the most out of AI going forward.

这实际上是企业中 AI agents 面临的头号问题。随着我们从 agentic coding 走向知识工作 agents 的世界，context（上下文）问题会变得更加尖锐——在 agentic coding 里，大量 context 都在代码库中，而用户也足够懂技术，能比较容易地把其余信息提供给 agent。我们在 Box 与客户的日常接触中每天都能看到这一点。对于现有的数字化知识，它往往分散在各类 legacy systems（遗留系统）中，或者存在于那些无法很好配合 agents 的环境里；同时，这些环境的访问控制也常常无法映射到实际工作真正需要完成的方式，这就成为让 agents 获得所需 context 的巨大障碍。这些内容都必须迁移到现代化、安全的云环境中。但除此之外，很多公司其实还没有把 agents 开展工作所需的一些关键信息捕捉下来并完成数字化。决策、流程和工作流往往存在于人的脑子里，或者存在于 tribal knowledge（组织内的隐性经验）中，而这些都需要被转化为 agents 可用的 unstructured data（非结构化数据）。这其实是 applied AI companies 的一个最大杠杆点之一，因为它们可以专门去解决如何把 agents 真正需要的信息和领域专长准确地提供给它们。但这也是为什么 FDEs 以及新型 system integrator 模式在当下会如此有效的原因之一。那些把这个问题解决好的公司，未来将能够从 AI 中获得最大的收益。

Again, maybe counterintuitive, but in the majority of conversations I have with CIOs, CTOs, and CEOs in large enterprises, they are either growing due to AI (in new job functions like FDEs, engineering, etc.) or at a minimum reinvesting efficiency savings back into the business in new areas (sales, marketing, etc.). David Solomon, CEO of Goldman Sachs, articulated this perfectly in a NYTimes OpEd last week. The AI boom is both creating all new jobs in the build out of AI systems and the implementation across sectors, but also freeing up dollars to invest in areas that have been underfunded or have more demand now because of AI. Most businesses have been constrained by how much software they can produce at a given cost, how many sales reps they can hire, how many marketing campaigns they can run, how they can do outbound customer success motions with enough tailoring, how they can find more risk in their business and prevent it, and 100s of other things. When AI makes it possible to do more of this, investment goes back into the business. The companies that better serve their customers win over the long run, and those that just try and find savings end up doing worse.