中文改写

人们习惯把以太坊叫作“区块链”，好像它和中本聪 2009 年创造的比特币本质上是同一种技术。Vitalik Buterin 认为，在很多方面确实如此；即便是正在按 Strawmap 路线图建设的未来“精简以太坊”（lean Ethereum），也保留了区块链的核心特征。但这项技术在过去十五年里已经演化了很多，未来三年还会演化得更多，到那时，把以太坊正在走向的东西称为一种性质不同的系统，是说得通的。

今天的以太坊有通用计算、权益证明、使用零知识证明的链上应用，以及提供扩容和隐私的 L2。明天的以太坊会有一个可调的计算旋钮，在极致规模和完全通用之间取舍；会有多种形式的多方共同出块；会有高度优化的权益证明；零知识证明会内置到基础层，并承担关键角色。

这篇文章要梳理的，是 2010 年的区块链和 2030 年的区块链之间最重要的几项根本差异，既从技术角度看，也从用户能期待哪些性质的角度看。

拿比特币白皮书逐节对照

他先把比特币白皮书逐节过了一遍，在截图上用三种颜色标出以太坊在 2015 年（黄色）、2025 年（绿色）和 2030 年（蓝色）的对应情况，再在旁边写批注。

五张截图里的批注要点如下（文字取自图中批注）：

• 第 2 节“交易”：白皮书把电子币定义为一串数字签名，每个所有者通过对上一笔交易的哈希签名来转移币。批注指出，状态和历史现在已经明确分开；签名将在链下聚合，每个区块在链上只留一个，而且很多时候会换成零知识证明。
• 第 4 节“工作量证明”：以太坊自 2022 年起改用权益证明，用验证者签名代替寻找 nonce。交易不再作为一整块数据打包，签名、证明等依赖会被拆出来，在内存池里聚合。白皮书隐含的假设是区块由一个参与者构建；现实中这一点已经被大幅削弱，比如权限有限的矿池、协议内的提议者与构建者分离，而 FOCIL 及之后可能的扩展，会把多方分权出块更深地写进协议。
• 第 5 节“网络”：白皮书缺少的部分包括内存池层面的聚合（EIP-8288 之后，只有入口节点能看到单个签名和证明）、发送者匿名（混合网络等）和分布式出块。有了 PeerDAS，每个节点只需下载每个区块的一小部分，目前覆盖 blob，以后会覆盖完整区块内容。以太坊的权益证明里，出块和分叉选择规则大体上是分开的，分叉选择主要依据证明投票而不是区块；投票是并行到来的，所以共识延迟大幅降低。
• 第 7 节“回收磁盘空间”：这一点已经做到，方式是把历史（交易）和状态明确分开。新的节省空间策略是，验证不再需要存储状态，改用 SNARK；很多情况下也不需要保存哈希，尤其是有了 SNARK 和 BAL 之后。白皮书还缺少对不同存储类型（数据库、平面文件等）及其优化取舍的明确考虑，以及分布式状态存储。
• 第 10 节“隐私”：白皮书的思路是公开所有交易，同时保持公钥匿名。批注指出，使用 ZK-SNARK 的隐私协议已经打破了“必须公开所有交易”这一前提，FOCIL 和 EIP-8288 会让这类交易更像一等公民；早在十年前就已经很清楚，有了现代数据分析工具，白皮书那种程度的隐私不够用。缺少的还有读取隐私（防止 RPC 查询暴露用户）和广播隐私（可以用混合网络等解决）；可编程隐私的最佳方案仍是开放问题。

几乎每一节都有大的变化。为了更紧凑，他把这些变化整理成了一张表：

区块链的核心属性几乎都在变

在他看来，区块链的几乎每一项核心属性，要么已经根本改变，要么即将根本改变：

• 验证：下载并重新执行 → 用 PeerDAS 抽样，再验证一个 SNARK
• 共识：PoW → PoS → 优化得多的 PoS
• 出块权：单个矿工出块 → 多方共同构建区块

他套用 AI 的口吻打趣说，唯一“诚实”的结论（好吧，是“诚实的笑点”）是：像完成 Lean 升级之后的以太坊这样的现代密码学网络，之所以还叫“区块链”，很大程度上是历史原因。实际上它是两股脉络合成的混合结构：

1. 中本聪的核心思想
2. 学术界五十年积累下来的强大密码学工具，它们在 2009 年还不存在，或者还不成熟

原文接着用一张时间线图回答“加密货币里到底有多少密码学”，对比 2009、2020 和 2030 年。图的左侧是学术成果：哈希与签名（1970 年代末）、椭圆曲线（1985）、部分同步条件下的 DLS 共识（1988）、PCP 定理（1990）、私有信息检索（1995）、洋葱路由（1995），以及 2010 年代初第一批可用的 SNARK 和 2010 年代末第一批可用的 STARK。右侧是实际系统：BitTorrent（2001）、可复用工作量证明（2004）、比特币（2009）、以太坊（2013）、Zcash（2016）、2017 至 2018 年定型的现代区块链 PoS 设计，以及标为“未来”的终点，左侧多项成果都有箭头指向它。

密码学并不是唯一重要的学科。形式化验证、数据库理论、点对点网络理论、信息论、经济学等等同样重要。但这些学科都和那套基本内核兼容：每个人都尝试用有效的工作量证明做出下一个区块，一个人成功并广播，其他人下载并重新执行，循环往复。密码学带来的变化与这套内核不兼容。

用户面对的取舍正在改变

最重要的结论是，用户面对的那套取舍正在发生根本变化。他用一张表对比了 2015 年和 2030 年以太坊的优点与缺点：

构建应用时，计算的结构开始变得很重要。在简单的区块链里，1 字节就是 1 字节，1 gas 就是 1 gas。在未来的架构里，同样多的计算，如果全塞进一笔难以拆解、串行执行的交易，成本会高得多；如果拆成封装良好、可以并行处理或剪枝的依赖，最好在交易进入最终区块之前就处理掉，成本会低得多。这会改变开发者的激励，也会逐渐改变所有以太坊应用的结构。他的推测是：长期来看，编程模式也许会收敛到这样一种形态：只有直接描述不可交换的状态变化和顺序的信息上链，其余一切在进入区块之前就被聚合掉。

示意图的图注是：对计算做结构化，能让区块链更有效地专注于它的本职工作。

去中心化开始带来性能

也许最重要的转变是，网络的去中心化，正从单纯为安全和稳健付出的代价，变成至少在少数有限场景下、连性能角度都算优势的东西。去中心化网络能并行存储更大量的数据，能并行完成大量计算，很多时候就在内存池里完成。少数情况下它还能提升隐私，因为只有去中心化网络才能有效隐藏元数据，比如数据和请求来自哪里。

回到 2010 年代中期，这曾是以太坊早期的一个梦想：去中心化不只是为了稳健，也是为了扩大规模。中心化系统能把工作分给不同参与者来提升性能，以太坊也应该能。当时做不到，主要原因只有一个：缺少验证。把工作拆开，就得验证每一份工作都做对了。早期设计尝试用随机抽样的委员会来补这个洞，但都卡在同一个瓶颈上：委员会搭建复杂、成本高，还会大幅增加延迟；而且委员会一旦失灵，没有任何补救手段。如今有了现代密码学，这个问题已经解决，这套方案的额外开销还在逐月下降。

去中心化有望改善性能的另一个方向是延迟。以太坊本身的延迟永远比不上服务器，但围绕它搭建的基础设施有可能做到。

总体来看，在用户和链之间建一层更强的去中心化中间层，而这一层本身不是链，可以让以太坊变得非常强大，同时不损害链的任何基本属性。

更远处的混淆技术

再往后，以太坊可能还会经历一次转变：混淆技术（iO）的兴起，这是他此前三篇文章讨论过的方向。这里的终极目标是，可用的混淆技术能消除隐私和通用性之间的取舍：在完全安全、加密的形式下进行完全通用的计算，参与方数量不受限制，而且可以异步参与。即使是较弱版本的混淆，也有很多用途，比如加密内存池。不过他强调，这篇文章的所有结论，在这些技术可用之前很久就已经成立。

什么是“密码学世界计算机”

这就是他所说的“密码学世界计算机”：以太坊从一个只能把计算和数据不加区分地丢上去执行的账本，转向一种把区块链与密码学隐私、密码学验证以及强大的去中心化链下组件结合起来的架构。

要完整建成这套设计，挑战还很多。让零知识证明足够高效、足够安全很难，但这属于封装起来的复杂性，而且已经在借助 AI 工具大力优化。更难、系统复杂度也更高的部分，可能是管理并并行化对海量状态的访问。这方面已经有不少思路，但还需要打磨，尤其要随着对未来应用的了解加深而调整。

按 Strawmap，明年计划中的分叉 Hegota 很可能是以太坊最后一次“常规”分叉：它的功能、技术和开发流程，2015 年的人都还认得出来。之后的一切都涉及递归 STARK、自动化形式化验证、高度优化的共识算法，以及让这一切抗量子。以太坊从 PeerDAS 开始，就已经从单纯的区块链转向更强大的东西；Hegota 之后，这场转型会成为以太坊的主线故事。最终结果是：比上一代技术单靠自身所能做到的，便宜得多、可扩展得多、也私密得多的高安全性计算。这就是密码学世界计算机。

文中对以太坊应用开发者最直接可用的一条是：把签名、证明这类可以并行处理或提前聚合的部分拆出来，只把描述顺序和不可交换状态变化的信息留在链上，按他的判断，这样的结构在未来的架构里成本会低得多。

原文

原文的九张配图以内嵌 data URI 形式发布，没有独立图片地址；本站逐字节转存到 /media/ 并在上方媒体区展示，正文中以 [Figure N] 标出原位置。原文 Markdown 取自作者博客仓库 vbuterin/blog 的 posts/the_cryptographic_world_computer.md。

The cryptographic world computer

We talk about Ethereum as “a blockchain”, as though it is fundamentally the same technology as the Bitcoin that Satoshi Nakomoto created back in 2009. In many ways it is, and in some ways even the “lean Ethereum” of the future that is being built along the lines of the Strawmap retains the core trappings of a blockchain. But at the same time, the technology has evolved greatly over the last fifteen years, and is poised to evolve even more over the next three - to the point where it becomes fair to call the thing that Ethereum is moving towards a qualitatively different kind of system.

The Ethereum of today has general-purpose computation, proof of stake, onchain applications using zero-knowledge proofs and L2s that provide scaling and privacy. The Ethereum of tomorrow will have computation with a tunable dial between extreme scale and full generality, multiple forms of multi-participant block construction, a highly optimized form of proof of stake, and zero-knowledge proofs baked in playing key roles at the base layer.

This post will walk through some of the most important fundamental differences between blockchains in 2010 and blockchains in 2030, both from the point of view of the technology and from the point of view of what properties you can expect to have as a user.

To start off, let’s go through the original Bitcoin whitepaper, section by section, and see how this compares to Ethereum - in 2015 (yellow), 2025 (green) and 2030 (blue).

[Figure 1: inline image in the original post, shown in the media section above]

[Figure 2: inline image in the original post, shown in the media section above]

[Figure 3: inline image in the original post, shown in the media section above]

[Figure 4: inline image in the original post, shown in the media section above]

[Figure 5: inline image in the original post, shown in the media section above]

There are large changes to basically every section. To make things more compact, let’s make it a table:

[Figure 6: inline image in the original post, shown in the media section above]

Pretty much every core property of what it means to be a blockchain either has fundamentally changed or soon will fundamentally change:

• Verification: download and re-execute → PeerDAS sample and verify a SNARK
• Consensus: PoW → PoS → much better optimized PoS
• Block construction authority: single miner makes block → multi-party block construction

As an AI might say, the only honest conclusion (ok fine, the honest punchline) is that a modern cryptographic network like the post-Lean-upgrade Ethereum is called a “blockchain” to a large extent for historical reasons. In reality it’s a hybrid construction that merges together two strands:

1. Core Satoshian ideas
2. New powerful cryptographic machinery that emerged out of 50 years of academia and did not even exist (or was not mature) in 2009

[Figure 7: inline image in the original post, shown in the media section above]

How much crypto is in crypto? 2009 vs 2020 vs 2030

Cryptography is not the only science that matters. Also important are: formal verification, database theory, peer-to-peer networking theory, information theory, economics and much more. But all of those things are compatible with the fundamental core of “everyone tries to make the next block with valid PoW, one guy succeeds, broadcasts it, everyone else downloads it and re-executes, repeat”. The cryptographic changes are not.

Now, what does this mean for users?

The most important conclusion is that the set of tradeoffs for users is radically changing:

[Figure 8: inline image in the original post, shown in the media section above]

When building applications, structure of computation is starting to matter a lot. In a simple blockchain, 1 byte = 1 byte and 1 gas = 1 gas. In the architectures of the future, the same amount of computation will cost you much more if you shove it all into one inscrutable serially-executed transaction, and much less if you put it into well-encapsulated dependencies that can be parallelized or pruned, ideally before the transaction even enters the final block. This affects the incentives of developers, and will over time affect the structure of all applications using Ethereum: perhaps in the long run, we will converge toward programming patterns where information directly related to describing non-commutative state changes and ordering is posted onchain, and everything else is aggregated before it even gets included into a block.

[Figure 9: inline image in the original post, shown in the media section above]

Structuring computation lets the blockchain more effectively focus on its job.

Perhaps the most important shift is that the decentralized nature of the network is shifting from being purely a burden incurred in the name of safety and robustness to at least sometimes, in a few limited cases, being a strength even from a performance perspective. The decentralized network allows larger volumes of data to be stored in parallel. It allows a high volume of computation to happen in parallel, in many cases inside the mempool. In a few cases, it increases privacy, because only decentralized networks can effectively hide metadata (eg. where data and requests are coming from).

Back in the mid-2010s, this was an early dream of Ethereum: that we could decentralize not just for robustness but also to increase scale. If centralized systems can split up work between different participants to increase their performance, so should we. Back then, this was not viable for one primary reason: the missing ingredient was verification. If you split up work, you have to verify that each unit of work was done correctly. Early designs attempted to patch this with randomly sampled committees, but they all ran into the same bottleneck: first, committees are complicated to set up and expensive and greatly increase latency, second, there is no recourse if the committee fails. Now, with modern cryptography, this problem is solved, and the overhead factor of that solution is decreasing month by month.

Another area to watch for where decentralization can potentially improve performance properties is latency. Ethereum itself will never have latency that competes with servers, but infrastructure built around it could.

In general, creating a stronger decentralized layer in the middle between users and a chain, that is not itself a chain, could make Ethereum very powerful without compromising any of the chain’s fundamental properties.

Later in the future, there is the possibility that Ethereum will go through one further shift - the possible rise of obfuscation (iO). The holy grail here is that viable obfuscation can eliminate the tradeoff between privacy and generalization: you can have fully generalized computation involving an unbounded number of (asynchronous) participants in a fully secure and encrypted form. Even weak versions of obfuscation have many applications in eg. encrypted mempools. But all of the conclusions in this post will apply long before any of that becomes available.

This is the "cryptographic world computer”: a shift from Ethereum as solely a ledger onto which you can indiscriminately dump computation and data that gets executed, to an architecture that combines blockchains with cryptographic privacy and verification, and powerful decentralized off-chain components.

There are still many challenges in fully building out this design. Making zero knowledge proofs efficient enough and safe enough is difficult, but it is encapsulated complexity and is already being heavily optimized with AI tools. The more difficult, and systemically complex, piece will likely be managing and parallelizing access to very large amounts of state. There are already many ideas on the horizon for how to handle this, though they need to be refined especially as we learn more about what applications will be running in the future.

If you look at the Strawmap, Hegota - the fork planned for next year - is likely to be Ethereum’s last “normal” fork, with features and technology and development process that would be recognizable to someone in 2015. Everything after that involves recursive STARKs, automated formal verification, highly optimized consensus algorithms, and making it all quantum-safe. With PeerDAS, Ethereum began its transition from being just a blockchain to being something much more powerful. Starting after Hegota, this transformation becomes Ethereum’s primary story. The final outcome of this: much more cheap, scalable and private high-security computation than anything that could be done with the previous era’s technology alone. The cryptographic world computer.

原文链接：https://vitalik.eth.limo/general/2026/09/27/the_cryptographic_world_computer.html

相关链接：https://strawmap.org/

中文改写

2025 年 9 月，Apollo Global Management 发过一张图，标题是“大公司 AI 采用率呈下降趋势”。数据来自美国人口普查局，看上去显示美国企业的 AI 采用在放缓。Simon Willison 当时用这张图演示过 GPT-5 的数据分析能力，后来又拿它测试 Claude 新出的 Code Interpreter。

一年后，他想用新数据再跑一遍，看看过去一年发生了什么变化。结果他发现，人口普查局的问卷受到了第 14434 号行政令的影响，也就是他口中那份“荒唐的”超级智能行政令。

这份调查怎么做

人口普查局每两周做一次商业趋势与展望调查（Business Trends and Outlook Survey，简称 BTOS）。每次发给 120 万家企业中的 20 万家；被调查的这组企业叫一个“面板”，轮换安排，所以单个企业每十二周才填一次问卷。

最新数据隔周的周四上午 10 点发布。

新图：大企业的比例从约 12% 到约 40%

他用 2026 年 10 月 22 日发布的最新数据，借助 GPT-6 Astra Extra High 重画了这张 AI 采用率图（对话记录在原文中有链接），时间窗口更新到最近十二个月，并且把人口普查局改过的问法也反映了出来。按原文配图的说明文字，这张图画的是“过去两周内报告使用了 AI 的雇主企业占比”，按企业员工规模分成 1 至 4 人、5 至 9 人、10 至 19 人、20 至 49 人、50 至 99 人、100 至 249 人、250 人以上七条线，每个点是连续六期调查的平均值。

图的左半部分对应旧问法“在生产商品或提供服务中使用 AI”：2023 年底各组都在 3% 到 5%，到 2025 年 9 月缓慢升到 7% 到 13%，250 人以上组最高。中间是一段空白，标注为“2025 年 11 月问题范围放宽”和“序列不可比”。右半部分对应新问法“在任何业务职能中使用 AI”：各条线从 2026 年 2 月起在更高的位置重新开始，并稳步上升，最新一期分别是：250 人以上 40.7%，100 至 249 人 33.5%，50 至 99 人 27.9%，20 至 49 人 23.8%，1 至 4 人 23.3%，10 至 19 人 22.1%，5 至 9 人 21.1%。

图下的脚注交代了算法：六期调查估计值的算术平均，按两种问法分别计算，不对断点做衔接处理；日期取调查参考期的结束日，和原图一致；最新一期参考期截至 2026 年 10 月 4 日，在 10 月 5 日至 18 日收集；新问法下第一个六期均值落在 2026 年 1 月 25 日。中间的空白既包括 2025 年的政府停摆，也包括新问法下凑齐六期数据所需的等待时间。图上没有显示调查误差。原文还附了大图链接。

作为对照，他也贴出了去年用 GPT-5 画的旧版图。按配图说明，那张图从 2023 年 11 月开始，各规模企业的比例在 3% 到 5% 之间，到 2025 年 8 月都在上升；250 人以上组一直最高，2025 年 7 月达到约 14% 的峰值，到 8 月回落到约 12%，其他几组也有轻微回落，但幅度更小。

Simon 把两组数字摆在一起：2025 年 9 月，250 名员工以上的大企业在“生产商品或提供服务”这个问题上回答“是”的比例约为 12%；换成“任何业务职能”的新问法后，同一批大企业到 2026 年 2 月约有 30% 回答“是”，到 2026 年 10 月升到约 40%。

图中数据来自人口普查局数据下载页的“Employment size class”（按员工规模分类）XLSX 文件，Simon 在原文里放了自己所用文件的副本。发给企业的问卷原件也可以查看，他同样存了一份 346KB 的 PDF 快照。

问卷措辞的第一次变化

这张图本身已经说明，人口普查局的问题在 2025 年 11 月 17 日发生了实质性变化。在那之前，问题是这样问的（中文为编辑译文，加粗为 Simon 所加）：

过去两周内，该企业是否在生产商品或提供服务中使用了人工智能（AI）？（AI 示例：机器学习、自然语言处理、虚拟代理、语音识别等）

2025 年 12 月 3 日，人口普查局宣布对这个问题做了重大修改，新问题是：

过去两周内，该企业是否在其任何业务职能中使用了人工智能（AI）？（AI 示例：机器学习、自然语言处理、虚拟代理、语音识别等）

“生产商品或提供服务”变成了“任何业务职能”。Simon 指出，11 月 17 日前后的答案因此不能直接比较。前面那组从 12% 到 40% 的数字，就跨过了这个断点。

从 2026 年 10 月 19 日起，问的是“超级智能”

Simon 去最新的 XLSX 文件里找更新后的问题措辞，没有找到原来的问法，找到的是这一段（中文为编辑译文，加粗沿用原文）：

过去两周内，该企业是否在其任何业务职能中使用了超级智能（Super Intelligence，SI）（以前称为人工智能或 AI）？（SI 使用示例包括：SI 驱动的数据分析；用于营销、产品设计或编程的生成式 SI；用于客户服务的虚拟 SI 代理；SI 驱动的预测或定价；用于质量控制的 SI 驱动计算机视觉等）

BTOS 的说明页交代了原因：从 2026 年 10 月 19 日那一轮收集开始，问卷措辞因第 14434 号行政令而更新。

目前已发布的数字仍然对应旧的“人工智能”措辞。第一批反映“超级智能”措辞的结果，按原文的说法会在“周四”发布；按原文发布日 2026 年 11 月 2 日推算，是 11 月 5 日。

改名会不会影响调查结果

Simon 认为，最直接的悬而未决的问题，是这次措辞变化会不会影响调查结果。

回答问卷的企业上一次看到这个问题是在十二周前，那时它还叫 AI。他希望新说法不会让企业困惑，好在问卷特意写明了“以前称为人工智能或 AI”。

不过他也坦言，看到“SI 驱动的数据分析；用于营销、产品设计或编程的生成式 SI；用于客户服务的虚拟 SI 代理”这样的写法，还是忍不住觉得荒唐可笑。

用 BTOS 做企业 AI 采用率的时间序列，至少要按问卷措辞切成三段：2025 年 11 月 17 日和 2026 年 10 月 19 日是两个断点。Simon 的做法是每种问法各自计算六期均值，不跨断点拼接，图上也直接标出“序列不可比”。

原文

原文的两张配图已移到上方媒体区，正文中以 [Image N] 标出原位置；配图的替代文字保存在媒体信息中。

The U.S. Census Bureau and “Super Intelligence”

I was reminded today of AI Adoption Rate Trending Down for Large Companies, a chart produced by Apollo Global Management in September 2025, which appeared to show a slowdown in U.S. corporate AI adoption based on numbers collected by the U.S. Census Bureau. At the time I used that chart to demonstrate GPT-5 data analysis, and then to experiment with Claude’s new Code Interpreter. I decided to try running those numbers again to see what had changed over the past year, and in doing so learned that the Bureau has fallen foul of Executive Order 14434, aka the asinine Super Intelligence one.

First, the updated chart

Every two weeks the U.S. Census Bureau runs a survey called the Business Trends and Outlook Survey (aka BTOS). Each time it goes to 200,000 businesses out of 1.2 million total—the surveyed group (a “panel”) cycles such that an individual business only completes the survey every twelve weeks.

The latest data is published on alternating Thursdays at 10am.

I used the most recent numbers (published October 22, 2026) to create a new version (GPT-6 Astra Extra High transcript) of the AI adoption chart, updated to reflect both the last twelve months and a change that was made to the survey question sent out by the Bureau:

[Image 1: updated AI adoption chart, shown in the media section above]

Here’s a link to the larger image.

For comparison, here’s that AI adoption chart from last year (my GPT-5 version): [Image 2: the 2025 GPT-5 version of the chart, shown in the media section above]

In September 2025 the largest (250+ employees) companies answered “yes” around 12% of the time to the question about whether they used AI in “producing goods or services”.

Under the new “any business function” wording, by February 2026, those same large companies answered “yes” around 30% of the time, and by October 2026 that had increased to around 40%.

The data in the above chart comes from the “Employment size class” XLSX file, available from here. Here’s a copy of the file I used.

You can view the survey document that is sent out to businesses—here’s my snapshot copy of that 346KB PDF.

The changing language of the questions

As the chart demonstrates, the question asked by the Bureau changed in a material way on November 17th, 2025. Prior to that date, the question they asked was this (emphasis mine):

In the last two weeks, did this business use Artificial Intelligence (AI) in producing goods or services? (Examples of AI: machine learning, natural language processing, virtual agents, voice recognition, etc.)

On December 3rd, 2025, the Bureau announced a significant change to that question. Here’s the replacement:

In the last two weeks, did this business use Artificial Intelligence (AI) in any of its business functions? (Examples of AI: machine learning, natural language processing, virtual agents, voice recognition, etc.)

No more “producing goods or services”—now it’s “in any of its business functions”! This means that the answers to the question before and after November 17th are not directly comparable.

From October 19, 2026, they ask about Super Intelligence

When I went looking for the updated question language in the latest XLSX file, I couldn’t find it. I found this instead:

In the last two weeks, did this business use Super Intelligence or SI (formerly referred to as Artificial Intelligence or AI) in any of its business functions? (Examples of SI use include: SI-powered data analysis; generative SI for marketing, product design, or coding; virtual SI agents for customer service; SI-powered forecasting or pricing; SI-driven computer vision for quality control, etc.)

The BTOS about page explains what happened:

Beginning with the October 19, 2026 collection, instrument wording was updated due to Executive Order 14434.

The current numbers reflect the old Artificial Intelligence survey wording—the first survey results reflecting the Super Intelligence wording are due for publication on Thursday.

Is this going to affect the survey?

The obvious open question here is whether this wording change is going to affect the survey results.

The businesses answering this survey will have last seen that question twelve weeks ago, when it was still called AI. Hopefully the new language isn’t going to confuse them, especially since the survey helpfully spells out “formerly referred to as Artificial Intelligence or AI”.

I can’t help but laugh at the absurdity of “SI-powered data analysis; generative SI for marketing, product design, or coding; virtual SI agents for customer service” though.

原文链接：https://simonwillison.net/2026/Nov/2/census-bureau-super-intelligence/

相关链接：https://www.census.gov/hfp/btos/data

中文改写

Meta AI 发布了 Muse Ember，和 Claude Haiku、GPT Luna 同属一档的小模型。Meta 的官方描述是：这是他们最小、最高效的模型，面向复杂推理的执行，以及 Agent 整天重复在做的那些活：工具调用、编程子任务、路由、信息抽取和长文档解析。

Simon Willison 在博客里把它和两个同档对手放在一起，按价格和每任务成本分别算了一遍。

短提示同价，长提示按门槛加价

提示较短时，Muse Ember 的定价和 GPT-6 Luna、Claude Haiku 5.5 完全一样：每百万输入 token 0.10 美元，每百万输出 token 0.50 美元。同一天发布的另一款大模型 Claude Fable 5.2 定价是 8 美元和 40 美元，是这一档的 80 倍。

提示变长后就没这么简单了。Luna、Ember 和 Haiku 都支持最多一百万输入 token，但三家都设了一个长度门槛，超过之后单价上调。把门槛算进去，Ember 的价格落在另外两家中间：

从表里能直接读出三点：Haiku 的门槛最低，10 万 token 就开始加价，加价后也最贵；Luna 的门槛最高，到 27.2 万 token 才加价；Ember 的缓存输入单价在三者里最低，短提示时每百万 token 只要 0.004 美元。

每 token 单价已经不够用

Simon 的看法是，现在按每 token 单价比较模型意义不大。推理 token 的存在，让不同模型解决同一个问题时消耗的 token 数量可能相差很多，单价低的模型未必总成本低。

Simon Willison 给出的第一个参照，是 Artificial Analysis 的每任务成本柱状图。按原文配图的替代文字，六个配置的合计成本从低到高是：Claude Haiku 5.5（low）约 0.04 美元，GPT-6 Luna（不推理）0.07 美元，GPT-6 Luna（xhigh）0.10 美元，Muse Ember（xhigh）0.12 美元，Muse Ember（max）0.14 美元，Claude Haiku 5.5（xhigh）0.25 美元。两个 Haiku 的数字在图上带星号，原文没有解释星号含义。

他认为更有用的是同一网站的“智能指数对每任务成本”散点图。按配图说明，Muse Ember（xhigh）得分约 34，成本约 0.12 美元；Muse Ember（max）得分约 36，成本约 0.14 美元，两者都落在图中的 Pareto 线上。Claude Haiku 5.5（xhigh）得分约 34.5，和 Ember（xhigh）接近，但每任务成本约 0.25 美元，位置在 Pareto 线右侧很远。GPT-6 Luna（xhigh）得分约 23.5，成本约 0.10 美元；Haiku 5.5（low）约 19 分、0.044 美元；Luna（不推理）约 9.5 分、0.07 美元。

Pareto 线指的是这样一条边界：线上的点想在一个维度上变好，只能在另一个维度上让步。Meta 在发布文章里也专门提到了这一点，原话的意思是：Muse Ember 位于 Artificial Analysis 完整排行榜的效率前沿上，没有哪个模型能用更低的每任务成本拿到更高的分数。这是 Meta 的自述，依据是 Artificial Analysis 的榜单数据。

Simon 还顺带提到，图中绿色区域中间、同样落在 Pareto 线上的那个灰点是 GPT-6.1 Terra（low），他觉得自己该多关注一下 Terra 系列。

六档推理画出的鹈鹕

Muse Ember 有六档推理强度：minimal、low、medium、high、xhigh 和 max。Simon 让每一档各画一张“骑自行车的鹈鹕”SVG。high 档没能返回完整的 SVG；其余几张让他意外的是，推理档位高低和画面质量并没有清晰的对应关系。他贴出的是 max 档那一张，配图说明写的是：自行车车架还行，车筐里有鱼，鹈鹕戴着一顶奇怪的蓝帽子。

他的评价是，Ember 画出的鹈鹕和他之前用 Muse Wonder 画的质量相当，而 Muse Wonder 的价格是 Muse Ember 的 40 倍。

在这一档小模型之间做选择时，按 Simon 的思路，先看每任务成本和得分，再看单价；如果业务里的提示经常超过 10 万 token，还要把三家不同的加价门槛一起算进去。

原文

原文的三张配图已移到上方媒体区，正文中以 [Image N] 标出原位置；价格表在原网页中用合并单元格排版，这里每个模型的第二行留空模型名。

Introducing Our Smallest, Most Efficient Model: Muse Ember

Introducing Our Smallest, Most Efficient Model: Muse Ember. Meta AI just launched their entry in the Claude Haiku/GPT Luna category: Muse Ember, which they describe as:

[...] our smallest and most efficient model designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.

For shorter prompts it's priced the same as GPT-6 Luna and Claude Haiku 5.5 - both are $0.10/million input tokens and $0.50/million output. Today's other big model release, Claude Fable 5.2, is 80x that price at $8/$40.

The pricing is a little less straightforward as prompts get longer, though. Luna, Ember, and Haiku all support up to a million input tokens, but they also all have a threshold after which the price goes up. Taking that threshold into account Ember is priced between the two others:

But comparing price-per-token doesn't really make sense these days, when reasoning tokens mean that different models might use dramatically different numbers of tokens to solve the same problem.

For that, one measure that's useful is the Artificial Analysis.

[Image 1: stacked bar chart, shown in the media section above]

And even more useful than that is their Intelligence Index vs. Cost per Task chart, which shows the following for those three models:

[Image 2: Intelligence Index vs. Cost per Task scatter plot, shown in the media section above]

The Muse Ember models sit right on that Pareto line - the line at which you can only improve one dimension by giving up another. Meta call this out in their announcement:

It also sits on the efficiency frontier of the full Artificial Analysis leaderboard: no model scores higher for less per task.

(That grey dot on the Pareto line in the middle of the green area is GPT-6.1 Terra (low) - clearly I should be paying more attention to the Terra family!)

Muse Ember has six reasoning levels - minimal, low, medium, high, xhigh, and max - so I rendered six pelicans riding bicycles. The "high" reasoning level failed to return a full SVG, and the other SVGs were surprising in that the reasoning level didn't cleanly map to the visual quality of the response.

Here's the "max" one, but the others are worth a look too:

[Image 3: the "max" pelican SVG, shown in the media section above]

The Ember pelicans are comparable in quality to those I got from Muse Wonder, which is 40x the price of Muse Ember.

原文链接：https://simonwillison.net/2026/Nov/10/muse-ember/

相关链接：https://research.meta.ai/blog/muse-ember