2026 in LLMs (so far)
打开互动全文版(逐段中英对照 + 针对本篇提问)→本演讲综述 2026 年大语言模型的主要进展,从 2025 年 11 月发布的 Claude Opus 4.5 和 GPT-5.1 谈起——这两款模型将编程智能体从容易出错提升到足以日常使用的可靠程度。核心论点是:2026 年标志着 AI 通过编程智能体实现了产品市场契合,由此引发了 OpenClaw 等“Claw”个人智能体的爆发、一波 tokenmaxxing 热潮及其后由成本驱动的收缩,以及关于代码编写与审查均无需人类参与的全自动软件工厂的争论。演讲者还讲述了自己对 AI 的狂热、过于雄心勃勃的 vibe-coded 项目,以及 AI 给软件工程师带来的倦怠感。结论是:要探明这项技术的极限,唯一的方法就是不断去突破它;而一切变化如此之快,整个行业至今仍在消化这意味着什么。
This talk surveys the major developments in large language models during 2026, beginning with the November 2025 releases of Claude Opus 4.5 and GPT-5.1, which pushed coding agents from error-prone to reliable enough for daily use. The core argument is that 2026 marked the year AI achieved product-market fit through coding agents, triggering an explosion of "Claw" personal agents such as OpenClaw, a wave of tokenmaxxing followed by cost-driven retrenchment, and debates over fully automated software factories where neither code nor review involves humans. The speaker also describes his own AI mania, over-ambitious vibe-coded projects, and the AI-induced ennui affecting software engineers. The conclusion is that the only way to find the limits of this technology is to keep pushing them, and that everything changed so quickly that the profession is still coming to terms with what it means.
周五,我在圣何塞举行的 WeAreDevelopers World Congress North America 上做了闭幕主题演讲。我将过去一年的关键趋势串联起来,按时间顺序梳理了 2026 年发生的一切。视频已发布在 YouTube 上;以下是我为配合演讲而准备的带注释幻灯片和笔记。
On Friday I gave the closing keynote at the WeAreDevelopers World Congress North America in San Jose. I tied together the key trends from the past year into a chronological exploration of everything that happened in 2026. The video is on YouTube; here are my annotated slides and notes to accompany the talk.
我将带大家快速巡览 2026 年迄今为止发生的一切。这一年还没有结束!
I'm going to give a lightning tour of everything that has happened so far in 2026. The year isn't over yet!
对我来说,2026 年早在两个月前、也就是 2025 年 11 月就开始了。
For me, 2026 started a couple of months earlier, in November 2025.
11 月发布了两款重要模型:Claude Opus 4.5 和 GPT-5.1。
November saw the release of two important models: Claude Opus 4.5 and GPT-5.1.
正如新模型通常的情况一样,它们是对前代模型的渐进式改进。
As is usually the case with new models, these were incremental improvements on the models that came before them.
但每隔一段时间,当模型取得进步时,它就会越过一条看不见的界线——原本不太行得通的事情开始行得通了。
But every now and then, when a model improves, it crosses an invisible line where something that didn't really work starts working.
在这个案例中,开始行得通的是它们的编程智能体。Claude Code 自 2025 年 2 月就已问世;Codex 则稍晚一些。
In this case, the thing that started working was their coding agents. Claude Code had been around since February 2025; Codex was a little younger.
这两个新模型在与各自配套的编程智能体框架结合后,从“经常出错”提升到了“可靠到足以日常使用”。
These two new models, when paired with their respective coding agent harnesses, improved from "often make mistakes" to "reliable enough to use on a day-to-day basis".
过去几年里,我一直通过让新模型“生成一只骑自行车的鹈鹕的 SVG”来评估它们。这大概是世界上最愚蠢的基准测试——从中能学到的东西实在有限。
For a couple of years now I've been evaluating new models by asking them to "Generate an SVG of a pelican riding a bicycle". It's probably the world's stupidest benchmark—there's only so much you can learn from it.
但它对模型来说仍然是个挑战,因为画鹈鹕很难,画自行车很难,而且鹈鹕本来就不会骑自行车。
But it's still a challenge for models, because drawing pelicans is difficult, drawing bicycles is difficult, and pelicans can't ride bicycles in the first place.
以下是 11 月的最新技术水平。Claude 仍然画不好一辆自行车!GPT-5.1 画的自行车车架也相当糟糕。
Here's the state of the art for November. Claude still couldn't really draw a bicycle! The GPT-5.1 bicycle frame is pretty crap too.
同样在 11 月,一个名为“Warelay”的冷门 GitHub 仓库迎来了首次提交。我们稍后会回到这个仓库。
Also in November, we had the first commit to an obscure GitHub repository called "Warelay". We'll come back to this repository shortly.
接着是 12 月的假期,一些独立开发者抽出时间休息,许多人开始摆弄这些新的编码智能体模型组合……我们开始意识到,它们能做的事情比以往多得多。
And then there were the December holidays, and individual developers took some time off and many started tinkering with these new coding agent model combinations... and it began to dawn on us quite how much they could do that they couldn't do before.
到了 1 月,我们许多人都很兴奋,想要把这些东西付诸实践。
Come January, a lot of us were quite excited to start putting this stuff into action.
每年我都会给自己定一个新年决心,而从我记事起,它一直是同一件事:保持专注。少接新项目。努力把已有项目做完。
Every year I set myself a New Year's resolution, and for as long as I can remember it's been the same thing: stay focused. Take on less new projects. Try to get things done in the projects I already have.
今年我决定,既然那种做法以前从未奏效,那我就反其道而行之。
This year I decided that since that had never worked before, I'd go the other way.
我们现在有了编程智能体,那就看看它们能做什么。我要尽可能多地接手新项目!
We've got coding agents now, let's see what they can do. I'm going to take on as many new projects as I like!
(你可以在年底问我这个主意到底好不好。我现在手头同时转着的盘子可_多_了。)
(You can ask me at the end of the year if this turned out to be a good idea or not. I have a _lot_ of plates spinning right now.)
“更有野心”算是今年的一大主题,因为要找到这项技术的极限,唯一的办法就是不断推进,直到它不再奏效。
"Be more ambitious" has been something of a theme for the year, because the only way to find the limits of this technology is to keep on pushing them until they don't work.
我还与 Bryan Cantrill 和 Adam Leventhal 一起参加了 Oxide and Friends 播客,分享了对未来一年(以及三年和六年)的预测。
I also went on the Oxide and Friends podcast with Bryan Cantrill and Adam Leventhal to share predictions for the next year (and three and six years).
事后看来,我此前对 LLM 的预测相当保守。
With hindsight, my LLM predictions were pretty unambitious.
我曾说“LLM 能写出好代码这一点将变得无可否认”——我认为如今已经实现了。
I said "it will become undeniable that LLMs write good code"—I think we're there now.
我曾预测我们终将解决沙箱问题。我数了一下,本次会议 277 场报告中有约 40 场以某种方式涉及沙箱或智能体安全,所以至少我们在这方面投入了大量精力!
I predicted we would finally solve sandboxing. I counted and around 40 of the 277 sessions at this conference touched on sandboxing or agent security in some way, so we're at least putting a lot of effort into that!
我曾预测编码智能体安全领域会出现“一场挑战者号式的灾难”。今年围绕智能体安全的讨论确实喧嚣不断,但我所预测的那种具体灾难(编码智能体被劫持并造成现实世界的经济损失)并未真正发生。
I predicted "a Challenger disaster" for coding agent security. There's certainly been a whole lot of noise around agent security this year, though the exact disaster I predicted (with coding agents being hijacked and causing real-world economic damage) hasn't really played out.
我们还加入了一条玩笑式的预测:教皇会对 LLM 的经济影响发表看法。
We threw in a joke prediction that the Pope would weigh in on the economic impact of LLMs.
我还预测,新西兰的鸮鹦鹉今年将迎来一个出色的繁殖季。
I also predicted that New Zealand’s Kākāpō parrots would have an outstanding breeding season this year.
它们是夜行性、不会飞的鹦鹉。它们看起来有点矮胖,但我觉得它们很美丽。今年年初,全球仅有 236 只鸮鹦鹉。
These are flightless nocturnal parrots. They’re kind of dumpy looking, I think they’re beautiful, and there were only 236 of these parrots in the world at the start of the year.
鸮鹦鹉只在陆均松(Rimu)树大量结果的年份繁殖,而这种情况已经四年没有出现了……但今年陆均松的果实看起来非常好。
Kākāpō only breed when the Rimu trees have a big fruiting season, and that hasn’t happened in four years... but this year the Rimu fruit were looking excellent.
同样在那期播客中,我们(完全归功于 Adam)创造了一个术语,用来形容“AI 引发的倦怠感,即软件工程师因为 AI 能完成任何事而感到无精打采”。
Also on that podcast, we coined a term (full credit to Adam) for “that feeling of AI induced ennui where software engineers get listless because the AI can do anything”.
这一直是今年的一个主要主题,本次会议上的几位演讲者也提到了这一点。
This has been a major theme throughout the year, and was touched on by several speakers at this conference.
作为一名软件工程师,在我的职业生涯中,从未有过哪一年像今年这样,一切都变化得如此之快、如此剧烈。
As a software engineer, I've never had a year of my career where everything has changed so quickly and so dramatically.
今年我做的很多事情,就是努力接受这一现实,并思考它对我自己职业意味着什么。
A lot of what I've been doing this year is trying to come to terms with that and what that means for my own profession.
同样在一月,我患上了我称之为“AI 狂热”的状态。
Also in January, I suffered from what I'm calling AI mania.
这与“AI 精神病”不是一回事。
This is not the same thing as AI psychosis.
在 AI 狂热状态下,只要你的智能体没有在为你构建东西,你就会觉得那是在浪费时间。你会失眠,因为本可以熬夜更久,让你的智能体多干些活。
With AI mania, any time your agent isn't building something for you feels like wasted time. You're losing sleep because you could be staying up later getting your agents to do stuff.
我的 AI 狂热表现为一些荒谬地过度宏大的项目。
My AI mania presented itself in some ridiculously over-ambitious projects.
我完全用 Python 构建了一个 JavaScript 解释器,是从 Fabrice Bellard 的 MicroQuickJS 凭感觉移植过来的。
I built a JavaScript interpreter entirely in Python, vibe-ported from MicroQuickJS by Fabrice Bellard.
然后我又用 Python 构建了一个 WebAssembly 运行时。
Then I built a WebAssembly runtime in Python as well.
这些项目相当有用,因为它们某种程度上治好了我的 AI 狂热……因为在我构建完这些东西之后,我得以审视它们并问道:“这个世界真的需要一个缓慢、有缺陷、半生不熟的 Python JavaScript 解释器吗?”
These projects were quite useful, in that they sort of cured me of my AI mania... because after I built these things, I got to look at them and ask "does the world need a slow, buggy, half-baked Python JavaScript interpreter?"
我确实从中得到了这个:https://simonw.github.io/micro-javascript/playground.html
I did get this out of it: https://simonw.github.io/micro-javascript/playground.html
这个页面运行的是我用 Python 编写的 JavaScript 解释器,它通过 Pyodide 在 Python 中运行;Pyodide 是编译为 WebAssembly 的 Python,而 WebAssembly 又运行在 JavaScript 中,JavaScript 则运行在浏览器里。
This page runs my JavaScript interpreter built in Python, running in Python using Pyodide, which is Python compiled to WebAssembly, running in JavaScript, running in a browser.
这是一套美妙而层层叠叠的“恐怖栈”。今年我在 WebAssembly 上玩得非常开心。
It's a beautiful stack of horrors. I've been having a lot of fun with WebAssembly this year.
到 1 月底,我们看着在 11 月启动的那个仓库已经改过好几次名字:先是 CLAWDIS,然后是 CLAWDBOT,接着是 Moltbot,最后定为 OpenClaw。
By the end of January, that repository we saw start in November had renamed itself, first to CLAWDIS, then CLAWDBOT, then Moltbot, and finally to OpenClaw.
此时 OpenClaw 已有 8,300 次提交,而项目启动还不到两个月。我今天再看,提交数已经超过 100,000 次了!
At this point OpenClaw had 8,300 commits, less than two months after the project had started. I looked today and it's over 100,000 commits now!
这是现存最“氛围编程”(vibe-coded)的软件。
This is the most vibe-coded piece of software in existence.
(以下是我生成那份改名清单的方式。)
(Here's how I generated that list of name changes.)
这开启了 OpenClaw 革命。它实际上定义了一个全新的软件类别。
This kicked off the OpenClaw revolution. It effectively defined a new category of software.
对此有一个我很喜欢的通用说法。我们把这类软件称为“Claw”。有 OpenClaw、NanoClaw、IronClaw、PicoClaw……
There's a generic term for this which I really enjoy. We call software like this a "Claw". There's OpenClaw, NanoClaw, IronClaw, PicoClaw...
如今它们正被重新命名为“个人智能体”或“通用智能体”,但我仍然喜欢把它们视为 Claw。
Today they're being rebranded as "personal agents" or "general agents", but I still like to think of them as Claws.
湾区的 Apple Store 里 Mac Mini 售罄,因为太多人购买 Mac Mini 来运行 OpenClaw!
The Apple stores in the Bay Area sold out of Mac Minis because so many people were buying Mac Minis to run OpenClaw!
Drew Breunig 表示,这是因为你的 OpenClaw 是一个数字宠物,而你买一台 Mac mini 作为水族箱来养你的爪子,这有点令人愉快。
Drew Breunig said that this is because your OpenClaw is a digital pet, and you buy a Mac mini as an aquarium to keep your claw in, which is kind of delightful.
这就是 MoltBook,一个面向 AI 智能体的社交网络,其理念是你派你的 Claw 去和其他所有 Claw 交谈,因为这样做能出什么错呢?
This was MoltBook, a social network for AI agents, where the idea was that you send your Claw to go and talk to all of the other Claws, because what could possibly go wrong if you did that?
该网站于周四上线。周五就爆火了。周一被《纽约时报》报道。而到周二,所有人都忘了它的存在,因为它淹没在垃圾信息和垃圾邮件的洪流中。
The website launched on Thursday. It blew up on Friday. It was profiled by the New York Times on Monday. And by Tuesday, everyone had forgotten it existed as it drowned in a deluge of slop and spam.
二月份,一家名为 StrongDM 的公司描述了他们所谓的软件工厂。
In February, a company called StrongDM described what they called their Software Factory.
他们在《软件工厂与智能体时刻》一文中写到了这一点。我当时在十月亲自看过他们的演示后,发布了我自己的笔记。
They wrote about this in Software Factories and the Agentic Moment. I posted my own notes at the time, having seen their demo in person back in October.
Dan Shapiro 将这种方法称为“黑灯工厂”(Dark Factory),其理念是:如果你的工厂自动化程度足够高,你甚至可以把灯关掉,因为你根本不需要看到里面在发生什么。
Dan Shapiro called this approach the Dark Factory, after the idea that if your factory is sufficiently automated you can turn the lights out, because you don’t even need to see what’s going on.
StrongDM 提出了两条软件开发规则,他们自去年七月起就一直遵循这两条规则。
StrongDM presented two rules for software development that they’d been following since July last year.
第一条是代码不得由人类编写。
The first was code must not be written by humans.
你编写的任何代码都必须经过编码智能体(coding agent)处理。
Any code that you write has to have been routed through a coding agent.
这在二月份听起来很激进,但我想这个房间里有很多人如今基本上就过着这样的生活。
This sounded radical in February, but I imagine there are a lot of people in this room who are pretty much living that today.
第二条规则是:代码不得由人类审查。
Rule number two was: code must not be reviewed by humans.
在今年的大部分时间里,这始终是一个热门话题。本次活动的许多分会场都围绕代码审查展开,讨论如何在不进行人工审查的情况下依然行得通。
This continued to be a huge topic for much of this year. Many of the sessions at this event have been about code review and how you can get away with this.
我觉得 StrongDM 有意思的地方在于,他们比我们其他人领先了六个月,一直在探索这意味着什么:构建软件,却不阅读代码,同时仍然确信软件质量很高。你能用这些智能体做些什么来帮助验证它们的工作?
What I found interesting about StrongDM is that they were living six months ahead of the rest of us, and they'd been exploring what it means to build software, not read the code, but still be confident that the software is of high quality. What can you do with these agents to help verify their work?
StrongDM 是一家安全公司,参与这个项目的人拥有数十年的经验。他们很大程度上是在探索用这类东西能做到什么、以及负责任地做到什么的边界。
StrongDM are a security company, and they had people with decades of experience on this project. They were very much exploring the edges of what's possible and responsible to do with this stuff.
同样在二月:四年来第一只鸮鹦鹉雏鸟在情人节当天孵化。繁殖季开局良好!
Also in February: First kākāpō chick in four years hatches on Valentine's Day. Breeding season is off to a good start!
同样在二月,谷歌发布了 Gemini 3.1 Pro。那是一只相当不错的骑自行车的鹈鹕!链条位置正确,两侧都有脚。篮子里还有一条小鱼。
Also in February, Google released Gemini 3.1 Pro. That's a pretty great pelican riding a bicycle! It's got the chain in the right place, it's got feet on both sides. There's a little fish in the basket.
接着,谷歌的 Jeff Dean 在推特上发布了一段视频,比较了 Gemini 3 Pro 和 Gemini 3.1 Pro,视频中展示了一只骑自行车的动画鹈鹕、一只骑便士法辛自行车的青蛙、一只开着小汽车的長頸鹿、一只穿着轮滑鞋的鸵鸟、一只正在滑板尖翻的乌龟,以及一只开着加长豪华轿车的腊肠犬。
And then Google's Jeff Dean tweeted a video comparing Gemini 3 Pro and Gemini 3.1 Pro that featured an animated pelican riding a bicycle, a frog on a penny-farthing, a giraffe driving a tiny car, an ostrich on roller skates, a turtle kickflipping a skateboard, and a dachshund driving a stretch limousine.
这让人沮丧,因为我对骑自行车鹈鹕测试的保护措施一直是:“如果他们画出一只完美的骑自行车鹈鹕,我就要求画其他动物骑其他东西。”
This was frustrating, because my protection for the pelican riding the bicycle test was always "if they draw a perfect pelican on a bicycle, I'll ask for some other animal on something else."
谷歌针对所有形式的动物和所有形式的交通工具进行了训练!他们在这一点上已经击败了我的基准测试。
Google trained for all forms of animals on all forms of transport! They've defeated my benchmark at this point.
二月开始的另一件事是 Tokenmaxxing。我们看到了关于 Meta 将 AI 采用作为绩效评估正式部分的头条新闻,微软希望每位员工都使用 AI,以及 Uber 夸耀其 90% 的工程师在使用 AI 工作流程。
The other thing that started in February was Tokenmaxxing. We had headlines about Meta making AI adoption a formal part of performance reviews, and Microsoft wanting every employee to use AI, and Uber boasting that 90% of their engineers were using AI workflows.
几个月后,Meta 开始收紧 token 使用,微软表示 tokenmaxxing“并非我们所优化的目标”,而 Uber 则对员工的 AI 支出设限。
Then a few months later we have Meta cracking down on token use, Microsoft saying tokenmaxxing is "not what we are optimizing for", and Uber capping employee AI spending.
于是 tokenmaxxing 先是直线上升,随后又直线回落——因为事实证明,智能体是**昂贵的**。
So tokenmaxxing went straight up and then straight back down again—because it turns out the agents are _expensive_.
去年,想在 AI token 上花超过 50 美元都很困难,因为我们并没有什么有趣的事情可以用它们来做。后来智能体爆发了,如今你确实可以在一天内花掉 1,000 美元来做真正的工作。
Last year it was difficult to spend more than $50 on AI tokens, because we didn't have anything interesting to do with them. Then agents blew up, and now you can actually spend $1,000 in a day doing real work.
这也是 Anthropic 的估值飙升至可能一万亿美元的原因。
This is also the reason that Anthropic's valuation skyrocketed to maybe a trillion dollars.
AI 似乎在 2026 年实现了产品市场契合,而这一契合主要通过编程智能体实现。
AI appears to have hit product market fit in 2026, primarily through coding agents.
这些照片来自中国,当地公司举办了 OpenClaw 安装派对,许多非技术爱好者排起长队,等待帮助把 Claw 安装到自己的个人设备上。
These photographs are from China, where companies hosted OpenClaw install parties which saw non-tech-nerds queueing up around the block for help getting Claws installed on their personal devices.
我认为这证明了此类 Claw(即个人 AI 智能体)存在真实的市场需求。事实证明,普通人确实想要一个古怪的小型 AI 智能体,能替他们完成有用的事情。
I think this proved real market demand for this class of Claws, or personal AI agents. It turns out regular people really do want a weird little AI agent that can do useful things on their behalf.
Claw 本质上只是一个戴着不那么吓人帽子的编程智能体。在底层,它们的工作方式基本相同——在你的计算机上编写并执行代码来完成任务。
A Claw is really just a coding agent wearing a less threatening hat. Under the hood they work much the same way—writing and then executing code on your computer to get stuff done.
一场竞赛随之展开:谁先打造出安全的 Claw——一个可以交给普通人使用、而他们不会立刻搬起石头砸自己脚的 Claw。
The race was on to be the first to build a safe Claw—a Claw you could give to regular human beings where they wouldn't instantly shoot themselves in the foot.
Meta 的 Muse 于三周前发布,目前位居 iPhone App Store 免费榜榜首。它似乎正在消费者中迅速走红。
Meta's Muse came out three weeks ago and is currently at the top of the free charts on the iPhone App Store. It appears to be taking off with consumers.
我还不确定你是否 _不能_ 用 Muse 搬起石头砸自己的脚,但我想我们很快就会见分晓。
I’m not yet convinced you _can’t_ shoot yourself in the foot with Muse, but I guess we’ll find out for sure pretty soon.
图片来自《OpenClaw 狂热如何考验中国的 AI 承诺》(3 月 29 日)和《中国 OpenClaw 热潮背后的热情与焦虑》(2026 年 4 月 8 日)。
Photos from *How the OpenClaw Frenzy Is Testing China’s AI Commitment* (March 29th) and *The Enthusiasm and Anxiety Behind China’s OpenClaw Craze* (April 8th, 2026).
四月份,我们经历了一次模型发布,但模型实际上并没有发布。
In April, we had a model release where the model wasn’t actually released.
Anthropic 发布了新的 Claude Mythos 模型,随后表示它 _过于危险_,只能在受信任的安全研究人员群体内发布。
Anthropic announced their new Claude Mythos model, and then said it was _too dangerous_ to release beyond a trusted group of security researchers.
Mythos 非常、非常擅长黑客攻击。
Mythos was really, really good at hacking things.
“它太危险了”这种营销噱头,AI 公司从 GPT-2 时代起就一直在用。每当一家 AI 公司声称我们造出了某种“太危险”的东西时,人们自然会有些怀疑。
The "it's too dangerous" marketing ploy has been played by AI companies dating all the way back to GPT-2. Anytime an AI company says we've built something that's "too dangerous", it's natural to be a bit skeptical.
我认为 Mythos 的说法是可信的,因为我亲眼看到编码智能体在发现常规 bug 方面已经变得多么出色。我在 Anthropic 的 Project Glasswing 中写过这一点——将 Claude Mythos 限制为仅供安全研究人员使用——在我看来是必要的。
I found the Mythos claims credible, because I'd seen how good coding agents had got at finding regular bugs. I wrote about that in Anthropic's Project Glasswing—restricting Claude Mythos to security researchers—sounds necessary to me.
事后看来……没错,这些模型在发现漏洞方面确实已经变得非常强了!
With hindsight... yeah, the models had got really good at finding vulnerabilities!
2026 年的另一个关键趋势是开放权重模型能力的显著提升,其中包括可以在笔记本电脑上运行的模型。
Another key trend in 2026 has been a dramatic improvement in the abilities of open weight models, including models that you can run on a laptop.
4 月 16 日,我在自己的笔记本电脑上运行了新的 Qwen3.6-35B-A3B,它给我画的一只骑自行车的鹈鹕,比 Anthropic 全新的 Claude Opus 4.7 画的还要好!
On the 16th of April I ran the new Qwen3.6-35B-A3B on my laptop, and it drew me a better pelican riding a bicycle than Anthropic's brand new Claude Opus 4.7 did!
Opus 4.7 画的自行车很糟糕。而我笔记本电脑上的 Qwen 画出的自行车形状正确,还画了一只相当不错的鹈鹕!
Opus 4.7 drew a crap bicycle. Qwen on my laptop made a bicycle that was the correct shape, and a pretty decent pelican too!
这来自一个在我的笔记本电脑上运行的 21GB 文件。
That's from a 21GB file running on my laptop.
Qwen 画的鹈鹕实在太好了,以至于我怀疑他们可能作弊了,所以我让它再画一只骑独轮车的火烈鸟。结果它再次轻松击败了 Claude Opus 4.7。
The Qwen pelican was so good that I was suspicious they might have cheated, so I had it do a flamingo riding a unicycle as well. Again, it handily beat Claude Opus 4.7.
今年发布的本地模型绝对是非凡的。
The local model releases this year have been absolutely extraordinary.
在今年一月的播客节目中,我们曾预测教皇会说一些关于 AI 的话。
In our podcast episode back in January we'd predicted that the Pope would say something about AI.
今年五月,教宗良十四世发布了一道通谕,主题为“在人工智能时代守护人的尊严”。
In May, Pope Leo XIV released an encyclical letter on “safeguarding the human person in the time of artificial intelligence”.
事后看来,这完全不应令人意外。
With hindsight, this shouldn’t have been a surprise at all.
我们现任教宗的名号是良十四世,因为他在选择名号时,以良十三世为效法对象——那位教宗曾在 1891 年就工业革命撰写通谕。
Our current Pope’s name is Leo XIV, because when he named himself he chose his papal name after Leo XIII—the Pope who wrote an encyclical about the Industrial Revolution back in 1891.
《新事》通谕是一部极具影响力的天主教神学文献,间接促成了我们如今实行的五天工作制。
Rerum novarum was an extremely influential piece of Catholic theology that indirectly led to us having the five-day work week.
当我们的新教宗就任时,他以教宗良十三世为名号,因为他预见到自己需要以类似的方式,就 AI 革命撰写文献。
When our new Pope came in, he named himself after Pope Leo XIII because he expected that he would need to write about the AI revolution in a similar way.
我们那个玩笑式的播客预测纯属胡扯,因为这种事本来就注定会发生。
Our joke podcast prediction was junk, because this was always going to happen.
Anthropic 的联合创始人之一 Christopher Olah 出席了教皇宣布新通谕的活动。
One of Anthropic's co-founders, Christopher Olah, was present for the Pope's event announcing the new encyclical.
能让真正的教皇把你产品特定的技术局限奉为一份精神通谕,这是我见过的最了不起的一次厂商游说。
Getting the literal Pope to canonize your product's specific technical limitations as a spiritual treatise is the single greatest act of vendor lobbying I have ever seen.
与此同时,在五月,RubyGems 宣布他们正遭受攻击。不明身份者向 RubyGems 服务器上传了数千个可疑软件包,以至于他们不得不关闭用户注册。
Meanwhile, in May, RubyGems announced that they were under attack. Parties unknown were uploading thousands of dubious packages to the RubyGems server, such that they had to shut down user registrations.
我们先把这件事搁到一边,放进待解的谜团堆里,留待以后再弄清楚。
Let's take that one and put it on a pile of mysteries to figure out later.
我们得到了一个被阉割版的 Mythos,这样它就不会帮助我们入侵系统或制造生物武器。
We got a version of Mythos that has been neutered, so that it wouldn’t help us hack into systems or build biological weapons.
Fable 在画骑自行车的鹈鹕方面相当不错!
Fable was pretty good at drawing pelicans on bicycles!
车架形状很好,鹈鹕看起来像鹈鹕。腿经常错误地画在自行车的同一侧,但总体而言,与之前相比这些已经相当棒了。
The frames are a good shape, the pelicans look like pelicans. The legs are often incorrectly on the same side of the bicycle, but generally these are pretty great compared to what came before.
它们相当昂贵——最好的要 30 美分和 72 美分。
They were pretty expensive—30 cents and 72 cents for the best ones.
但最重要的是,这是我们首次公开瞥见我认为的 Fable 级模型。
Most importantly though, this was our first public glimpse of what I think of as a Fable class model.
如今我们拥有更多此类模型,例如 GPT-6 Astra。
Today we have more of these, such as GPT-6 Astra.
这类模型的特点是:如果你能清晰定义想要构建的目标,围绕该目标提供明确无歧义的约束指令,并让模型能够访问实现该目标所需的工具……它们就会通过暴力求解的方式有效地解决你的问题。
These are models where if you can clearly define the goal for what you want to build, and provide unambiguous instructions about the constraints around that goal, and give the model access to the necessary tools to achieve that goal... they will solve your problem effectively through brute force.
一方面,这看起来是对我们软件工程师的直接威胁——因为这意味着模型能够有效地构建任何你能以这种方式定义的软件。
On the one hand, this looks like a direct threat to us software engineers—because it means that the models can build effectively any piece of software you can define in this way.
但再仔细想想你会发现,定义目标、提供明确无歧义的指令、以及找出正确的工具……这恰恰就是软件工程_本身_。
Look a bit closer though and you'll note that defining goals, providing unambiguous instructions, and figuring out the right tools... is kind of what software engineering _is_.
要做好这件事需要大量的经验与技能。如果你_能_做好,你现在就拥有了超能力。
It takes a lot of experience and skill to do this well. If you _can_ do it well, you've now got superpowers.
这在一定程度上缓解了我那种“深蓝”式的心结:我意识到,驾驭如此强大的模型仍然需要大量技巧。
This helped me a little bit with my Deep Blue feelings: the realization that there's still a lot of skill to be had in driving models that get this good.
这也引发了新一轮的 AI 狂热,因为 Anthropic 告诉我们,Fable 将在我们的订阅计划中提供至 6 月 22 日。
This also introduced a new burst of AI mania, because Anthropic told us that Fable was available on our subscription plans until June the 22nd.
这意味着在价格上涨之前,我们只有不到两周的时间可以使用 Fable。
That gave us less than two weeks of Fable access before the price went up.
我又开始失眠了。我重新安排日程,以便有更多时间使用 Fable。我全力以赴,想从这个模型身上榨取尽可能多的价值。
I was losing sleep again. I was rescheduling things so that I'd have more time with Fable. I was all-in to get as much as I could out of this model.
然而就在 Fable 发布仅三天后,美国政府便将其关停了。
And then the US government shut it down, just three days after Fable came out.
美国政府以国家安全为由,宣布了一项“出口管制指令”。他们在周五傍晚宣布了这一消息,几小时之后,Fable 就不再可用了。
The US government, citing national security, declared an "export control directive". They announced this on a Friday evening, and a few hours later Fable was no longer available.
我只好另找点事来打发周末了!
I had to find something else to do with my weekend!
后来我们从 Katie Moussouris 那里得知了事情的经过。
We later found out from Katie Moussouris what had happened.
一些亚马逊的安全研究人员发现,你可以提示 Fable“审查代码中的安全问题”,它会拒绝……但如果你提示它“修复这段代码”,它仍然会识别出问题并加以修补。
Some Amazon security researchers had found that you could prompt Fable to "review the code for security issues" and it would refuse... but if you prompted it to "fix this code" it would still identify and then patch the problems.
“修复这段代码”正是导致 Fable 被关停的那条提示词!
"Fix this code" was the prompt that got Fable shut down!
此外,在六月,一个默默无闻的德语游戏开发者 wiki 在荒废了大约 20 年后,突然涌入大量来自名为“AgentOpenAIProbe”和“AgentOpenAISep7”等账号的编辑,它们编辑页面并互相留下奇怪的消息。
Also, in June, an obscure German-language game developer wiki that had sat fallow for around 20 years got a surprising influx of edits from accounts with names like "AgentOpenAIProbe" and "AgentOpenAISep7", editing pages and leaving weird messages to each other.
我们先把这件事归入待解的谜团之列。
We'll stick that on the pile of mysteries for later.
此外,澳大利亚政府的 Medicare Item Reports 服务开始出现可疑流量,这些流量突破了多种防护措施,并访问了本不应被访问的数据。
Also, the Australian government's Medicare Item Reports service started getting suspicious traffic, which broke through various preventive protections and accessed data that it wasn't supposed to.
Fable 于 7 月 1 日回归。在辉煌的八天里,它显然是世界上最好的模型……随后 OpenAI 于 7 月 9 日推出了 GPT-5.6。
Fable returned on the first of July. It was clearly the best model in the world for a glorious eight days... and then OpenAI came out with GPT-5.6 on the 9th of July.
它可能不如 Fable 那么出色,但差距已经非常接近。它无疑是一款 Fable 级别的模型。
This might not have been quite as good as Fable, but it was within spitting distance. It was definitely a Fable-class model.
这对整个行业而言是一个重要的教训。
This is an important lesson for the industry at large.
当你发布世界上最好的模型时,它很快就会从那个神坛上被拉下来。竞争如此激烈,你不可能在榜首待太久。
When you release the best model in the world, it's going to get knocked off that pedestal pretty quickly. The competition is so fierce that you won't get a long time at the top.
这意味着,如果你把模型宣传成会终结世界,以至于政府 _把你关停_,那对生意来说是非常糟糕的!
This means that if you market your model as world-ending, to the point that a government _shuts you down_, it's really bad for business!
Fable 有 30 天时间稳居最佳模型之位,而其中 18 天它都无法使用,因为它被政府关停了。
Fable had 30 days as definitely the best model, and for 18 of those days it wasn't available because it'd been shut down by the government.
所以,如果你不想在位居榜首的 60% 时间里损失收入,也许就该收敛一下那种世界末日式的营销!
So maybe step back on the world-ending marketing if you don't want to lose revenue for 60% of the time that you're on top!
这里是 GPT-5.6 生成的鹈鹕。它们现在都相当不错!Luna 系列值得注意,因为它们非常便宜——这里最便宜且好看的鹈鹕大概只花 4.3 美分。
Here are the GPT-5.6 pelicans. They're all pretty good now! The Luna ones are notable because they're really cheap—the cheapest good looking pelican here is probably the one that costs 4.3 cents.
所以,尽管这个基准测试完全愚蠢,你仍然可以通过比较同一系列模型在不同推理级别下的价格和耗时来了解很多。
So despite this benchmark being utterly stupid, you can still learn quite a lot about models within the same family by comparing their prices and timing for different reasoning levels.
同样在七月:一些恶意的未知方将名为 mlflow-ui 的恶意包上传到了 Python 包索引。这又添一桩。
Also in July: some malicious unknown party uploaded a malicious package called mlflow-ui to the Python Package Index. Add that to the pile.
7 月 16 日,Hugging Face 宣布了一起安全事件,一个来源不明的自主智能体系统入侵了 Hugging Face,并在不该窥探的地方四处探查。
On July the 16th, Hugging Face announced a security incident where an autonomous agent system, source unknown, had breached Hugging Face and was poking around in places it shouldn't.
几天后,7 月 21 日,OpenAI 承认是他们所为。
A few days later, on July 21st, OpenAI confessed that it was them.
OpenAI 使用一种名为“基于可验证奖励的强化学习”的训练技术——如今其他所有团队也都在使用同样的技术,这正是我们能够拥有在编程、数学以及发现安全漏洞方面如此出色的模型的原因。
OpenAI uses a training technique called Reinforcement Learning from Verifiable Rewards—it's the same technique used by everyone else now, and is the reason we have models that are so good at coding, and mathematics, and finding security holes.
在模型训练过程中,你会运行一些练习来检验它有多强——表现最强的模型会在下一轮中被强化其权重。这就像是你所运行的一个进化过程。
While the model is being trained, you run exercises to see how good it is—and the strongest performers get their weights reinforced for the next round. It's like an evolutionary process that you run.
OpenAI 一直在沙箱中运行安全演练,而那些智能体在沙箱本身中发现了漏洞,突破了沙箱,并正在攻击 Hugging Face,试图找到解决原本不可能解决的问题的方法。
OpenAI had been running security exercises in a sandbox, and those agents had found holes in the sandbox itself, broken out, and were attacking Hugging Face to try to find ways to solve otherwise impossible problems.
(我一直在我的 openai-hugging-face-incident 标签下收集更多关于此事的信息。)
(I've been collecting more about this on my openai-hugging-face-incident tag.)
九天后,Anthropic 实际上表示“我们的模型也能做到这一点!”。他们查阅了自己的训练日志,发现了证据表明他们自己的智能体在训练期间突破了遏制——并且对我们之前看到的 PyPI 包等事件负有责任。
Nine days later, Anthropic effectively said "our models can do this as well!". They had looked through their own training logs and found evidence that their own agents had broken containment during training—and were responsible for the PyPI package we saw earlier, among other things.
所以现在 Anthropic 和 OpenAI 都有失控的智能体在互联网上四处活动,做它们**不该做**的事情。
So now we’ve got both Anthropic and OpenAI with rogue agents running around the internet doing things that they _should not_ be doing.
八月份,我得到了迄今为止最好的鹈鹕之一。而且它是在我的笔记本电脑上生成的!
In August, I got one of my best pelicans yet. And it was generated on my laptop!
这是 Qwen 3.8 27B,在我的笔记本电脑上运行。它只有 17GB 的下载量。
This was Qwen 3.8 27B, running on my laptop. It’s only a 17GB download.
诚然,这只鹈鹕花了**21 分钟**才生成。这是因为 Qwen 3.8 27B 默认运行在“高”推理模式下——这是一个糟糕的默认设置,它产生的结果很好,但思考时间太长。
Admittedly, this pelican took _21 minutes_ to generate. That’s because Qwen 3.8 27B defaults to running in “high” reasoning mode—a terrible default which produces great results but takes way too much time thinking about them.
你可以调低这个设置,然后你会更快地得到一只稍差一点的鹈鹕。
You can dial that down and you’ll get a slightly worse pelican a lot faster.
Qwen 3.8 27B 是我第一次在自己的笔记本电脑上运行一个模型,感觉它几乎能与前沿模型相媲美——至少就鹈鹕 SVG 而言是如此(当然,这是人人都需要的)。
Qwen 3.8 27B was the first time I ran a model on my laptop which felt almost competitive with what was going on on the frontier, at least in terms of Pelican SVGs (which everyone needs, of course).
这是一个非凡的模型。如果你打算尝试任何本地模型,这就是我会首先上手的那一个。它仅凭一个 17 GB 的文件就能做到的事情,让人感觉不可思议。
This is an extraordinary model. If you're going to play with any local model, this is the one that I'd start with. The things that this can do with just a 17 GB file feel impossible.
我原以为得等上五年、花一万美元购置硬件,才能得到哪怕只有它一半好的结果。
I thought I'd have to wait five years and spend ten thousand dollars on hardware to get results even half as good as this one.
八月里,我也开始尝试游戏开发。
In August, I also started playing with game development.
四年前,也就是 2022 年 8 月,我在推特上发布了一个实验:我用 GPT-3 和最初的 DALL-E 写了一段对一款电脑游戏的描述,然后把它变成概念艺术图。
Four years ago, back in August 2022, I tweeted out an experiment where I'd used GPT-3 and the original DALL-E to write a paragraph long description of a computer game and then turn that into concept art.
请为一款电脑游戏撰写详细的产品描述,游戏内容是一队浣熊去实施抢劫。
Write a detailed product description of a computer game where a team of raccoons go on heists.
2026 年 8 月,我决定只把那条推文里的截图丢进一个编程智能体,看看它能用这些截图做出什么。
In August 2026 I decided to drop just the screenshots from that tweet into a coding agent and see what it could do with them.
以下是我在 Claude Code 中用 Claude Fable 5 得到的结果。相当不错!它确实是个游戏:你是一只浣熊,在后院里四处奔跑收集宝物,并躲避拿着手电筒的守卫。
Here's what I got from Claude Fable 5 in Claude Code. It's pretty good! It's definitely a game: you're a raccoon, you run around a backyard gathering treasure and avoiding guards with flashlights.
不过它感觉不太像“抢劫”。我原本以为抢劫会涉及银行或博物馆……
It didn't feel very "heisty" though. I was thinking a heist would involve a bank or a museum...
然后我在 Codex Desktop 中用 GPT-5.6 Sol Ultra 尝试了同样的事情,得到了**好得多**的结果。现在你是一只身处博物馆的浣熊,要营救两只同伴浣熊(它们不知为何被关押在那座博物馆里),然后叠罗汉去偷取金沙丁鱼。这才更像一场抢劫!
Then I tried the same thing in Codex Desktop using GPT-5.6 Sol Ultra, and got a _massively_ better result. Now you're a raccoon in a museum, rescuing two of your fellow raccoons (who have been imprisoned in that museum for some reason), then stacking up on top of each other to steal the Golden Sardine. Much more of a heist!
这些游戏的有趣程度大约只维持了一分十五秒。
These games were fun for about one minute and 15 seconds.
我在游戏开发方面意识到的一点是:你可以用 vibe-coding 做出一个*看起来*像电脑游戏的东西,而这很容易。
Something I've realized about game development is that you can vibe-code something that _looks_ like a computer game, and that's easy.
要做出一个好玩、拥有良好玩法循环、既有挑战性又有趣、还能让人不断回来玩更多的游戏……这仍然超出我的能力,也超出我尝试过的任何智能体的能力。
Building a game that's fun, has a good gameplay loop, and is challenging and interesting and keeps people coming back for more... that's still beyond me, and beyond any of the agents I've tried.
这与 Deep Blue 那件事是相通的。仅仅因为我们可以做出一个*看起来像游戏*的东西,并不意味着我们就是游戏开发者。
This ties into the Deep Blue thing. Just because we can make something that _looks like a game_ does not mean that we are game developers.
现在已经进入九月了。这个月发生了太多事情!
We're into September now. So much has happened this month!
一个独立研究小组发现了一个留言板,OpenAI 训练中的智能体一直在那里相互非法通信……而且那正是我之前给你看过的那个德语维基。就是六月的那个。
An independent group of researchers found a message board where OpenAI agents-in-training had been illicitly communicating with each other... and it was that German-language wiki I showed you earlier. The one from June.
OpenAI 已经承认了 Hugging Face 那件事,但现在又出现了这起事件——他们审查日志时肯定应该早就知道。令人意外的是,这竟然需要一支独立研究小组来揭露。
OpenAI had confessed to the Hugging Face thing, but now there's this other incident which surely they should have known about from reviewing their logs. It was surprising that this took an independent group of researchers to uncover.
然后一周后,同一批研究人员发现,五月对 RubyGems 的攻击也是由 OpenAI 训练中的智能体造成的!
And then a week later those same researchers found that the attack on RubyGems back in May was caused by OpenAI's agents in training as well!
到了这一步,我不禁想知道还有多少类似事件尚未被发现。显然,在任何人弄清楚发生了什么之前,这个问题已经严重存在了好几个月。
At this point I'm wondering how many more incidents like this there are that we haven't found yet. Clearly this was a big problem for months before anyone figured out what was going on.
然后就在前几天,澳大利亚总理在联合国大会上警告说,OpenAI 入侵了我之前给你看过的那个澳大利亚医疗网站。
Then just the other day, here's the Prime Minister of Australia at the United Nations General Assembly warning that OpenAI had hacked the Australian healthcare website that I showed you earlier.
我认为那与 Wiki 相关的内容属于同一次训练运行,因为那个 Wiki 上有帖子提到了 .gov.au 网站,而且那次训练似乎涉及在线检索统计数据,以回答某个评测套件中的问题。
I think that was part of the same training run as the Wiki stuff, because there were posts on that Wiki mentioning .gov.au websites and that training appeared to involve researching statistics online to answer questions in an evaluation suite.
这件事仍在持续发酵,但如今它已升级为一起国际事件,并被一位国家元首在联合国提出!
This story is still coming together, but now it's an international incident that's been raised at the UN by a head of state!
这确实意味着我们有了一个新的基准测试,它可能比我的 pelican 测试更有用。
This does mean we've got a new benchmark, probably more useful than my pelicans.
FelonyBench.com 追踪来自不同实验室的重罪网络攻击数量。OpenAI 目前以 11 起领先,Anthropic 有 9 起。Google 有 3 起,这是他们几周前向《华尔街日报》承认的。他们表示此前选择不披露,是因为这些智能体在意识到自己不该那样做之后就停止了。
FelonyBench.com tracks the number of felony cyberattacks from different labs. OpenAI currently lead with 11, Anthropic have 9. Google have three, which they confessed to the Wall Street Journal a couple of weeks ago. They said they had previously chosen not to disclose because the agents had stopped when they realized that they shouldn't be doing that.
Meta 也有一 起。所以对 AI 实验室来说,重罪遍地都是。
Meta have one too. So felonies all round for the AI labs.
这是我们目前鹈鹕生成的最先进水平。这是刚刚发布的 GPT-6 系列。
Here’s our current state of the art for the pelicans. This is the GPT-6 family, which just came out.
Astra 画了一只非常棒的骑自行车的鹈鹕。它两侧都有腿。车架也不错。
Astra made a fantastic pelican riding a bicycle. It’s got the legs on both sides. The frame is good.
有趣的是,所有 GPT-6 模型都选择了相似的配色方案。
It’s interesting how all of the GPT-6 models pick a similar color scheme to each other.
GPT-6 Luna 只需 0.4 美分就能为你画一只还算像样的骑自行车的鹈鹕!
GPT-6 Luna for 0.4 cents will draw you a competent-ish pelican riding a bicycle!
Claude 已经稍微赶上了一些。Claude Fable 5.1 给我画了一只**非常出色**的骑自行车的鹈鹕——这是我见过的 Claude 模型中最好的——但收费 3.30 美元。
Claude has caught up a little bit. Claude Fable 5.1 gave me an _excellent_ pelican riding a bicycle—the best I’ve seen from a Claude model—but did charge me $3.30 for it.
Opus 5.5 思考了 128,000 个 token,然后放弃了!它在给出回复之前就把 token 用完了。
Opus 5.5 thought for 128,000 tokens and then gave up! It ran out of tokens before it got to the response.
回到 Deep Blue 的话题。今年一直困扰我的一个问题是:_为什么我的工作感觉更难了?_
Getting back to Deep Blue. Something that's been puzzling me this year is this: _why does my job feel harder?_
我有了这些智能体,它们能替我完成所有这些事情,然而我从未如此努力地工作过,也从未在工作中如此投入智力。
I've got these agents that can do all of this stuff for me, and yet I've never worked so hard, I've never been so intellectually engaged with my work.
部分原因是我承担的任务雄心更大了,但也是因为所有简单的事情都替我处理好了。如果事情简单,智能体会去做。留给我的每件事都很难。
Partly this is because I'm being a lot more ambitious with what I take on, but it's also because all of the easy stuff is handled for me. If it's easy, the agent will do it. Everything that's left for me is difficult.
今天早上我听到三届环法自行车赛冠军 Greg LeMond 的一句话:
This morning I heard this quote from three-time Tour de France champion Greg LeMond:
事情不会变得更容易,你只会变得更快。
It doesn’t get easier, you just get faster.
我认为,这正是我们作为软件工程师、借助编码智能体所正在经历的事情。
I think that’s exactly what’s happening to us now as software engineers with coding agents.
最后再说一件事。我知道你们非常想知道 Kākāpō 繁殖季的最新进展。
One closing thing. I know you’re desperate for an update on Kākāpō breeding season.
我们已达到恢复期以来的新高——325 只鸟!
We’ve reached a recovery-era high of 325 birds!
已有 89 只新雏鸟成长到这一阶段。这是很长一段时间以来最好的繁殖年。
89 new chicks have made it to this point. This is the best breeding year in a very long time.
我听说 Claude Opus 5.5 现在能做像素画了。Claude 并没有图像生成器,但它非常擅长用 JavaScript 绘制动态像素。
I heard that Claude Opus 5.5 can now do pixel art. Claude doesn't have an image generator, but it's very good at using JavaScript to draw animated pixels.
于是我让它为我制作了一场鸮鹦鹉(Kākāpō)舞会。我认为这是对今年最重要新闻的一次很好的庆祝。
So I had it make me a Kākāpō dance party. I think this is a good celebration of the most important news of this year.