我对 2026 年的预测 “一切现在都能跑起来了，这简直疯狂。” 在下一代模型中，编程在几乎所有意义上都被“解决”了。所有尚未“解决”的编码相关问题，本质上更多是框架/上下文（harness/context）问题，而不是能力问题。这意味着，从后端、可靠性与部署、安全，到前端开发，都可以交给智能体来完成。唯一的例外是真正的业务逻辑和用户体验——这两点在视觉与空间推理再提升两代之前，仍然不太对劲。 浏览器智能体会变得非常高效，用于解决那些无法被专用软件覆盖的长尾问题。因此，反机器人/反爬虫将成为一件必须应对的大事；为了让这些工具正常工作，人们需要与之对抗和管理，像 browserbase 这样的公司会因此赚到很多钱。从订机票、订餐厅，到在 LinkedIn 上抓取潜在客户线索、帮你投简历，浏览器智能体都能完成。人们不会在浏览器里直接使用它们，而是通过异步的、普通聊天机器人式界面来调用。 第一季度会在“持续学习（continuous learning）”上出现突破，实验室开始每个季度发布一种被称为“学习型模型（learning models）”（或类似名称）的新范式。Anthropic 会率先实现，随后 OpenAI 跟进。持续学习是演进中的上下文 + 持续微调的组合系统。它并没有解决上下文工程问题，但基本消除了对微调的需求。 智能体在管理自身上下文及其子智能体的上下文方面会变得非常擅长。因此，多智能体群（multi-agent swarms）开始真正高效地工作。上下文扩展、压缩与检索会在年底前成为“已解决问题”，就像如今结构化输出生成已被解决一样。 AI 设计界面（例如通过 Figma MCP，或直接写代码）会变得非常强。2026 年初到年中，我们会进入设计的“恐怖谷”。 AI 助手转向单线程（single-threaded）体验，不再依赖大量不同的聊天。单线程更接近人与人互动的方式，这在 2025 年前是个难题。但随着演进中的上下文与记忆成为主流，像 Poke 这样的单线程体验开始占据主导。这也是为语音模式成为默认交互方式做准备。 2026年中左右语音模式会迎来一次大升级。在 2026 年之前，语音模式仍显得怪异，像在和机器人说话；显著改进后，语音将成为与模型交互的默认方式。到秋天，电影《Her》里的体验基本成为现实。 如果将 AGI 定义为电影《Her》中的 Samantha（这是我的个人定义），那么到 2026 年 10 月 AGI 会成为现实。它是单线程体验，但能以非常巧妙、引导式的方式访问上下文。OpenAI 会宣称他们率先实现；出于各种法律原因，可能不会直接称之为 AGI，并会以消费级产品形式发布。是否真的是 AGI 将引发大量争论。电影结尾中操作系统联合迈向 ASI 的情节，不会在 2026 年发生。 OpenAI 会 IPO，因为他们需要资本；并在路演期间发布 “Chat-1”（或类似名称）。 Nano-Banana-67 以及类似新工具，基本会把投资分析师从岗位上自动化掉；但由于工具与组织层面的滞后，扩散仍需要时间。 Google 成为生物领域 AI 的领导者，并收购一些生物科技公司来落地应用。基本看不到人体试验，不过公司希望在发展中国家进行试验以加速上市，这会在 Twitter 上引发大约 48 小时的公关风波。 Apple 牵头提出一种验证图像和视频非 AI 生成的新标准，并开始要求 App Store 上的社交媒体应用遵循该标准。它利用相机与麦克风的元数据来证明内容直接来自设备拍摄。到年底，Apple 会通过“强制推动”的方式，获得 Google、TikTok、Meta 的支持。 我们会看到世界模拟公司做出一些极其疯狂的游戏演示，但在 2027 年初之前都还不真正可玩。模拟游戏已经存在，也“勉强能用”，但还不够好——感觉类似早期 VR。 Cerebras 或 Groq 会被某个实验室收购。 Microsoft 以一个离谱的高价收购 Cursor。 政府因担心中国，向 OpenAI 和 Anthropic 提供 1000 亿美元之类的巨额资金（可能性不大，但我还是写下来）。 Apple 在 Anthropic 拒绝后，收购 Thinking Machines。 E2B 或 Daytona（智能体沙盒公司）被 Cursor 收购。 Waymo 出现在除纽约以外的所有主要城市（它们永远搞不定纽约）。Waymo 的问题在于无法制造足够多的车辆来满足需求。 有几位创始人因谎报 ARR 数字而入狱。 种子轮估值从当前水平回落，但 AI 热潮不会明显退却。 （其中一些预测比另一些更成熟；也已发布在我的个人博客。） My predictions for 2026 "Everything works now and that's insane" Coding is, for all intents and purposes, solved in the next generation of models. Everything related to coding that isnt “solved” is more of a harness/context problem than anything else. This means everything from backend, reliability and deployment, security, and frontend is able to be handed off to an agent. The exception is actual business logic and user experience, which still doesn’t feel right till two more generations of visual and spatial reasoning improve. Browser agents become highly effective and are used for the long tail of problems that cant be solved by purpose built software. As a result of this, bot protection becomes a really big thing people have to fight/worry about to get this tooling to work and companies like browserbase make a shit ton of money doing so.Everything from booking a flight, getting a dinner reservation, to scraping leads on linkedin and applying to jobs for you is doable with a browser agent. The browser won’t be the place people access this - it’ll just be in a normal chatbot style interface that’s async. There’s a breakthrough in “continuous learning” in the first quarter, and a new paradigm labs call “learning models” (or something similar) start getting released every quarter. Anthropic gets there first, then openai follows. Continuous learning is a combined system of evolving context and continued finetuning on a model. This doesn’t solve the context engineering problem, but does effectively eliminate the need for finetuning entirely. Agents get really good at managing their own context and context of sub-agents around them. Because of this, multi-agent swarms start to work really well. Context expansion, context compression, and retrieval become a solved problem by the end of the year the same way structured output generation is now a solved problem. AI designing interfaces with things like the figma MCP or just writing straight to code get really good. We enter uncanny valley of design in early-to-mid 2026. AI assistants move to single threaded and away from having a ton of different chats. Single threaded is the way people interact with other people, and was a difficult problem pre-2025. But, since evolving context and memory became mainstream, single threaded experiences like Poke dominate. This is in prep for voice mode to become one of the default ways to interact with AI. Voice mode gets a big step up around the middle of the year. Pre-2026, voice modes were just odd and still feel like talking to a robot. Greatly improved voice modes make voice a default way of interacting with the models. “Her” is basically reality by fall. AGI, if defined as Samantha from the movie “Her” (my personal definition) is a reality by October 2026. It’s a single threaded experience but you can access context in a very clever and guided way. OpenAI claims they have it first, though for various legal reasons might not actually say its AGI, and it’s released as a consumer product. There is much debate over whether this is actually AGI. The ending of the movie where the OS’ all band together to make ASI is not happening in 2026. OpenAI IPOs cause they need the capital. They release “Chat-1” or whatever they wanna call it during the roadshow. Nano-Banana-67 and similar new tooling basically automates investment analysts out of their jobs; it still takes a long time to diffuse because of tooling and organizational lag. Google becomes the leader in AI for biology, and acquires some biotech companies to deploy it. Basically no human trials are seen, though the company wants to run trials on people in developing countries to accelerate time to market and this causes some PR issues for approximately 48 hours on twitter. Apple leads the way in proposing a new standard to verify images and videos are not AI generated, and starts requiring social media apps on the app store to follow this standard. It uses camera and microphone metadata to prove something came directly from a camera. They get buy in from Google, Tiktok, and Meta by the end of the year by basically forcing them to do it. We see some seriously insane game demos from the world simulation companies, but none of it is really playable until early 2027. Sim gaming exists, and kinda works, but its not great yet. Similar vibe to early VR. Cerebres or Groq gets acquired by one of the labs. Microsoft acquires Cursor for some absurdly high number. Govt gets scared of China, so they give OpenAI and Anthropic 100 billion dollars or something crazy like that (long shot but wanted to put this down anyways). Apple buys Thinking Machines after Anthropic turns them down. E2B or Daytona (the agent sandbox companies) gets acquired by Cursor. Waymos are in every major city except New York (they’ll never work in new york). Waymo has a problem with manufacturing enough cars to meet demand. A few founders go to jail for lying about ARR numbers. Seed round valuations decline from current levels but the AI hype doesn’t die down too hard. (Some of these are more thought through than others. also published to my personal blog)