The AI landscape shifts fast, but most of the noise obscures the signal. Strip away the product launches and funding rounds, and three structural changes are reshaping what AI systems can actually do in 2026. They are not incremental improvements to chatbots. They are paradigm shifts in how machines perceive, understand, and act in the world — and they are converging.
This post examines each trend: what it is, why it matters, and which projects exemplify it. The thread connecting all three is a single trajectory — from AI that talks to AI that inhabits the physical world alongside us.
2026年重塑AI格局的三大趋势
AI领域变化迅猛，但大多数喧嚣掩盖了真正的信号。抛开产品发布和融资新闻，2026年有三项结构性变革正在重塑AI系统的实际能力。它们不是聊天机器人的渐进式改良，而是机器感知、理解和行动方式的范式转移——而且这三大趋势正在交汇。
本文将逐一剖析每个趋势：它是什么、为什么重要、以及哪些项目代表了这一方向。贯穿三者的主线是一条清晰的演进轨迹——从会"说话"的AI，走向与我们共同"栖居"于物理世界的AI。
1. Full-Duplex Streaming Interaction — From "Turn-Based Chat" to "Real-Time Co-presence"
For years, interacting with AI has felt like using a walkie-talkie. You press a button (or hit Enter), send your message, and wait. The model processes, generates a response, and sends it back. Then it is your turn again. This is half-duplex communication: information flows in one direction at a time.
The limitation runs deeper than mere latency. Half-duplex interaction means the model cannot listen while you speak. It cannot notice you hesitating mid-sentence and adjust. It cannot interrupt you when you are heading in the wrong direction, or interject with a relevant thought before you finish yours. It is, fundamentally, a correspondence system — not a conversation.
In 2026, this is changing. A new generation of AI systems operates in full-duplex streaming mode: simultaneously receiving audio, video, and text input while producing real-time output. The model can listen and speak at the same time. It can observe your facial expressions while formulating its response. It can be interrupted mid-sentence and seamlessly redirect. It can even proactively interject when it detects something important.
This is not a UI improvement. It is a shift from asynchronous message exchange to real-time co-presence — the kind of fluid, overlapping, interruptible interaction that characterizes human conversation.
Thinking Machines Lab Interaction Models
In early 2026, Thinking Machines Lab introduced a family of Interaction Models designed from the ground up for continuous, real-time engagement. Unlike traditional models that process a discrete prompt and emit a discrete response, these models maintain a persistent percept stream: audio, video, and text flow in continuously, and responses flow out continuously. The system can perceive a user's tone shift, notice a gesture, or detect an environmental change — all while it is in the middle of speaking. This is the first production-grade demonstration of full-duplex simultaneous I/O in a general-purpose AI system.
MiniCPM-o 4.5
On the open-source front, MiniCPM-o 4.5 — developed jointly by Tsinghua University and ModelBest — represents a landmark achievement. At approximately 9 billion parameters, it is a full-duplex omni-modal model that runs on edge devices with less than 12GB of RAM. Its architecture, called Omni-Flow, unifies audio, visual, and textual streaming into a single bidirectional pipeline. The model can watch a live camera feed, listen to speech, read on-screen text, and respond vocally — all simultaneously, all in real time. The significance is practical: full-duplex interaction is no longer confined to cloud-scale infrastructure. It can run on a laptop.
1. 全双工流式交互——从"回合制聊天"到"实时共在"
多年来，与AI交互的体验一直像使用对讲机。你按下按钮（或敲回车），发送消息，然后等待。模型处理、生成回复、发送回来。然后又轮到你。这就是半双工通信：信息一次只能单向流动。
这个局限远不止延迟问题。半双工交互意味着模型无法在你说话的同时倾听，无法察觉你句子中间的犹豫并做出调整，无法在你方向错误时打断你，也无法在你话还没说完时就插话提出相关想法。从根本上说，这是一个信件往来系统——而非对话。
2026年，这一切正在改变。新一代AI系统运行在全双工流式模式下：同时接收音频、视频和文本输入，同时产生实时输出。模型可以边听边说，可以在组织回复的同时观察你的面部表情，可以在话说到一半被打断后无缝转向，甚至可以在检测到重要信息时主动插话。
这不是UI的改进，而是从异步消息交换到实时共在的转变——那种流动的、交叠的、可中断的交互方式，正是人类对话的本质特征。
Thinking Machines Lab 交互模型
2026年初，Thinking Machines Lab推出了一系列从底层设计就面向持续实时交互的交互模型（Interaction Models）。与处理离散提示并输出离散回复的传统模型不同，这些模型维持一条持续的感知流：音频、视频和文本持续流入，回复持续流出。系统可以在说话过程中感知用户语气的变化、注意到手势、或检测到环境变化。这是通用AI系统中首次达到生产级水准的全双工同步I/O演示。
MiniCPM-o 4.5
在开源领域，清华大学与面壁智能（ModelBest）联合开发的MiniCPM-o 4.5是一座里程碑。约90亿参数的全双工全模态模型，可在不到12GB内存的边缘设备上运行。其名为Omni-Flow的架构将音频、视觉和文本流统一为单一双向管线。模型可以同时观看实时摄像头画面、聆听语音、阅读屏幕文字并以语音回应——全部同步进行，全部实时处理。其实际意义重大：全双工交互不再局限于云端基础设施，它已经可以在一台笔记本电脑上运行。
2. Omni-Modal Fusion — From "Multi-Modal Input" to "Real-World Understanding"
The first wave of multi-modal AI was essentially a stapler. A language model was trained on text. Then, separately, a vision encoder was trained on images. Then, a speech recognizer was trained on audio. Finally, the outputs of all these modules were fed into the language model as text tokens. The result could accept multiple modalities, but it did not understand the world through them. It understood text about images, text about audio, text about video — but never the raw, interconnected fabric of reality itself.
True omni-modal models are different. They learn a unified representation space where text, image, audio, video, code, and even action sequences exist as points in the same geometric landscape. A spoken sentence, a facial expression, a line of code, and a physical gesture are not translated into a common language — they already are the same language, encoded in the same embedding space.
This matters because understanding the world requires more than recognizing individual modalities. It requires grasping spatial relationships (objects exist in 3D space and occlude each other), temporal continuity (events unfold in sequence and have cause and effect), physical laws (gravity pulls, light reflects, solid objects collide), causal structures (actions have consequences), and human intent (people act for reasons). These are not features you can bolt on. They emerge — or fail to emerge — from the architecture itself.
Google Gemini Omni
The clearest illustration of this shift is Google Gemini Omni. This system combines reasoning and generation into a single model that emphasizes world understanding. Its headline capability is transforming arbitrary combinations of image, text, video, and audio inputs into unified, coherent video output. But generating video is not the point — understanding enough to generate is the point.
Consider what video generation actually requires. A person turning around must maintain consistent body structure throughout the rotation. A dropped ball must accelerate downward at a rate consistent with gravity. A camera panning across a room must preserve spatial relationships between objects. Multiple people interacting must exhibit causal relationships — one person's action triggers another's response. Get any of these wrong, and the video looks uncanny, dreamlike, obviously synthetic.
Gemini Omni's ability to produce physically plausible video is evidence that the model has internalized something resembling a world model — not just statistical patterns of pixels, but an implicit understanding of the spatial, physical, and causal logic that governs how the real world looks when it moves. This is a step from "multimodal model" to world generative model: a system that generates pixels that obey real-world physics.
World Labs and the Spatial Intelligence Frontier
A parallel effort is underway at World Labs, the venture founded by Stanford's Li Fei-Fei. World Labs pursues spatial intelligence — the ability for AI to understand and reason about 3D environments, object relationships, and physical interactions. The goal is not merely to label objects in a scene but to understand the scene as a navigable, manipulable space. This represents the same directional shift: from AI that sees to AI that understands what it sees in physical, spatial terms.
2. 全模态融合——从"多模态输入"到"真实世界理解"
第一波多模态AI本质上是一台订书机。语言模型在文本上训练；然后，另外训练一个视觉编码器处理图像；再训练一个语音识别器处理音频；最后，所有这些模块的输出作为文本token喂给语言模型。结果可以接受多种模态，但并不通过它们理解世界。它理解的是关于图像的文本、关于音频的文本、关于视频的文本——而非现实本身那种原始的、相互交织的纹理。
真正的全模态（omni-modal）模型截然不同。它们学习一个统一表征空间，其中文本、图像、音频、视频、代码甚至动作序列都作为同一片几何地貌中的点存在。一句话语、一个面部表情、一行代码和一个肢体手势，并不需要被翻译成共同语言——它们本身就是同一种语言，编码在同一个嵌入空间中。
这之所以重要，是因为理解世界需要的远不止识别单独的模态。它需要把握空间关系（物体存在于三维空间并相互遮挡）、时间连续性（事件按序展开且有因果关系）、物理定律（引力下拉、光线反射、固体碰撞）、因果结构（行动产生后果）和人类意图（人的行为有其动因）。这些不是可以后期加装的模块，它们从架构本身中涌现——或者无法涌现。
Google Gemini Omni
这一转变最清晰的例证是Google Gemini Omni。该系统将推理与生成整合进单一模型，强调对世界的理解。其标志性能力是将图像、文本、视频和音频的任意组合输入转化为统一连贯的视频输出。但生成视频不是重点——理解得足够深入才能生成才是重点。
想想视频生成实际需要什么。一个人转身时，必须在整个旋转过程中保持一致的身体结构。一个掉落的球必须以符合重力的加速度下落。镜头平移扫过房间时，必须保持物体之间的空间关系。多人互动必须展现因果关系——一个人的动作引发另一个人的反应。任何一个环节出错，视频就会看起来诡异、梦幻、明显是合成的。
Gemini Omni生成物理合理视频的能力，证明模型已经内化了某种类似世界模型的东西——不仅仅是像素的统计模式，而是对真实世界运动时所遵循的空间、物理和因果逻辑的隐性理解。这是从"多模态模型"到世界生成模型的一步跨越：一个生成的像素服从真实世界物理规律的系统。
World Labs与空间智能前沿
另一条平行的探索来自斯坦福大学李飞飞教授创办的World Labs。World Labs追求空间智能——AI理解和推理3D环境、物体关系和物理交互的能力。其目标不仅仅是标注场景中的物体，而是将场景理解为一个可导航、可操控的空间。这代表着同一个方向的转变：从会"看"的AI，到以物理空间方式"理解所见"的AI。
3. World Action Model (WAM) — From Semantic Understanding to "Predict and Act"
The first two trends give AI the ability to perceive in real time and understand the world it perceives. But perception without action is observation, not agency. The third trend closes the loop.
Traditional world models have a well-defined job: given the current state of the world, predict the next state. Show the model a frame of a ball mid-air, and it predicts the next frame — the ball a bit lower. This is powerful for simulation and planning, but it is fundamentally passive. The model watches the world change; it does not ask how its own actions would change the world.
A World Action Model (WAM) extends this with a critical question: "After I take action X, how will the world change? And to reach goal Y, what sequence of actions should I take?" The model no longer just predicts world states — it predicts world states conditioned on its own actions and then generates those actions. World state prediction and action planning become a single, unified process.
The distinction is crisp: a world model predicts "how the world changes." A WAM predicts "how my actions change the world." The first is a spectator. The second is an agent.
NVIDIA WAM
NVIDIA's WAM is a robot AI model pretrained on large-scale video data. It jointly learns to predict future world states and the robot actions that produce those states. By training on video — which inherently contains both visual state transitions and the implicit actions that caused them — the model learns the coupling between "what I do" and "what happens." This is not imitation learning (copying demonstrated actions) and not reinforcement learning (trial-and-error in a simulator). It is a third path: learning the action-consequence structure of the physical world directly from observation.
DreamZero
Published in 2026, DreamZero is a 14-billion-parameter robot foundation model built on an image-to-video diffusion backbone. Given a single image of a robot and its environment, DreamZero simultaneously predicts the robot's next actions and generates a video of the resulting future state. The two predictions — action and visual outcome — are generated jointly, each informing the other. The robot does not first decide what to do and then visualize the result; it reasons about doing and seeing as one integrated process.
Zhiyuan Genie Envisioner
Released at WAIC 2025, the Zhiyuan Genie Envisioner is an action-driven world model platform with a distinctive architecture. Its foundation is GE-Base, a multi-view video diffusion model that can imagine how a scene will look from multiple camera angles. Layered on top is GE-Act, an action decoder that translates predicted visual futures into concrete motor commands. Together, they enable a closed loop: the system can imagine a future state (making a sandwich, pouring tea), verify that the imagined state is physically plausible, and then act to realize it. This "imagine, verify, act" cycle is a concrete demonstration of what WAM makes possible — not just predicting the future, but planning and executing actions to bring a desired future into being.
3. 世界行为模型（WAM）——从语义理解到"预测与行动"
前两大趋势赋予了AI实时感知和理解所感知世界的能力。但没有行动的感知只是观察，而非能动性。第三个趋势闭合了这个回路。
传统的**世界模型（world models）**有明确的职责：给定世界当前状态，预测下一个状态。给模型看一帧球在空中的画面，它预测下一帧——球的位置稍低一些。这对于仿真和规划很有用，但本质上是被动式的。模型观察世界变化，并不追问自身行动会如何改变世界。
**世界行为模型（WAM）**在此基础上增加了一个关键问题："在我执行动作X之后，世界会如何变化？为了达到目标Y，我应该采取怎样的动作序列？"模型不再仅仅预测世界状态——它预测以自身动作为条件的世界状态，然后生成那些动作。世界状态预测与动作规划成为单一统一的过程。
区别很清晰：世界模型预测**"世界如何变化"，WAM预测"我的行动如何改变世界"**。前者是旁观者，后者是行动者。
NVIDIA WAM
NVIDIA的WAM是一个在大规模视频数据上预训练的机器人AI模型。它联合学习预测未来世界状态和产生这些状态的机器人动作。通过在视频上训练——视频天然包含视觉状态转换和引发这些转换的隐含动作——模型学习到"我做什么"与"发生什么"之间的耦合。这不是模仿学习（复制示范动作），也不是强化学习（在模拟器中试错），而是第三条路径：直接从观察中学习物理世界的行动-后果结构。
DreamZero
2026年发表的DreamZero是一个140亿参数的机器人基础模型，构建在图生视频扩散模型骨架之上。给定一张机器人及其环境的单张图像，DreamZero同时预测机器人的下一步动作并生成 resulting 未来状态的视频。两种预测——动作和视觉结果——联合生成，彼此相互影响。机器人不是先决定做什么再可视化结果；它将"做"和"看"作为一个整合过程来推理。
智元精灵视界（Genie Envisioner）
在WAIC 2025上发布的智元精灵视界是一个以动作为驱动的世界模型平台，架构独具特色。其基座是GE-Base，一个多视角视频扩散模型，能从多个摄像机角度想象场景的未来面貌。其上叠加了GE-Act，一个将预测的视觉未来转化为具体电机指令的动作解码器。两者结合形成闭环：系统可以想象一个未来状态（做三明治、倒茶），验证想象的状态是否物理合理，然后行动以实现它。这个"想象-验证-行动"循环具体展示了WAM的可能性——不仅仅是预测未来，而是规划和执行动作，将期望的未来变为现实。
Convergence — From Chatbots to Embodied Intelligence
These three trends do not exist in isolation. They are converging, and their convergence points somewhere specific.
Full-duplex interaction provides the communication layer — the ability for AI to engage with humans and environments in real time, with the fluidity of natural conversation rather than the rigidity of command-response protocols. Omni-modal fusion provides the understanding layer — a unified representation of the world that integrates sight, sound, language, and physical intuition into a coherent model of reality. World Action Models provide the agency layer — the ability to not just perceive and understand, but to predict the consequences of actions and plan sequences of behavior to achieve goals.
Stack these three together and you get something qualitatively different from a chatbot. You get a system that can perceive the world in real time through multiple senses, understand what it perceives in physical and causal terms, plan actions to achieve objectives, and communicate with humans fluidly throughout the process. You get, in short, the architecture for embodied general intelligence.
The trajectory is visible in the industry. Tesla's Optimus program, Figure AI's humanoid robots, and the wave of robotics startups emerging from both Silicon Valley and Shenzhen are all converging on architectures that resemble WAM-like world understanding combined with real-time perception and action. The question is no longer whether AI will move from screens into the physical world. The question is how fast, and which of these converging paradigms will prove most tractable at scale.
The chatbot era was the prologue. The embodied era is beginning.
趋势交汇——从聊天机器人到具身智能
这三大趋势并非孤立存在，它们正在交汇，而交汇之处指向一个明确的方向。
全双工交互提供通信层——让AI能够以自然对话般的流畅性，而非命令-响应式的僵硬协议，与人类和环境进行实时交互。全模态融合提供理解层——将视觉、声音、语言和物理直觉整合为对现实的连贯模型，形成对世界的统一表征。世界行为模型提供能动性层——不仅能感知和理解，还能预测行动后果并规划行为序列以达成目标。
将这三者叠加，你得到的东西与聊天机器人在质上截然不同。你得到的是一个能够通过多种感官实时感知世界、以物理和因果逻辑理解所感知内容、规划行动以达成目标、并在整个过程中与人类流畅沟通的系统。简而言之，你得到的是具身通用智能的架构。
这条轨迹在产业界已经清晰可见。特斯拉的Optimus项目、Figure AI的人形机器人，以及从硅谷到深圳涌现的一波波机器人创业公司，都在向类似WAM式世界理解结合实时感知与行动的架构收敛。问题已不再是AI是否会从屏幕走入物理世界，而是速度有多快，以及这些交汇的范式中哪一种将在规模化时被证明最为可行。
聊天机器人时代是序章。具身时代正在开启。