State of AI with Nathan Benaich - The case for intelligence beyond language
Summary
在2026年Rise峰会上，Google DeepMind研究副总裁Raya Hadsell提出核心观点：只要用足够的数据训练一个大模型来学习某个复杂系统，它就能掌握该系统的底层规律；语言只是众多此类系统之一，生物、合成世界与物理运动同样如此。她批评整个领域把语言当成全部游戏，认为这是一种浪费。她以刚发布的Diffusion Gemma为例，说明扩散式文本生成能一次生成整块内容、双向推理并自我纠错，因而在数独等需要全局推理的任务上优于逐词生成的自回归模型，还能带来更快推理和自适应计算。她进一步阐述世界模型（从Genie 1到Genie 3）如何提供无限丰富的训练环境，被DeepMind视为通往AGI的关键，并在教育等场景中充满潜力。在机器人方面，她坦言这是最难的部分，DeepMind将问题拆分为「大脑」与「脊椎」，并用Gemini Robotics与视觉-语言-动作模型协同控制。她指出前沿在于高效学习，世界模型可生成以假乱真的交互数据来训练机器人，同一套基础模型配方最终既能理解世界又能训练机器人。她的结语呼应开场的不满：应停止只盯着语言，未来十年的进步在于世界、机器人、生物、天气等更丰富的系统，而与把聊天机器人做得更大关系不大。
Highlights
-
The most important thing Google DeepMind has learned in 20 years, Raya Hadsel argued at this year's Rise, is that a large model trained on enough data from a complex system learns the underlying patterns of that system.
Raya Hadsell在今年的Rise峰会上认为，Google DeepMind二十年来学到的最重要的一课是：一个在某个复杂系统的足够数据上训练出来的大模型，能够学会该系统的底层规律。
States the central thesis of the whole talk -
It is the reason, she said, that a left-to-right model is structurally ill-suited to Sudoku, which demands seeing the whole grid and reasoning across it at once.
她说，这正是为什么从左到右的模型在结构上不适合解数独——数独要求同时看到整个网格并对其进行全局推理。
Concrete, surprising limitation of autoregressive models -
Prompt a country lane in Kent, England, ad rain, and the model fills in the person walking, the puddles, and a pair of yellow boots she's fairly sure she does not own.
给它一个提示——英格兰肯特郡的一条乡间小路，再加上雨——模型就会补全走路的人、水洼，以及一双她相当确定自己并不拥有的黄色靴子。
Vivid, memorable demonstration of world models -
World models now sit in that chain. They can generate novel interaction data at high enough fidelity to train the robot model. And Hadsell said, you cannot tell the real robot logs from the simulated ones.
世界模型如今处在这条链条中。它们能以足够高的保真度生成全新的交互数据来训练机器人模型。Hadsell说，你无法区分真实机器人日志和模拟出来的日志。
Striking claim that simulated data is indistinguishable from real -
Almost all the field's attention sits on language, and she thinks that is a waste. The same recipe that produced today's language models can be pointed at far richer systems. Worlds, robots, biology, weather.
整个领域几乎所有注意力都集中在语言上，而她认为这是一种浪费。造就今天语言模型的同一套配方，可以对准远比语言丰富得多的系统——世界、机器人、生物、天气。
Bold closing opinion challenging the field's focus
Full transcript
The case for intelligence beyond language with Raya Hadsel, VP of research at Google DeepMind at Rise 2026. The most important thing Google DeepMind has learned in 20 years, Raya Hadsel, argued at this year's Rise, is that a large model trained on enough data from a complex system learns the underlying patterns of that system. Language is one such system, whether biology, synthetic worlds, and physical movement are others.
and her frustration is that the field keeps treating language as the whole game. Hadsell spoke at the first research and applied AI summit in 2017 and returned to open the 10th. She is now VP of research at DeepMind, where she co-leads a frontier AI unit of roughly a thousand researchers and engineers across about a hundred projects, and she did not start in machine learning. Her undergraduate degree was in philosophy. Her talk was in part a corrective.
The case against one token at a time. Less than 48 hours before she spoke, DeepMind released Diffusion Gemma, a new open model in the Gemma 4 family, under an Apache 2.0 license. It is a 26 billion parameter mixture of experts model, and it generates text in a way most language models do not. Instead of producing one token after another, it borrows the recipe behind image and video generation. Take a sample, add noise, and learn to denoise it.
Applied to text, that means generating a whole block at once, rather than committing to each word in sequence. The mechanism buys two things. Inference is faster because the model is not walking left to right. And because it writes and revises a whole block, it can reason in both directions and correct itself. Hadsell showed it solving a problem while displaying its work. After the first denoising step, it commits to the wrong answer. By the second, it reconsideres.
And by the third, the whole block is consistent and right. An autoregressive model, having locked in each token as it went, cannot do that. It is the reason, she said, that a left-to-right model is structurally ill-suited to Sudoku, which demands seeing the whole grid and reasoning across it at once. Diffusion also delivers adaptive computation as a side effect. Easy questions converge in a couple of steps. Harder ones take more. In practice, it is fast enough to keep up with you in real time.
writing, running, and testing code as quickly as you can describe what you want. Still, Diffusion Gemma is an experimental release rather than a replacement for the standard autoregressive Gemma 4 models, and Text Diffusion carries trade-offs of its own. DeepMind shipped the fine-tuning code alongside the model. Better agents need better worlds. Hadsel traced DeepMind's world model work back to a simple idea. You get better agents by building better environments, ones rich enough to keep generating new situations to learn from.
A world model simulates the dynamics of an environment, and DeepMind sees these as central to the path to AGI because they supply unlimited, rich training worlds for agents. The line runs from Genie 1, trained in 2023 on 2D platformer games, and published in 2024 to Genie 2, which moved to latent diffusion and produced interactive 3D environments that were still synthetic and not real-time to Genie 3, which is high resolution.
works across domains, and runs in real time for several minutes while staying consistent. You feed it a text prompt and get back a world you can move through. Prompt a country lane in Kent, England, ad rain, and the model fills in the person walking, the puddles, and a pair of yellow boots she's fairly sure she does not own. You can change the world midstream by typing a new prompt into it. Last month, the team grounded Genie in Street View. They take one of Street View's static panoramas, turn it into a short video, prompt in changes such as winter weather or a character and feed that in as the opening frames of a genie world. Enough to walk across the Golden Gate Bridge and then dive underwater while the model keeps the bridge's geometry intact. Hadsel is most excited about education. Being present at a historical event instead of reading about it or following a red blood cell through the circulatory system with the model tied to a textbook chapter so that, in her words, no dragons will come in if we don't want them to.
Robotics is still the hard part. Hadcel has worked on robots for more than a decade and is blunt that the field is hard, an interplay of hardware, software, and understanding with the goal of general purpose machines that can drop into any task a human can do. DeepMind splits the problem the way a body does, into a brain and a spine. The spine handles the fast, fine motor control that should not have to rut through the brain. The brain handles understanding and planning. Two models carry this.
Gemini Robotics ER 1.6 is a Gemini model tuned for spatial understanding and tool use, able to pick individual objects out of cluttered scenes. A separate vision language action model takes the visual input and the reasoner's instructions and outputs motor control. Working with Boston Dynamics, they found tasks that needed more than spatial sense. Getting spot to read a small gauge required tool calls to magnify the dial and reasoning to interpret it. The frontier, she said, is efficient learning. The ways to collect robot data run from expensive, narrow tele-operation, through physics simulators like Mujoko, to the prize of learning directly from Internet video. World models now sit in that chain. They can generate novel interaction data at high enough fidelity to train the robot model. And Hadsell said, you cannot tell the real robot logs from the simulated ones. Pool that with cross-embodiment learning, where data collected on one robot transfers to others.
and the performance gain comes for free. This is where the talk's argument closes on itself. The same foundation model recipe pointed at worlds becomes the engine that trains the robots. Stop staring at language. Her closing note returned to the frustration she opened with. Almost all the field's attention sits on language, and she thinks that is a waste. The same recipe that produced today's language models can be pointed at far richer systems. Worlds, robots, biology, weather, and the diversity of Earth's biosphere at planetary scale, the data already exists. The models have barely been built. The next decade of progress on her account runs through those systems and has little to do with making chatbots bigger.