Author: Techub News Compilation

Introduction

In September 2025, Demis Hassabis, co-founder and CEO of Google DeepMind, appeared on the well-known tech podcast All-In Podcast for an extensive dialogue rich in information. As a key driver of milestone AI projects such as AlphaGo and AlphaFold, as well as a co-recipient of the 2023 Nobel Prize in Chemistry, Hassabis has always been a beacon in the global AI field. In this conversation, he not only shared behind-the-scenes stories of receiving the Nobel Prize but also systematically explained the latest advancements at Google DeepMind, his thoughts on the path to AGI (Artificial General Intelligence), and how AI is reshaping the infrastructure of science, entertainment, and even human society. In the context of a global AGI wave and fierce competition among tech giants, Hassabis's insights are undoubtedly of high reference value.

Summary

• The world model is a key step toward AGI: Models like Genie learn “intuitive physics” through vast amounts of video, generating dynamic interactive worlds without pre-setting a physics engine, which is vital for robotics and smart glasses assistants.
• AGI may still need 5-10 years: Hassabis believes current AI still lacks true creativity, continuous learning, and consistent reasoning, requiring one or two key breakthroughs to reach true general intelligence.
• AI will usher in a “new Renaissance in science”: From drug discovery (Isomorphic Labs) to material design and fusion control, AI will become the ultimate scientific tool, with long-term contributions far exceeding its current energy consumption.
• Humanoid robots have unique value: For general robots that need to integrate into human environments (homes, offices), humanoid forms may be more practical and adaptable than specialized forms.
• The duality of AI empowering creation: It greatly lowers the threshold for ordinary people to create while super-empowering top professionals, giving rise to new entertainment forms of “co-creation.”

From AlphaFold to Nobel Prize: The Starting Point of AI-Driven Science

The interview began with Hassabis's extraordinary experience of receiving the Nobel Prize in Chemistry. He described receiving a call from Sweden as “like a dream,” a moment every scientist dreams of. During the award week in Stockholm, what shocked him most was signing his name next to greats like Feynman, Curie, Einstein, and Niels Bohr in a signed book preserved for 120 years. Hassabis admitted that despite rumors about AlphaFold possibly winning the prize, the Nobel committee's extreme confidentiality requirements and the award's greater emphasis on the actual impact of scientific breakthroughs (which often requires decades of accumulation) made everything uncertain until the phone rang. AlphaFold's solution to the protein structure prediction problem was a perfect example of Hassabis's childhood dream—using AI to accelerate scientific discovery. This Nobel medal not only recognizes him and his team but also signifies that AI has officially become the core engine of foundational scientific discoveries.

Speaking of Google DeepMind’s positioning within Alphabet, Hassabis likened it to “the engine room of the entire Google and Alphabet.” Since the merger of all AI efforts within Google (including the original DeepMind) into Google DeepMind a few years ago, the team has grown to about 5,000 people, over 80% of whom are engineers and PhD researchers. Gemini is the core model they have built, along with various AI systems including video models and interactive world models. These models have been deeply integrated into nearly all of Google's products and services, and billions of users interact with their AI technology daily, from AI Overview and AI modes to Gemini applications. This combination of “cutting-edge research” and “billion-scale deployment” constitutes Google DeepMind's unique advantage.

Genie and World Models: Enabling AI to Understand the Physical World

The conversation quickly shifted to a revolutionary class of models recently released by Google DeepMind—the “world models,” represented by Genie. Hassabis demonstrated a video generated by Genie on-site: users can create a fully interactive 3D environment with just a text prompt (such as “a person in a chicken suit” or “jet ski”). Users can control the perspective in real-time using arrow keys and the spacebar, exploring this world. The key point is that all pixels are generated in real-time, not existing in any pre-set video or game. When a character turns to look in one direction, the scenery on that side is created instantaneously while maintaining physical consistency (e.g., graffiti on the previously visible wall will remain).

Hassabis explained that the Genie model reverse-engineered “intuitive physics” by watching millions of YouTube videos and others on the internet. It is not based on traditional 3D rendering engines (like Unity or Unreal), nor is it pre-programmed with physical laws. Instead, it learns the complex laws governing how the world operates purely from sequences of 2D images, including light reflection, material flow, and object behavior. As someone who developed game engines and physics systems in the 90s, Hassabis expressed being “extremely shocked” that Genie could accomplish all of this “so effortlessly.”

Why are world models so important? Hassabis pointed out that to build true AGI, systems must understand the physical world around us, not just the abstract worlds of language or mathematics. This is crucial for applications like robotics and smart glasses assistants. An intelligent glasses assistant that helps in daily life must understand the physical environment its user is in and the underlying physical laws. The breakthrough of world models equips AI systems with the "senses" and "common sense" to comprehend the physical world. He predicts that as models like Genie evolve (Gen 4, Gen 5…), they will get increasingly close to a profound simulation of the real world.

Robots, AGI Timeline, and Missing Puzzle Pieces

The natural extension of world models is robotics technology. Hassabis introduced their “Gemini Robot Model,” which is fine-tuned based on Gemini. In the demonstration, a desktop robot with two robotic arms could understand natural language instructions such as “put the yellow object in the red bucket” and translate them into precise sequences of actions. This illustrates the power of multimodal models—they can integrate the understanding of the real world with interactions with robots rather than relying solely on specialized, narrow models for robots.

Regarding the form of robots, Hassabis's viewpoint has evolved. He previously believed that task-specific robots needed dedicated forms, but now he thinks that humanoid robots hold significant importance for general robots that must work in human-designed environments. Stairs, door handles, tools—our physical world is designed for human bodies. Instead of redesigning the whole world, it is better to make robots adapt to our existing environment. Of course, in specific scenarios like industrial production lines or laboratories, specialized robot forms remain the optimal solution.

On the timing for the arrival of AGI (Artificial General Intelligence), Hassabis provided a relatively cautious estimate. He believes that current top AI models (including those from his own company) are far from “doctoral-level intelligences.” They may exhibit doctoral student-level capabilities on certain tasks, but lack generality and consistency, sometimes making basic errors in high school-level math or simple counting. True AGI should fully achieve high-level performance.

In his view, there are still several core puzzle pieces missing on the road to AGI: first, true creativity, meaning the ability to make “intuitive leaps” from known knowledge, proposing entirely new hypotheses or theories, similar to how Einstein presented the special theory of relativity; second, continuous learning ability, to learn new knowledge online and adjust behavior; and third, a stronger capability for reasoning and consistency. Hassabis predicts that it may take “one or two key breakthroughs,” with a time scale of about 5 to 10 years. This contrasts with some more radical industry predictions (like within a few years).

Scientific Revolution and Energy Challenges: The Long-Term Returns of AI

As a scientist, Hassabis firmly believes that AI's highest mission is to accelerate scientific discovery and address significant challenges faced by humanity. In addition to AlphaFold, Google DeepMind's AI systems have been applied in various scientific fields such as material design, plasma control of fusion reactors, weather forecasting, and solving Olympiad-level mathematics problems. The Isomorphic Labs he leads (spun off from DeepMind) is focused on breakthroughs based on AlphaFold, aiming to overhaul the drug discovery process with the goal of shortening the new drug development cycle from several years or even a decade to a few weeks or days. They have currently collaborated with pharmaceutical giants like Eli Lilly and Novartis and expect to have candidate drugs enter preclinical stage next year.

Regarding the current hot topic of AI energy consumption, Hassabis believes there are “two truths.” On one hand, Google DeepMind has invested heavily in optimizing model efficiency due to the need to provide AI services efficiently to hundreds of millions of users daily (like AI Overview), achieving efficiency improvements of 10 to 100 times over the past two years for models with similar performance through techniques such as distillation. On the other hand, to push the frontier towards AGI, research teams still need to continuously train new models with larger datasets and computation power, which will drive up total demand. He optimistically believes that in the long run, AI’s contributions in optimizing power grids, designing new materials and renewable energy, and addressing climate change will far exceed the energy consumed by its own operations. AI is a means to solve problems, not just a source of problems.

Democratization of Creation and Future Entertainment Forms

Beyond hardcore technology, Hassabis also envisions the impact of AI on culture and entertainment. Top video generation models like Imagen and Veo, along with image generation tools like “Nano Banana” that possess excellent instruction-following consistency, are enabling “democratization of creation.” Ordinary people no longer need to grind through thick Photoshop tutorials as before; they can create by simply describing ideas with language.

At the same time, AI is also “super-empowering” top professionals. Hassabis mentioned that they are collaborating with renowned directors like Darren Aronofsky and other top creators. These tools enable professional creatives to try various ideas at unprecedented speeds, increasing productivity tenfold or even a hundredfold, ultimately achieving their most perfect visions faster. He envisions a new form of entertainment or art type that integrates elements of “co-creation”: top creative visionaries build captivating dynamic story worlds, while thousands of users can co-create parts of the content within, with the creators serving as “editors-in-chief” of this world.

The Next Decade: A “New Scientific Renaissance”

When asked about his outlook for the next decade, Hassabis acknowledged the challenge of predicting even ten weeks in the AI field, let alone ten years. However, he firmly believes that if we achieve true AGI within the next decade, it will usher in a "new Renaissance in science" or "golden age". Every aspect of society, from energy to human health, will benefit from breakthroughs brought by this ultimate scientific tool. This is not just a technological upgrade but a profound revolution reshaping the boundaries of human civilization's cognition and capabilities. Demis Hassabis and the Google DeepMind he leads stand at the forefront of this revolution.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。