The Intersection of Creativity, Science, and Artificial General Intelligence

Demis Hassabis, CEO of Google DeepMind, sat down to discuss the rapid advancements in AI, the path to Artificial General Intelligence (AGI), and the profound impact these technologies are having on creativity and scientific discovery. The conversation touched upon the current anxieties surrounding AI, the evolution of creative tools, and the fundamental connection between simulation, imagination, and intelligence.

The Path to AGI and Navigating AI Risks

The current frenzy around AI, particularly text-based models capable of generating code and identifying vulnerabilities, has sparked widespread concern, even leading to calls for bans in some sectors. Hassabis acknowledges these anxieties, stating that as we approach AGI, a more systematic approach is crucial. He views the current challenges, such as those seen with cyber threats, as "warning shots" for humanity, highlighting potential future risks in areas like bio and nuclear technology.

"We've got to get ready for that, and I think we need a more systematic way to deal with the issues," Hassabis emphasized, suggesting the need for an international standards body to test frontier AI systems for robustness and sufficient guardrails.

Regarding the technical pathways to AGI, Hassabis champions a multi-pronged research strategy. Google DeepMind's history, he notes, is built on a broad and deep research bench, responsible for many of AI's significant breakthroughs, from Transformers to AlphaGo. Their current approach continues to bet on multiple avenues, including scaling their multimodal foundation model, Gemini, pushing hard on coding, and developing generative media models like Omni and VEO.

"We think that it's important to give the these models understanding of the world around us, the context around us," Hassabis explained. "And I think in the end, to have a full AGI system, you need to be able to also understand the physical world around you." This understanding, he believes, is essential for applications like robotics and smart glasses.

Talent, Competition, and the Evolution of AI Research

The AI landscape has become ferociously competitive, with major players vying for top talent. Hassabis acknowledges the movement of talent between leading labs but asserts that Google DeepMind possesses the largest and broadest research bench. He reflects on the early days of DeepMind, founded in 2010 when AI was largely dismissed as a dead end, even in academia.

"We just felt the small band of us felt that that actually with the right ideas and using learning systems, reinforcement learning and betting on neural networks that a lot of fast progress could be made," he recalled. This foresight has now led to a global awakening to AI's potential, drawing in every major company.

AI's Impact on Creativity and Creative Industries

In the context of an advertising conference, Hassabis highlighted the transformative power of AI tools for creators. He noted that the pace of improvement is staggering, with monthly advancements in generative models. A significant leap over the past year has been the ability to "live edit" the output of generative models.

"You know, if you it was part of the creative process obviously is you generate the first idea, the first concept, but you like some of it but not other parts of it," Hassabis described. "You don't want to have to regenerate the whole thing... You want to be able to describe in natural language ideally as you would to a designer like, 'Okay, keep that part the same but change this to something else.'" This fine-grained control, coupled with relentless quality improvements, has been a game-changer.

The conversation also addressed the ongoing debate about disclosure and the use of AI in creative work. Hassabis stressed the importance of dealing with misinformation and deepfakes, a concern they foresaw years ago. This led to the development of SynthID, a robust, imperceptible digital watermarking system embedded in all their generative models to detect AI-generated content. He hopes this will become a regulatory standard, also aiding in intellectual property rights.

However, regarding the disclosure of AI use in creative processes, Hassabis is less certain. He likened AI tools to other advanced tools like Photoshop, suggesting that their use might simply be an evolution of the creative toolkit rather than something that always requires explicit disclosure, provided the output's synthetic origin is clear.

Hassabis sees a two-fold change in creativity driven by AI:

1. Democratization: AI tools lower the barrier to entry, allowing more people to experiment with their ideas quickly and easily. This can lead to new creators finding paths into industries and less gatekeeping.
2. Enhancement for Professionals: For established creators, AI can act as a powerful amplifier, enabling them to explore more ideas, iterate faster, and achieve more with less expense and time.

He cautioned that like any tool, AI can be used lazily, diminishing creativity, or innovatively, enhancing it. The creative industries are still figuring out the best ways to leverage these new capabilities. He expressed optimism that AI could spark entirely new genres of games, much like graphics and AI did in the 1990s.

The Question of Compensation and New Economic Models

The training of AI models on vast datasets of human-created content raises questions about compensation and auditability. Hassabis acknowledged that a new economic model is likely needed, drawing parallels to the evolution of streaming services and content ID systems in the music and video industries.

"It's going to be difficult to kind of agree on objectively like what that is," he admitted, referring to the challenge of attributing specific percentages of influence from training data. He also pointed out that human creators themselves are products of their experiences and exposure to other art forms and creators, making the concept of pure originality complex.

The Inseparability of Science and Creativity

Hassabis's career has consistently bridged creativity and science. His early work in video games, his neuroscience research on the hippocampus and creativity, and now his leadership in AI development all underscore this connection. He views AlphaFold, a groundbreaking AI system for protein folding, as a highly creative scientific approach.

He argues that the capabilities needed for AI to analyze scientific data, such as visual data from cells or proteins, are fundamentally the same as those required to analyze YouTube videos or process general visual input from a camera.

"So, a lot of these capabilities are general purpose and you develop them for one thing, but really that's just a means to an end for another thing," Hassabis explained. He recounted how early DeepMind work in games like Go and Atari was not an end in itself but a means to develop challenging tasks that served as a "ladder" to tackle real-world problems in science, such as protein folding and drug discovery.

"That's personally what I spend my time using these AI systems for is AI for science. That's my my always been my main passion and the main reason why I'm building these AI tools," he stated.

Recreating the Brain's Imaginative Power

Hassabis's 2007 neuroscience paper, which linked the hippocampus to creativity by studying patients who lost memory and the ability to picture the future, is a foundational piece of his thinking. He theorized that if memory is a reconstructive process, then imagination should use the same brain mechanisms, but with a different goal: creating novel outputs from component parts. This hypothesis was validated through his research.

He sees parallels between how video models generate worlds from prompts and the brain's imaginative processes. While not suggesting direct implementation copying, he believes understanding the brain's algorithmic principles and representations can inspire AI development. Neuroscience friends are actively comparing AI model outputs to human fMRI studies, even decoding imagined images and recreating them with AI.

The Einstein Test and the Need for Visual Understanding

Hassabis proposed the "Einstein Test" as a benchmark for true creativity: giving an AI all the data Einstein had up to a certain point and seeing if it could independently derive theories like relativity. While text models might seem applicable, Hassabis emphasizes the visual and imaginative aspects of Einstein's thought experiments.

"I think you're going to need to access and at least understand and certainly if you're going to propose new experiments or have to do new experiments in order to test your hypothesis to develop it further," Hassabis stated. "So, it's not just a a theory, then I think you're going to need to have an understanding of the world of atoms, not just the world of bits or the world of logic."

Simulations: The Engine of Imagination and Discovery

Reflecting on his ambitious early video game project, "Republic: The Revolution," which aimed to simulate a former Soviet republic, Hassabis sees a full-circle moment with today's AI capabilities. He believes that in a few years, systems might be able to generate such complex virtual worlds, realizing the vision he had decades ago.

The connection between simulations and AI is fundamental. Hassabis defines imagination as a type of simulation, and simulations are crucial for exploring possibilities and selecting optimal paths, as demonstrated by AlphaGo's strategy.

He highlights the potential of simulations in areas like economics, where controlled, large-scale reruns of economic scenarios could lead to more rigorous, scientific decision-making, unlike the current trial-and-error approach with interest rates.

"Most of the time the things we want to simulate... we don't actually understand well enough how the system really works on a mathematical level," Hassabis noted. "So, then an AI system could learn it that that simulation from the data." This ability for AI to learn simulations from data represents a significant advancement towards understanding and modeling complex systems.

Hassabis concluded by emphasizing the interconnectedness of creativity and science within AI, offering inspiration for the audience to apply these principles in their own work.