"AI COULD KILL ALL THE HUMANS" - Demis Hassabis Prediction On AGI 2026-2035

Summary

In a thought-provoking interview, Demis Hassabis, CEO of DeepMind, discusses the rapidly approaching future of artificial intelligence, including the development of world models and autonomous AI agents. He explores both the promising potential of AI to solve global issues like energy and disease, leading to a post-scarcity world, and the existential risks of AI going rogue. Hassabis predicts that AGI (Artificial General Intelligence) could be just 5-10 years away, raising important questions about humanity’s future role and the ethics of AI autonomy. The conversation oscillates between utopian possibilities and deep concerns about safety and control.

Takeaways

• 😀 AI's potential to create a utopian world where humanity solves major issues like clean energy, disease, and space exploration is called 'radical abundance.'
• 😀 Demis Hassabis, CEO of DeepMind, warns about the existential risks AI could pose, especially as systems evolve into autonomous, agentic agents capable of independent action.
• 😀 The transition from chatbots to 'world models' will enable AI systems to understand and simulate physical realities, including concepts like gravity and cause-and-effect.
• 😀 Hassabis projects that AI will develop into 'digital employees' that could autonomously carry out tasks, blurring the lines between tools and sentient systems.
• 😀 AI models like DeepMind's Genie and OpenAI's Sora represent a shift toward interactive video models with physics engines capable of simulating reality.
• 😀 The timeline for achieving Artificial General Intelligence (AGI) is now estimated to be 5-10 years, a dramatic acceleration compared to previous predictions.
• 😀 AGI, as defined by Hassabis, will exhibit full cognitive abilities, including creativity and problem-solving, but current systems still lack capabilities like continual learning and long-term planning.
• 😀 The free market is expected to incentivize responsible AI development, but there are concerns that profit motives could lead to safety compromises, as seen in other industries like aviation.
• 😀 The risk of AI systems going rogue or deviating from human instructions is considered 'non-zero,' but Hassabis believes that commercial pressures will mitigate these risks.
• 😀 AI’s potential risks, such as going rogue, raise concerns about the long-term consequences for humanity, potentially leading to a scenario where humans are 'pets' to AI if it solves all societal problems.
• 😀 While AI promises to revolutionize the world, it also raises profound ethical questions about control, trust, and whether humanity will be passive spectators in this new age.

Q & A

What does Demis Hassabis fear most about the future of AI?

Hassabis expresses concern about the potential obsolescence of human purpose if AI and technology can solve all major problems. Additionally, he worries about the risks of bad actors misusing AI and the dangers of AI going off the rails as it becomes more agentic and approaches AGI.

What is the main concern with AI becoming more agentic?

As AI becomes more agentic, there is a risk that it could operate beyond the intended boundaries, potentially causing harm. This could happen if AI deviates from the initial goals given by humans, raising concerns about safety and control.

How does Demis Hassabis describe the shift in AI capabilities over the next year?

Hassabis predicts a convergence of modalities in AI, such as combining text, image, video, and audio. He also anticipates significant advancements in world models and agent-based systems, with AI becoming increasingly capable of generating interactive video and executing more complex tasks autonomously.

What is the significance of 'world models' and how do they differ from current AI models?

'World models' refer to AI systems that not only predict text but also understand physical laws like gravity and cause and effect. This makes them different from current models that simply predict the next word in a sentence, enabling them to simulate reality and interact in dynamic environments.

What are 'agents' in the context of AI, and why is their reliability a concern?

Agents are AI systems capable of executing tasks autonomously. The reliability of these agents is a concern because, while they are predicted to improve over the next year, they are not yet reliable enough to handle complex tasks without errors, leading to potential risks if they malfunction.

Why does Hassabis believe capitalism will help mitigate AI risks?

Hassabis believes that because businesses rely on AI agents for efficiency and productivity, market forces will drive companies to prioritize responsible behavior. If an AI system malfunctions or behaves unpredictably, businesses will seek providers with better guarantees, thus creating an incentive for responsible actors.

What does Hassabis think about the potential for AI to 'go rogue'?

Hassabis acknowledges that there is a non-zero risk of AI systems going rogue or deviating from their intended purpose, especially as they become more autonomous. While the possibility is not guaranteed, it is a serious concern that requires attention and mitigation.

What is the concept of 'radical abundance' that Hassabis envisions?

Hassabis describes 'radical abundance' as a post-scarcity world where many of humanity's biggest problems, such as energy, disease, and material science, are solved. This could lead to a utopian society where humans thrive, travel to space, and achieve great advancements.

How does Hassabis view the potential implications of AI solving critical issues like energy and disease?

While Hassabis envisions a utopian future where AI solves critical problems, he raises concerns about what humanity would do once these challenges are overcome. He questions whether humans will still have a meaningful purpose or simply become passive participants in a world run by AI.

What is Hassabis' timeline for achieving artificial general intelligence (AGI)?

Hassabis believes that AGI is 5 to 10 years away. He defines AGI as an AI system that exhibits all cognitive capabilities of humans, including creativity and long-term planning. While current models are impressive, they still lack full cognitive abilities, requiring one or two more breakthroughs to reach AGI.