Interview
Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold
- Demis Hassabis predicts the emergence of AGI-like systems within the next decade, noting that the trajectory aligns with the original 20-year roadmap set for DeepMind in 2010.
- He anticipates that systems will likely not be controlled by a single private company long-term, advocating for broad international collaboration involving civil society, academia, and governments.
- The path to AGI is expected to combine large multimodal models (serving as world models) with planning mechanisms similar to AlphaZero, rather than relying on pure reinforcement learning or LLMs alone.
- Multimodal systems ingesting audiovisual data will enable AI to better understand the physics of the real world, improving "grounding" beyond text-based abstraction.
- Scaling laws have proven surprisingly effective, though Hassabis notes it is an empirical question whether this will hit an asymptote or a "brick wall" in the near future.
- DeepMind intends to publish "responsible scaling laws" and related safety policies publicly over the coming months to establish clear development guardrails.
- To address the data bottleneck, the industry will increasingly rely on synthetic data generated through realistic simulations and self-play, moving beyond reliance solely on human-generated text.
- Hassabis observes that while current models show asymmetric improvements in specific domains (like coding or math), they also demonstrate surprising general transfer effects in reasoning, mirroring human learning.
- The field is moving away from the "move fast and break things" mentality toward a "bold and responsible" approach, emphasizing that AGI is too consequential for uncontrolled experimentation.
- Security for frontier model weights is being addressed through Google's enterprise-grade firewall, hardened sandboxes, and strict access controls to prevent theft by rogue actors or nation-states.
- Open-sourcing foundational, general-purpose AI models is viewed as potentially dangerous by Hassabis due to the risk of bad actors repurposing them, necessitating a balance between open science and security.
- Gemini's development required managing distributed computing challenges and iterative hyperparameter tuning, as scaling laws do not hold perfectly when extrapolating across orders of magnitude in compute.
- Current models are not yet capable of asking the right scientific questions or formulating hypotheses, but they can accelerate scientific discovery by searching combinatorial spaces (e.g., protein folding).
- Hassabis expects the interface with AI to evolve from text chat to true fluid multimodality, incorporating video, voice, environmental context via cameras, and eventually robotics and touch.
- The integration of Google Brain and DeepMind has pooled compute resources and engineering talent, resulting in Gemini as the first major collaborative output of this unified structure.
- Safety evaluations must evolve to detect root node traits like deception and the ability to exfiltrate code, potentially using specialized narrow AI to audit broader systems.
- Hassabis remains optimistic about AI's potential to solve major global challenges like disease, climate change, and material science, independent of waiting for full AGI.
- The current surge in public interest and VC funding has created a chaotic environment, requiring the field to maintain scientific rigor and foresight to manage risks associated with rapidly advancing capabilities.
- Neural network "grounding" is hypothesized to arise from two sources: the inherent structure of language and the grounded feedback from human raters in RLHF systems.
- Future AI systems will likely require episodic memory and personalization to become reliable, daily-use assistants for general tasks, moving beyond current static context windows.
- Robotics progress is being accelerated by transferring skills learned in general multimodal models to the physical world, addressing the historical data scarcity in robot training.