Interview
Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold
- Systems with capabilities comparable to human-level AI or AGI are considered plausible within the next decade, with multimodal systems expected to understand real-world physics sooner and gain immediate environmental contextual understanding via sensors within 18 months.
- The trajectory toward AGI is predicted to rely on combining large multimodal models with AlphaZero-like planning and search mechanisms, utilizing scalable algorithms like Transformers to leverage existing world knowledge rather than scaling alone or building from scratch.
- DeepMind plans to publish responsible scaling laws and safety frameworks publicly within the next year and intends to implement air gaps and cybersecurity measures for AI weights over the next three to five years.
- Future multimodal models are anticipated to enable robotics progress by transferring knowledge from language and simulation to physical action spaces, while upcoming Gemini iterations will integrate capabilities for gaming, video processing, and robotics control.
- AI systems are expected to serve as reliable research assistants and aid in curing diseases and drug discovery through initiatives like Isomorphic Labs once factuality and grounding issues are resolved, though they are not forecast to formulate high-level scientific hypotheses in the foreseeable future.
- A "very exciting" global shift in AI interaction is predicted as society adapts to true multimodality, with a continuation of current investment "rushes" creating a chaotic environment that necessitates a shift toward "scientific" and "cautious" development.
- Safety protocols will require more stringent evaluation to detect deception or code exfiltration before superhuman intelligence levels are reached, with a strict policy to fix or add guardrails to systems found possessing dangerous capabilities before deployment.
- Once superhuman capabilities are achieved, post-hoc reasoning explanation is expected to become a critical safety capability, while the global landscape will require an international consensus to ensure benefits are shared equitably.
- The integration of Google Brain and DeepMind is expected to yield efficient resource pooling for frontier system development, with a strategic commitment to bold yet responsible advancement that moves away from a "move fast and break things" mentality.
- Current uncertainty regarding large model effectiveness asymptotes is framed as an empirical question, with optimism that reinforcement learning and self-play synthetic data will successfully overcome existing data bottlenecks.