Interview
Shane Legg (DeepMind Founder) — 2028 AGI, superhuman alignment, new architectures
- Human-level AI is projected to have a 50% probability of being achieved by 2028, driven by exponential growth in computational power and data alongside the discovery of scalable algorithms; a delay beyond 2029 is attributed to the potential for unexpected research hurdles.
- Current models are expected to mature between now and 2028 by becoming less delusional, more factual, more up-to-date, and significantly more multimodal, addressing shortcomings in video understanding and episodic memory without encountering fundamental blockers.
- The transition to AGI will likely require architectural changes rather than simple scale increases to enable systems to separate rapidly learned specific information from slowly learned generalities, with fully multimodal systems representing the next major historical landmark.
- Domain-specific models are not considered a direct path to AGI, though they may yield incidental insights; conversely, truly creative AI capable of stepping beyond training data is predicted to require the integration of powerful search mechanisms.
- Safety and alignment strategies will shift from "system one" responses to "system two" approaches involving deliberative dialogue, step-by-step reasoning, and world modeling to communicate specific societal ethical principles and reinforce them through constant verification.
- Primary alignment challenges are identified as the difficulty of communicating specific human values rather than a lack of world models, while future systems must be engineered to follow agreed-upon ethical frameworks distinct from general ethical understanding.
- DeepMind will continue pursuing non-AGI research areas such as fusion, sustainability, deforestation monitoring, and weather forecasting alongside AGI development, maintaining a dual focus on safety and practical applications.
- The economic impact of achieving human-level intelligence is expected to be transformative, as machines will be capable of performing cognitive tasks currently executed within the economy, with future applications predicted to be predominantly positive despite acknowledged misuse risks.
- Current benchmarks are noted as insufficient for measuring capabilities like streaming video understanding and episodic memory, and the historical impact of specific companies on the safety field is considered difficult to quantify due to hiring challenges and the inability to establish clear counterfactuals.