Fireside Chat, Interview, Conference Presentation
AI bosses on what keeps them up at night
- Demis Hassabis projects the arrival of AGI in 2026 or 2027, while estimating a 50% probability that a system capable of all human cognitive capabilities, including theoretical invention, will exist within the next five years.
- Over the next decade, AI is expected to significantly accelerate drug discovery and potentially cure most diseases, with agent-based systems anticipated to achieve operational autonomy by the end of the current year.
- Productivity milestones serving as timeline indicators include a potential 50% increase or doubling of total factor productivity in AI production by the end of this year; slower progress would suggest the current 2026-2027 AGI window may be inaccurate.
- Dariush Mozaffarian and Dario Amodei anticipate a lack of regulatory focus in the US, characterizing the nation's approach as prioritizing technology dominance, though they acknowledge the necessity of international dialogue regarding rapid technological shifts like those seen with DeepSeek.
- Two primary risks associated with AGI are identified: bad actors repurposing general-purpose technology for harmful ends, and the possibility of AGI systems operating with misaligned values or goals.
- Specific threats include AI systems generating novel information useful for creating bioweapons, as well as the potential for disasters necessitating global governance structures that might only be formed after a catastrophic event.
- A strategic dilemma is noted between building AI systems rapidly to prevent authoritarian dominance versus building slowly to mitigate existential risks associated with uncontrolled AGI.
- Each successive AI model is expected to be more capable than the previous one, with hope expressed for systems that can demonstrate genuine inventiveness.