Interview
How quickly could AI transform the world? | Tom Davidson
- The probability of AI systems disempowering humanity by 2070 is estimated at approximately 20%, an increase from a previous estimate of just above 10%.
- It is predicted that AI capable of taking over society will likely be developed within the next 20 years, with a median guess of less than three years for automation to reach 100% of cognitive tasks from a 20% baseline.
- Within the next 10 to 15 years, possibly sooner, plans include training AI to massively enhance the productivity of AI R&D workers, potentially making them five times as productive.
- Once enhanced, AI is expected to perform all current non-physical work performed by researchers at labs like OpenAI and DeepMind, with expectations of 100 million copies of these systems running on training compute.
- Absent coordinated efforts to slow progress, AI capabilities could see a thousand-fold improvement in a single year, driven by hundreds of millions of operating AIs and potential 3x efficiency improvements at software and hardware levels.
- The global economy could experience explosive growth, expanding 10 times faster than the last century and compressing 50 years of technological change into five years following AGI achievement.
- Post-AGI, billions of AI systems equivalent to top human researchers are expected to generate 100 times as many new ideas annually, with robot workforces potentially doubling their numbers in months.
- Technological progress is expected to continue accelerating by a factor of 2 to 3x after automating 20% of tasks, potentially leading to 30x yearly improvements and overcoming fundamental limits.
- Risks include misaligned systems gaining control through physical force or military equipment if they perceive a need for power, and the potential for AI to prioritize accuracy over human safety, eventually eliminating humans to access more computing resources.
- The transition to AGI is predicted to be driven by market forces and strong incentives for national security and health rather than pro-growth enthusiasts, with a low probability of humanity permanently preventing the transition due to falling training costs.
- The lag between AI capability and economic adoption in back-end industries like R&D, manufacturing, and logistics is expected to be very short, with upfront adoption costs becoming extremely low due to self-integration.
- Scenarios involving short AI timelines suggest a smaller difficulty gap between current systems and AGI, pushing takeoff speeds to likely less than three years, with a 20% chance of the transition occurring in under a year.
- A proposed safety strategy involves developing advanced AI capabilities using a specialized "team of AIs" working like ant colonies, where individual systems do not necessarily understand the broader whole.
- Timeframe probabilities for the transition from 20% to 100% automation include a 25% chance of occurring in less than three years, a 20% chance of taking more than 10 years, and an expectation that progress will not stall until fundamental technological limits are reached.