牛津大学：AI超越人类编年史_21页_1mb

Summary of "When Will AI Exceed Human Performance? Evidence from AI Experts"

Core Content

This study presents a large-scale survey of AI experts to forecast when artificial intelligence (AI) will surpass human performance in various tasks and occupations. The survey, conducted among researchers who published at the 2015 NIPS and ICML conferences, collected data on AI progress timelines, social impacts, and safety considerations.

Main Predictions

• High-Level Machine Intelligence (HLMI):

  • The aggregate forecast gives a 50% chance of HLMI occurring within 45 years and a 10% chance within 9 years.
  • Asian respondents expected HLMI to arrive 30 years earlier than North American respondents (30 vs. 74 years).
  • Full automation of labor is expected to occur in 122 years (50% chance) and 20 years (10% chance).

• AI Milestones:

  • 32 milestones were considered, with the expectation that 20 of them will be achieved within 10 years (mean 50% probability).
  • Examples of milestones include:
    • Translating languages (by 2024)
    • Writing high-school essays (by 2026)
    • Driving a truck (by 2027)
    • Working in retail (by 2031)
    • Writing a bestselling book (by 2049)
    • Working as a surgeon (by 2053)

Key Findings

1. Regional Differences in Predictions

• Asian respondents generally expected AI milestones and HLMI to arrive much earlier than North American respondents.
• The median gap between Asian and North American predictions for full automation was +64.4 years.
• For specific milestones like truck driver and surgeon, the gap was +9.6 and +10.2 years, respectively.

2. Acceleration of AI Progress

• 67% of respondents believed that AI progress has accelerated in the second half of their careers.
• The median time for AI to perform vastly better than humans in all tasks was 10% (i.e., two years after HLMI is achieved).
• The median probability of an intelligence explosion (rapid acceleration in AI development) was 20%.

3. Long-Term Impacts of HLMI

• Respondents assigned 25% to a "good" long-term impact and 20% to an "extremely good" impact.
• The probability of a bad outcome was 10%, and of an extremely bad outcome (e.g., human extinction) was 5%.
• 48% of respondents believed that AI safety research should be prioritized more than the status quo.

4. Survey Methodology and Demographics

• The survey included 352 researchers (21% of 1634 contacted).
• 82% of respondents worked in academia, and 21% in industry.
• The sample included researchers from 43 countries, with the largest numbers from Asia and North America.
• Non-response bias was minimal, as differences in demographics (e.g., citations, seniority) between respondents and non-respondents were small.

5. Elicitation of Beliefs

• Two question framings were used: fixed-probability and fixed-years, which influenced the responses.
• The combined framing (averaging both) was used in the final analysis.
• The mean of individual CDFs was used to generate the aggregate forecast for HLMI and milestones.

Conclusion

The study highlights the divergence in expectations among AI experts, particularly between Asian and North American researchers, and underscores the need for more accurate forecasting of AI progress to inform public policy and safety research. It also emphasizes the uncertainty in AI timelines and the importance of prioritizing AI safety to mitigate potential risks. The results suggest that while AI may achieve significant milestones within the next decade, the full realization of HLMI and its global impact remain uncertain and far off.