2017年-数据局_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 machine learning researchers to predict the timeline for AI to exceed human performance in various tasks and occupations. The survey aimed to provide insights for policymakers to anticipate and manage AI-related trends.

Main Predictions

• High-Level Machine Intelligence (HLMI):
  Researchers predict a 50% chance of HLMI occurring within 45 years and a 10% chance within 9 years.

  • Asian respondents expect HLMI to arrive 30 years earlier than North Americans (74 years).
  • For full automation of labor, the 50% chance is 122 years from now, and the 10% chance is 20 years.

• AI Milestones:
  Respondents expect 20 out of 32 AI milestones to be achieved within the next ten years.

  • Examples of milestones include: translating languages (by 2024), writing high-school essays (by 2026), driving a truck (by 2027), working in retail (by 2031), and writing a bestselling book (by 2049).

Key Findings

1. Regional Differences:

   • Asian respondents consistently predict earlier arrival of HLMI and automation compared to North Americans.
   • For example, the median estimate for full automation is 104.2 years in Asia and 168.6 years in North America.

2. Career Progression:

   • 67% of researchers believe that AI progress has accelerated in the second half of their careers.
   • Only 10% think it accelerated in the first half.

3. Intelligence Explosion:

   • The median probability of an intelligence explosion (rapid AI advancement after HLMI) is 10% within 2 years of HLMI.
   • The median probability of dramatic global technological progress after HLMI is 20% within 2 years.

4. Long-Term Impact of HLMI:

   • Respondents believe there is a 25% chance of a "good" long-term impact and 20% chance of an "extremely good" impact.
   • Conversely, a 10% chance of a "bad" impact and 5% chance of an "extremely bad" impact (e.g., human extinction).

5. AI Safety Research Prioritization:

   • 48% of respondents believe AI safety research should be prioritized more than the status quo, while only 12% think it should be prioritized less.

Survey Methodology

• Population: All researchers who published at the 2015 NIPS and ICML conferences (total of 1634 authors).
• Response Rate: 21% (352 respondents).
• Question Framing: Two methods were used to elicit probability forecasts—fixed-probability and fixed-years.
• Data Analysis:
  • Aggregate forecasts were calculated by fitting a Gamma CDF to individual responses.
  • Confidence intervals were generated using bootstrapping.
  • Regression analysis was used to assess the influence of demographic factors on HLMI predictions.

Demographics

• Respondents: 82% work in academia, 21% in industry.
• Regions: Respondents came from 43 countries, with the majority from Europe and North America.
• Gender: 91.9% of respondents identified as male, 5.4% as female, and 2.7% as not applicable.
• Citations and Seniority: Respondents had fewer citations and less seniority than non-respondents, suggesting a potential non-response bias.

Statistical Analysis

• Regression Results (Table S2):

  • Region significantly affects HLMI predictions (e.g., North American respondents expect HLMI in 74 years, while Asian respondents expect it in 30 years).
  • Question framing (fixed-probability vs. fixed-years) also influences predictions, with fixed-probability leading to earlier estimates.

• Demographic Comparison (Table S3):

  • The survey found no significant differences in the distribution of undergraduate regions, gender, or other variables between respondents and non-respondents.
  • However, respondents had lower citation counts and less seniority than non-respondents, indicating possible selection bias.

Implications

• The study highlights the uncertainty in AI timelines and the divergence in expectations across regions.
• It emphasizes the importance of AI safety research and the potential for transformative AI to reshape society.
• The results suggest that AI experts are optimistic about future progress but cautious about its long-term consequences.

Conclusion

The survey provides valuable insights into the timing and societal impact of AI advancements, informing both researchers and policymakers. The regional differences in predictions and the emphasis on AI safety underscore the need for diverse perspectives and proactive planning in AI development.