www.lesswrong.com/posts/BEtzRE2M5m9YEAQpX/there-s-no-fire-alarm-for-artificial-g...
1 correction found
Grace et al. (2017) surveyed the predictions of 352 presenters at ICML and NIPS 2015. Respondents’ aggregate forecast was that the proposition “all occupations are fully automatable” (in the sense that “for any occupation, machines could be built to carry out the task better and more cheaply than human workers”) will not reach 50% probability until 121 years hence. Except that a randomized subset of respondents were instead asked the slightly different question of “when unaided machines can accomplish every task better and more cheaply than human workers”, and in this case held that this was 50% likely to occur within 44 years.
This reverses the Grace et al. survey framing. The paper says most respondents were asked the “unaided machines can accomplish every task” HLMI question (50% in 45 years), while only a subset got the “all occupations are fully automatable” wording, which yielded a much later 122-year forecast.
Full reasoning
The post describes the Grace et al. survey backwards.
In the paper, the main HLMI question was the one defined as “when unaided machines can accomplish every task better and more cheaply than human workers.” The authors report that this question got an aggregate 50% estimate in 45 years.
Only a subset of respondents were instead asked the employment-framed question about “when all occupations are fully automatable”, and that version produced a much later 50% estimate in 122 years.
So the post incorrectly says that the automation-of-all-occupations question was the main result and that a randomized subset was instead asked the unaided machines question. The paper says the opposite. It also slightly misstates the paper’s headline numbers: the arXiv/published version reports 45 years and 122 years, not 44 and 121.
The post also says the 352 respondents were “presenters,” but the paper says they were researchers who published at the 2015 NIPS and ICML conferences, with 352 respondents to the survey invitation.
2 sources
- When Will AI Exceed Human Performance? Evidence from AI Experts
The paper says: “High-level machine intelligence (HLMI) is achieved when unaided machines can accomplish every task better and more cheaply than human workers.” It then reports a 50% chance “within 45 years.” It adds: “While most participants were asked about HLMI, a subset were asked ... when all occupations are fully automatable,” which was “122 years from now.”
- [1705.08807] When Will AI Exceed Human Performance? Evidence from AI Experts
The abstract says researchers believe there is a 50% chance of AI outperforming humans in all tasks in 45 years and of automating all human jobs in 120 years.