【】

It's not just your imagination. AI researchers themselves are having their minds blown by the sheer pace of AI development, too, a new survey has found.

A wide-ranging survey of AI experts released this week backs up the perception that AI development really is accelerating at a dizzying pace — at least from the point of view of experts in the field. It also helps quantify the infamous divide in tech world sentiment between die-hard AI fans, and AI "doomers" who supposedly preach caution because they fear some sort of AI apocalypse scenario. 

In spite of the divide, there seem to be slightly more die-hards, and — if you read between the lines — they seem to be perceived as winning.

The paper on the survey is a pre-publication release from AI Impacts, a San Francisco-based research firm that receives funding from billionaire and Facebook co-founder Dustin Moskovitz's grant-making entity Open Philanthropy. 

By averaging survey responses from 2,778 AI researchers who met the authors' own standard for notability, and comparing them to a previous similar survey, the authors found that broadly speaking, AI experts perceive a sense of acceleration across the board. The authors note that on average, when it came to questions about 32 different AI-related tasks, "the 50th percentile year they were expected to become feasible shifted 1.0 years earlier," between 2022 and 2023.

In less technical speak, the average AI prediction shifted a year earlier in time at some point between the 2022 survey and this one from 2023. This is a much more powerful finding than if the average expert had said "yes" to a question like, "Do you think things are accelerating in the AI world?" because it shows that the experts actually revised numerous time estimates about that acceleration on a year-over-year basis.

Perhaps the marquee findings in the study are the downright drastic shifts in respondents' aggregate forecasts for two key concepts: High-Level Machine Intelligence (HLMI) and Full Automation of Labor (FAOL) when compared to similar forecasts made in 2022. HLMI, in particular, showed an estimated arrival time that had dropped by 13 years between 2022 and 2023. Meanwhile, the forecast for FAOL decreased by 48 yearsover that same period. 

This document is a downright remarkable shift in perception. Over the course of a single, mind-bending year, AI experts came to believe that the point at which "for any occupation, machines could be built to carry out the task better and more cheaply than human workers" would arrive nearly a half-century sooner than they had the previous year.

Given how fast these experts think these material consequences will arrive, it's telling to read their stated beliefs about whether AI should develop faster, the opinion held by the so-called "effective accelerationists," or slower, the opinion held by the AI doomers. The apparent contingent of hardcore doomers, or at least those who want AI to develop "much slower," was the tiniest group of respondents, at 4.8 percent. Meanwhile, the apparent accelerationists — those whose response was "much faster" — absolutely obliterated the doomers with 15.6 percent. 

But the "somewhat slower" group of respondents to this question actually won the plurality, with 29.9 percent of responses, followed by "current speed" at 26.9, and "somewhat faster" at 22.8. This muddy middle, made up of the three more status-quo-leaning answers, accounted for 79.6 percent of all responses.

However, it's worth dwelling on an important distinction noted in the survey: the respondents only have expertise in AI, as opposed to expertise in forecasting, either generically or about AI itself. They might therefore lack, "skills and experience, or expertise in non-technical factors that influence the trajectory of AI," the authors write. Actually, this scholarly word of caution is worth keeping in mind just about any time you read about AI experts opining about the future in any context.

But these findings aren't irrelevant just because AI researchers lack psychic powers. These are some of the people who drive this technology forward, and a window into their subjective beliefs about their own area of expertise gives us a hint about what, on average, these people want, fear, and see on the horizon: they think an AI-driven automated world is coming more quickly than ever, and taken as a group, they're mostly on the fence about whether the pace of AI change is good. 

But rather unsettlingly, those who want to put rockets on this already accelerating freight train significantly outnumber those who want to slam on the breaks.

最新评论