Since the inception of the field, the goal of Artificial Intelligence (AI) research has been to develop a machine-based system that can perform the same general-purpose reasoning and problem-solving tasks humans can. While computers have surpassed humans in many information-processing abilities, this "general" intelligence has remained elusive.
AI, and particularly machine learning (ML), is advancing rapidly, with previously human-specific tasks such as image and speech recognition, translation and even driving, now being successfully tackled by narrow AI systems.
But there is a stunning diversity of opinion about when general AI may arrive, according to published expert surveys. For example this study finds 50% of AI researchers accord a 50% probability to "High level machine intelligence" (HLMI) by 2040, while 20% say that 50% probability will not be reached until 2100 or later. Similarly, this survey finds an aggregated probability distribution with a 25%-75% confidence interval (comparable to Metaculus sliders below) ranging from 2040 to well past 2100.
It would be nice to tighten these probability intervals considerably, so we ask of the Metaculus community:
When will the first AGI be first developed and demonstrated?
One issue is that AGI is rather difficult to precisely define. A separate question addresses a similar issue by asking about human-machine intelligence parity in a particular adversarial test. Here we'd like a definition that connects more closely with established benchmarks for various capabilities; it also sets an arguably somewhat lower bar.
For these purposes we will thus define "an artificial general intelligence" as a single unified software system that can satisfy the following criteria, all easily completable by a typical college-educated human.
Able to reliably pass a Turing test of the type that would win the Loebner Silver Prize.
Be able to score 75th percentile (as compared to the corresponding year's human students; this was a score of 600 in 2016) on all the full mathematics section of a circa-2015-2020 standard SAT exam, using just images of the exam pages and having less than ten SAT exams as part of the training data. (Training on other corpuses of math problems is fair game as long as they are arguably distinct from SAT exams.)
Be able to learn the classic Atari game "Montezuma's revenge" (based on just visual inputs and standard controls) and explore all 24 rooms based on the equivalent of less than 100 hours of real-time play (see closely-related question.)
By "unified" we mean that the system is integrated enough that it can, for example, explain its reasoning on an SAT problem or Winograd schema question, or verbally report its progress and identify objects during videogame play. (This is not really meant to be an additional capability of "introspection" so much as a provision that the system not simply be cobbled together as a set of sub-systems specialized to tasks like the above, but rather a single system applicable to many problems.)
Resolution will be by direct demonstration of such a system achieving the above criteria, or by confident credible statement by its developers that an existing system is able to satisfy these criteria. In case of contention as to whether a given system satisfies the resolution criteria, a ruling will be made by a majority vote of the question author and two AI experts chosen in good faith by him. Resolution date will be the first date at which the system (subsequently judged to satisfy the criteria) and its capabilities are publicly described in a talk, press release, paper, or other report available to the general public.