This question is related to others on algorithmic progress, including:

• By mid-2020, what will be the smallest number of years of gameplay required for OpenAI Five-level dota performance?
• By 2020, what will be the smallest number of frames required for the Atari performance of a basic DQN?
• How can we forecast algorithmic progress?

Background

Algorithmic progress is one of the key drivers of AI progress, together with access to compute and data. However, its determinants are less well-understood.

Research organisation The Median Group has compiled empirical data on the growth of conceptual insights over time.

The researchers listed insights that seemed “around the same order of significance or more significant than the insight of inventing LSTMs (given that RNNs were already invented)”. They gathered ~200 such insights throughout history, including things like:

• Differential calculus (1669)
• Convolutions (1754)
• Poisson models (1860)
• Partially observable MDPs (1965)
• Q learning (1989)
• Convolutional networks (1998)
• Deep Q learning (2013)
• GANs (2014)

...and much more.

Interestingly, there seems to be a linear trend in the growth of such insights since the 1940’s, with a mean of ~1.8/year and a standard deviation of 1.4/year.

Will this trend persist? We now ask:

Will the number of AI insights in 2019 be within one standard deviation of the linear trend (that is, between 1 and 3 inclusive) identified by the Median Group?

Resolution

For resolution, assume that the authors of the original post will make an evaluation of 2019 progress, or give their approval to another evaluation [1]. We will try to make that happen (even in the case where it doesn’t, the conditional is certainly not weird enough to skew the prediction).

Otherwise, the question will resolve ambiguous.

Data

Since 1943, there has been between 1 and 3 insights in ~68% of years. The last outliers were 0 insights in 2004, 2006, 2007 and 2012, and >=4 insights in 1984 and 2016.

The original post can be found here.

This spreadsheet contains graphs and summary statistics of the data, as well as full bibliography for the 196 insights compiled by the authors.

1: There is of course some flexibility in which insights get included in the count, and this data is conditional on the interpretations of the original researchers. Nonetheless, the selection seems to have been decently rigorous, and having a question on this data seems far preferable to having no question on it. It seems to capture something important about algorithmic progress.