This question was suggested by zakeeulhaq, who also did extensive background research.

A recent paper by DeepMind (Saxton et al., 2019) analyzes the mathematical reasoning ability of neural models; using the Transformer model (Vaswani et al., 2017). Their best mark was 35% in a high school mathematics exam (questions sampled from a UK GCSE Edexcel Higher Tier paper, 2012-2013). The set of questions covered algebra, arithmetic, calculus, polynomial factorization and elementary probability theory.

See here for an accessible video introduction to the paper. More in-depth background is given in a comment below, citing zakeeulhaq’s notes.

We now ask:

By Jan 1st 2021, will there EITHER…

1. … be a credible report of a neural model with a score of >=70% on the task suite used in the 2019 DeepMind paper…

2. OR be judged by a council of experts that it’s 95% likely such a model could be implemented, were a sufficiently competent lab to try… [1]

3. OR be a neural model with performance on another benchmark judge by a council of experts to be equally impressive, with 95% confidence?

Resolution

For purposes of resolution, these terms are interpreted in accordance with the Technical AI Forecasting Resolution Dictionary v0.2.0. Any term whose interpretation deliberately departs from this standard has been marked with a ^."

In particular, a credible report is defined as any item in:

• Blog-post co-authored by at least one of the researchers

• Pre-print posted to arXiv or similar

• Peer-reviewed (published) paper

• Second-hand report from a trusted source providing at least as much information as the median acceptable blog post

[1] The point of this counterfactual resolution condition is as follows. Not all trajectories to advanced AI pass by solving a high school maths exam. For example, it is possible that DeepMind were to impressively solve symbolic reasoning in full generality, and yet drop this particular line of research. In such a world, there might be a point at which it is clear that the feat in the question is achievable, even though no one has actually bothered to implement the experiment.

This is similar to how the release of the AlphaZero agent (playing chess, Go and shogi) should give us >=95% confidence that it is possible to play Othello at superhuman level without domain-specific knowledge, even though no one (to the author's knowledge) ran that experiment.

The exact composition and structure of the council will be announced soon, we’re currently recruiting members here.