GPT-2 was a major step forward for language models, but it's still not at human level performance in meaningful text generation and reasoning (for discussion on the nuances of this point, see this and this).

One of the top benchmarks for this type of reasoning is the Winograd Schema challenge.

Rather than base the test on the sort of short free-form conversation suggested by the Turing Test, the Winograd Schema Challenge (WSC) poses a set of multiple-choice questions that have a particular form.

Q. The trophy would not fit in the brown suitcase because it was too big (small). What was too big (small)?

• Answer 0: the trophy
• Answer 1: the suitcase

GPT-2 scored a 70.70% on the 140 Q&A dataset, while human performance is at 92%+. It beat the previous record of 63.7%, but it's still a long way from human level performance.

When will an unsupervised language model get to human level performance (>=92%) on the Winograd Schema Challenge?

Resolution

The date of the model will be set as the date of the first credible public release (e.g. blog post, pre-print, peer-reviewed paper). The question will resolve retroactively one week prior to this date.

An unsupervised language model will be defined as one that has not been trained specifically to maximise WSC score. (For example, GPT-2 was trained to predict the next word in long text sequences.) If the models parameters have been pre-trained on another dataset/objective function, but are then fine-tuned to fit the WSC challenge, the model does not count as unsupervised.