Your submission is now in Draft mode.

Once it's ready, please submit your draft for review by our team of Community Moderators. Thank you!

Submit Essay

Once you submit your essay, you can no longer edit it.


This content now needs to be approved by community moderators.


This essay was submitted and is waiting for review.

GPT-3 Outperforms Fourth Graders


Machine intelligence has been steadily progressing since the invention of the digital computer, but this progress has arguably been accelerating of late, with widespread deployment of machine learning systems and dramatically increased funding of artificial intelligence research.

GPT-3 is a machine learning language model developed and announced by OpenAI in May 2020. It is a 175 billion-parameter model trained on a large corpus of text, and has proven surprisingly capable.

GPT-3 has not definitely not reached intelligence parity with humans, as defined, for example, by this question. But has GPT-3 achieved (text-based) intelligence parity with a human 4th grader?

Assume that prior to 2021, a generalized intelligence test will be administered as follows. A team of three experts will collaboratively craft a set of 20 questions to be put to both GPT-3 and three humans.

The humans will be relatively typical 4th grade (aged 9-10) students with above average grades while not being particularly precocious (but being willing to take a 20-question test) chosen independently of the questioners. The exam will be scored blindly by a disinterested group of three graders, averaged over the three graders.

GPT-3 will be run with settings (chosen below) intended to optimize its performance. Each question will be answered three times by GPT-3, and the scores of the three trials averaged.

Question resolves positively if the total of averaged GPT-3 scores exceeds the total of the averaged human scores on such a test prior to 2021.

Resolution is ambiguous if no such text is administered prior to 2021-01-01.

Some fine print:

  • All questions will be given in text form, advantaging GPT-3.

  • The 20 questions will be chosen to advantage the humans. All topics and question types are fair game, but:

  • This is not a Turing test meant to distinguish GPT-3 from 4th graders, but rather a test of who/what answers questions correctly. That is, questions should be such that both GTP-3 and the 4th graders would be easily and objectively outperformed by, say, human graduate students. Questions will be scored (0-2) as incorrect, partial credit, or correct.

  • Judges will be requested not to be prejudiced one way or another even if they can discern who/what is giving which answers. To help, scoring will be done with the identities of the respondents scrambled for each question (rather than all answers from one respondent scored at once.)

  • GPT-3's questions will be given through a "wrangler" who is experienced with and has access to GPT-3 and can choose settings and "prime" the system using at most 2048 characters of text (determined prior to knowing any of the questions), and supplement each question with up to 32 characters of non-leading text that is the same for each question and chosen in advance. (For example "Teacher: " to form a student-teacher dialogue.)

  • The 4th graders can have as much time as they like to type their replies, but cannot get help or look online for answers, nor use external aids like pencil and paper.

Make a Prediction


Note: this question resolved before its original close time. All of your predictions came after the resolution, so you did not gain (or lose) any points for it.

Note: this question resolved before its original close time. You earned points up until the question resolution, but not afterwards.

Current points depend on your prediction, the community's prediction, and the result. Your total earned points are averaged over the lifetime of the question, so predict early to get as many points as possible! See the FAQ.