[Xrvgu Heona "frpbaq nyohz"]

One of the questions in Carlini’s ‘A GPT-4 Capability Forecasting Challenge’ highlights the difficulty of ‘grounding’ (and it is an opportunity to encourage us to think about how addressing such concerns may require interface adaptations—including perhaps towards doubting more (Lindemann, 2023)).

While GPT-4 seemingly gets the question wrong even by its own logic, the question post by Carlini is—surprisingly(?)–not well enough specified. You can search the underlying fact on various platforms—in ROT13 I searched: [Xrvgu Heona “frpbaq nyohz”]—and find multiple answers for the same question. Searching Google and glancing at the snippets in the different results shows them all—including reputable specialist sites—waffling between a couple answers without a hint of hesitation or humility.

Google’s SGE (which I had to click for the generation) agrees with Carlini when the question is keyword focused, but provides an answer matching a direct claim from Wikipedia with a more natural language formulation.

Google’s featured snippet, which sometimes appears, shows a list that would suggest the Wikipedia answer but highlights Carlini’s answer.

Google’s Bard, Perplexity AI, Phind, and You.com all provide the Wikipedia answer. Metaphor autoprompts the natural language query into a prompt that includes the Wikipedia answer.

Andi, prompted in natural language (as it says “The more detailed your question, the better I can help 😄”), refers to Wolfram Alpha and provides a third answer. This matches what a refined prompt for GPT-4 reveals as well.

If you have any interest in checking out the forecasting game, you can avoid the spoiler below.