HN user
I'm not saying its objective or quantitative, but I do think its an interesting task because it would be challenging for most humans to come up with a good design of a pelican riding a bicycle.
also: NITPICKER ALERT
Love seeing this benchmark become more iconic with each new model release. Still in disbelief at the GPT-5 variants' performance in comparison but its cool to see the new open source models get more ambitious with their attempts.
If accuracy is a major concern, then it's probably guaranteed better to go with the HTML documents. Otherwise, I've heard Docling is pretty good from a few co-workers.
I've got to know these 2-3 more exercises that you speak of. Also what is the more intelligent way to emulate that.
Is that Dwarf Fortress I see on your GitHub profile picture :)