Goalposts

Hacker News has set AI a lot of challenges over the years. Which ones has it met?

About whether ChatGPT fails to understand causation.

My litmus test question for LLMs is about negative mutual information. Its a weird quirk of probability theory where pairwise variables always have non-negative MI, but a set of three variables can have negative MI.

Its both a non-intuitive result and one where there is a tooon of text about pairwise MI and little about negative MI.

So far, nothing has passed the test.

An LLM correctly answers a question about how three variables can have negative mutual information.

Has this happened?

Yes 25 (23%)Not sure 79 (71%)No 7 (6%)

Votes cast 1–2 October 2026: 100,590 votes from 9,694 people.