Retrieval ≠ Discovery. Finding an answer isn't the same as creating one.

The buzz the past few days has been around the Paul Erdos questions. I actually have a funny story about that, from my Carnegie Mellon days.

Austin, Michael (Mike), and I were walking back from a party one night. We were all part of the same fraternity, and it was a long walk, so we started brainstorming creative ways to raise money. These two, being insufferable math majors, throw out (ironically): "We should just solve a couple of Erdos problems."

For context — Paul Erdos was a legendary mathematician who posed thousands of problems in combinatorics and number theory, many with small cash prizes attached. Basically, solve a theorem, win $100.

Fast-forward to this week: OpenAI researchers claimed GPT-5 had "solved" a batch of Erdos problems. Turns out… not quite. Mathematicians pointed out the answers were already in the literature — GPT-5 had retrieved known solutions, not discovered new ones. Posts were frantically edited… and the internet had its fun.

Two good reminders from all of this:

1. Retrieval ≠ Discovery. Finding an answer isn't the same as creating one.

2. Data lineage matters. If you don't know where information came from, you can't trust what it means. This is especially difficult (and ignoring it, even more dangerous) with foundational models.